Citation
Developing Semantic Digital Libraries Using Data Mining Techniques

Material Information

Title:
Developing Semantic Digital Libraries Using Data Mining Techniques
Creator:
KIM, HYUNKI ( Author, Primary )
Copyright Date:
2008

Subjects

Subjects / Keywords:
Archives ( jstor )
Conceptual hierarchies ( jstor )
Databases ( jstor )
Datasets ( jstor )
Digital libraries ( jstor )
Information search ( jstor )
Mathematical vectors ( jstor )
Metadata ( jstor )
Mining ( jstor )
Subject terms ( jstor )

Record Information

Source Institution:
University of Florida
Holding Location:
University of Florida
Rights Management:
Copyright Hyunki Kim. Permission granted to the University of Florida to digitize, archive and distribute this item for non-profit research and educational purposes. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder.
Embargo Date:
5/31/2007
Resource Identifier:
436098749 ( OCLC )

Downloads

This item is only available as the following downloads:


Full Text

PAGE 1

DEVELOPING SEMANTIC DIGITAL LIBRARIES USING DATA MINING TECHNIQUES By HYUNKI KIM A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLOR IDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2005

PAGE 2

Copyright 2005 by Hyunki Kim

PAGE 3

To my family

PAGE 4

ACKNOWLEDGMENTS I would like to express my sincere gratitude to my advisor, Dr. Su-Shing Chen. He has provided me with financial support, assistance, and active encouragement over the years. I would also like to thank my committee members, Dr. Gerald Ritter, Dr. Randy Chow, Dr. Jih-Kwon Peir, and Dr. Yunmei Chen. Their comments and suggestions were invaluable. I would like to thank my parents, Jungza Kim and ChaesooKim, for their spiritual support from thousand miles away. I would also like to thank my beloved wife, Youngmee Shin, and my sweet daughter, Gayoung Kim, for their constant love, encouragement, and patience. I sincerely apologize to my family for having not taken care of them for so long. I would never have finished my study without them. Finally, I would like to thank my friends, Meongchul Song, Chee-Yong Choo, Xiaoou Fu, Yu Chen, for their help. iv

PAGE 5

TABLE OF CONTENTS page ACKNOWLEDGMENTS.................................................................................................iv LIST OF TABLES...........................................................................................................viii LIST OF FIGURES...........................................................................................................ix ABSTRACT.......................................................................................................................xi CHAPTER 1 INTRODUCTION........................................................................................................1 1.1 Motivation...............................................................................................................1 1.2 Objective.................................................................................................................2 1.3 Approach.................................................................................................................2 1.4 Research Contributions...........................................................................................4 1.5 Outline of Dissertation............................................................................................5 2 BACKGROUND..........................................................................................................6 2.1 Digital Libraries......................................................................................................6 2.1.1 Digital Objects..............................................................................................8 2.1.2 Metadata.......................................................................................................9 2.1.3 Interoperability in Digital Libraries............................................................12 2.2 Federated Search...................................................................................................12 2.3 OAI Protocol for Metadata Harvesting.................................................................15 2.4 Data Mining..........................................................................................................17 2.4.1 Document Preprocessing............................................................................18 2.4.2 Document Classification............................................................................23 2.4.3 Document Clustering..................................................................................23 3 DATA MINING AND SEARCHING IN THE OAI-PMH ENVIRONEMENT.......26 3.1 Introduction...........................................................................................................26 3.2 Self-Organizing Map............................................................................................28 3.3 Data Mining Method using the Self-Organizing Map..........................................30 3.3.1 The Data.....................................................................................................31 v

PAGE 6

3.3.2 Preprocessing: Feature Extraction/Selection and Construction of Document Input Vectors..................................................................................32 3.3.3 Construction of a Concept Hierarchy.........................................................35 3.4 System Architecture..............................................................................................38 3.4.1 Harvester.....................................................................................................39 3.4.2 Data Provider..............................................................................................41 3.4.3 Service Providers........................................................................................42 3.5 Integration of OAI-PMH and Z39.50 Protocols...................................................48 3.5.1 Integrating Federated Search with OAI Cross-Archive Search.................48 3.5.2 Data Collections.........................................................................................49 3.5.3 Mediator.....................................................................................................50 3.5.4 Semantic Mapping of Search Attributes.....................................................50 3.5.5 OAI-PMH and Non-OAI-PMH Target Interfaces......................................51 3.5.6 Client Interface...........................................................................................51 3.6 Discussion.............................................................................................................52 3.7 Summary and Future Research.............................................................................54 3.7.1 Summary.....................................................................................................54 3.7.2 Future Research..........................................................................................55 4 AUTOMATED ONTOLOGY LINKING BY ASSOCIATIVE NAIVE BAYES CLASSIFIER..............................................................................................................56 4.1 Introduction...........................................................................................................56 4.2 Related Work........................................................................................................58 4.2.1 Document Classification............................................................................58 4.2.2 Frequent Pattern Mining.............................................................................62 4.3 Gene Ontology......................................................................................................64 4.4 Associative Nave Bayes Classifier......................................................................65 4.4.1 Definition of Class-support and Class-all-confidence................................67 4.4.2 ANB Learning Algorithm...........................................................................70 4.4.3 ANB Classification Algorithm...................................................................71 4.5 Experiments..........................................................................................................74 4.5.1 Real World Datasets...................................................................................75 4.5.2 Preprocessing and Feature Selection..........................................................77 4.5.3 Experiments................................................................................................82 4.6 Summary and Future Research.............................................................................87 4.6.1 Summary.....................................................................................................87 4.6.2 Future Research..........................................................................................87 5 DATA MINING OF MEDLINE DATABASE..........................................................90 5.1 Introduction...........................................................................................................90 5.2 Data Mining Method for Organizing MEDLINE Database.................................93 5.2.1 The Data.....................................................................................................93 5.2.2 Text Categorization....................................................................................93 5.2.3 Text Clustering using the Results of MeSH Descriptor Categorization.....94 5.2.4 Feature Extraction and Selection................................................................95 vi

PAGE 7

5.2.5 Construction of a Concept Hierarchy.........................................................96 5.2.6 Experimental Results..................................................................................96 5.3 User Interfaces......................................................................................................97 5.3.1 MeSH Major Topic Tree View and MeSH Term Tree View.....................97 5.3.2 MeSH Co-occurrence Tree View...............................................................98 5.3.3 SOM Tree View.........................................................................................99 5.4 Discussion.............................................................................................................99 5.5 Summary.............................................................................................................100 6 CONCLUSIONS......................................................................................................101 APPENDIX GENE ONTOLOGY TERMS DISCOVERED IN MEDLINE CITATIONS.................103 REFERENCES................................................................................................................106 BIOGRAPHICAL SKETCH...........................................................................................114 vii

PAGE 8

LIST OF TABLES Table page 2-1 DC elements.............................................................................................................11 2-2 Operation types in the OAI-PMH............................................................................16 3-1 Statistical data for the number of harvested records................................................31 4-1 An example transaction database.............................................................................69 4-2 Support, class-support, all-confidence, class-all-confidence, bond, and class-bond values using the transaction database of Table 4-1.........................................69 4-3 Description of datasets.............................................................................................75 4-4 Number of citations with N classes (1 <= N <= 4)...................................................76 4-5 Top 20 GO terms......................................................................................................77 4-6 Top 10 words for small, medium, and large datasets...............................................82 4-7 Confusion matrix......................................................................................................83 4-8 Performance summary of classifiers........................................................................84 4-9 Number of k-itemsets (k >= 2) mined by LB and ANB algorithms.........................85 4-10 Train and test times of classifiers (Time: CPU Seconds).........................................86 4-11 Test time and precision of ANB with ISM-B (Time: CPU Seconds)......................86 4-12 Total numbers of vertices, edges, and unique noun phrases mine for molecular function, biological process and cellular component...............................................88 viii

PAGE 9

LIST OF FIGURES Figure page 2-1 Self-organizing process of digital libraries................................................................7 2-2 Digital object..............................................................................................................8 2-3 Federated searching approaches...............................................................................13 2-4 Data mining process.................................................................................................18 3-1 Self-organizing map.................................................................................................29 3-2 Overview of text data mining processes..................................................................31 3-3 Preprocessing steps..................................................................................................34 3-4 Document matrix......................................................................................................35 3-5 Top-tier SOM...........................................................................................................37 3-6 Construction of conceptual hierarchies....................................................................38 3-7 System architecture of the integrated DL system.....................................................39 3-8 Harvester architecture..............................................................................................40 3-9 Federated search interface........................................................................................44 3-10 Search results page...................................................................................................44 3-11 Interface of top-level concept hierarchy...................................................................46 3-12 Browsing interface of the leaf node.........................................................................46 3-13 Interface of concept summarizer..............................................................................47 3-14 Federated search architecture...................................................................................48 3-15 Search interface........................................................................................................52 3-16 Search results interface.............................................................................................52 ix

PAGE 10

4-1 A binary classification in SVM................................................................................61 4-2 Gene Ontology.........................................................................................................65 4-3 ANB learning algorithm...........................................................................................71 4-4 ANB Classification Algorithm.................................................................................74 4-5 Preprocessing steps..................................................................................................79 4-6 A sample list of stop words......................................................................................80 4-7 Graph representation of a GO term..........................................................................88 5-1 MeSH tree structures................................................................................................91 5-2 An example of MEDLINE record............................................................................93 5-3 Interface of MeSH major topic tree view.................................................................98 5-4 Interface of SOM tree view......................................................................................99 x

PAGE 11

Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy DEVELOPING SEMANTIC DIGITAL LIBRARIES USING DATA MINING TECHNIQUES By Hyunki Kim May 2005 Chair: Su-Shing Chen Major Department: Computer and Information Science and Engineering We define the semantic digital libraries as the digital libraries that can discover hidden, useful information from large amounts of stored data using data mining techniques such as clustering, classification, association rule mining, and visualization techniques. To build a semantic digital library, we first propose an integrated digital library system that provides multiple viewpoints of harvested metadata collections by combining search and data mining technologies. This system provides three value-added services: (1) the cross-archive search service provides a term view of harvested metadata, (2) the concept browsing service provides a subject view of harvested metadata, and (3) the collection summary service provides a collection view of each metadata collection. We also propose a text data mining method using a hierarchical self-organizing map algorithm to build concept hierarchies from Dublin Core metadata. We then present a new classification method, called Associative Nave Bayes (ANB), to associate MEDLINE citations with Gene Ontology (GO) terms. We define the xi

PAGE 12

concept of class-support to find frequent itemsets and the concept of class-all-confidence to find interesting itemsets. In the training phase, ANB finds frequent and interesting itemsets and estimates the class prior probabilities and the probabilities of itemsets for all classes. Once the frequent and interesting itemsets are discovered in the training phase, new unlabeled examples are classified by the classification algorithm by incrementally choosing the most interesting itemset. Empirical test results on three MEDLINE datasets show that ANB is superior to both nave Bayesian classifier and Large Bayes. The results also show that ANB is more scalable than Support Vector Machines. Finally, we present a text mining method that uses both text categorization and text clustering for building concept hierarchies for MEDLINE citations. The approach we propose is a three-step data mining process for organizing MEDLINE database: (1) categorizations according to Medical Subject Headings (MeSH) terms, MeSH major topics, and the co-occurrence of MeSH descriptors; (2) clustering using the results of MeSH term categorization; and (3) visualization of categories and hierarchical clusters. The hierarchies automatically generated may be used to support users in browsing behavior and help them identify good starting points for searching. xii

PAGE 13

CHAPTER 1 INTRODUCTION 1.1 Motivation Digital libraries (DLs) have emerged from the building of information infrastructures on the Internet (Chen, 1998). The frame of DLs contains a collection of networked, heterogeneous information bases that may be databases, knowledge bases, information repositories, and other information resources. The union of all the information bases is called the information space of a digital library. The users of a digital library may browse and navigate through the information space to access information and to solve problems using the accessed information. The research field of digital libraries must be viewed as a union of subfields from a variety of domains combined with new research issues in order to realize its full potential (Nuernberg et al., 1995). The field of digital libraries presents a set of complex research issues, which are interoperable digital library system architecture (Chen, 1998; Liu, 2002; Lagoze and Sompel, 2001a; Lagoze and Sompel, 2001b), digital copyright/rights management (Chen, 2003), relaxing or overcoming the information overload problem created by new information technologies (e.g., Internet) and the continuing explosion of data (Kim et al., 2003), preservation of digital records (Chen, 2003), security issues (Chen et al., 2004), and so on. Thus, solutions to these problems will require a blending of approaches from a variety of fields. In this dissertation, we concentrate on two fundamental research issues, which are interoperability and overcoming information overload, in digital libraries. 1

PAGE 14

2 1.2 Objective The objective of this dissertation is to build a semantic digital library. We define the semantic digital libraries as the extension of digital libraries that can automatically discover hidden, useful information from large amounts of data stored in heterogeneous digital libraries. To achieve this goal, the Open Archives Initiative Protocol for Metadata Harvesting (OAI-PMH) and data mining techniques such as clustering (Kim et al., 2003) classification (Lewis, 1998; Meretakis et al., 1999; Yang and Liu, 1999), association rule mining (Agrawal et al., 1993; Zaki et al., 1997) and visualization are used. The OAI-PMH provides an application-independent interoperability framework to develop and promote interoperability solutions in heterogeneous digital libraries. Data mining techniques are used to 1.3 Approach Interoperability plays a central role in digital libraries towards better accessibility, organization, retrieval, and dissemination of scholarly information. The community of digital libraries has invested a substantial amount of effort and cost in collection of a broad range of rich digital resources and providing access to them to the public. To fully utilize these resources, the issue of accessibility and visibility of the digital resources is of prime concern. However, most digital libraries have been built in isolation utilizing different technologies, protocols, and metadata in terms of syntax and semantics (Liu, 2002). To promote interoperable solutions in heterogeneous digital libraries, the Open Archives Initiative (OAI) recently established a metadata harvesting protocol (Lagoze and Sompel, 2001a; Lagoze and Sompel, 2001b). However, there are still challenging

PAGE 15

3 issues in the OAI environment such as metadata incorrectness, poor quality of metadata, and vocabulary inconsistency that have to be solved in order to create various high-quality services (Kim et al., 2003). To solve those problems this research proposes an integrated digital library system in the OAI-PMH environment that provides cross-archive search and multiple viewpoints of harvested metadata collections by combining search and data mining technologies. With the explosion of data on the Internet, users’ inability of expressing their information needs might become more serious, unless users have either a precise knowledge in a domain of their interest or an understanding of collection. Thus, a better mechanism is needed to fully exploit structured metadata, to organize information more effectively, and to help users explore within the organized information space. To tackle those problems, we present two text data mining methods: a MeSH (Medical Subject Headings)-regulated hierarchical clustering and an association-based classification method. The MeSH-regulated clustering method using a hierarchical self-organizing map (SOM) algorithm (Kohonen, 2001) can be used to build concept hierarchies from Dublin Core metadata (Dublin, 1999) contained in MEDLINE citations. The concept hierarchies generated by using the SOM is used to help users in browsing behavior, and to help them understand the contents of collection as a way of choosing good collections for their search. Providing multiple viewpoints of a document collection and allowing users to move among these viewpoints will enable both inexperienced and experienced searchers to more fully exploit the information contained in a document collection. User interfaces with multiple viewpoints for this underlying system are also presented.

PAGE 16

4 We also present a new classification method, called Associative Nave Bayes (ANB), to associate MEDLINE citations with Gene Ontology (GO) terms (Gene Ontology Consortium). We define the concept of class-support to find frequent itemsets and the concept of class-all-confidence to find interesting itemsets. In the training phase, ANB finds frequent and interesting itemsets, employing an extended Eclat algorithm (Zaki et al., 1997; Zaki, 2000), and estimates the class prior probabilities and the probabilities of itemsets for all classes. Once the frequent and interesting itemsets are discovered in the training phase, new unlabeled examples are classified by the classification algorithm by incrementally choosing the most interesting itemset. Empirical test results on three MEDLINE datasets show that ANB is superior to nave Bayesian classifier. The results also show that ANB outperforms the state of the art Large Bayes (LB) classifier (Meretakis et al., 2000; Meretakis and Wuthrich, 1999) and is more scalable than Support Vector Machines (Joachims, 1999; Joachims, 2001). 1.4 Research Contributions The contributions of this dissertation are three-fold. First, we propose an integrated digital library system using the metadata harvesting method towards digital library interoperability. Second, we integrate cross-archive search with data mining and thus provide multiple viewpoints of metadata collections. Most research in the OAI-PMH framework has concentrated on cross-archive search, reference linking and citation analysis, peer-review, and componentized Open Digital Libraries (Chen and Choo, 2002; Liu et al., 2001; Liu et al., 2002a; Suleman and Fox, 2002). From the viewpoint of data mining, less research has been conducted to extract nuggets of information. Finally, we develop two data mining algorithms, the MeSH-regulated hierarchical clustering

PAGE 17

5 algorithm and the associative nave Bayes classification algorithm, for building semantic digital libraries. 1.5 Outline of Dissertation The rest of this dissertation is organized as follows. Chapter 2 gives a brief introduction to digital libraries, including digital objects, metadata, digital library interoperability, and then describes federated search, OAI-PMH, and data mining process and techniques. Chapter 3 discusses the design and implementation of an integrated digital library based on the OAI-PMH. We also present a text data mining method employing the self-organizing map and the integration of the OAI-PMH protocol with Z39.50 protocol Chapter 4 starts by introducing document classification methods, which are nave Bayes, Large Bayes, and Support Vector Machine, frequent patter mining, and Gene Ontology. We then present a new classification method to automatically classify MEDLINE citations with the GO terms. Finally, we compare our algorithm with other algorithms and explain the experimental results. Chapter 5 describes a MeSH-regulated hierarchical clustering method that uses both text classification and text clustering for building a concept hierarchy for MEDLINE citations. Chapter 6 summarizes the results presented in this dissertation and gives conclusions.

PAGE 18

CHAPTER 2 BACKGROUND In this chapter, we describe a brief outline of digital libraries, metadata, federated search techniques, Open Archives Initiative Protocol for Metadata Harvesting (OAI-PMH), data mining methods that include document classification and document clustering. 2.1 Digital Libraries Digital library is a relatively new term being used extensively today especially in the publishing and education community. Different people have different views of what a digital library actually is and of what it consists (Chen, 1998). Digital library can be thought of as the architecture that houses and delivers information which can be accessed, searched and managed by the user. Digital libraries serve as the window of information for the user to disseminate particular information of interest and provide a powerful knowledgebase system through networking technologies. A digital library system may be thought of as mediating certain kinds of interactions among people and computing systems. Digital libraries are often characterized by the large volume of information accessible to the user through network infrastructures and should be considered as the life cycle of information. Digital libraries provide an architecture for the acquisition, collection, indexing, utilization, dissemination and organization of the digital elements that it houses (Chen, 1998). A definition of digital libraries is given by Chen (1998, page 18): 6

PAGE 19

7 In our perspective, digital libraries organize knowledge whether it belongs to traditional libraries or on the Internet. Thus digital libraries transcend traditional libraries. The least common denominator of digital libraries should be a networked information system providing users uniform access to distributed information repositories locally and remotely. Furthermore digital technologies enables users to do much more – to access, organize, disseminate and utilize all the repositories at their desktops via networks. Therefore digital libraries should be more than networked information systems and considered as the life cycle of information. In the following Figure 2-1, we describe the self-organizing nature of digital libraries. This spiral process is the circle in the figure interconnecting basic library functions, each of which may move to any other directly. Collection Acquisition Indexing Utilization Dissemination Organization Collection Acquisition Indexing Utilization Dissemination Organization Figure 2-1. Self-organizing process of digital libraries In the life cycle of information, the basic library operations are information acquisition and collection, information indexing and organization, and information dissemination and utilization. There are also secondary operations which are distinctive to digital technology and are included in the basic library operations. The secondary operations are information conversion and transformation, information communication and transmission, information brokerage and integration and information delivery and presentation (Chen, 1998).

PAGE 20

8 2.1.1 Digital Objects Information content in digital libraries is encapsulated in digital objects. Digital objects are essentially the basic building blocks for digital libraries (Chen, 1998). Digital objects vary in types such as tutorials, documents, news, and software. The media formats used for digital objects that may include images, text, audio, video, 3D renderings, graphs, and combinations of these media formats are diverse but are not limited. Digital objects are dynamic, interactive, adaptive, and extensible (Chen, 1998). Figure 2-2 depicts the structure of a digital object. Object Handle Metadata Objects Information Contents Interaction with users/applications Figure 2-2. Digital object As shown in Figure 2-2, the digital object consists of an object handle, metadata objects, and information contents. The object handle is a unique identifier to distinguish one digital object from another object. Metadata objects are used to describe annotations and labeling of the digital object such as title, author, created date, and so on. Information contents contain concepts or knowledge on a particular topic. Digital objects are dynamic in that they may be formed from networked sources. Therefore, a digital object might be a collection of digital objects but presented to the user as a single digital object. Such kinds of digital objects have complex internal

PAGE 21

9 structures and are called compound or complex objects (Chen, 1998). For example, an Hypertext Markup Language (HTML) page might contain text and images. The images might be acquired from other digital object sources on another server through the network. However, the HTML page is a digital object by itself but is formed from multiple networked digital objects. The dynamic nature of digital objects is made available through the use of handles to distinguish one digital object from another. The handles are mostly created automatically for the digital objects based on a predetermined handle syntax and format for that repository. It can be an internal database identifier or an identifying name formed from the unique directory and object (file) name of that digital object. Other forms of handles that are commonly used to distinguish among digital objects are Universal Resource Locater (URL) and Universal Resource Name (URN). Both URL and URN are a subset of Universal Resource Identifier (URI). 2.1.2 Metadata Metadata are usually defined as “data about data” and consist of structured, standardized descriptions of resources, whether digital or physical, that aid in the discovery, retrieval and use of those resources. Libraries have been creating metadata for centuries, in the form of book catalogs, card catalogs and, more recently, online catalogs. Metadata allow the precise description of resources (and the sharing of such descriptions) in relatively small and discrete packages of information called metadata records, without the necessity of involving the resources themselves in the transaction. For example, in the networked environment of the Web, metadata records describing resources useful in education can be gathered up (harvested) from geographically distributed resource repositories without affecting the location of the resources they describe. These harvested metadata records can be assembled in metadata repositories

PAGE 22

10 where they can function as an "online union catalog" of distributed teaching and learning resources. The Dublin Core (DC) metadata standard defines a simple set of elements for describing a wide range of networked resources (Dublin, 1999). The DC standard comprises of fifteen elements all of which are optional and repeatable. The purpose of the DC standard is to support wide area searching with a small set of elements. DC standard uses descriptive metadata that can be used for indexing, discovery, and identification of a resource. Table 2-1 explains 15 DC elements and their definitions. The Dublin Core element set has been kept as small and simple as possible to allow a non-specialist to create simple descriptive records for information resources easily and inexpensively, while providing for effective retrieval of those resources in the networked environment. The DC metadata is commonly used for the following reasons: 1. Simplicity of creation and maintenance: The DC element set has been kept as small and simple as possible to allow a non-specialist to create simple descriptive records for information resources easily and inexpensively, while providing for effective retrieval of those resources in the networked environment. 2. Commonly understood semantics: Discovery of information across the vast commons of the Internet is hindered by differences in terminology and descriptive practices from one field of knowledge to the next. The Dublin Core can help the a non-specialist searcher finds his or her way by supporting a common set of elements, the semantics of which are universally understood and supported. For example, scientists concerned with locating articles by a particular author, and art scholars interested in works by a particular artist, can agree on the importance of a "creator" element. Such convergence on a common, if slightly more generic, element set increases the visibility and accessibility of all resources, both within a given discipline and beyond. 3. International scope: The Dublin Core Element Set was originally developed in English, but versions are being created in many other languages, including Finnish, Norwegian, Thai, Japanese, French, Portuguese, German, Greek, Indonesian, and Spanish. The DCMI Localization and Internationalization Special Interest Group are coordinating efforts to link these versions in a distributed registry.

PAGE 23

11 Table 2-1. DC elements Element Description Comment Title A name given to the resource Typically, title will be a name by which the resource is formally known Creator An entity primarily responsible for making the content of the resource Examples of Creator include a person, an organization, or a service. Typically, the name of a Creator should be used to indicate the entity Subject A topic of the content of the resource Subject will be expressed as keywords, key phrases or classification codes that describe a topic of the resource Description An account of the content of the resource Examples of Description include: an abstract, or table of contents Publisher An entity responsible for making the resource available Examples of publisher include a person, an organization, or a service. Contributor An entity responsible for making contributions to the content of the resource Examples of contributor include a person, an organization, or a service. Typically, the name of a Contributor should be used to indicate the entity Date A date of an event in the lifecycle of the resource Typically, date will be associated with the creation or availability of the resource Type The nature or genre of the content of the resource Type includes terms describing general categories, functions, genres, or aggregation levels for content Format The physical or digital manifestation of the resource Format may include the media-type or dimensions of the resource Identifier An unambiguous reference to the resource within a given context Recommended best practice is to identify the resource by means of a string or number conforming to a formal identification system Source A reference to a resource from which the present resource is derived The present resource may be derived from the Source resource in whole or in part Language A language of the intellectual content of the resource Language defines twoand three-letter primary language tags with optional subtags Relation A reference to a related resource Relation is to identify the referenced resource by means of a string or number conforming to a formal identification system Coverage The extent or scope of the content of the resource Coverage will include spatial location (a place name or geographic coordinates), temporal period or jurisdiction Rights Information about rights held in and over the resource Rights will contain a rights management statement for the resource, or reference a service providing such information

PAGE 24

12 2.1.3 Interoperability in Digital Libraries Interoperability in digital libraries plays a fundamental role towards better accessibility, organization, retrieval and dissemination of scholarly information. D libraries have invested substantial amount of effort and cost in collection of a broad range of rich digital resources and providing access to them to the public. To fully utilize these resources, the issue of accessibility and visibility of the digital resources is of prime concern. According to the NSDL community, digital library interoperability is defined as (NSDL, 2002): Interoperability requires cooperation at three levels: technical, content, and organizational. Technical agreements cover formats, protocols, and security systems so that messages can be exchanged, etc. Content agreements cover the data and metadata, and include semantic agreements on the interpretation of the messages. Organization agreements cover the ground rules for access, for changing collections and services, payment, authentication, etc. The main challenge in providing interoperability among digital library collections is the issue of decentralization and heterogeneity among digital libraries. Decentralization often means a diversity of query languages, information retrieval protocols, capabilities, attributes and organizational structures. Seamless interoperability of digital libraries involves reconciling heterogeneity and integrating the digital libraries at several levels. Cross-archive search towards digital library interoperability is an ongoing important issue especially in the escalating cost and effort required in setting up and managing multiple distributed digital resource collections. 2.2 Federated Search Distributed information access is essential to digital library operations because digital libraries are distributed across networks. Distributed information access means

PAGE 25

13 that users are able to browse and search heterogeneous distributed digital libraries jointly without facing search gaps. A distributed information retrieval problem arises when the documents are spread across heterogeneous distributed digital libraries. In general, there are two ways to implement a coherent set of digital services across heterogeneous digital repositories: a distributed information retrieval approach and a harvesting approach (Liu et al., 2002b) as shown in Figure 2-3. User DL Remote DL1 Remote DLn QueryResults QueryQueryResultsResults User DL QueryResults HarvestedMetadata Remote DL1 Remote DLn QueryQueryResultsResults (A) (B) Figure 2-3. Federated searching approaches (A) Distributed information retrieval approach (B) Harvesting approach Distributed information retrieval has been widely researched as the method for federating DLs. Distributed information retrieval is cast in three parts: server selection (database selection), query processing, and results merging (Powell et al., 2000). 1. Server selection: A subset of servers to which the query will be forwarded is selected based on server selection methods. 2. Query processing: Each selected server processes the query and returns search results. 3. Results merging: Search results returned from multiple search servers are merged into a single results list and the results list is presented to the user.

PAGE 26

14 An experimental study on the distributed information retrieval shows that distributed searching in a large number of heterogeneous nodes does not easily provide a federated search service, since the distributed search system may be vulnerable to its weakest component (Powell et al., 2000). Furthermore, most research on distributed information retrieval usually assumes cooperation of search servers. This means that the majority of methods proposed, and evaluated in simulated environments of homogeneous cooperating servers, are never applied in practice (Craswell, 2000). In the harvesting approach, distributed information retrieval can be emulated after harvesting metadata from repositories and a building cross-archive search on the harvested metadata (Chen and Choo, 2002). This can cause data duplication and thus cross-archive searching does not always provide fresh data. Good retrieval performance, however, can be achieved in the harvesting approach compared to that of distributed information retrieval regardless of a number of nodes. Liu et al. (2002a) introduced a cross-archive searching interface, which allow users to search and select specific fields with data constructed from the harvested metadata, and an interactive searching interface for the subject field to solve the problem of metadata variability. Suleman and Fox (2002) proposed an Open Digital Library (ODL) architecture by extending the Open Archives Initiative Protocol for Metadata Harvesting to provide simple and reusable component models. By separating the retrieval of result sets from indexing and storing them, as in the harvesting approach, a light-weight application can access a larger number of repositories. Less storage space is required and the data is guaranteed to be fresh as every user query is routed to the target repositories. However, performance is an issue with

PAGE 27

15 federated search as it is determined by the slowest target site. Integration of the heterogeneous targets is difficult as different targets are largely incompatible in search options and result set objects (Liu, 2002). To facilitate querying multiple target repositories, a homogeneous structure or protocol is often assumed. Information retrieval protocols such as Z39.50 (ANSI/NISO, 1995), STARTS (Gravano et al., 1997), Dienst (Davis and Lagoze, 2000), and Search/Retrieve Web Service (SRW/U) can be used to provide interoperability among repositories. The Z39.50 (ANSI/NISO, 1995) protocol was designed for client-server access and adapted to federated searching, whereby a system performing a search operation on multiple repositories could send the query to all of them in a standardized format and then process the returned results as appropriate. The Dienst protocol and the STARTS protocol both implemented variations of federated search methods, where queries are sent to remote sites in real time. Some federated search projects may use proprietary protocols or data archive specific interfaces to provide interoperability among data archives. However, there is great reliance on the result set object formatting returned from the respective targeted sites and will need constant monitoring and modifications in order to keep up with the most recent version of the result set object formatting. 2.3 OAI Protocol for Metadata Harvesting The Open Archives Initiative is an organization formed by a broad range of librarians, publishers, researchers, and archivists. Its goal is to create simple standards to support interoperability among heterogeneous digital libraries. The OAI-PMH (Lagoze and Sompel, 2001a) provides an application-independent interoperability framework. Table 2-2 shows the operation types (verbs) in the OAI

PAGE 28

16 PMH. OAI-PMH requests are expressed as Hypertext Transfer Protocol (HTTP) requests such as HTTP GET or POST methods. All responses to OAI-PMH requests must be well-formed eXtensible Markup Language (XML) instance documents. The returned XML record has three parts: Header – information common to all records and necessary for the harvesting process, Metadata – metadata elements of returned records, and About – optional container to hold data about the metadata part of the record. Table 2-2. Operation types in the OAI-PMH Operation Description Parameters GetRecord Retrieve an individual metadata record from a repository. identifier: object for which to return metadata metadataPrefix: format to use Identify Retrieve information about a repository (e.g. name of the repository, protocol version, administrator’s email address). ListIdentifiers Retrieve only headers rather than records. set: the set to list identifiers from from: starting date until: ending date resumptionToken: token to get next batch of identifiers metadataPrefix: format to use ListMetadataFormats Retrieve the metadata formats available from a repository identifier: object for which to list metadata formats ListRecords Harvest records from a repository. Optional arguments permit selective harvesting of records based on set membership and/or datestamp. set: the set to list identifiers from from: starting date until: ending date resumptionToken: token to get next batch of records metadataPrefix: format to use ListSets Retrieve the set structure of a repository, useful for selective harvesting. resumptionToken: token to get next batch of records

PAGE 29

17 In the OAI framework, there are two classes of participants: data providers and service providers (Lagoze and Sompel, 2001a). Data providers adopt the OAI technical framework as a means of exposing the content of metadata. The requirement for metadata interoperability is addressed by requiring that all OAI data providers supply metadata in a common format, the Dublin Core Metadata Element Set (DCMES) (Lagoze and Sompel, 2001a; Lagoze and Sompel, 2001b). Service providers harvest metadata from data providers using the OAI-PMH and use the harvested metadata as the basis for building value-added services. These archives would then act as a federation of repositories, by indexing documents in a standardized way so that multiple collections could be searched as though they form a single collection (Lagoze and Sompel, 2001b). This service is called cross-archive search. While current Web search engines usually deal with semi-structured data, cross-archive search engines using the OAI-PMH framework should exploit structured metadata describing the core information properties. 2.4 Data Mining To improve retrieval efficiency and effectiveness, data mining uses document classification, document clustering, machine learning, and visualization technologies. Data mining is defined as the nontrivial extraction of implicit, previously unknown, and potentially useful information from data (Chakrabarti, 2000; Frawley et al., 1992). Data mining has been popularly treated as a synonym of knowledge discovery in databases (KDD), although some researchers view data mining as an essential step of knowledge discovery. Knowledge discovery process consists of computational theories and a variety of software systems to extract useful knowledge from large volumes of data. In general, knowledge discovery process consists of an iterative sequence of seven stages (Han and Kamber, 2000).

PAGE 30

18 1. Data cleaning is the removal of noise, errors, the correction of incorrect data from data sources, and the resolution. 2. Data integration combines data from multiple sources into a coherent data warehouse. 3. Data reduction and transformation obtains reduced representation in volume but produces the same or similar analytical results and find useful features through normalization and aggregation. 4. Data selection is the selection of relevant attributes to the analysis. 5. Data mining is search for patterns of interest in one or more representational forms based on selected data mining algorithms. 6. Pattern evaluation is the identification of interesting mined patterns. 7. Visualization interprets the mined patterns and highlight most relevant mined patterns to users in a comprehensible manner. Figure 2-4 shows this data mining process. Data Target DataPreprocessedData -----------------------------------------Selection Preprocessing TransformedData Transformation Data MiningPatterns Interpretation/Evaluation Knowledge Figure 2-4. Data mining process 2.4.1 Document Preprocessing To apply data mining techniques, the documents should be converted into numerical data for further analyses. This step is generally called document preprocessing.

PAGE 31

19 Document preprocessing involves feature extraction, feature selection and document representation. Feature Extraction The first step in document preprocessing is feature extraction. The general task in this phase is to generate a list of unique words, called indexing terms, from the documents. Feature extraction subsumes tokenization, stop word removal, and stemming. The goal of tokenization is to separate a stream of the characters in the document into a stream of tokens, in such a way that each individual word as well as every punctuation will be a different token. Tokenizer maps all words in the document into lowercase, and separates punctuations from the preceding and/or following tokens. It also expands verb contractions. The goal of stop word removal is to remove high frequency words, which are called stop words. Stop words usually comprise 40% to 50% of the text words (Salton, 1983) and are pronouns, prepositions, conjunctions, and so on. Since stop words have poor discriminating power and cannot possibly be used by themselves to identify the document content, they are eliminated by consulting a list of predefined stop words. For instance, “a,” “about,” “above,” and “according” are regarded as stop words in general. Stemming, also called morphological normalization, is used to reduce different words with the same stem into the word root form. This process reduces the number of unique words. A commonly used word stemming algorithm is the Porter stemming algorithm (Porter, 1980).

PAGE 32

20 Feature Selection Feature selection is a major problem related to data mining. In general, traditional feature extraction methods select the n highest weighted terms as the features. The problem of feature selection is to find a subset of features for optimal classification or clustering. For clustering, when the processing task is to partition a given document collection into clusters of similar documents, a choice of good features along with good clustering algorithms is of paramount importance (Dhillon et al., 2002). For classification, a critical part of feature selection is to rank features according to their importance. Much research has been done to solve the feature selection problem of the high dimensional data (Christianini and Shawe-Taylor, 1999; John et al, 1994; Yang and Pedersen, 1997; Youn, 2004). We give a brief introduction on the following four feature selection methods. Document Frequency Thresholding Document frequency is the number of documents in which a word occurs in a dataset. In document frequency thresholding, all words occurring in less than n documents in the dataset are eliminated. Document frequency thresholding is based on the assumption rare terms are either non-informative, or not influential in global performance (Yang and Pederson, 1997). Since it is a simple method, it scales upto a very large collection of documents. Document frequency thresholding is an unsupervised feature selection method. Mutual Information Mutual information (Cover and Thomas, 1991) is commonly used in statistical language modeling of word associations and related applications. Mutual information

PAGE 33

21 measures the entropy difference provided by considering two random variables w and c together instead of individually. {,}{0,1}(,)()(|)(,) (,)log()()ywIcwHcHcwPcwPcwPyPw Here, the variable c is the class label assigned to a document and the variable w describes whether a particular word occurs in the document or not. A weakness of mutual information is that the score is strongly influenced by the marginal probabilities of terms (Yang and Pederson, 1997). Odds Ratio The odds ratio of word w (Van Rijsbergen et al., 1981) can be computed as follows: (|)(1(|)()log(1(|))(|)PwpositivePwnegativeOddsRatiowPwpositivePwnegative where P(w | positive) is the probability of the documents that contains word w in positive example and P(w | positive) is the ration of the documents containing word w in negative example. In this measure, features frequently occurring in positive example have higher scores. Chi-square (2) Test The chi-square test (Mood et al., 1974) measures the association between the term t and the category c. It is defined to be 2N(P(t,c)(,)(,)(,))chi-square(t,c)=()()()()PtcPtcPtcPtPtPcPc 1()((,))niichisquaretavgchisquaretc

PAGE 34

22 where N is the number of document in the dataset. The value of the metric shows the dependence between term t and class c and is used as the ranking criterion. Document Representation In Information Retrieval (IR), there are two popular methods, a Boolean model and a vector-space model, for representing documents. In either case, the documents d can be represented by term vectors of the from 123,,,...,ndwwww where wi identifies a term assigned to document d. Document representation based on the Boolean model considers whether index terms are present or absent in a document. The document representation of the Boolean model is given by 123,,,...,, where {0,1}nidwwwww Vector-space model assign non-binary weights to index terms and the definition of this model is given by 123,,,...,, where 0nidwwwww The term weighting scheme in vector-space model is commonly based on the term frequency (TF) or the term frequency combined with the inverse document frequency (TF-IDF). _()(,)(iitermweightwTFwdIDFw )i where term frequency TF(wi, d) is the number of times word wi occurs in document d and inverse document frequency IDF(wi) is the inverse of document frequency. Document frequency DF(wi) is defined as the number of documents in which word wi occurs in a

PAGE 35

23 dataset. Inverse document frequency is the measure of how rare word wi in the dataset. IDF(wi) is given by ||()log()iiDIDFwDFw where |D| is the number of document in the dataset. Inverse document frequency shows that words appearing in many documents are not useful for distinguishing a relevant document from a non-relevant one. 2.4.2 Document Classification Classification is the process of automatically assigning a document to a predefined one or more subject categories. Classification tends to be used interchangeably with categorization. A classifier is a function that maps an input vector, 123,,,...,ndtttt , to a confidence that the input belongs to a class (i.e., ()()fdconfidenceclass ). Classification is a two-step process (Han and Kamber, 2000). A model describing a predetermined set of categories is constructed in the first step. The data used to construct the model is called the training data and each individual record of the training data referred to as a training sample or a training example. In the second step, the model is used to classify a new example into one or more predetermined categories. 2.4.3 Document Clustering Document clustering is used to group similar documents into a set of clusters (Baeza-Yates and Ribeiro-Neto, 1999). To improve retrieval efficiency and effectiveness, related documents should be collected together in the same cluster based on shared features among subsets of the documents.

PAGE 36

24 Document clustering methods are in general divided into two ways: hierarchical and partitioning approaches (Vesanto and Alhoniemi, 2000). The hierarchical clustering methods build a hierarchical clustering tree called a dendrogram, which shows how the clusters are related. There are two types of hierarchical clustering: agglomerative (bottom-up) and divisive (top-down) approaches (Vesanto and Alhoniemi, 2000). In agglomerative clustering, each object is initially placed in its own cluster. The two or more most similar clusters are merged into a single cluster recursively. A divisive clustering initially places all objects into a single cluster. The two objects that are in the same cluster but are most dissimilar are used as seed points for two clusters. All objects in this cluster are placed into the new cluster that has the closest seed. This procedure continues until a threshold distance, which is used to determine when the procedure stops, is reached. Partitioning methods divide a data set into a set of disjoint clusters. Depending on how representatives are constructed, partitioning algorithms are subdivided into k-means and k-medoids methods. In k-means, each cluster is represented by its centroid, which is a mean of the points within a cluster. In k-medoids, each cluster is represented by one data point of the cluster, which is located near its center. The k-means method is minimizing the error sum of squared Euclidean distances whereas the k-medoids method is instead using dissimilarity. These methods are either minimizing the sum of dissimilarities of different clusters or minimizing the maximum dissimilarity between the data point and the seed point. Partitioning methods are usually better than hierarchical ones in the sense that they do not depend on previously found clusters (Vesanto and Alhoniemi, 2000). On the other

PAGE 37

25 hand, partitioning methods make implicit assumptions on the form of clusters and cannot deal with the tens of thousands of dimensions (Vesanto and Alhoniemi, 2000). For example, the k-means method needs to define the number of final clusters in advance and tends to favor spherical clusters. Hence, statistical clustering methods are not suitable for handling high dimensional data, for reducing the dimensionality of a data set, nor for visualization of the data. A new approach to addressing clustering and classification problems is based on the connectionist, or neural network computing (Chakrabarti, 2000; Kohonen, 1998; Kohonen, 2001; Roussinov and Chen, 1998; Vesanto and Alhoniemi, 2000). The self-organizing map is an artificial neural network algorithm is especially suitable for data survey because it has prominent visualization and abstraction properties (Kohonen, 1998; Kohonen, 2001; Vesanto and Alhoniemi, 2000).

PAGE 38

CHAPTER 3 DATA MINING AND SEARCHING IN THE OAI-PMH ENVIRONMENT 3.1 Introduction A main requirement of digital libraries (DLs) and other large information infrastructure systems is interoperability (Chen, 1998). The Open Archives Initiative (OAI) is an experimental initiative for the interoperability of digital libraries based on metadata harvesting (Lagoze and Sompel, 2001a; Lagoze and Sompel, 2001b). The goal of OAI is to develop and promote interoperability solutions to facilitate the efficient dissemination of content. However, simply harvesting metadata from archives tends to generate a great deal of heterogeneous metadata collections whose diversity may grow exponentially with the proliferation of the Open Archives Initiative Protocols for Metadata Harvesting (OAI-PMH) data providers. Furthermore, there are still challenging issues such as metadata incorrectness (e.g., XML encoding or syntax errors), poor quality of metadata, and vocabulary inconsistency that have to be solved in order to create a variety of high-quality services. Even though the OAI is explicitly purposed for coarse granularity resource discovery (Lagoze and Sompel, 2001b), to provide users with high quality services, we believe that both data providers supply high-quality metadata with richer metadata schemes and services providers offer a wide variety of value-added services. The most common value-added services that have been developed in the OAI-PMH framework are cross-archive search services, reference linking and citation analysis services, and peer-review services to date. 26

PAGE 39

27 One of challenging problems for service providers originates from vocabulary inconsistency among the OAI-PMH archives (Kim et al., 200; Liu et al., 2002; Sompel et al., 2000). This is because the OAI-PMH uses the simple (unqualified) Dublin Core that uses no qualifiers as the common metadata set to encourage metadata suppliers to expose Dublin Core metadata according to their own needs at the initial stage (Lagoze and Sompel, 2001b). The Dublin Core metadata standard (Dublin, 1999) consists of fifteen elements and each element is optional and repeatable. It only provides a set of semantic vocabularies for describing the information properties, not restricting their content any further. Thus, the use of Dublin Core differs significantly among the individual archives. Content data for some elements may be selected from a controlled vocabulary, a limited set of consistently used and carefully defined terms, and data providers try to use controlled vocabularies in several elements. But controlled vocabularies are only widely used in some elements not relating to the content of item, such as type, format, language, and date elements, among OAI-PMH archives (Liu et al., 2002). Without basic terminology control among data providers, vocabulary inconsistency can profoundly degrade the quality of service providers. Since most of the cross-archive search engines are based on keyword search technology in Information Retrieval, cross-archive keyword search of metadata that is harvested often results in relatively low recall (only a proportion of the relevant documents are retrieved) and poor precision (only a proportion of the retrieved documents are relevant). In addition, users sometimes cannot exactly articulate their information needs. The users’ inability of expressing their information needs might become more serious, unless users have either a precise knowledge in a domain of their

PAGE 40

28 interest or an understanding of collection. Thus, a better mechanism is needed to fully exploit structured metadata, to organize information more effectively, and to help users explore within the organized information space (Chen, 1998; Kim et al., 2003). To provide users with powerful methods for organizing, exploring, and searching collections of metadata harvested, we propose an integrated digital library system with multiple viewpoints of harvested metadata collections by combining cross-archive search and data mining methods. This system provides three value-added services: (1) the cross-archive search service provides a term view of harvested metadata, (2) the concept browsing service provides a subject view of harvested metadata, and (3) the collection summary service provides a collection view of each metadata collection. We also propose a text data mining method using a hierarchical self-organizing map (SOM) algorithm with two input vectors to build concept hierarchies from Dublin Core metadata. The concept hierarchies generated by using the SOM can be used to help users in browsing behavior, to help them understand the contents of collection as a way of choosing good collections for their search, and to classify the harvested metadata collections for the purpose of selective harvesting. User interfaces for this underlying system are also presented. Providing multiple viewpoints of a document collection and allowing users to move among these viewpoints will enable both inexperienced and experienced searchers to more fully exploit the information contained in a document collection (Powell and French, 2000). 3.2 Self-Organizing Map The self-organizing map (Kohonen, 2001) is an unsupervised-learning neural network algorithm for the visualization of high-dimensional data. In its basic form, it produces a similarity graph of input data. It converts the nonlinear statistical relationships

PAGE 41

29 between high-dimensional data into simple geometric relationships of their image points on a low-dimensional display, usually a regular two-dimensional grid of nodes. The SOM may be described formally as a nonlinear, ordered, smooth mapping of the high-dimensional input data manifolds onto the elements of a regular, low-dimensional array. The SOM, thereby, compresses information while preserving the most important topological and/or metric relationships of the primary data elements on the display. It may also be thought to produce some kind of abstraction. These two characteristics of visualization and abstraction can be utilized in a number of ways in complex tasks such as process analysis, machine perception, control, communication, and data mining (Kohonen, 2001). Figure 3-1 shows an example of SOM. Figure 3-1. Self-organizing map The SOM algorithm is a recursive regression process and works as follows: 1. Initialize network: For each node i, set the initial weight vector wi(0) to be random. 2. Present input: Present x(t), the input pattern vector x at time t (0
PAGE 42

30 sequentially. The input pattern vector x may be chosen at random or cyclically from the training data set. 3. Compute winning node: Calculate winning node c with smallest distance between weight vector and input vector ()()min{()()}cii x twtxtwt hence argmin{()()}iicxtw t)c 4. Update weights: Update weights for c and nodes within neighborhood Nc(t) (1)()()()(),((1)(),()iiciiiicwtwthtxtwtifiNtwtwtifiNt where hci is a sclar kernel function. 5. Present next input: Decrease so that cih (1)()cicihtht . Reduce the neighborhood set so that for all i. Repeat from step 2 choosing a new unique input vector (1)()iNtNt i (1)(), x txjj t until all iterations have been made(. )tn 3.3 Data Mining Method using the Self-Organizing Map Data mining is used to extract implicit, previously unknown, and potentially useful information from data. The approach we propose here is a three-step data mining process for organizing metadata: (1) MeSH-regulated document clustering using the hierarchical SOM, (2) deduction of the concept hierarchy, and (3) visualization of the concept hierarchy. Visualization of the concept hierarchy is described in section 3.4.3. Figure 3-2 illustrates the overview of our data mining method for the clustering of harvested metadata records.

PAGE 43

31 1.Construction of document vectors-Preprocess text-Construct document vector as weighted term vectors 2.Construction of SOM-Initialize SOM-Present each document in order-Train SOM Repeat many times 3. Construction of user interface-Select labels automatically to characterize map regions-Create user interface (e.g. HTML, JSP files) DocumentsDocument vectorsSOMUser Interface Figure 3-2. Overview of text data mining processes 3.3.1 The Data For the following experiments we harvested a collection of 21,022 metadata elements from 5 OAI registered data providers. The harvested metadata describes either electronic theses and dissertation materials or technical reports in scientific fields. Table 3-1 shows statistical figures related to the harvested metadata. Table 3-1. Statistical data for the number of harvested records OAI Repository ID Repository Name Number of Harvested Records caltechcstr Caltech Computer Science Technical Reports 395 LSUETD LSU Electronic Thesis and Dissertation Archive 350 HKUTO Hong Kong University Theses Online 8,598 N/A M.I.T. Theses 7,803 VTETD Virginia Tech Electronic Thesis and Dissertation Collection 3,876

PAGE 44

32 3.3.2 Preprocessing: Feature Extraction/Selection and Construction of Document Input Vectors Document Representation The Dublin Core Metadata Element Set (DCMES) contains 15 elements, which can be refined to add richness of description (Dublin, 1999). DCMES describes content, intellectual property, and instantiation of the resource (Hillmann, 2001). Coverage, description, type, relation, source, subject, and title elements are used to describe the content of the resource. Especially, description, title, and subject elements give the account, name, and topic of the content of the resource, respectively. We consider these three elements important to be indexed for the clustering of metadata by employing the SOM. To produce a concept hierarchy employing the SOM, documents must be represented by a set of features. To do this, we use full-text indexing to extract a list of terms and then weight these terms using the vector space model in Information Retrieval (Baeza-Yates and Riberto-Neto, 1999; Salton and Buckley, 1988). In the vector space model, documents are represented as term vectors using the product of the term frequency (TF) and the inverse document frequency (IDF). Each entry in the document vector corresponds to the weight of a term in the document. We use a normalized TFIDF term weighting scheme, the best fully weighted scheme (Salton and Buckley, 1988), so that longer documents are not unfairly given more weight and all values of a document vector are distributed in the range of 0 to 1. Thus, a weighted word histogram can be viewed as the feature vector describing the document (Kohonen et al., 2000). The best fully weighted scheme is the popular TFIDF representation with Euclidian length normalization and is defined as (Salton and Buckley, 1988):

PAGE 45

33 2||(,)log()||(,)log()iiijjjDTFwdDFwxDTFwdDFw During preprocessing we make two input vector sets for the SOM: subject and description feature vectors. The subject input vector constructed by indexing the subject elements of the metadata collection is used to construct a subject-specific top-tier concept map. The description input vector built by indexing the title and description elements of the metadata collection is used to build a sub-layer concept hierarchy after constructing the top-tier concept map. The preprocessing procedure is divided into two stages: noun phrase (feature) extraction and term weighting. In the noun phrase extraction phase, we first take the description, subject, and title elements from the database and tokenize these elements based on the Penn Treebank tokenization scheme (http://www.cis.upenn.edu/~treebank/tokenization.html) to detect sentence boundaries and to separate extraneous punctuations from the input text. We then automatically assign part of speech tags to words reflecting their syntactic category by employing the rule-based part of speech tagger (Brill, 1992; Brill, 1994). After recognizing the chunks that consist of noun phrases and chunking them from the tagged text, we extract noun phrases. We then count occurrences of all terms in the whole corpus for each input set. From the corpus constructed by indexing the title and description elements, we eliminate some terms occurring less than the predefined frequency threshold (in this experiment less than 10 times) and common terms by consulting a list of 903 stop words. Finally, we weight the indexed terms using the best fully weighted scheme and assign corresponding

PAGE 46

34 term weights to each document. Thus, the weighted term vector can be used as the input vector of the SOM. Tokenization Part-of-speech Tagging Noun Phrase Chunking The/ DTNN/IN/JJ/JJ/NN/NNSVBZ/VBNINDTNN/IN/JJ/NNNN performance/ofk-aryn-cubeinterconnectionnetworksis/analyzedunder/the/assumption/ofconstantwirebisection/./. The performance of k-aryn-cube interconnection networks (e.g., tori) is analyzed under the assumption of constant wire bisection. [ The/DT performance/NN ][ k-ary/JJn-cube/JJ interconnection/NN networks/NNS ] [ the/DT assumption/NN ] [ constant/JJ wire/NN bisection/NN ] of/IN is/VBZ analyzed/VBN under/INof/IN./. [1]: performance[2]: k-aryn-cube interconnection network[3]: [4]: tori[5]: assumption[6]: constant wire bisectionInput text The performance of k-aryn-cube interconnection networks is analyzed under the assumption of constant wire bisection ( e.g. , tori). e.g. Stop-word Removal [1]: performance (Term frequency: 1)[2]: k-aryn-cube interconnection network (Term frequency: 1)[3]: tori(Term frequency: 1)[4]: assumption (Term frequency: 1)[5]: constant wire bisection (Term frequency: 1) •Term Frequency: #occurrences of a term in a text•Document Frequency: #documents that a term appears in corpus DT: determinerNN: singular nounIN: prepositionJJ: adjectiveNNS: plural nounVB: verb Document Frequency Thresholding(e.g., 5) [1]: abortion (Document frequency : 1) [2]: absorption (Document frequency: 6)[3]: accelerometers (Document frequency: 1)[4]: accent and accentuation (Document frequency: 12)[5]: accounting (Document frequency: 7) [1]: absorption (Document frequency: 6)[2]: accent and accentuation (Document frequency: 12)[3]: accounting (Document frequency: 7) Frequency Counting Document matrix Figure 3-3. Preprocessing steps After the feature selection, a collection of documents can be represented by a m n term-document matrix. An entry in the matrix corresponds to the frequency of a term in the document. Figure 3-4 depicts an example of document matrix. A value of 0 in Figure

PAGE 47

35 3-x means that the corresponding term has no significance in the document or it simply does not occur in the document. T1T2. TtD1d11d12d1tD2d21d22d2t: : : ::: : :Dndn1dn2dntDocument matrix airart . zooD10 12 25D20 17 21: : : ::: : :Dn43 7 32Document matrix example Figure 3-4. Document matrix In our experiment, we identified 1,760 terms by indexing the subject elements in the 14,029 metadata set. We also indexed 1,996 terms from the description and title elements of metadata set, after removing 141,617 terms, which were either common terms or terms having a low term frequency. Although all documents of the metadata set had the title elements, 6,960 and 16,478 documents did not have the subject and description elements, respectively. 3.3.3 Construction of a Concept Hierarchy We then constructed a concept hierarchy by extending the multi-layered SOM algorithm (Roussinov and Chen, 1998), permitting unlimited layers of SOM, with two input vector sets. The vector size of the subject feature vector was 1,674 and the vector size of the description feature vector was 1,871. The SOM defines a mapping from the input data space onto a usually two-dimensional array of nodes. Every node i is represented by a model vector, also called reference vector, mi = , where n is input vector dimension.

PAGE 48

36 The following algorithm describes how to construct a subject-specific concept hierarchy using two different input vector sets for Dublin Core metadata. 1. Initialize network by using the subject feature vector as the input vector: Create a two-dimensional map and randomly initialize model vectors mi in the range of 0 to 1 to start from an arbitrary initial state. 2. Present input vector in sequential order: Cyclically present the input vector x(t), the weighted input vector of an n-dimensional space, to all nodes in the network. Each entry in the input vector corresponds to the weight of a noun phrase in the document; zero means the term has no significance in the document or it simply doesn’t exist in the document. 3. Find the winning node by computing the Euclidean distance for each node: In order to compare the input and weight vectors, each node computes the Euclidean distance between its weight vector and the input vector. The smallest of the Euclidean distance identifies the best-matching node that is chosen as the winning node for that particular input vector. The best-matching node, denoted by the subscript c, is }{miniicmxmx . 4. Update weights of the winning node and its topological neighborhoods: The update rule for the model vector of node i is )()()()()()1(tmtxthttmtmiciii where t is the discrete-time coordinate, (t) is the adaptation coefficient, and hci(t) is the neighborhood function, a smoothing kernel centered on the wining node. 5. Repeat steps 2-4 until all iterations have been completed. 6. Label nodes of the trained network with the noun phrases of the subject feature vectors: For each node, we determine the dimension with the greatest value, and label the node with a corresponding noun phrase for that node, and aggregate nodes with the same noun phrase into groups. Thus, a subject-specific top-tier concept map is generated. 7. Repeat steps 1-6 by using the description feature vector as the input vector for each grouped concept region: For each grouped concept region containing more than k documents (e.g. 100), recursively create a sub-layer SOM and repeat steps 1-6 by using the description feature vector as the input vector. At this point new input feature vector of the sub-layer SOM is dynamically created by selecting only those items that belong to the concept region represented by its parent SOM from the description feature vector. Thus, different sets of feature vectors are used for different clusters and this reduces the training time significantly.

PAGE 49

37 Figure 3-5 depicts the result of a 20 20 SOM by using the subject feature vector as the input vector after iterating 100 times. There are 35 concept regions in the map. Note that more important concepts occupy larger regions, and similar concepts are grouped in the same neighborhood (Lin et al., 1999). We inserted this information into the MySQL database to build an interactive user interface. Figure 3-5. Top-tier SOM After the construction of the top-level SOM, we recursively create sub-layer SOMs, by repeating the steps 1-6 using the description feature vector as the input vector. Figure 3-6 shows this procedure.

PAGE 50

38 Subject feature vectorDescription feature vector Figure 3-6. Construction of conceptual hierarchies 3.4 System Architecture The proposed integrated DL system is designed to serve both as a data and a service provider. The architecture of the system consists of three main components as shown in Figure 3-7: the harvester, the data provider, and a set of the service providers. The harvester that issues OAI-PMH requests to archives and collects metadata encoded in XML from archives supports selective and automatic harvesting. Selective harvesting allows harvesters to limit harvest requests to portions of the metadata available from a repository (Lagoze and Sompel, 2001a). Although the harvester component can be viewed as a part of a service provider, we have distinguished it to better clarify the importance of this component in the integrated DL architecture. The data provider can expose the harvested metadata, reformat harvested metadata to other metadata formats,

PAGE 51

39 and correct inaccurate metadata before it is exposed to service providers. The service providers provide users with three value-added services: a cross-archive search service, an interactive concept exploration service, and a collection summarizing service. We use MySQL database as a storage system through Java DataBase Connectivity (JDBC). The integrated DL system is highly extensible and the components of the system are reusable for rapid development because of its object-oriented design and implementation. DigestedMetadata DigestedMetadataHarvestedMetadata HarvestedMetadataService Providers’Data Service Providers’DataHarvester HarvesterDataProvider DataProviderServiceProvider 1 ServiceProvider 1ServiceProvider n ServiceProvider nDatabase Interface layer Database Interface layer Integrated Digital Library System DataProvider 1 DataProvider 1 InternetDataProvider n DataProvider n ServiceProvider 1 ServiceProvider 1 ServiceProvider n ServiceProvider n User 1 User 1 User n User n Figure 3-7. System architecture of the integrated DL system 3.4.1 Harvester The harvester collects metadata from various OAI-PMH compliant archives as shown in Figure 3-8. The web accessible graphical user interface that is password protected allows the system administrator to harvest metadata from a data provider. In this step the administrator can selectively harvest metadata based on (1) the date the

PAGE 52

40 metadata was added or modified, (2) the metadata prefix (a string to specify the metadata format in OAI-PMH requests), (3) the archive predefined set, and (4) the individual records. The latter parameter for selective harvesting allows the administrator to specify which records are worth harvesting and digested into the integrated DL. DL Server HarvesterComponent DP1 Harvester GUI:-URL to harvest-Selective harvesting parameters Harvest APIParametersData ProvidersHarvested Metadata SelectiveRecord Filter DP2 DPn... Digested Metadata OMI-PMHRequest/Response Figure 3-8. Harvester architecture The harvester component also has the capability to do automatic harvesting on a timely basis (daily, weekly or monthly). This service is controlled by the administrator based on how often new metadata is inserted, and the rate the metadata is updated for a given data provider. In the automatic harvesting, only newly added or updated metadata from the original data provider is requested and digested into the integrated DL. Automatic harvesting based on statistical data gathered from the data providers is currently being investigated. Such statistical data would be useful for automatic adaptation harvesting without human intervention and it is also needed to minimize the problem of propagation change (Suleman and Fox, 2002). Propagation change happens when the harvester waits a certain amount of time between harvest operations; and thus

PAGE 53

41 becomes out of synchronization with the data provider that it is harvesting. This problem is further propagated to the next harvester harvesting the previously stated harvester. 3.4.2 Data Provider The data provider can expose the harvested metadata to service providers. It can also reformat the harvested metadata into other metadata formats and correct the inaccurate metadata if possible before the metadata is exposed to service providers. The data provider can expose single or combined harvested metadata sets to service providers. This provides a one-stop DL for service providers to harvest various metadata sets from different data providers separately or in a single request. The set field in the OAI-PMH is used to group the metadata from different data providers into different sets by using the OAI repository ID as the setSpec parameter. If an OAI repository ID is not available, the repository’s name is used instead. By specifying more than one set group in the OAI request, service providers can harvest metadata sets from different data providers in a single request. For example, service providers who want to harvest LSUETD and VTETD metadata sets can specify the OAI-PMH set parameter as “set=LSUETD,VTETD” (or “set=LSUETD%2CVTETD” after encoded for transmission). This is an extension of the OAI-PMH in our part to serve integrated metadata sets to service providers. However, the service provider will not be able to differentiate between the metadata sets unless the individual record identifiers are descriptive enough (e.g. oai:VTETD:etd-11). Metadata sets can also be grouped, based on terms or concepts. By employing the SOM described in section 3.1, hierarchical organization of sets can be built with tens of top-level nodes, each of which is itself a set. One example is to group the metadata sets into different top-level subject-specific domains such as “Chemical Engineering” or “Civil Engineering”. Each top-level set may

PAGE 54

42 have sub-layer sets of the sub-subjects or related subjects such as “Computer Science” and “Algorithm.” When necessary, metadata digested into the DL can be converted to other metadata formats before being exposed to service providers. For example, metadata harvested from VTETD as Dublin Core can be reformatted to MAchine-Readable Cataloging (MARC) format on the fly before being exposed to service providers. This allows the service providers to focus their services on the metadata harvested instead of having to build harvesters for different metadata formats. We are also able to expose metadata sets from data providers that have XML syntax or encoding errors in their metadata. Not all data providers strictly implement the OAI-PMH as with our experiments. By relaxing our harvesting approach, we are able to obtain metadata sets from these data providers and correct the syntax or encoding errors where possible before the metadata is exposed to service providers. The system can be an alternative data provider site for service providers. Based on our harvesting experiments, a significant number of data providers could not be accessed at certain periods of time. The availability of metadata and also the success of an OAI harvest operation are dependent on the accessibility of the data provider. Our DL thus can provide a replication service for other data providers. 3.4.3 Service Providers Service Providers use the metadata harvested via the OAI-PMH as a basis for building value-added services (Lagoze and Sompel, 2001a). A key design principle of the service providers is to provide multiple viewpoints of harvested metadata. Powell and French (1998) introduced the concept of multiple viewpoints, representation schemes of several organizations of a collection. They demonstrate that providing multiple

PAGE 55

43 viewpoints of a document collection and allowing users to move among these viewpoints during a search/browse session will facilitate the location of useful documents (Powell and French, 1998). To provide multiple viewpoints, this service provider layer offers three value-added services. First, a cross-archive search interface provides a term view of harvested metadata. Second, a concept browsing interface gives a subject view of harvested metadata. Finally, a collection summarizing interface provides a collection view of each collection. Cross-archive Searching The cross-archive search provides users with a way to search on part or all of the archived metadata from the archives harvested, thus emulating a federated search service. The search service is based on full-text search on the DC elements from the harvested metadata collection. A simple, keyword-based search is provided as shown in Figure 3-9, where the title and description metadata elements from all harvested archives are queried. The resulting records returned are ranked based on the relevance to the user’s query term. This simple search provides a quick and efficient way for users to query the harvested metadata collection. We also provide an advanced search for users who want more refined results for their search process. The query fields provided are based on DC elements and users can sort the resulting records based on the results relevance to the query term, the date the record was created or updated, and the original archive name. Boolean operators (AND/OR) can be used in the query fields. Users are also provided with the option to choose the archives that they want to query. Users will have more benefit from the advance search compared to the simple search, because the query fields are more specific and defined, thus, offering higher recall and precision rates. Users are also able to choose

PAGE 56

44 related and specialized archives, which might be more relevant to the users’ search process and interest. Figure 3-9. Federated search interface Figure 3-10. Search results page The result set from the search process can be further refined with the filter elements. Users can filter the result set based on the recentness of the records (e.g., return

PAGE 57

45 only records with date on or later than specified), record type (e.g., journal records) and record format (e.g., MPEG). Users are also able to retrieve a single record by specifying the OAI identifier for that record. Figure 3-10 shows the search results page. Although having a cross-archive search service across heterogeneous data providers is feasible, we feel that providing users with metadata collections related to or emphasizing certain subject matter is more useful. Currently our cross-archive search service is focused on metadata harvested from data providers related to education. Interactive Concept Browsing After training the SOM and inserting data of the concept hierarchies into the database, we built an interactive concept browsing interface for organizing and browsing the harvested metadata collection. A key design concept of the concept browsing interface is to provide both a tree view of the concept hierarchy and a map view of the concept hierarchy to satisfy various users’ preferences. Each user has its own research focus and interest area. This fact has an important implication that different users may want the system to function differently to fit their different interests. We represented the concept hierarchy as a tree structure by alphabetically linearizing the concept space of the SOM. Figure 3-11 shows the user interface. The concept tree shown on the left pane displays the clustered information space. In each level of hierarchy, concepts are listed in alphabetical order, along with the number of documents that contain each concept. If the user clicks on a concept label, the associated document set is displayed on the right pane. Users may see the detailed information of a document by clicking on the title of the document shown on the right pane as shown in Figure 3-12.

PAGE 58

46 Figure 3-11. Interface of top-level concept hierarchy Figure 3-12. Browsing interface of the leaf node

PAGE 59

47 Collection Summarizing It is possible to achieve good document retrieval performance by carefully selecting a small number of good collections (Powell et al., 2000). To help users understand the contents of collections as a way of choosing good collections for their search, we provide a collection summarizing interface by clustering each harvested collection with the SOM. Figure 3-13 shows the interface of collection summarizer. After choosing a collection of their interest, users may select the map view or the tree view. This interface also provides information about a collection such as repository identifying information (repository name, earliest date stamp, harvesting granularity, and description of repository) and the set structure of the collection by issuing the OAI-PMH requests. Figure 3-13. Interface of concept summarizer

PAGE 60

48 3.5 Integration of OAI-PMH and Z39.50 Protocols In this section, we present our effort and experience in building cross-archive search employing two very different approaches: (1) federated search and (2) cross-archive search based on harvesting approach. We have developed cross-server federated search on DL as web services. The first is the federated search system, which mediates searches across distributed biological databases. As the OAI-PMH plays an important role in the National Science, Technology, Engineering and Mathematics Education Digital Library (NSDL) architecture, we have developed the integrated DL system with cross-server search using the OAI-PMH. In this section, we propose a combination of federated search of OAI and non-OAI repositories. 3.5.1 Integrating Federated Search with OAI Cross-Archive Search Meta-DL Client Interface Meta-DL M ed i at o r OAI Target Interface Non-OAI Target Interface Database In t e rf ace Web Interface Pubmed Nucleotide Protein Digested Metadata Tropicos BioMed Central Z39.50 Server OAI Non-OAI Direct Figure 3-14. Federated search architecture

PAGE 61

49 To offer a richer scenario of discovery and use for our digital library users, we have integrated both federated search of biological databases from the DLLSL project with the OAI harvested data providers. The integration of differing standards minimizes the need to master different search and retrieve protocols. By offering a myriad of data collections to be searched upon in a single user interface, this Meta-DL system offers a simulating search environment with more records to be queried in hope that it will provide more relevant documents to our users. Figure 3-14 shows the Meta-DL federated search architecture. 3.5.2 Data Collections Our Meta-DL collection consists of four biomedical databases (literature and sequence databases) and a botanical database. The Meta-DL system uses the federated searched biological data archives: PubMed, Entrez Protein and Entrez Nucleotide. PubMed was developed by the NCBI at the National Library of Medicine (NLM). PubMed provides access to citations from biomedical literature which includes Medline journals. The Protein database, available via the NCBI Entrez retrieval system, contains sequence data from the translated coding regions from DNA and protein sequences. This database includes GenBank, European Molecular Biology Laboratory (EMBL), DNA Database of Japan (DDBJ), Protein Information Resource (PIR), SWISSPROT, Protein Research Foundation (PRF) and Protein Data Bank (PDB). The Nucleotide database, also available through the NCBI Entrez retrieval system, has about 20 billion bases from GenBank, EMBL, DDBJ and Genome Sequence Database (GSDB). It also includes sequence data from U.S. Patent and Trademark Office (USPTO) and from other international patent offices.

PAGE 62

50 The data from OAI harvested data providers are stored in a local relational database. These archives consist of the Tropicos and the BioMed Central collections. Tropicos is a botanical database at the Missouri Botanical Garden (MOBOT) and has over 1.5 million specimen records with full locality data, coordinates and elevation information. Tropicos contains 851,000 name records for plants and associated information on bibliography, types, nomenclature, usage, distribution and morphology. Our local Tropicos data contains 73,800 records. BioMed Central (BMC) is an open access publisher that provides peer-reviewed biomedical research articles. BMC supports PubMed Central and other databases as well as self-archived articles by authors. Our local BMC data contains 1,651 records. 3.5.3 Mediator The Mediator provides a unified query and result set retrieval infrastructure to the differing search protocols in order to communicate with the heterogeneous target archives. The Mediator listens for client requests through the client interface and spawns separate processes to translate and distribute the user datagram to the archives. These processes will then wait for the result set returned from the queried archives before being collated by the mediator. This normalized result set view is then presented to the client interface for the user. The Mediator is able to referee searches to the different archives by collating the different search attributes into a single common mapping list as described next. 3.5.4 Semantic Mapping of Search Attributes In order to search across differing search protocols and attributes, we have mapped a subset of the OAI harvested archives Dublin Core metadata elements as search parameters (e.g., keyword, title, and author) to the search attributes from the PubMed and

PAGE 63

51 Entrez databases. This semantic mapping of search attributes across different target repositories provides a standard query attribute set as well as a basis for future inclusion of repositories as it is based on Z39.50 Bib1 attribute set values (ANSI/NSIO, 1995). 3.5.5 OAI-PMH and Non-OAI-PMH Target Interfaces Separate target interfaces are needed for the OAI harvested archives and the non-OAI archives as both types differ in their query and retrieve protocols. These target interfaces allow the Mediator to send queries and receive result set objects from the different repository types. The target interface goes through a conversion process to transform the user query to the target databases query syntax. The target interface to the OAI archives is constructed using MySQL database interface. This interface directly queries our local OAI harvested collections. Structured Query Language (SQL) queries are constructed on-the-fly and the target interface will then collect the result set for further examination by the user. The target interface for the PubMed and Entrez databases are constructed using HTTP GET and POST methods to the E-Utilities web query interface at NCBI. The target interface for the non-OAI repositories has timeout and retry values as querying these distributed collections are susceptible to network problems. 3.5.6 Client Interface The Meta-DL client user interface provides a single unified search and retrieval view to the user. The different search attributes and target repositories to query can be chosen by the user as shown in Figure 3-15. The Meta-DL client user interface is built using Java Server Pages (JSP) and Servlets and can be accessed through a web browser. Result lists from the user query are displayed in a new browser window and the full text record can be accessed by the record link as shown in Figure 3-16.

PAGE 64

52 Figure 3-15. Search interface Figure 3-16. Search results interface 3.6 Discussion Most research in the OAI framework has concentrated on cross-archive search, reference linking and citation analysis, peer-review, and componentized Open Digital

PAGE 65

53 Libraries (Chen and Choo, 2002; Kim et al., 2003; Liu et al., 2002b; Suleman and Fox, 2002). Recently, a scalable architecture and repository synchronization in the OAI framework have been studied (Liu et al., 2002a). Less research has been conducted to extract nuggets of information from the point of data mining. Our work is different from other work in that we integrate cross-archive search with data mining and thus provide multiple viewpoints of metadata collections. We have also proposed the multi-layered self-organizing map algorithm for building a subject-specific concept hierarchy using two input vector sets constructed by indexing the harvested metadata collection. The proposed SOM algorithm is different from other SOM-variant algorithms. First, it uses two different input vectors to cluster Dublin Core metadata more meaningfully. Second, after constructing the top-level concept map and aggregating nodes with the same concept on the map into a group, it dynamically reconstructs input vector by selecting only those items that are contained for each concept region from input vector of the higher level to generate the sub-layer map. Thus, new input vector would reflect only the contents of the region and not the all collection for each SOM. The concept hierarchy generated by the SOM can be used for two purposes: building an interactive concept browsing service with multiple viewpoints and automatic classification of harvested metadata for the purpose of selective harvesting. We, however, encountered several problems during our experiment. The problem of metadata variability limits the effectiveness of our data provider component as well as our services offered. Some metadata harvested do not have title and description elements and the date, type and format elements show great variation among archives. Metadata variability occurs due to the use of unqualified (simple) DC.

PAGE 66

54 Many data providers could not be harvested due to various reasons: connection refused, service temporarily unavailable, and XML encoding errors at the data provider server (Kim et al., 2003; Liu et al., 2001). Some registered data providers only contain a few metadata records. These data providers may be experimental OAI archives. Metadata quality offered by data providers is also questionable. Different archives present different metadata qualities and even a single archive might have different metadata qualities for their records. This is due to the subjective interpretation of the records by the metadata authors. Metadata quality affects the usefulness and value of services offered on top of the metadata. 3.7 Summary and Future Research 3.7.1 Summary We have proposed the integrated DL system architecture with OAI and self-organizing capabilities. The proposed DL system can serve both as a data provider and a service provider. To provide the users with powerful methods for organizing, exploring, and searching a collection of harvested metadata, we combined cross-archive search with data mining technologies. Our approach of harvesting metadata from various OAI registered archives, providing cross-archive search, concept browsing, and collection summarizing services, presents an integrated and centralized server beneficial to the learning community and other OAI service providers. Our research will be useful to promote accessibility, reusability and interoperability of metadata among the wide user community. Finally, we have proposed the integration of federated search (non-OAI-PMH) and harvesting (OAI-PMH) methods to do cross-archive search. By mixing two different

PAGE 67

55 approaches to cross-archive searching, we are able to combine the advantages from both approaches and at the same time, lessen the impact of their flaws. 3.7.2 Future Research To be more efficient, our system may be improved in several directions. Metadata synchronization between data provider and service provider is very important (Liu et al., 2002a; Liu, 2002; Liu et al., 2002b). Our system is able to selectively harvest metadata from data providers. However, the SOM processing for new or modified metadata is not feasible, since we have to recalculate input vectors of SOM and retrain the SOM with new input vectors. A Further limitation when using the SOM is that the size and lattice type of the map should be determined in advance. It is difficult to choose optimal parameters for the SOM without having the knowledge of the type and organization of the documents. Therefore, to obtain the best SOM with the minimum quantization error, we have to repeat training procedures several times with different parameter settings. This process is often very time-consuming for large collections of documents.

PAGE 68

CHAPTER 4 AUTOMATED ONTOLOGY LINKING BY ASSOCIATIVE NAVE BAYES CLASSIFIER 4.1 Introduction With the exponential growth of biomedical data, life science researchers have met a new challenge how to exploit systematically the relationships between genes, sequences and the biomedical literature (Yandell and Majoros, 2002). Usually most of known genes are found in the biomedical literature and MEDLINE is a worthy database for this kind of information. MEDLINE, developed by the U.S. National Library of Medicine (NLM), is a database of indexed bibliographic citations and abstracts (National Library of Medicine). It contains over 4,600 biomedical journals. MEDLINE citations and abstracts are searchable via PubMed (http://www.ncbi.nlm.nih.gov/entrez/query.fcgi) or the NLM Gateway (http://gateway.nlm.nih.gov/gw/Cmd). The biomedical literature has much to say about gene sequence, but it also seems that sequence can tell us much about the biomedical literature. Currently, highly trained biologists read the literature and select appropriate Gene Ontology (GO) terms to annotate the literature with GO terms. Gene Ontology database has more recently been created to provide an ontological graph structure for biological process, cellular component, and molecular function of genomic data (Smith et al., 2003). McCray et al. (2002) show that the GO is suitable as a resource for natural language processing (NLP) applications because a large percentage (79%) of the GO terms is passed the NLP parser. They also show that 35% of the GO terms were found in 56

PAGE 69

57 a corpus collected from the MEDLINE database and 27% of the GO terms were found in the current edition of the Unified Medical Language System (UMLS). Another recent research work completed in April 2003 shows that about 110,000 MEDLINE abstracts can be linked to the Gene Ontology (Smith and Cleary, 2003). Although some research works demonstrate that NLP is applicable to GO and MEDLINE database can be linked to the GO terms (McCray et al., 2002; Smith and Cleary, 2003; Raychaudhuri et al., 2002; Yandell and Majoros, 2002), there are inherent challenging issues to fully exploit both MEDLINE and GO databases. One of them is that there are too many class categories (i.e. GO terms) in the GO because the GO is a large, complex graph in itself. For example, in the GO database released as of February 2005, there were a total of 17,593 terms (Gene Ontology Consortium). Furthermore, GO grows in coverage and evolves in a monthly cycle. Finally, MEDLINE contains over 12 million article citations. Beginning in 2002, it began to add over 2,000 new references on a daily basis (National Library of Medicine). In order to arrange the MEDLINE contents in a useful way, we believe text classification and text clustering should be used extensively. Text classification is a boiling down of the specific content of a document into a set of one or more pre-defined labels (Hearst, 1999). Text clustering can group similar documents into a set of clusters based on shared features among subsets of the documents (Kohonen, 1998; Chen et al., 1996). In this chapter, we present an association-based classification method to classify MEDLINE citations with GO terms. By linking the Go terms with MEDLINE citations, our objective is to develop a knowledge management system that will integrate two ontologies: MeSH and GO, with potentially sequence patterns.

PAGE 70

58 The remainder of chapter is organized as follows. We briefly introduce document clustering, document classification and frequent pattern mining methods in section 4.2. We also briefly explain the Gene Ontology in section 4.3. We explain an association-based classification method in section 4.4 and then describe experiments in section 4.5. Conclusions are given in section 4.6. 4.2 Related Work In this section, we give a brief outline of document classification, and frequent pattern mining techniques. 4.2.1 Document Classification Given a fixed set of classes C = {c1, c2,, cn}, a training instance is a pair (di, ci), where di is a document represented as a vector of m attribute values = , depicting m measurements made on the vocabulary V = {W1, W2,,Wm}, and ci id C is a class label associated with di for a given instance. A training set S is a set of labeled instances S = {(d1,ci), (d2,c2),, (dm,cm)}. The goal in document classification is to infer a classification rule from the training set S so that it classifies new examples with high accuracy (Joachims, 2001). Nave Bayes Classifier The naive Bayes (NB) classifier is a probabilistic classification method. NB is based on the Bayes’ theorem and the nave Bayes independence assumption. Bayes’ theorem says that to achieve the highest classification accuracy, a document d should be assigned to the class ci for which P(ci|d) is highest. The nave Bayes independence assumption states that the probability of word wi is independent from any other word wj given that the class is known. Although this assumption is clearly false, it allows the easy

PAGE 71

59 estimation of the conditional probability P(Wj|ci). In the learning phase, NB estimates the class prior probabilities P(ci) and the conditional probability of each attribute wi given the class label ci P(Wj|ci). The estimated P(ci) is defined as ()iicPc D where |ci| denotes the number of training documents in class ci and |D| is the total number of training documents. Given a new document d = , NB predicts the class as the one with the highest probability of ci: '1(,)(|)||(',)jijiiWVTFWcPWcVTFW c where |V| is the total number of attributes in V and TF(W,ci) is the overall number of times word w occurs within the documents in class ci. At training time, NB requires linear time both to the number of training documents and to the number of features and thus its computational requirements are minimal. At classification time, a new example can be also classified in line time both to the number of features and to the number of classes. NB is particularly well suited when the dimensionality of the inputs is high and can often outperform more sophisticated classification methods due to its simplicity and effectiveness (Mitchell, 1997). Large Bayes Classifier Recently the use of frequent pattern mining techniques for classification has been researched (Liu et al., 1998; Meretakis, 1999; Meretakis, 2000). Large Bayes (LB) classifier extends the NB by employing frequent pattern mining techniques (Meretakis, 1999; Meretakis, 2000). LB represents each document as a transaction consisting of a set

PAGE 72

60 of items that express the presence or absence of specific words. At the learning phase, LB employs an Apriori-based association miner (Agrawal et al., 1993; Agrawal and Srikant, 1994) to discover associations among items, also called itemsets. Each such itemset l is labeled with the conditional probabilities for its joint occurrence with each class label P(l, ci). The learning procedure of LB discovers itemsets that are frequent in the training data and are also interesting. An itemset l is defined to be frequent, if its support (see 4.2.3) is above the user-defined minimum support. The interestingness of an itemset l is defined in terms of the error when estimating P(l, ci) using subsets of l. Let l be an itemset of size |l| and lj , lk be two (|l|-1)-itemsets obtained from l by omitting the jth and kth item respectively (Meretakis, 1999; Meretakis, 2000). Such itemsets lj, lk can be used to produce an estimate, Pj,k(l, ci) of P(l, ci): ,(,)(,)(,)(,)(,) j ikiestijkijki P lcPlcPlcPlcPllc To classify a new document, some of the generated itemsets are selected and their conditional probabilities P(l,ci) are used to compute the probability that the new document belongs to a certain class. The learning time of LB mainly depends on the performance of the association miner used for the discovery of the itemsets. Apriori-based association mining methods scale up linearly with the number of examples performing several passes over the data but exponentially in the worst case with the number of features (Meretakis, 2000). Support Vector Machines Support Vector Machine (SVM) is a promising classification method for a binary classification problem (Joachims, 2001). SVM maps a given set of n dimensional input vectors nonlinearly into a high dimensional feature space and separate the two classes of

PAGE 73

61 data with a maximum margin of hyperplane. In general, there could be several hyperplanes that separate the two classes of data without error. From those multiple hyperplanes, SVM selects the one with the largest margin . In its simplest linear form, the margin is defined by the distance of the hyperplane to the nearest of the positive and negative examples. The positive/negative examples that are nearest to the hyperplane called support vectors. Figure 4-1 depicts a binary classification problem in SVM. In Figure 4-1, the positive example is marked with a solid circle, while the negative example is marked with a dashed circle. HyperplaneSupport vectors Figure 4-1. A binary classification in SVM For the multi-class classification problem, a binary SVM is generated for each class ci in general. Each SVM is trained for each binary classification problem. Given a new document d to be classified, each SVM estimates P(ci|d). The example is classified into the class ci for which the corresponding P(ci|d) is highest. This reduction of a multi-class problem into m binary tasks is called a one-versus-all method (Joachims, 2001).

PAGE 74

62 4.2.2 Frequent Pattern Mining The frequent pattern mining task, first introduced by Agrawal et al. (1993), can be stated as follows. Let be a set of distinct items. A transaction T=(TID, X) has a unique identifier TID and contains a set of items X, called an itemset, such that. A transaction database D is a set of transactions. A set of itemsets is also called an itemset. An itemset that has k items is referred to as a k-itemset. The support of an itemset X, support(X), is the number of transactions that contains the itemset X in D. An itemset X is called frequent if support(X) is no less than a user-specified minimum support threshold, min_support. In probability terms, support(X) is the probability of observing itemset X given D, and we can write },...,,{21miiiI IiiiXk},...,{2,1 ()()() supportXPX f requencyXD where frequency(X) is the number of transactions containing X and D is the total number of transactions in D. An association rule is a conditional implication of the form X -> Y, where , and IYX, Y X 0. The confidence of this rule, confidence(X -> Y), is a conditional probability that a transaction having X also contains Y. In probability terms, we can write this as ()(|)() ()confidenceXYPYXsupportXYsupportX Given a transaction database D, a user-specified minimum support, and a user-specified minimum confidence, the problem of frequent pattern mining is to generate all

PAGE 75

63 frequent itemsets. Confidence possesses no closure property, while support possesses downward closure property: if a set of items has support, than all its subsets also have supports (Brin et al., 1997a). Instead of support and confidence, several interest measures have been proposed for finding the most interesting association rules. Brin et al. (1997a) introduce lift and conviction. Lift, originally called interestness, measures the cooccurrence of an itemset {X, Y} and is defined as ()()() ()()() ()()() ()liftXYliftYXPXYPXPYsupportXYsupportXsupportYconfidenceXYsupportY Since lift is symmetric and is not downward closed, lift is susceptible to noise in small databases. Conviction is a measure of implication since it is a directed measure (Brin et al., 1997a). Conviction measures the probability that X occurs without Y with the actual frequency of the occurrence of X without Y. It is defined as ()()()()1() 1(PXPnotYconvictionXYPXnotYsupportYconfidenceXY ) A chi-square (2) test measures the dependence or independence of items (Brin et al., 1997b). Since it is based on statistical theory and is not possessed the downward closure property, some approximate similarity measures should be used to determine positive and negative correlations between items.

PAGE 76

64 4.3 Gene Ontology The word “ontology” is sometimes used in many different senses. A definition of ontology is given by Gruber (1993, pp.1): An ontology is an explicit specification of a conceptualization. The term is borrowed from philosophy, where an ontology is a systematic account of Existence. For AI systems, what "exists" is that which can be represented. When the knowledge of a domain is represented in a declarative formalism, the set of objects that can be represented is called the universe of discourse. This set of objects, and the describable relationships among them, are reflected in the representational vocabulary with which a knowledge-based program represents knowledge. Thus, in the context of AI, we can describe the ontology of a program by defining a set of representational terms. In such an ontology, definitions associate the names of entities in the universe of discourse (e.g., classes, relations, functions, or other objects) with human-readable text describing what the names mean, and formal axioms that constrain the interpretation and well-formed use of these terms. Formally, an ontology is the statement of a logical theory. The Gene Ontology (GO) is the result of an effort to create a controlled vocabulary for representing gene functions and products. GO is a community database resource developed by the GO consortium, a group of scientists who work on a variety of model organism databases. GO includes three structured ontologies to describe molecular functions, biological processes, and cellular components. As of February 2005, GO contains 17,593 terms. GO terms are organized in parent-child structures called directed acyclic graphs (DAGs), which differ from hierarchies in that a 'child' (more specialized term) can have many 'parents' (less specialized terms). In GO, a child term can have one of two different relationships to its parent(s): is_a or part_of. An is_a relationship means that the term is a subclass of its parent and a part_of relationship usually means that wherever the child exists, it is as part of the parent. Figure 4-2 illustrates the structure of GO.

PAGE 77

65 Figure 4-2. Gene Ontology 4.4 Associative Nave Bayes Classifier We present a new classification method, called Associative Nave Bayes (ANB), to associate MEDLINE citations with Gene Ontology (GO) terms. We define the concept of class-support to find correlated itemsets and the concept of class-all-confidence to find interesting itemsets from correlated itemsets. In the training phase, ANB finds frequent and interesting itemsets and estimates the class prior probabilities and the probability itemsets. Once the frequent and interesting itemsets are discovered in the training phase,

PAGE 78

66 new unlabeled examples are classified by the classification algorithm by incrementally choosing the largest, most interesting itemset. Empirical test results on three MEDLINE datasets show that ANB is superior to nave Bayesian classifier. The results also show that ANB outperforms the state of the art Large Bayes (LB) classifier and is more scalable than SVM in terms of training time. The nave Bayes independence assumption states that the word probabilities for one text position are independent of the words occurring in the other positions. Although the independence assumption is clearly wrong, the nave Bayes classifier performs surprisingly well in many text classification domains (Mitchell, 1997; Lewis, 1998). It has been shown that it is possible to improve the nave Bayes classifier when the independence assumption is relaxed or violated as long as the ranks of the conditional probabilities of classes given an example are correct. To relax this assumption, we propose an Associative Nave Bayes classifier (ANB). The simple use of frequent pattern mining to document classification may cause some problems. If min_support is set to a low value, a huge number of frequent itemsets will usually be generated. Many of which are of no interest. If min_support is set to a high value, infrequent items that may be interesting to users will not be generated. This problem is called the rare item problem (Liu et al., 1999). To measure the correlation of an association, several interest measures have been proposed instead of support. For an itemset {X,Y}, lift measures how many times often X and Y occur together. The 2 metric is used to determine the dependence between X and Y. However, these two measures do not have the downward closure property (Brin et al., 1997a; Brin et al., 1997b). Recently Omiecinski (2003) introduced two interest measures,

PAGE 79

67 all-confidence and bond. Both have the downward closure property and experimental results on real data sets show that these measures are appealing since they reduce significantly the number of itemsets mined and the itemsets are coherent (Lee et al., 2003; Omiecinski, 2003). The all-confidence of a set of items X is the smallest confidence of any rule for the set of items, X. The bond of a set of items X is the ratio of the cardinality of the set of transactions containing all items in X and the cardinality of the union of transactions containing any item of X. Since LB uses min_support to find frequent itemsets, LB suffer from the rare item problem. LB also uses the inverse of lift to measure the interestness of an itemset. Since lift does not have the downward closure property and is susceptible to noise in small databases, LB cannot prune the itemset search space. Finally, the learning algorithm of LB cannot be applied to depth-first search manner association mining algorithms such as FP-growth frequent pattern mining algorithm (Han et al., 2000), since, to compute the interest of an itemset l, it needs the estimated conditional probability P(l’|ci) of all subsets l’ of l with size |l|-1 (see section 4.2.2). To overcome those problems, we adopt an interest measure, all-confidence (Omiecinski, 2003), to find interesting rare itemsets as well as frequent itemsets. 4.4.1 Definition of Class-support and Class-all-confidence In the context of document classification, a transaction database D in frequent pattern mining corresponds to a training set, where each transaction T is a training instance. To employ a frequent pattern mining method as a subroutine of feature selection for classification, each itemset should have an associated statistical significance for each class label ci. For this, we formally define the class-support of an itemset, how many times the itemset occurs given a class ci in D.

PAGE 80

68 Definition 1. The class-support of an itemset X, class_support(Xci), is the ratio of the number of transactions that contain an itemset X and are labeled with class ci in D and the total number of transactions in D. DdXcdDddcXPXsupportclassiici|||)|()(_ Class-support is the probability that a randomly chosen itemset in class ci will be the itemset X. The sum of all class-supports of an itemset X is equal to the support of an itemset X in D. iciXsupportclassXsupport)(_)( We define class-all-confidence to find frequent and interesting itemsets. We let P(X) be the power set of X, i.e. the set of all subsets of a set X. Definition 2. The class-all-confidence of an itemset X, class_all_conf(Xci), is the smallest confidence of any rule for the set of items X in class ci. That is, all rules produced from this item set would have a confidence greater than or equal to its class_all_conf value. }|{|{})((||)(__dxDddiXxxXPxxiMAXcXdXDddXconfallclassici The denominator is the maximum cardinality of transactions containing the power set of X, except the empty set and the improper subset. The maximum value in the denominator will occur when the subset of X consists of a single item. Hence, instead of computing the power set of X, 1-itemsets are considered to calculate the maximum value.

PAGE 81

69 Definition 3. The class-bond of an itemset X, class_bond(Xci), is the ratio of the count of the set of transaction in class ci containing all items in X and the count of the union transactions containing any item of X. |_()||((){})iciddDXdXcclassbondXddDxxPXxxd | Table 4-1. An example transaction database Items Transaction A B C D E Class T1 1 1 0 0 0 C1 T2 1 1 1 0 0 C1 T3 0 0 1 1 0 C1 T4 0 0 1 1 0 C1 T5 0 0 0 0 1 C1 T6 0 0 0 0 1 C2 T7 0 1 1 0 0 C2 T8 0 1 1 0 0 C2 T9 0 0 0 0 0 C2 T10 1 0 0 1 0 C2 Table 4-2. Support, class-support, all-confidence, class-all-confidence, bond, and class-bond values using the transaction database of Table 4-1 Class-support Class-all-confidence Class-bond Itemset Support C1 C2 All-confidence C1 C2 Bond C1 C2 A 3/10 2/10 1/10 1 2/3 1/3 1 2/3 1/3 B 4/10 2/10 2/10 1 1/2 1/2 1 1/2 1/2 C 5/10 3/10 2/10 1 3/5 2/5 1 3/5 2/5 D 3/10 2/10 1/10 1 2/3 1/3 1 2/3 1/3 E 2/10 1/10 1/10 1 1/2 1/2 1 1/2 1/2 AB 1/5 1/5 0 1/2 1/2 0 2/5 2/5 0 AC 1/10 1/10 0 1/5 1/5 0 1/6 1/6 0 AD 1/10 0 1/10 1/3 0 1/3 1/5 1/5 0 BC 3/10 1/10 2/10 3/5 1/5 2/5 1/2 1/3 2/3 CD 1/5 1/5 0 2/5 2/5 0 1/3 1/3 0 ABC 1/10 1/10 0 1/5 1/5 0 1/7 1/7 0 Table 4-1 shows an example transaction database preprocessed, which contains 10 transactions with their corresponding items and class labels. The support, class-support,

PAGE 82

70 all-confidence, class-all-confidence, bond, and class-bond values for all itemsets with nonzero supports using the transaction database of Table 4-1 are shown Table 4-2. 4.4.2 ANB Learning Algorithm Associative Nave Bayes classifier (ANB) consists of two parts: a frequent and interesting itemset generator (ANB-IG) based on the Eclat algorithm (Zaki et al., 1997; Zaki, 2000) and a classifier (ANB-CL). Eclat carries out a bottom-up search on the subset lattice and determines the support of itemsets by intersecting transaction lists. It is one of the best frequent pattern mining algorithms in terms of the time and space complexity (i.e., the maximum run time and memory required by an algorithm) (Zaki et al., 1997). The main task of ANB-IG, shown in Figure 4-3, is to generate the frequent and interesting itemsets that have both class-support above min_class_support and class-all-confidence above min_class_all_conf. The input to the algorithm is a set of frequent and interesting k-itemsets. Initially, k is set to 1 and we assume that all 1-itemsets are interesting. Each itemset is associated with a list of all the transaction identifiers (tidlist) that contain the itemset. This kind of data layout is called the vertical (or inverted) data layout. Each itemset also has a set of class counters to compute class-supports for each class ci. The algorithm outputs the frequent and interesting itemsets F with their class-supports. Input: the set of frequent and interesting itemsets of length i Output: the Frequent and interesting itemsets F with their class-supports ANB-IG(Li) Definition: Li = the set of frequent and interesting i-itemsets li = an individual itemset contained in Li (i.e. Li = l1l2l3ln) 1) L1 = {frequent and interesting 1-itemsets with their tidlists and class counts}; 2) for all itemsets li Li do begin 3) Ti = {}; 4) for all itemsets lj Li, with j > i do begin

PAGE 83

71 5) r = li lj; 6) r.tidlist = li.tidlist lj.tidlist; 7) for all transactions t r do begin 8) i = class of r; 9) r.class_counti++; 10) end 11) for all classes ci C do begin 12) class_support(rci) = r.class_counti / |D|; 13) class_all_conf(rci) = r.class_counti / MAX(L1.counti[r[1]], L2.counti[r[2]],, Lk.counti[r[k]]); 14) if class_support(rci) min_class_support and class_all_conf(rci) min_class_all_conf then 15) Ti = Ti {r}; F|r| = F|r| {r}; 16) end 18) end 19) end 20) for all Ti {} do ANB-IG(Ti); 21) return F with their class-supports; Figure 4-3. ANB learning algorithm The algorithm works as follows. Candidate itemsets are generated in line 5 and their tidlists are produced by intersecting the tidlists of for all distinct pairs of li and lj in line 6. In line 7-10, the class counts of the candidate itemset are counted. The for loop in lines 11-16 selects only those itemsets that are frequent and interesting from the candidate itemsets. In line 19, the algorithm is called recursively with those itemsets found to be frequent and interesting at the current level until all frequent and interesting itemsets have been enumerated. 4.4.3 ANB Classification Algorithm After finding the frequent and interesting itemsets, new unlabeled documents are classified by the algorithm ANB-CL. An item set of a new document X={l1, l2, , ln} is used to generate a set of n discrete itemsets {|()}LllFlPX . The generated itemsets are then used to compose a product approximation. The product approximation is defined to be an approximation to a higher order distribution made up of a product of several of its lower order component distributions, such that the product is an extension of the lower order distribution (Lewis, 1959). For instance, given X={a, b, c} and F={

PAGE 84

72 {a}, {b}, {c}, {d}, {a, b}, {b, d}, {a, b, d} }, we have L={ {a}, {b}, {c}, {a, b} }. All of the product approximations P(X)=P(a, b, c) to L are listed as follows: 1. P(a)P(b)P(c) 2. P(a, b)P(c) 3. P(a)P(c)P(b|a) 4. P(b)P(c)P(a|b) A product approximation P(a, b)P(c)P(a), however, is not valid because each new itemset should contain at least one item not contained in the previous itemsets (i.e., the third itemset is already covered by the first itemset). We denote a set of the already discovered items as C, where C is a subset of X. To quickly approximate a probability distribution, a selection algorithm, selectUncoveredItemset shown in Figure 4-4, is used to incrementally choose a locally optimal itemset from L until all items are discovered. Given a class label ci, a set of the discovered items C, and a set of candidate itemsets L, the selection algorithm, called an itemset selection method A (ISM-A), finds an itemset satisfying the following conditions in the given order of priority: **1. 'argmax {| for all }lLLlClLlClClL '****2. ''argmax {'| for all '}lLLllLlllL ''******3. '''argmax(,) {''|(,)(,) for all ''}ilLiiLPlclLPlcPlclL A set of itemsets L' that includes the largest number of uncovered items is selected by the condition 1 and L' should not be an empty set. If the set L' has more than one

PAGE 85

73 member, then a set of the longest itemsets L is selected from L' by the condition 2. Among the members of the set L', given class ci, a set of itemsets L with the highest class support is then found by the condition 3. A randomly chosen itemset from L is used to incrementally build a product approximation. For the above example, in our algorithm, the longest itemset {a, b} is selected first, and then the itemset {c} is selected. '' ' ''' ''' A product approximation, like any other function, can be approximated by a number of different procedures (Chow and Liu, 1968). Another selection property we consider is choosing the itemsets containing the smallest, uncovered items with the highest class support from L to add as many as itemsets to the product approximation. This is equivalent to maximizing the number itemsets used for minimizing the conditional independence assumptions (Meretakis and Wuthrich, 1999). This selection method, called an itemset selection method B (ISM-B), is used by LB and is defined as follows: **1. 'argmin {| for all }lLLlClLlClClL '****2. ''argmax {'| for all '}lLLllLlllL ''******3. '''argmax(,) {''|(,)(,) for all ''}ilLiiLPlclLPlcPlclL We compare those two itemset selection methods in section 4.5.3 in terms of the classification time and the classification performance. By default, we use the first selection method. Figure 4-4 shows the pseudo code of ANB-CL. Input: The discovered itemset F and a test instance X

PAGE 86

74 Output: the predicted class of X ANB-CL(F, X) 1) C = ; 2) denominator = ; 3) nominator = ; 4) L = {|; ()}llFlPd 5) while (C X) do begin 6) li = selectUncoveredItemset(C, L, ci); 7) denominator = denominator {l}; 8) nominator = nominator {l C}; 9) C = C l; 10) L = L – l; 11) end 12) return the class c, where atordenoliatornoliiCcclpclPcPcimin'min),'(),()(maxarg Figure 4-4. ANB Classification Algorithm Line 4 selects the discrete candidate itemsets from X by selecting only those itemsets that belong to F from the power set of X. The while loop in line 5-11 incrementally constructs the product approximation of X by adding one itemset at a time until no more itemsets can be added. The procedure selectUncoveredItemset in line 6 selects the longest, undiscovered itemset l. 4.5 Experiments In this section, we report our experimental results on the performance of Associative Nave Bayes (ANB) in comparison with Nave Bayes (NB), Large Bayes (LB), and Support Vector Machine (Joachims, 1999; Joachims, 2001). Meretakis et al. (1999) showed that LB outperforms several state-of-the-art algorithms such as NB, C4.5 (Quinlan, 1993), Tree-Augmented Nave Bayes (Friedman, 1997), and Classification Based on Association (Liu et al., 1998). The experimental results show that ANB outperforms both NB and LB and, in terms of training time and classification accuracy. ANB is efficient and scalable on large dataset compared to LB and SVM. Experiments

PAGE 87

75 were performed on a 2.8GHz Pentium IV PC with 1GB of memory in Linux environment. Algorithms were coded with GNU C/C++. 4.5.1 Real World Datasets Table 4-3. Description of datasets Dataset GO Term GO ID #Train #Test Acrosome GO:0001137 824 354 Endosome GO:0003596 942 405 Fermentation GO:0003973 844 363 Glial fibrillary acidic protein GO:0013861 913 393 Intermediate filament GO:0003464 854 368 Nuclear membrane GO:0003460 914 393 Oogenesis GO:0007709 910 391 Phosphorylation GO:0003982 847 365 Proliferating cell nuclear antigen GO:0001690 917 395 Small Spetrin GO:0005892 834 359 Ciliary or flagellar motility GO:0000944 7,229 3,100 DNA replication GO:0000660 8,534 3,659 Endocytosis GO:0004992 8,043 3,449 Locomotion GO:0012740 7,303 3,132 Menstrual cycle GO:0000949 7,465 3,210 Nucleotide-excision repair GO:0000635 8,212 3,521 Peptide cross-linking via an oxazole or thiazole GO:0009882 7,627 3,261 Sex chromosome GO:0000748 7,106 3,047 Synaptic transmission GO:0000894 7,684 3,295 Medium Wound healing GO:0001321 6,995 2,999 Cell differentiation GO:0000885 16,354 7011 Collagen GO:0003403 14,579 6,250 Cytochrome GO:0003314 16,226 6,955 Drug resistance GO:0007018 15,620 6,696 Homeostasis GO:0001116 14,919 6,396 Intercellular junction GO:0003722 17,893 7,670 Memory GO:0005781 15,085 6,466 Synapsis GO:0005240 15,113 6,479 Tissue regeneration GO:0012976 17,542 7,519 Large Tumor antigen GO:0005995 16,871 7,232 In this study, we constructed three kinds of datasets (small, medium, and large datasets) to evaluate the performance of ANB algorithm. We used the holdout method

PAGE 88

76 (Devroye, 1996) to randomly divide each dataset into two parts: a training set and a test set. We randomly assigned a 70% of data to the training set and the remaining 30% of each data to the test set for each dataset. Table 4-3 lists the detailed information on the data sets used in this experiment and each data set contains 10 classes. We investigated how many documents are contained in multiple relevant classes in our datasets. Table 4-4 lists the number of citations in each dataset and the number of documents with N classes (1 <= N <= 4). Table 4-4. Number of citations with N classes (1 <= N <= 4) #Citations with N classes Dataset 1 2 3 4 #Citations Training 8,741 29 0 0 8,799 Small Test 3,750 18 0 0 3,786 Training 74,988 605 0 0 79,198 Medium Test 31,961 356 0 0 32,673 Training 129,364 14,780 418 6 160,202 Large Test 55,634 6,234 188 2 68,674 To construct training and test data, we first surveyed how many GO terms are contained in MEDLINE citations. For each GO term, we made a query statement, limiting the results to the Medical Subject Heading (MeSH) major topic field and to citations with abstracts in English (National Library of Medicine). For example, we formulated a query statement “cell adhesion”[MeSH Major Topic] AND has abstract[text] AND English[Lang] to retrieve the MEDLINE citation that contains the term “cell adhesion.” After submitting all query statements to PubMed, we found that a total number of 564 out of 17,593 GO terms found in MEDLINE citations. Table 4-5 lists the top 20 most frequently occurring GO terms in MEDLINE. We list all GO terms found in MEDLINE citations in appendix.

PAGE 89

77 Table 4-5. Top 20 GO terms GO ID GO Term #Citations GO:0001469 Protein 1,383,991 GO:0000417 Cell 762,860 GO:0004433 Peptide or protein amino-terminal blocking 681,065 GO:0004443 Peptide or protein carboxyl-terminal blocking 681,065 GO:0005741 Physiological process 239,942 GO:0003091 Plasma protein 201,665 GO:0003382 Nucleic acid 198,735 GO:0001124 Behavior 193,769 GO:0004886 Cell growth and/or maintenance 162,424 GO:0003067 Peptide hormone 154,067 GO:0000007 Reproduction 140,163 GO:0005779 Learning and/or memory 125,908 GO:0009012 Cell surface antigen 120,105 GO:0003395 DNA 118,704 GO:0008758 Vitamin or cofactor transport 116,905 GO:0000153 Vitamin or cofactor transporter activity 116,905 GO:0007524 Embryonic development 105,202 GO:0000246 Cytoplasm 103,436 GO:0000209 Intracellular 103,436 GO:0008214 Channel or pore class transporter activity 95,821 4.5.2 Preprocessing and Feature Selection Preprocessing Before performing classification, each document should be represented as a feature vector. For this, we use a bag-of-words representation. The bag-of-words representation is the most popular approach for representing text documents in classification. In the bag-of-words approach, all the distinct words, called indexing terms, constituting the documents are encoded as feature vectors of fixed dimensionality, where each word is represented as a vector and each value of the vector represents a term weight in a document. There are two term weighting methods, a binary indexing method and a weighted indexing method. In the binary indexing scheme, a term that occurs in a document is assigned a weight of 1, while a term that does not occurs in the document is

PAGE 90

78 assigned a weight of 0. In the weighted indexing approach, a term weight may be used to reflect term importance by using the term frequency TF(w, d), the number of times the term w occurs in document d. To weight terms in the document collection, the binary indexing scheme is used for both ANB and LB, while the weighted indexing scheme is used for SVM. Each document in the collection contains only the title and abstract from MEDLINE bibliographic database. The reason why we do not use the full texts of the documents is that the use of additional utilization of the full texts of the documents appears to produce very little improvement over titles and abstracts alone in most subject areas (Salton and Lesk, 1968). A substantial advantage of word-based representation is its simplicity and efficiency, although it ignores syntactic and semantic structures of document. However more complex document representations such as multi-word level, semantic level representations have not yet shown consistent and substantial improvements (Joachims, 2001, page 14). While more expressive representations can capture more of the meaning of the document, their increased complexity degrades the quality the quality of statistical models based on them. After preprocessing, each document di is represented as a document vector 12,,...,iiiidwww m where each wij is the term weight that is assigned to the jth term for document di. Thus, a collection of documents is represented as an m n term-document matrix, where each of the m distinct terms in the collection is assigned to a column in the matrix and each of the n documents in the collection is assigned to a row in the matrix. The preprocessing steps are described in Figure 4-5.

PAGE 91

79 Input: a set of documents D Output: an mn term-document matrix M Preprocessing (D) Definition: V: a set of indexing terms 1) for all documents d D do begin 2) tokenize a stream of characters in document d into a set of tokens; 3) identify all the unique terms C in document d from the set of tokens; 4) remove stop words from the set of terms C; 5) for all indexing terms w C do begin 6) reduce a term w to word stem form w’; 7) compute a term frequency TF(w’, d) in document d; 8) if w’ V then 9) V = V {w’}; 10) end 11) end 12) construct a matrix M using V and D; 13) return M; Figure 4-5. Preprocessing steps In line 2, a tokenizer converts a stream of characters in the document into a set of tokens, mapping all words in the document into lowercase, separating punctuations, and expanding verb contractions. After the identification of all the unique terms in the document in line 3, stop words are removed from the set of identified terms. Stop words are the high-frequency, insignificant words that appear in the document (i.e., pronouns, prepositions, conjunctions, and so on) but need to be eliminated. We use the stop word list developed by Gerard Salton and Chris Buckley for the SMART information retrieval system at Cornell University and it consists of 571 stop words. A sample list of stop words is listed in Figure 4-6. In line 5, each indexing term is reduced to its root form. Stemming is a process for removing the common morphological and inflectional endings from words in English (Porter, 1980) and the Porter stemming algorithm is used. This projects various words like “computation,” “computer,” “computability,” “computing,” and “compute” onto the same root form “comput.” The Porter stemming algorithm is

PAGE 92

80 available at http://www.tartarus.org/~martin/PorterStemmer/ maintained by the author. A term frequency TF(w, d) of term w in document d is calculated in line 6 and if the term w is not in the vocabulary V, then it is added into V. Finally, a term-document matrix is constructed in line 12. a a's able about above according accordingly across actually after afterwards again against ain't all allow allows almost alone along already also although always am among amongst an and another Figure 4-6. A sample list of stop words After the stemming and stop word removal, for the training datasets, we obtained a vocabulary of 95,622 out of 109,150 unique words for small dataset, a vocabulary of 469,827 out of 532,941 distinct words for medium dataset, a vocabulary of 739,128 out of 841,608 unique words for large dataset, respectively. For the test datasets, we identified a vocabulary of 47,436 out of 54,738 unique words for small dataset, a vocabulary of 217,872 out of 247,490 distinct words for medium dataset, a vocabulary of 357,953 out of 406,618 unique words for large dataset, respectively. Feature Selection In general, the bag-of-words approach led to feature vectors that are large and sparse. A major difficulty of text classification problems is the high dimensionality of the feature space (Yang and Pedersen, 1997). To remove non-informative terms from the

PAGE 93

81 feature space, feature selection methods such as information gain (Cover and Thomas, 1991), a Chi-square (2) test (Mood et al., 1974), document frequency thresholding (Yang and Pedersen, 1997), mutual information (Cover and Thomas, 1991), and odds ratio (Van Rijsbergen et al., 1981) are used. Among them, Yang and Pedersen (1997) found that information gain and a 2 test are most effective in aggressive term removal without losing classification accuracy in their experiments. Therefore, we employ information gain as a feature selection method. Information gain is one of popular feature selection methods in classification (Joachims, 2001; Yang and Pedersen, 1997). The information gain of term w measures the entropy difference between the number of bits of information necessary to encode class ci and the number of bits of information that term w constitutes to encoding class ci independent of the other terms in the document. The information gain of term w is defined as 111()()(|) =()log()()(|)log(|) ()(|)log(|)iinniiiiiiniiiIGwEntropycEntropycwPcPcPwPcwPcwPwPctPct The terms with uniform distributions have high entropy and thus those n terms with the highest entropy are selected as features. Information gain is also called average mutual information (Fano, 1961). Given the term-document matrix M, we computed the information gain of each distinct term and selected a total of 200, 600, and 1000 terms (features) that have the highest average mutual information with the class variable as features for each dataset.

PAGE 94

82 Table 4-6 lists the top 10 words with their entropy selected by information gain for each dataset. Table 4-6. Top 10 words for small, medium, and large datasets Small Dataset Medium Dataset Large Dataset Rank Entropy Word Entropy Word Entropy Word 1 0.31324 nuclear 0.32877 dna 0.22711 collagen 2 0.29374 spectrin 0.25030 repair 0.21443 resist 3 0.28998 endosom 0.23136 ciliari 0.18350 synapt 4 0.28319 acrosom 0.20678 heal 0.18138 memori 5 0.27749 pcna 0.19123 replic 0.15077 cytochrom 6 0.27395 ferment 0.15801 chromosom 0.14654 cyp 7 0.25640 gfap 0.14780 wound 0.14610 antigen 8 0.25164 glial 0.11327 cycl 0.11062 synaps 9 0.22813 fibrillari 0.09983 synapt 0.08886 heal 10 0.20252 sperm 0.09889 gene 0.08851 neuron 4.5.3 Experiments Performance Measures For evaluating the performance of ANB, we use the standard precision, recall, and F1 measure. Precision (p) is the ratio of correct predictions by the classification system divided by the total number of the system’s predictions. Recall (r) is defined to be the ratio of correct predictions by the classification system divided by the total number of correct predictions. Precision and recall are defined as: TPpTPFP TPrTPFN where |TP| (True Positives) is the number of correct predictions to class ci, |FP| (False Positives) is the number of incorrect predictions to class ci, and |FN| (False Negatives) is the number of incorrect, rejected predictions from class ci. Assuming a binary classification, four classification outcomes are possible as shown in Table 4-7.

PAGE 95

83 Table 4-7. Confusion matrix Predicted class Yes No Yes True Positives False Negatives Actual class No False Positives True Negatives The F1 measure combines recall and precision into an equally weighted single measure as follows: 12(,)2 2rpFrprpTPTPFPFN Experimental Setup To make the evaluation results of LB comparable to the results of ANB, we implemented LB using the Eclat (Zaki et al., 1997; Zaki, 2000) instead of using the Apriori (Agrawal et al., 1993; Agrawal and Srikant, 1994), since the training time of LB is dependent on the run time of Apriori and Apriori-based itemset mining methods scale up exponentially in the worst case with the number of features (Meretakis et al., 2000). In all experiments, the parameters of LB are set to the standard values as suggested by Meretakis et al. (2000). The minimum support was set to 1% and the minimum interestness was set to 0.04. For ANB, we fix the minimum class-support to 0.5% and the minimum class all confidence to 20%, respectively. For SVM, we chose a linear SVM due to its popularity and fast training time compared to non-linear SVMs (e.g. polynomial, radial basis function, or sigmoid SVMs) in text classification (Joachims, 2001; Yang et al., 2003). It is important to select a good value of C, the amount of training error tolerated by the SVM, for the linear SVM. Among the possible values of C{0.05, 0.1, 0.5, 1.0, 5, 10, 1000}, we chose C=5, since the linear SVM with C=5 performed best on our datasets in terms of classification accuracy. We used the

PAGE 96

84 SVMmulticlass package developed by Joachims (http://svmlight.joachims.org/), which is an implementation of the multi-class SVM. Results Table 4-8 summarized the performance scores, precision (p), recall (r), and F1 measures of four algorithms on three datasets for vocabulary sizes of 200, 600, and 1000 words. For small dataset, ANB is comparable to SVM and is superior to other classifiers. SVM significantly outperforms NB, LB, and ANB for medium and large datasets. In all cases, ANB is superior to both NB and LB. Table 4-8. Performance summary of classifiers #Words: 200 #Words: 600 #Words: 1000 Dataset p r F1 p r F1 p r F1 NB 89.27 89.51 89.39 91.75 91.90 91.83 91.70 91.83 91.77 LB 92.54 92.79 92.66 93.20 93.19 93.11 92.26 92.38 92.32 ANB 93.24 93.39 93.31 93.85 93.96 93.90 93.32 93.39 93.35 Small SVM 93.65 93.78 93.71 94.08 94.22 94.15 94.03 94.16 94.09 NB 80.83 81.24 81.03 85.52 86.16 85.54 87.28 87.97 87.76 LB 85.07 85.75 85.41 86.95 87.68 87.31 88.09 88.80 88.44 ANB 85.33 85.93 85.63 87.21 87.89 87.55 88.46 89.14 88.80 Medium SVM 89.84 90.15 89.99 91.39 91.62 91.51 91.63 91.86 91.74 NB 71.75 73.16 72.45 76.58 78.40 77.48 78.37 79.80 79.08 LB 75.27 79.70 77.42 77.89 82.14 79.96 78.87 83.06 80.83 ANB 75.66 79.75 77.65 78.63 82.60 80.57 79.48 83.72 81.55 Large SVM 81.73 81.74 81.74 83.62 83.70 83.66 83.90 83.97 83.93 Table 4-9 shows the total number of itemsets mined by LB and ANB on three datasets for vocabulary sizes of 200, 600, and 1000 words. We observe that as the size of vocabulary increases, the number of itemsets mined by LB increases significantly, while the number of itemsets mined by ANB increases slightly. This is because LB uses support to find frequent itemsets. LB also uses the inverse of lift that does not have the downward closure property and thus it cannot prune the search space from the early stage of training.

PAGE 97

85 Table 4-9. Number of k-itemsets (k >= 2) mined by LB and ANB algorithms Dataset Classifier #Words: 200 #Words: 600 #Words: 1000 LB 2,812 8,543 12,555 Small ANB 391 442 445 LB 2,364 7,419 9,606 Medium ANB 200 220 221 LB 2,252 6,445 9,259 Large ANB 84 101 101 Table 4-10 shows the time to train and classify the algorithms for each dataset. The learning time of LB and ANB depends on the association mining algorithm and user-defined threshold values. Although LB and ANB use the same association mining algorithm in our experiment, ANB is faster than LB at the learning phase since ANB uses class-support and class-all-confidence and they both have the downward closure property. The classification time of LB and ANB mainly depends on the number of itemsets mined at the learning phase. In addition, LB tries to maximize the number of itemsets used in product approximation by choosing the smallest, uncovered items incrementally, while ANB tries to minimize the number of itemsets used in product approximation by selecting those itemsets that have the largest uncovered items in it. Due to those factors, training time and classification time of LB are slow compared to those of ANB. The training time of linear SVM tends to increase dramatically with an increasing training set size and feature set size, although a linear SVM can be trained much faster than a nonlinear SVM (Joachims, 2001). Thus, SVM would not be a feasible solution in such an environment where retraining is frequently required with an increasing size of training documents.

PAGE 98

86 Table 4-10. Train and test times of classifiers (Time: CPU Seconds) #Words: 200 #Words: 600 #Words: 1000 Dataset Classifier Train Test Train Test Train Test NB 0.11 0.49 0.22 0.94 0.28 1.26 LB 1.64 19.83 24.26 90.22 68.87 179.58 ANB 0.14 4.42 0.29 13.63 0.4 23.35 Small SVM 29.01 0.02 33.16 0.04 35.90 0.03 NB 0.85 3.57 1.66 6.97 2.13 9.02 LB 3.27 123.26 22.46 547.84 35.81 905.99 ANB 0.95 24.32 2.17 88.38 2.78 142.2 Medium SVM 4818.69 0.16 5115.62 0.28 5571.04 0.32 NB 1.69 7.18 3.35 14.53 4.4 19.35 LB 6.74 237.36 35.28 1074.5 65.05 2020.09 ANB 1.98 53.16 4.15 177.91 5.28 300.8 Large SVM 45776.46 0.35 50460.76 0.45 50344.57 0.69 We introduced two itemset selection methods, ISM-A and ISM-B, for building a production approximation in section 4.4.3. Table 4-11 shows the test time and precision of ANB with ISM-B. ANB with ISM-A consistently shows better performances those of ANB with ISM-B. This demonstrates that selecting those itemsets that contain the longest, uncovered items is better than selecting those itemsets that contains the smallest, uncovered items. Table 4-11. Test time and precision of ANB with ISM-B (Time: CPU Seconds) Dataset Performance measure #Words: 200 #Words: 600 #Words: 1000 Classification time 4.42 16.1 23.48 Precision 93.13 93.66 93.31 Recall 93.28 93.78 93.39 Small F1 93.20 93.72 93.35 Classification time 25.81 101.99 142.01 Precision 85.29 87.15 88.46 Recall 85.89 87.82 89.14 Medium F1 85.58 87.49 88.80 Classification time 57.45 199.27 301 Precision 75.58 78.55 79.48 Recall 79.71 82.43 83.72 Large F1 77.59 80.44 81.55

PAGE 99

87 4.6 Summary and Future Research 4.6.1 Summary To associate MEDLINE citations with GO terms, we have presented a new classification method. We have also proposed interest measures, class-support, class-all-confidence, class-bond, to find frequent and interesting itemsets for text classification. For a large amount of text dataset, it seems that ANB classifier may be a good choice although SVM had been applied to text data and it was reported to be successful (Joachims, 2001). Our empirical research indicates that ANB outperforms both NB and LB in all cases and, for small dataset, ANB is comparable to SVM. Although SVM is superior to other algorithms, SVM may not be a feasible solution to use more than tens of thousand training instances, because the large and uncertain memory requirements of this algorithm are coupled with the super-linear dependence of its execution time on the number of training instances (Dumais et al., 1998; Youn, 2004). Another issue is simplicity of implementation and ease of use. SVM is a complex algorithm and needs to tune parameters to get a best performance. 4.6.2 Future Research Since the GO database contains a large number of categories, it would be a good approach that compress the GO database into a manageable size using the graph-structured data mining method and then classify the MEDLINE citations with the compressed GO terms. The approaches to graph-based data mining are categorized into five groups, greedy search based approaches, inductive logic programming (ILP) based approaches, inductive database based approaches, mathematical graph theory based approaches, and kernel function based approaches (Washio and Motoda, 2003).

PAGE 100

88 To check how complex the GO is, we analyzed the GO database released in May 2004. We indexed a total of 5,541 unique synonyms of GO terms and a total of 15,121 unique noun phrases from the GO database. After representing GO as a labeled, connected graph, we found a total of 40,008 vertices, a total of 112,642 edges, and a total of 39,206 unique labels. Figure 4-7 illustrates the graph representation of a GO term used in this analysis. Table 4-12 shows the total numbers of vertices, edges, and unique noun phrases mined for three organizing principles of GO, molecular function, biological process and cellular component. GO Term keywordsnamesynonymsTerm type GO Term GO TermIs_apart_of Figure 4-7. Graph representation of a GO term Table 4-12. Total numbers of vertices, edges, and unique noun phrases mine for molecular function, biological process and cellular component Ontology #Vertices #Edges #Noun phrases Biological process 21,557 60,431 21,177 Cellular component 5,601 10,119 5,359 Molecular function 15,996 42,089 15,941 All ontologies 40,008 112,642 39,206 Our future research is to develop a graph structured data mining method that finds frequent, interesting substructures from the GO. Once discovered, the substructure can be

PAGE 101

89 used to compress the GO data by replacing instances of the substructure with a pointer to the discovered substructure.

PAGE 102

CHAPTER 5 DATA MINING OF MEDLINE DATABASE 5.1 Introduction MEDLINE, developed by the U.S. National Library of Medicine (NLM), is a database of indexed bibliographic citations and abstracts. It contains over 4,600 biomedical journals. MEDLINE citations and abstracts are searchable via PubMed (http://www.ncbi.nlm.nih.gov/entrez/query.fcgi) or the NLM Gateway (http://gateway.nlm.nih.gov/gw/Cmd). The NLM produces the MeSH (Medical Subject Headings) for the purposes of subject indexing, cataloging and searching journal articles in MEDLINE with an annual update cycle. MeSH consists of descriptors (or main headings), qualifiers (or subheadings), and supplementary concept records. It contains more than 19,000 descriptors which are used to describe the subject topic of an article. It also provides less than 100 qualifiers which are used to express a certain aspect of the concept represented by the descriptor. MeSH terms are arranged both alphabetically and in a hierarchical tree, in which specific subject categories are arranged beneath broader terms. MeSH terms provide a consistent way of retrieving information regardless of different terminology used by the authors in the original articles. By using MeSH terms, the user is able to narrow the search space in MEDLINE. As a result, by adding more MeSH terms to the query, retrieval performance may be improved (French et al., 2001). Figure 5-1 shows MeSH tree structures. 90

PAGE 103

91 Figure 5-1. MeSH tree structures However, there are inherent challenges, as well. There may be information overload (Pratt and Fagan, 2000), and users may be unable to express their information needs, in order to take fully advantage of the MEDLINE database. MEDLINE contains over 12 million article citations. Beginning in 2002, it began to add over 2,000 new references on a daily basis. Although the user may be able to limit the search space of MEDLINE with MeSH terms, keyword searches often result in a long list of results. For instance, when the user queries the term “Parkinson’s Disease” by limiting it to the MeSH descriptors, PubMed returns over 21,000 results. Here, there is a problem of information overload, with the user having difficulty finding relevant information. The inability of users to express information needs may become more serious, unless users have a precise knowledge in their area of interest, or an understanding of

PAGE 104

92 MeSH and its structure. The use of common abbreviations, technical terms, and synonyms in biomedical articles prevents users from articulating their information needs accurately. To avoid the vocabulary problem, MeSH may be used. However, it is difficult for an unfamiliar user to locate appropriate descriptors and/or qualifiers, since MeSH is a very complex thesaurus. Furthermore, new terms are added, some are modified, and others are removed each year as biomedical fields change. An imprecise query usually results in a long list of irrelevant hits. Under such circumstances, a better mechanism is needed to organize information in order to help users explore within an organized information space (Chen, 1998). In order to arrange the contents in a useful way, text categorization and text clustering have been researched extensively. Text categorization is a boiling down of the specific content of a document into a set of one or more pre-defined labels. Text clustering can group similar documents into a set of clusters based on shared features among subsets of the documents (Kohonen, 1998). In this chapter, we present a text data mining method that uses both text categorization and text clustering for building a concept hierarchy for MEDLINE citations. The approach we propose is a three-step data mining process for organizing MEDLINE database: (1) categorizations according to MeSH terms, MeSH major topics, and the co-occurrence frequency of MeSH descriptors, (2) clustering using the results of MeSH term categorization, and (3) visualization of categories and hierarchical clusters. The hierarchies automatically generated may be used to support users in browsing behavior as well as help them identify good starting points for searching. An interface for this underlying system is also presented.

PAGE 105

93 5.2 Data Mining Method for Organizing MEDLINE Database In this section, we will explain the data mining method proposed in detail. We use MySQL to store MEDLINE citations and additional data that are generated by the data mining process. 5.2.1 The Data For the following experiment, we extracted a total of 1,736 citations encoded in XML (eXtensible Markup Language) from the query “Secondary Parkinson Disease”, limiting the results to the MeSH major topic field and to citations with abstracts in MEDLINE. Figure 5-2 shows an example of MEDLINE data. UI -22378477PMID-12491210...TI -Recurrent Kawasaki disease-like syndrome in a patient with acquiredimmunodeficiency syndrome.AB -A review of Kawasaki disease (KD)-like syndromes (KDLS) in patients withhuman immunodeficiency virus (HIV) raised the question whether vasculitisin children and KDLS in immunocompromised adults might be etiologicallyrelated. We describe a 42-year-old white man with AIDS and Kaposi sarcomawho presented with KDLS, which was diagnosed on the basis of clinicalcriteria for KD......MH -AIDS-Related Opportunistic Infections/immunology/*metabolismMH -Acquired Immunodeficiency Syndrome/*immunologyMH -AdultMH -Case ReportMH -HumanMH -MaleMH -Mucocutaneous Lymph Node Syndrome/immunology/*metabolismMH -Plasmapheresis Figure 5-2. An example of MEDLINE record 5.2.2 Text Categorization Categorization refers to an algorithm or procedure which results in the assignment of categories to documents (Hearst, 1999). We chose the MeSH major topic, the MeSH descriptor and qualifier, and a co-occurrence frequency of MeSH descriptors as features to be used in categorization. To categorize the collection according to the selected

PAGE 106

94 features, we first parse the data collection encoded in XML using SAX (Simple API for XML). After extracting the MeSH major topics, the MeSH descriptors, and the co-occurrence frequency of MeSH descriptors for each citation, we insert the data into the corresponding MySQL tables. 5.2.3 Text Clustering using the Results of MeSH Descriptor Categorization Since many MeSH terms may be assigned to a citation and vice versa, categorization with the MeSH terms or the co-occurrence frequency of MeSH terms often results in a large list or hierarchy. Some categories may contain a large number of documents. Simply listing categories associated with documents is inadequate for organizing data (Hearst, 1999). To alleviate this problem, the approach we propose here is to cluster the results of MeSH descriptor categorization using the hierarchical self-organizing map. We choose only those MeSH descriptor categories whose document frequencies are over a predetermined threshold for clustering. Document frequency is the number of documents in which a term occurs. Terms are extracted and selected using category dependent document frequency thresholding from the categories chosen. There are two ways that document frequency is calculated: category independent term selection and category dependent term selection (Chen and Ho, 2000). In category independent term selection, document frequency of each term is computed from all the documents in the collection and the selected set of terms are used on each category. In category dependent term selection, document frequency of each term is calculated from only those documents belonging to that category. Thus, different sets of terms are used for different categories.

PAGE 107

95 After the feature extraction and selection, and the SOM clustering, a concept hierarchy is obtained, by relying on the MeSH descriptors for the top layer, and by using feature vectors extracted from the titles and abstracts for the sub-layer. 5.2.4 Feature Extraction and Selection To produce a concept hierarchy using the SOM, documents must be represented by a set of features. For this purpose, we use full-text indexing to extract a list of terms (words or phrases). The input vector is constructed by indexing the title and abstract elements of the collection. We then weight these terms using the normalized term frequency (TF) x inverse document frequency (IDF) term weighting scheme (Salton and Buckley, 1998). The preprocessing procedure is mainly divided into two stages: noun phrase extraction and term weighting. In the noun phrase extraction phase, we first fetch the MEDLINE identifier, the title and abstract elements from the collection and then tokenize the title and abstract elements based on Penn Treebank tokenization. The MEDLINE identifier is used as a document identifier. We then automatically assign part of speech tags to words reflecting their syntactic category by using the rule-based part of speech tagger (Brill, 1992; Brill, 1994). After recognizing the noun phrase chunks from the tagged text, we extract a set of noun phrases for each citation. At this stage, we remove common terms by consulting a list of 906 stop words. We computed document frequency of all terms using category dependent term selection for those MeSH descriptor categories whose document frequencies were over a predetermined threshold (in this experiment, greater than 100 times). We then eliminate terms from the feature space whose document frequencies are less than a predetermined threshold (in this experiment, less than 10 times). Finally, we weight the selected terms

PAGE 108

96 using the best fully weighted scheme, and assign corresponding term weights to each document for each category selected. Thus, the weighted term vector set can be used as the input vector set for the SOM. 5.2.5 Construction of a Concept Hierarchy We use the hierarchical SOM algorithm described in section 5.2 to construct a concept hierarchy from the data. Our algorithm is different from other SOM-variant algorithms, in that each sub-layer SOM dynamically reconstructs a new input vector from an upper-level input vector. For each MeSH descriptor category containing more than 100 documents, we generated a concept hierarchy using the SOM, limiting the maximum level of hierarchy to 3. We built a 10 x 10 SOM, and presented each input vector 100 times to the SOM. We then recursively built the sub-layer concept hierarchy by training a new 10 x 10 SOM with a new input vector, which is dynamically constructed by selecting only a document feature vector contained in the concept region from the upper-level feature vector. The concept hierarchy generated contains two kinds of information: category labels extracted from the MeSH descriptors for the top-level, and the concept hierarchy using the SOM for the sub-layer. We inserted this information into the MySQL database to build an interactive user interface. 5.2.6 Experimental Results For the results of categorization, we extracted 2,210 distinct MeSH descriptors, 70 distinct MeSH qualifiers, 269 distinct MeSH major topics, and 60,192 co-occurring MeSH descriptors from the collection. On average, each citation in the collection contains 14 MeSH descriptors, 10 MeSH qualifiers, and 4 MeSH major topics.

PAGE 109

97 For text clustering, we identified a total of 20,367 distinct terms from the collection after the stop word removal. A total of 22 categories containing more than 100 citations were identified from the results of MeSH descriptor categorization. After the category dependent document frequency thresholding, an average of 66 terms were selected per category, ranging from 14 terms for one category to 260 terms for another category. After the hierarchical SOM clustering, 193 distinct concepts were generated from 22 categories. 5.3 User Interfaces We provide four different views, three category hierarchies and one clustering hierarchy to users. We represent these hierarchy information as hierarchical trees to help users understand MeSH qualifiers and descriptors, so they could find a set of documents of interest, and locate good starting points for searching. 5.3.1 MeSH Major Topic Tree View and MeSH Term Tree View The MeSH term tree displays the categorized information space, arranged by first descriptors and then qualifiers. Figure 5-3 shows the interface of the MeSH term tree. In each level of hierarchy, MeSH terms are listed in alphabetical order, along with their document frequencies. When the user clicks on a category label that is either a descriptor or a qualifier on the left pane, the associated document set is displayed on the right pane. At this point, if the category is a descriptor, the associated qualifiers in the collection are also expanded as its children in the tree. Users can see more detailed information of a document by clicking on the title of a document that is shown on the right pane. To help users better understand the meaning of an ambiguous MeSH term, the corresponding descriptor data and context in the MeSH tree may be displayed by clicking on the link “MeSH Descriptor Data & Tree Structures” within each level of the tree.

PAGE 110

98 In some cases, the user may want to see the category arranged by only MeSH major topics. The MeSH major topic tree provides the same information as the MeSH term tree except that it shows the category hierarchy arranged by only MeSH major topics. Figure 5-3. Interface of MeSH major topic tree view 5.3.2 MeSH Co-occurrence Tree View The MeSH co-occurrence tree provides the co-occurrence relationship of MeSH descriptors, along with their co-occurrence frequency in the collection. Since an average of 14 MeSH descriptors are assigned to each citation in the collection, there are a large number of nodes in the co-occurrence tree. To better organize the co-occurrence tree, the interface allows the user to select the co-occurrence frequency range. Thus, the user can easily identify co-occurring semantic types in the collection.

PAGE 111

99 5.3.3 SOM Tree View The SOM tree is constructed for each MeSH descriptor whose document frequency is less than some predetermined threshold. Typically, 10 to 12 MeSH descriptors are assigned to each MEDLINE citation. Thus, some categories associated with a large number of citations do not characterize the information in a way that is of interest to the user (Hearst, 1999). To solve this problem, we further arrange those categories hierarchically using the SOM. In some cases, clustering seems useful in helping users filter out sets of documents that are clearly not relevant and should be ignored (Hearst, 1999). Figure 5-4 shows the interface for browsing the SOM tree. Figure 5-4. Interface of SOM tree view 5.4 Discussion We have also proposed the multi-layered self-organizing map algorithm for building a subject-specific concept hierarchy using two input vector sets constructed by

PAGE 112

100 indexing the MEDLINE citations. The proposed SOM algorithm is different from other SOM-variant algorithms. First, it uses two different input vectors to cluster MEDLINE database more meaningfully. Second, after constructing the top-level concept map and aggregating nodes with the same concept on the map into a group, it dynamically reconstructs input vector by selecting only those items that are contained for each concept region from input vector of the higher level to generate the sub-layer map. Thus, new input vector would reflect only the contents of the region and not the all collection for each SOM. The concept hierarchy generated by the SOM can be used for building an interactive concept browsing service with multiple viewpoints. 5.5 Summary We have proposed a three-step data mining process for organizing MEDLINE database: (1) categorizations according to MeSH terms, MeSH major topics, and the co-occurrence of MeSH descriptors; (2) clustering using the results of MeSH term categorization; and (3) visualization of categories and hierarchical clusters.

PAGE 113

CHAPTER 6 CONCLUSIONS We have proposed the integrated DL system architecture with OAI-PMH and self-organizing capabilities. The proposed DL system can serve both as a data provider and service provider. To provide the users with powerful methods for organizing, exploring, and searching collection of harvested metadata, we combined cross-archive searching with interactive concept browsing services. We have also proposed the multi-layered self-organizing map algorithm for building a subject-specific concept hierarchy using two input vector sets constructed by indexing the harvested metadata collection. The generated concept hierarchy can be used for two purposes: building an interactive concept browsing service and automatic classification of harvested metadata for the purpose of selective harvesting. Our approach of harvesting metadata from various OAI registered archives, providing cross-archive search, concept browsing, and collection summarizing services, presents an integrated and centralized server beneficial to the learning community and other OAI service providers. Our research will be useful to promote accessibility, reusability and interoperability of metadata among the wide user community. We have proposed a new classification algorithm, called Associative Nave Bayes (ANB) classifier. We formally define two interest measures, called class-support and class-all-confidence, to find frequent and interesting itemsets for text classification. Our experimental results performed on real data sets show that ANB outperforms the nave Bayes (NB) classifier and the Large Bayes (LB) classifier. For a large amount of text 101

PAGE 114

102 dataset, it seems that ANB classifier may be a good choice because of its scalability and accuracy. Finally, we have proposed a three-step data mining process for organizing MEDLINE database: (1) categorizations according to MeSH terms, MeSH major topics, and the co-occurrence of MeSH descriptors; (2) clustering using the results of MeSH term categorization; and (3) visualization of categories and hierarchical clusters. The proposed SOM algorithm is different from other SOM-variant algorithms. First, it uses the results of categorization. Second, after constructing the top-level concept map and aggregating nodes with the same concept on the map into a group, it dynamically reconstructs input vector by selecting only terms that are contained for each concept region from the input vector of the higher level and re-computing their weights to generate the sub-layer map.

PAGE 115

APPENDIX GENE ONTOLOGY TERMS DISCOVERED IN MEDLINE CITATIONS GO ID GO Term #Citations 1469 protein 1383991 417 cell 762860 4433 peptide or protein amino-terminal blocking 681065 4443 peptide or protein carboxyl-terminal blocking 681065 5741 physiological process 239942 3091 plasma protein 201665 3382 nucleic acid 198735 1124 behavior 193769 4886 cell growth and/or maintenance 162424 3067 peptide hormone 154067 7 reproduction 140163 5779 learning and/or memory 125908 9012 cell surface antigen 120105 3395 DNA 118704 8758 vitamin or cofactor transport 116905 153 vitamin or cofactor transporter activity 116905 7524 embryonic development 105202 246 cytoplasm 103436 209 intracellular 103436 8214 channel or pore class transporter activity 95821 3296 small-molecule carrier or transporter 91330 7740 red or far red light signaling pathway 90896 7610 red or far-red light photoreceptor activity 90896 3383 RNA 81921 4270 catabolism 78640 977 metabolism 78640 359 chromosome 63704 238 plasma membrane 58756 5780 learning 56619 5762 sensory perception 53290 15300 habituation 51596 9274 insertion or deletion editing 51339 261 signal transduction 48670 5729 aging 48133 3454 membrane 44818 5765 visual perception 44008 103

PAGE 116

104 5371 secretion 43665 1365 type I hypersensitivity 43613 5731 cell death 38313 182 nucleus 37266 874 growth 37003 1329 apoptosis 36085 1326 programmed cell death 36085 1284 cell communication 34839 5821 circulation 34382 5029 muscle contraction 34200 8936 insulin 33615 3379 lymphocyte antigen 32715 116 cell cycle 30670 7162 cytochrome 30501 12487 establishment and/or maintenance of actin cytoskeleton polarity 28685 12604 establishment and/or maintenance of apical/basal cell polarity 28685 5261 establishment and/or maintenance of cell polarity 28685 4286 establishment and/or maintenance of chromatin architecture 28685 12488 establishment and/or maintenance of cytoskeleton polarity 28685 13942 establishment and/or maintenance of epithelial cell polarity 28685 9128 establishment and/or maintenance of microtubule cytoskeleton polarity 28685 13941 establishment and/or maintenance of neuroblast cell polarity 28685 9133 establishment and/or maintenance of polarity of embryonic epithelium 28685 5873 establishment and/or maintenance of polarity of follicular epithelium 28685 9136 establishment and/or maintenance of polarity of larval imaginal disc epithelium 28685 4311 transcription 28125 4376 transcription, RNA-dependent 28125 4244 mutagenesis 27654 5725 pregnancy 27067 3563 mitochondrion 25610 3724 cell junction 25550 3722 intercellular junction 25550 12976 tissue regeneration 25061 5995 tumor antigen 24103 2289 inositol or phosphatidylinositol kinase activity 23903 1135 inositol or phosphatidylinositol phosphatase activity 23903 2295 inositol or phosphatidylinositol phosphodiesterase activity 23903 885 cell differentiation 23348

PAGE 117

105 3314 cytochrome P450 23162 7018 drug resistance 22294 5240 synapsis 21592 5781 memory 21545 13370 synapse 21514 1116 homeostasis 21301 3084 collagen 20821 3403 collagen 20821 9879 oxazole or thiazole biosynthesis 19993 9880 oxazole or thiazole metabolism 19993 1298 lymphocyte activation 19601 973 cell growth 19540 444 cytokinesis 19540 5219 cytokinesis 19540 9106 cytokinesis 19540 12681 social behavior 19240 234 motor activity 18555 823 cytoplasmic vesicle 18470 939 cytoskeleton 18447 5757 blood coagulation 17820 5758 hemostasis 17820 3354 lectin 17611 4885 transport 17519 17568 cognition 16034

PAGE 118

REFERENCES Agrawal, R., Imielinski, T., and Swami, A.N. (1993). Mining Association Rules between Sets of Items in Large Databases, In Proceedings of the 1993 ACM SIGMOD International Conference on Management of Data, pp. 207-216. Agrawal, R., and Srikant, R. (1994). Fast Algorithms for Mining Association Rules, In Proceedings of the 20th International Conference on Very Large Data Bases, pp. 487-499, Morgan Kaufmann Publishers Inc. ANSI/NISO. (1995). Information Retrieval (Z39.50): Application Service Definition and Protocol Specification (ANSI/NISO Z39.50-1995), Besthesda, MD: NISO Press. Baeza-Yates, R., and Ribeiro-Neto, B. (1999). Modern Information Retrieval, ACM Press: Addison Wesley. Barcala, F., Vilares, J., Alonso, M., Grana, J., and Vilares, M., (2002). Tokenization and Proper Noun Recognition for Information Retrieval, In Proceedings of the 13th International Workshop on Database and Expert Systems Applications (DEXA), pp. 246-250: IEEE Computer Society. Berry, M.W., and Browne, M. (1999). Understanding Search Engines: Mathematical Modeling and Text Retrieval, SIAM, Philadelphia, PA: Soc for Industrial & Applied Math. Brill, E. (1992). A Simple Rule-based Part of Speech Tagger, In Proceedings of the 3rd Conference on Applied Natural Language Processing, pp. 152-155, Trento, Italy: Association for Computational Linguistics. Brill, E. (1994). Some Advances in Transformation-based Part of Speech Tagging, In Proceedings of the 12th National Conference on Artificial Intelligence, pp. 722-727, Seattle, WA: American Association for Artificial Intelligence. Brin, S., Motwani, R., and Silverstein, C. (1997a). Beyond Market Baskets: Generalizing Association Rules to Correlations, In Proceedings of the ACM SIGMOD International Conference on Management of Data, 26(2), May 13-15, Tucson, Arizona, USA, pp. 265-276. Brin, S., Motwani, R., Ullman, J. D., and Tsur, S. (1997b). Dynamic itemset counting and implication rules for market basket data, In Proceedings of the ACM SIGMOD International Conference on Management of Data, 26(2), May 13-15, Tucson, Arizona, USA, pp. 255-264. 106

PAGE 119

107 Chakrabarti, S. (2000). Data Mining for Hypertext: A Tutorial Survey, ACM SIGKDD Explorations. 1(2), pp. 1-11. Charikar, M., Chekuri, C., Feder, T., and Motwani, R. (1997). Incremental Clustering and Dynamic Information Retrieval, In Proceeding of the 29th annual ACM symposium on Theory of computing, pp. 626-635. Chen, H., and Ho, T. K. (2000). Evaluation of Decision Forests on Text Categorization, In Proceedings of the 7th Conference on Document Recognition and Retrieval Vol. 3967, pp. 191-199. Chen, H., Schuffels, C., and Orwig, R. (1996). Internet Categorization and Search: A Self-Organizing Approach, In Journal of Visual Communication and Image Representation, 7(1), pp. 88-102. Chen, S. (1998). Digital libraries: The Life Cycle of Information, Better Earth Publisher. Chen, S. (2003). Preserving Digital Records and the Life Cycle of Information, Advances in Computers, Vol. 57, pp. 69-107, Elsevier Science. Chen, S., and Choo, C. (2002). A DL Server with OAI Capabilities: Managing the Metadata Complexity, In Proceedings of the 2002 Joint Conference on Digital Libraries, Portland, OR. Chow, C. K., and Liu, C. N. (1968). Approximating Discrete Probability Distributions with Dependence Trees, IEEE Transactions on Information Theory, 14, pp. 462-467. Christianini, N., and Shawe-Taylor, J. (1999). An Introduction to Support Vector Machines, Cambridge University Press, Boston, MA. Cover, T., and Thomas, J. (1991). Elements of Information Theory, Wiley. Craswell, N. (2000). Methods for Distributed Information Retrieval, Australian National University Ph.D. Thesis, http://pigfish.vic.cmis.csiro.au/~nickc/pubs/craswellthesis.pdf. Davis, J. R., and Lagoze, C. (2000). NCSTRL: Design and Deployment of a Globally Distributed Digital Library, In JASIS, 51(3), pp.273-280. Devroye, L., Gyorfi, L., and Lugosi, G. (1996). A Probabilistic Theory of Pattern Recognition, Springer-Verlag. Dhillon, I., Kogan, J., and Nicholas, C. (2002). Feature Selection and Document Clustering, In Michael Berry (Ed.), A Comprehensive Survey of Text Mining, Lecture Notes in Computer Science, Springer-Verlag.

PAGE 120

108 Dittenbach, M., Merkl, D., and Rauber, A. (2000). The Growing Hierarchical Self-Organizing Map, In Proceedings of the International Joint Conference on Neural Networks (IJCNN 2000), Vol. 6, pp. 15-19. Dublin Core Metadata Initiative. (1999). Dublin Core Metadata Element Set (DCMES) Version 1.1: Reference Description, http://www.dublincore.org/documents/1999/07/02/dces/. Dumais, S., Platt, J., and Heckerman, D. (1998). Heckerman, and M. Sahami. Inductive learning algorithms and representations for text categorization. In Proceedings of the Seventh international Conference on Information and knowledge Management, pp. 148-155. Fano, R. (1961). Transmission of Information, MIT Press, Cambridge, MA. Frawley, W., Piatetsky-Shapiro, G., and Matheus, C. (1992). Knowledge Discovery in Databases: An Overview, AI Magazine, 13(3), pp. 57-70. French, J. C., Powell, A. L., Gey, F., and Perelman, N. (2001). Exploiting a Controlled Vocabulary to Improve Collection Selection and Retrieval Effectiveness, In Proceedings Tenth International Conference on Information and Knowledge Management (CIKM), pp. 199-206, Atlanta, GA: ACM Press. Friedman, N., Geiger, D., and Goldzmidt M. (1997). Bayesian Network Classifiers, In Machine Learning, Vol 29, pp. 131-163. Gene Ontology Consortium, Gene Ontology, http://www.geneontology.org/. Gravano, L., Chang, K., Garcia-Molina, H., Lagoze, C., and Paepcke, A. (1997). STARTS: Stanford Protocol Proposal for Internet Retrieval and Search, In Technical Report: CS-TR-97-1580, Stanford University, http://www-db.stanford.edu/gravano/starts.html. Gruber, T. R. (1993). A translation approach to portable ontologies. Knowledge Acquisition, 5(2), pp. 199-220. Hammouda, K.M., and Kamel, M.S. (2003). Incremental Document Clustering using Cluster Similarity Histograms, In Proceedings of the IEEE/WIC International Conference on Web Intelligence (WI2003), pp 597-601. Han, J., and Kamber, M. (2000). Data Mining, Concepts and Techniques, Morgan Kaufmann Publishers. Han. J., Pei, J., and Yin, Y., (2000). Mining frequent patterns without candidate generation, In Proceedings of the 2000 ACM SIGMOD International Conference on Management of Data, pp. 1-12.

PAGE 121

109 Hearst, M. A. (1999). The Use of Categories and Clusters for Organizing Retrieval Results, Natural Language Information Retrieval, Vol. 7, pp. 333-374, Dordrecht, Kluwer Academic Publishers. Hillmann, D. (2001). Using Dublin Core, Dublin Core Metadata Initiative Recommendation. http://www.dublincore.org/documents/2001/04/12/usageguide/. Joachims, T. (1999). A Probabilistic Analysis of the Rocchio Algorithm with TFIDF for Text Categorization, In Proceedings of the 14th Internation Conference on Machine Learning, pp. 143-151, San Francisco, CA: Morgan Kaufmann Publishers. Joachims, T. (2001). Learning to Classify Text using Support Vector Machines, Kluwer Academic Publishers. John, R., Kohavi, R., and Pfleger, K. (1994). Irrelevant Features and the Subset Selection Problem, In Cohen, W. and Hrish, H. (Eds.), In Proceedings of the 11th International Conference on Machine Learning, pp. 121-129, San Mateo, CA: Morgan Kaufmann Publishers. Kim, H., and Chen, S. (2005). Ontology Search and Text Mining of MEDLINE Database, In Pardalos P.M. (Ed.), To be published in Biocomputing Book Series, Kluwer Academic Publishers. Kim, H., Choo, C., and Chen, S. (2003). An Integrated Digital Library Server with OAI and Self-Organizing Capabilities, In Proceedings of the 7th European Conference on Research and Advanced Technology for Digital Libraries (ECDL 2003), LNCS Vol. 2769, pp. 164-175, Trondheim, Norway: Springer-Verlag. Kohonen, T. (1998). Self-Organization of Very Large Document Collection: State of the Art, In Niklasson, L., Bodn, M., and Ziemke, T., (Ed.), In Proceedings of ICANN98, the 8th International Conference on Artificial Neural Networks, Vol. 1, pp. 65-74, London UK: Springer-Verlag. Kohonen, T. (2001). Self-Organizing Maps, 3rd Edition, Berlin, Germany: Springer-Verlag. Kohonen, T., Kaski, S., Lagus, K., Salojarvi, J., Honkela, J., Paatero, V., and Saarela, A. (2000). Self Organizing of a Massive Document Collection, IEEE Transactions on Neural Networks, Special Issue on Neural Networks for Data Mining and Knowledge Discovery, 11(3), pages 574-585. Lagoze, C., and Sompel, H. Van de (2001a). The Open Archives Initiative Protocol for Metadata Harvesting, Open Archives Initiative, http://www.openarchives.org/OAI/openarchivesprotocol.htm.

PAGE 122

110 Lagoze, C., and Sompel, H. Van de (2001b). The Open Archives Initiative: Building a low-barrier interoperability framework, In Proceedings of the First ACM/IEEE Joint Conference on Digital Libraries, pp 54-62, Roanoke, VA. Lee, Y.-K., Kim, W.-Y., Cai, Y. D., and Han, J. (2003). CoMine: Efficient Mining of Correlated Patterns, In Proceedings of 2003 Int. Conf. on Data Mining (ICDM'03), Melbourne, FL. Lewis, D. D. (1998). Naive (Bayes) at Forty: The Independence Assumption in Information Retrieval, In Proceedings of 10th European Conference on Machine Learning (ECML), Chemnitz, Germany, pp. 4-15, Springer-Verlag. Lewis, P. M. (1959). Approximating Probability Distributions to Reduce Storage Requirements, Information and Control, Vol. 2, pp. 214-225. Lin, C., Chen, H., and Nunamaker, J. F. (1999). Verifying the Proximity Hypothesis for Self-organizing Maps, In Proceedings of the Thirty-Second Annual Hawaii International Conference on System Sciences Vol. 1, pp. 1052, IEEE Computer Society. Liu, B., Hsu, W., and Ma, Y. (1998). Integrating Classification and Association Rule Mining, In Proceedings of the Fourth International Conference on Knowledge Discovery and Data Mining (KDD-98, Plenary Presentation), New York, USA, pp. 80-86. Liu, B., Hsu, W., and Ma, Y. (1999). Mining Association Rules with Multiple Minimum Supports, In Proceedings of the Knowledge Discovery and Data Mining Conference, pp. 337-341. Liu, X. (2002). Federating Heterogeneous Digital Libraries by Metadata Harvesting, Ph.D. Thesis, Old Dominion University. Liu, X., Brody, T., Harnad, S., Carr, L., Maly, K., Zubair, M., and Nelson, M. L. (2002a). A Scalable Architecture for Harvest-Based Digital Libraries The ODU/Southampton Experiments, In D-Lib Magazine, 8(11). Liu, X., Maly, K., Zubair, M., Hong, Q., Nelson, M. L., Knudson, F., and Holtkamp, I. (2002b). Federated Searching Interfaces Techniques for Heterogeneous OAI Repositories, In Journal of Digital Information, 2(4). Liu, X., Maly, K., Zubair, M., and Nelson, M. L. (2001). Arc An OAI Service Provider for Digital Library Federation, In D-Lib Magazine, 7(4). McCray, A.T., Browne A.C., and Bodenreider, O. (2002). The Lexical Properties of the Gene Ontology (GO), In Proceedings of AMIA Annual Symposium, pp. 504-508.

PAGE 123

111 Meretakis, D., Fragoudis, D., Lu, H., and Likothanassis, S. (2000). Scalable Association-based Text Classification, In Proceedings of the Ninth International Conference on Information and Knowledge Management (CIKM-2000), pp. 5-11, Washington, DC, USA: ACM Press. Meretakis, D., and Wuthrich, B. (1999). Extending Naive Bayes Classifiers Using Long Itemsets, In Proceedings of the fifth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD-99), pp 165-174, San Diego, CA, USA: ACM Press. Mitchell, T. (1997). Machine Learning, McGraw-Hill Companies, Inc. Mood, A., Graybill, F., and Boes, D. (1974). Introduction to the Theory of Statistics, McGraw-Hill, 3rd edition. National Library of Medicine, MEDLINE Fact Sheet, http://www.nlm.nih.gov/pubs/factsheets/medline.html. National Science Digital Library, http://nsdl.org/. NSDL (National Science Digital Library). (2002). Technical Infrastructure White Paper, Version 2.0, http://www.smete.org/nsdl/workgroups/technical/nsdl_tech_arch2.0.doc. Nuernberg, P.J., Furuta, R., Leggett, J.J., Marshall, C.C., and Shipman III, F.M. (1995). Digital Libraries: Issues and Architectures, In Proceedings of the Second Annual Conference on the Theory and Practice of Digital Libraries (Digital Libraries '95), Austin, Texas, USA. Omiecinski, E. R. (2003). Alternative Interest Measures for Mining Associations in Databases, In IEEE Transactions on Knowledge and Data Engineering, 15(1), pp. 57-69. Porter, M.F., (1980) An algorithm for suffix stripping, Program, 14(3), pp. 130-137. Powell, A., and French, J. C. (1998). The Potential to Improve Retrieval Effectiveness with Multiple Viewpoints, Technical Report CS-98-15, Department of Computer Science, University of Virginia. Powell, A. L., and French, J. C. (2000). Growth and Server Availability of the NCSTL Digital Library, In Proceedings of 5th ACM Conference on Digital Libraries, pp 264-265. Powell, A. L., French, J. C., and Callan, J. (2000). The Impact of Database Selection on Distributed Searching, In Proceedings of the 23rd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 232-239, ACM Press.

PAGE 124

112 Pratt, W., and Fagan, L. (2000). The Usefulness of Dynamically Categorizing Search Results, In Journal of the American Medical Informatics Association (JAMIA), 7(6), pp 605-617. Quinlan, J. R. (1993). C4.5: Programs for machine learning, Morgan Kaufmann. Raychaudhuri, S., Chang, J. T., Sutphin, P.D., and Altman R.B. (2002). Associating Genes with Gene Ontology Codes Using a Maximum Entropy Analysis of Biomedical Literature, Genome Research, 12(1), pp 203-214. Rennie, J. (2001). Improving Multi-class Text Classification with Niave Bayes, AI Technical Report 2001-004, Massachusetts Institute of Technology. Roussinov, D., and Chen, H. (1998). A Scalable Self-Organizing Map Algorithm for Textual Classification: A Neural Network Approach to Thesaurus Generation, In Communication Cognition and Artificial Intelligence, 15(1-2), pp 81-111. Salton, G., and Buckley, C. (1988). Term-Weighting Approaches in Automatic Text Retrieval, In Information Processing and Management, 24(5), pp 513-523. Salton, G., and Lesk, M. E. (1968). Computer Evaluation of Indexing and Text Processing, In Journal of the ACM, 25(1), pp. 8-36. Smith, B., Williams J., and Schulze-Kremer, S. (2003). The Ontology of the Gene Ontology, In Proceedings of the Annual Symposium of the American Medical Informatics Association, pp. 609-613, Washington DC. Smith, T. C., and Cleary, J. G. (2003). Automatically linking MEDLINE abstracts to the Gene Ontology, In Proceedings of the ISMB 2003 BioLINK Text Data Mining, Brisbane, Australia. Sompel, H. Van de, Krichel, T., and Nelson, M. L. (2000). The UPS Prototype: An Experimental End-user Service across E-print Archives, In D-Lib Magazine, 6(2). Suleman, H., and Fox, E. A. (2002). Beyond Harvesting: Digital Library Components as OAI Extensions, Technical Report TR-02-25, Dept. of Computer Science, Virginia Tech. Van Rijsbergen, C., Harper, D., and Porter, M. (1981). The Selection of Good Search Terms, Information Processing and Management, 17, pp. 77-91. Vesanto, J., and Alhoniemi, E. (2000). Clustering of the Self-Organizing Map, In IEEE Transactions on Neural Networks, 11(3), pp 586-600. Washio,T., and Motoda, H. (2003). State of the Art of Graph-based Data Mining, In ACM SIGKDD Explorations Newsletter, 5(1), pp. 59-68.

PAGE 125

113 Wong, W. C., and Fu, A. (2000). Incremental Document Clustering for Web Page Classification, In Proceeding of the Int. Conf. on Info. Society in the 21st Century: Emerging Technologies and New Challenges (IS2000), Japan. Yandell, M. D., and Majoros, W. H. (2002). Genomics and natural language processing, Nature Reviews Genetics, 3(8), pp. 601-610. Yang, Y., and Liu, X. (1999). A re-examination of text categorization methods, In Proceedings of the 22nd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 42-49, Berkeley, California, USA. Yang, Y., and Pedersen, J. O. (1997). A Comparative Study on Feature Selection in Text Categorization, In Proceedings of the 14th International Conference on Machine Learning (ICML-97), pp. 412-420. Yang, Y., Zhang, J., and Kisiel, B. (2003). A scalability analysis of classifiers in text categorization, In Proceedings of the 26th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 96-103. Youn, E. (2004). Feature Selection and Discriminant Analysis in Data Mining, Ph.D. Thesis, University of Florida. Zaki, M. J., Parthasarathy, S. Ogihara, M., and Li, W. (1997). New Algorithms for Fast Discovery of Association Rules, In Proceedings of the 3rd Int'l Conf. on Knowledge Discovery and Data Mining (KDD97), pp. 283-286: AAAI Press. Zaki, M. J. (2000). Scalable Algorithms for Association Mining, IEEE Transactions on Knowledge and Data Engineering, 12(3), pp. 372-390.

PAGE 126

BIOGRAPHICAL SKETCH Hyunki Kim received his B.E. degree in computer engineering from Chonbuk National University, Chonju, Republic of Korea, in 1994 and his M.S. degree in computer engineering from Chonbuk National University, Chonju, Republic of Korea, in 1996. Between 1996 and 2001, he worked for the Electronics and Telecommunications Research Institute, Taejon, Korea, as a member of the research staff. His research areas include digital libraries, text clustering, and text classification. 114