Group Title: Department of Computer and Information Science and Engineering Technical Reports
Title: CSI : a paradigm for behavior-oriented delivery services in mobile human networks
Full Citation
Permanent Link:
 Material Information
Title: CSI : a paradigm for behavior-oriented delivery services in mobile human networks
Alternate Title: Department of Computer and Information Science and Engineering Technical Report
Physical Description: Book
Language: English
Publisher: Department of Computer and Information Science and Engineering, University of Florida
Place of Publication: Gainesville, Fla.
Copyright Date: 2008
 Record Information
Bibliographic ID: UF00095714
Volume ID: VID00001
Source Institution: University of Florida
Holding Location: University of Florida
Rights Management: All rights reserved by the source institution and holding location.


This item has the following downloads:

2008446 ( PDF )

Full Text

CSI: A Paradigm for Behavior-oriented Delivery

Services in Mobile Human Networks

(Paper ID: 1569110782)

Abstract- We propose behavior-oriented services as a new
paradigm of communication in mobile human networks. Our
study is motivated by the tight user-network coupling in future
mobile societies. In such a paradigm, messages are sent to
inferred behavioral profiles, instead of explicit IDs. Our paper
provides a systematic framework in providing such services. First,
user behavioral profiles are constructed based on traces collected
from two large wireless networks, and their spatio-temporal
stability is analyzed. The implicit relationship discovered between
mobile users could be utilized to provide a service for message
delivery and discovery in various network environments. As an
example application, we provide a detailed design of such a
service in challenged opportunistic network architecture, named
CSI. We provide a fully distributed solution using behavioral
profile space gradients and small world structures.
Our analysis shows that user behavioral profiles are surpris-
ingly stable, i.e., the similarity of the behavioral profile of a user
to its future behavioral profile is above 0.8 for two days and
0.75 for one week, and remains above 0.6 for five weeks. The
correlation I,,tlit itm of the similarity metrics between a user
pair at different time instants is above 0.7 for four days, 0.62 for
a week, and remains above 0.5 for two weeks. Leveraging such
a stability in user behaviors, the CSI service achieves delivery
rate very close to the delay-optimal strategy (above 94%), with
minimal overhead (less than I of the optimal). We believe that
this new paradigm will act as an enabler of multiple new services
in mobile societies, and is potentially applicable in server-based,
heterogeneous or infrastructure-less wireless environments.

We envision future networks that consist of numerous ultra
portable devices delivering highly personalized, context-aware
services to mobile users and societies. Such scenarios elicit
strong, tight-coupling between user behavior and the network.
Users' mobility and on-line activities significantly impact
wireless link characteristics and network performance, and
at the same time, the network performance can potentially
influence user activities and behavior. Such a tight user-
network coupling provides a rich set of opportunities and poses
several challenges. On one hand, fundamental understanding
of the mobile user behavior becomes crucial to the design and
analysis of future mobile networks. On the other hand, novel
services can now be introduced and utilize such a coupling
to effectively navigate mobile societies, providing efficient
information dissemination, search and resource discovery.
In this paper, we propose a novel behavior-driven commu-
nication paradigm to enable a new class of services in mobile
societies. Current communication paradigms, including unicast
and multicast, require explicit identification of destination
nodes (through node IDs or group membership protocols),
while directory services map logical, interest-specific queries
into destination IDs where parties are then connected using

interest-oblivious protocols. The power and scalability of such
conventional paradigms might be quite limited in the context
of future, highly dynamic mobile human networks, where it
is desirable in many scenarios to support implicit membership
based on interest. In such scenarios, membership in interest-
groups is not explicitly expressed by users, it is rather implic-
itly and autonomously inferred by network protocols based
on behavioral profiles. This removes the dependence on third
parties (e.g. directory lookup), maintenance of group mem-
bership (e.g., in multicast) or the need to flood user interests
to the whole network, and minimizes delivery overhead to
uninterested users.
Applying such a behavior-driven paradigm in mobile net-
works poses several research challenges. First, how can user
behavior be captured and represented adequately? Second, is
user behavior stable enough to enable meaningful prediction
of future behavior with a short history? How can such services
be provided when the interest or behavior cannot be centrally
monitored and processed? And finally, can we design privacy-
preserving services in this context?
To address these questions we propose a systematic frame-
work with two phases 1) behavioral profile extraction by
analyzing large-scale empirical data sets, investigating the
stability of users in the behavioral space, and 2) leverage the
behavioral profiles for service design We use the implicit
structure in the human networks to guide message and query
dissemination given a target profile.
Specifically, we first analyze network activity traces and
design a summary of user behavioral profiles based on the
mobility preferences. The similarity of the behavioral profile
for a given user to its future profile is high, above 0.75 for eight
days and remains above 0.6 for five weeks. The surprising
observation is that, the similarity metric between a pair of users
predicts their future similarity reasonably well. The correlation
coefficient between their current and future similarity metrics
is above 0.7 for four days, and remains above 0.5 for fifteen
This phenomenon demonstrates that the behavioral profile
we design is an intrinsic property of a given user and a valid
representation of the user for a good period of time into the
future. We refer to this phenomenon as the stability of user
behavioral profiles, which can be used to map the users into
a high dimensional behavioral space. The behavioral space is
defined as a space where each dimension reflects a particular
interest. For example, when we consider mobility preferences,
each dimension represents the fraction of time spent at a given
location. The position of users in the behavioral space reflects
how similar they are with respect to the behavioral profile

we construct. We propose a new communication paradigm, in
which a target profile is used to replace network IDs to indicate
the intended receivers) of a message (i.e., those with matching
behavioral profile to the target profile chosen by the sender are
the intended receivers.). It is a Communication paradigm in
human networks based on the Stability of the user behavioral
profile to discover the receivers Implicitly, abbreviated as CSI.
We present two modes of operation under the over-arching
paradigm: the target mode (CSI:T) and the dissemination
mode (CSI:D). The target mode is used when the target profile
is specified in the same context as the behavioral profile (i.e.,
the target profile is in terms of mobility preferences). The
dissemination mode, on the other hand, is used when the target
profile is de-coupled from mobility preferences.
We show that our CSI schemes perform very close to
the delay-optimal schemes assuming global knowledge and
improve significantly over the baseline dissemination schemes.
For the CSI:T mode, comparing with the delay-optimal proto-
col, our protocol is close in terms of success rate (more than
* I' I and has less overhead (less than I to the optimal),
and the delay is about I"'. more. For the CSI:D mode,
our protocol features lower storage overhead than the delay-
optimal protocol with more than "I'. success rate CSI:D
uses a storage overhead less than i.1I'. of the delay-optimal
protocol, while the delay of CSI:D is about :-". more than
the optimal.
Our Contributions
(1) We introduce the notion of multi-dimensional behavioral
space, and devise a representation of user behavioral profiles
to map users into the behavioral space. Our study is the first
to establish conditions for stability of the relationship between
campus users in this space.
(2) We propose CSI, a new communication paradigm deliver-
ing message based on user profiles. The target profile in CSI
can even be independent of the context of behavioral profile
we use to construct the behavioral space.
(3) We design an efficient dissemination protocol utilizing the
stability of behavioral profiles and SmallWorld in mobile soci-
eties, then empirically evaluate and validate the efficacy of our
proposal using large-scale traces from university campuses.
The outline of the rest of the paper is as follows. We discuss
the related work in section II and important background in
section III. This is followed by an analysis to understand
the user behavioral pattern in section IV. We further discuss
the potential usages of this understanding in section V and
design our CSI schemes in section VI as an example. We
use simulations to evaluate the performance of CSI schemes
in section VII. Finally, we discuss some finer points in
section VIII and conclude in section IX.

We conduct the first detailed systematic study on the spatio-
temporal stability of user behaviors in mobile societies, a new
dimension that has not been considered before. We lay the
foundation of this work on a solid analysis of empirical user
behaviors, enabled by extensive collections of user behavioral
traces. Many of them can be found in the archives at [1],

[2]. Our effort on the extraction of behavioral profiles and
behavior-based user classification is related to the reality min-
ing project [16] and the work by Hsu et al. [4] and Ghosh et
al. [20]. We leverage the representation of mobility preference
matrix defined by Hsu et al. [4], which reveals more detailed
user behavior than the five categories representation used in the
reality mining [16] and the presence/absence encoding vector
used by Ghosh et al. [20].
In centralized trace analysis, the capability of classifying
users based on their mobility preferences [4] or periodic-
ity [19] could potentially lead to applications such as behavior-
aware advertisements or better network management. While
understanding user behavior for these applications has its
own merit, applications in centralized scenario (where user
behaviors are collected, processed and mined at an aggregation
point) are not our major focus in the paper.
The major application considered in this paper is to design a
message dissemination scheme in decentralized environments.
While several previous works exist in the delay tolerant
network field, most of them (e.g. [3], [5], [17], [6], [10])
consider one-to-one communication pattern based on network
identities. The one-to-many communication targeted at a be-
havioral group presented in this paper is a new paradigm
in decentralized environments. Some of the previous work
assume existing infrastructure: PeopleNet [18] uses specialized
geographic zones for queries to meet. The queries are delivered
to randomly chosen nodes in the corresponding zone through
the infrastructure. Others (e.g., [17], [10]) rely on persistent
control message exchanges (e.g., the delivery probability) for
each node to learn the structure of the network, even when
there is no on-going traffic. From the design point of view,
our approach differs from them by avoiding such persistent
control message exchanges to achieve better power efficiency,
an important requirement in decentralized networks.
The spirit of our design is more similar to the work by
Daly et al. [6], in which each node learns the structure of the
network locally and uses the information for message forward-
ing decisions. They use the SmallWorld network structure [7]
which often exists in human networks (as has been investigated
in [14], [9]) and push the message toward nodes with high
centrality to improve the chance of delivery. However, the
learning process still involves message exchanges about past
encounters, even in the absence of actual traffic. Our work,
on the other hand, relies on the intrinsic behavioral pattern of
individual nodes to p1sii .ln" themselves in the behavioral
space in a localized and fully distributed manner, without
exchanging encounter history between nodes. The use of user
behavioral profiles to understand the structure of the space
is similar to the mobility space routing by Leguay et al. [3]
and the utility-based routing by Aiklas et al. [8]. The major
differences between this work and [3], [8] are two fold: First,
we design the CSI:D mode, in which the target profile need not
be related to the behavioral profile based on which the message
dissemination decisions are made. Second, we also provide
a non-revealing option in our protocol, thus no node has to
explicitly reveal its behavioral pattern or interests to others,
as opposed to [3], [8]. The idea of merging similar users into
a group based on their behavior has also been proposed in a

Each row represents the
percentage of time spent at each Dorm Office
location for a day
o 04 0 1
An entry represents the W I
percentage of online time during
time day i at location

Each column corresponds to a location

Fig. 1. Illustration of the association matrix to describe a given user's location
visiting preference.

two-tiered routing structure [10].
Another related paper is the work by Hsu et al. [15] where
the authors focus on only sending messages to users with
similar behavioral profile to the sender. In this paper we
introduce the notion of the target profile to decouple the
behavioral profile of the sender from the destination profile
in the message This significantly enhances the capability of
the message dissemination schemes, by allowing the sender
to specify target behavioral profile (in CSI:T mode), or even
some target profiles that are orthogonal to the behavior based
on which we measure the similarity between users (in CSI:D

A. Mobility-based User Behavior Representation
We represent mobile user behavior of a given user using
the association matrix as illustrated in Fig. 1. In the matrix,
each row vector describes the percentage of time the user
spends at each location on a day, reflecting the importance
of the locations to the user1. In [4] it has been shown that the
location visiting preferences can be leveraged to classify users
of wireless networks on university campuses. For a given user,
the singular value decomposition (SVD) [21] is applied to its
association matrix M, such that

M = U VT (1)

where a set of eigen-behavior vectors, v1, v2, ..., vrank(v) that
summarize the important trends in the original matrix M
can be obtained from matrix V, with corresponding weights
W7v, w~*2, ..., **W .k(V) calculated from the eigen-values in
matrix E. This set of vectors are referred to as the behav-
ioral profile of the particular user, denoted as BP(M), as
they summarize the important trends in user M's behavioral
pattern. The behavioral similarity metric between two users
A and B is defined based on their behavioral profiles, vectors
a,'s and bj's and the corresponding weights, as

Sim(BP(A), BP(B))

rank(A) rank(B)

i WWb ji- bj, (2)
i=l j=l

which is essentially the weighted cosine similarity between
the two sets of eigen-behavior vectors.

'While there may be numerous other representations of user behavior, we
shall show that this representation possesses desirable characteristics for the
purposes of this study. Further investigation of other representations is a
subject of future work.


Trace source USC [12] Dartmouth [13]
Time/duration 2006 spring 2004 spring
of trace semester quarter
Start/End 01/25/06- 04/05/04-
time 04/28/06 06/04/04
Unique 137 b 545 APs/
2137 buildings
locations 162 buildings
Unique MACs analyzed 5,000 6,582

B. Traces
In this paper, we seek a realistic, deep understanding of
user behavior patterns by analyzing semester/quarter-long user
behavioral logs collected from operational campus networks
from public trace archives [1], [2]. We present results based
on two data sets from the University of Southern California
(USC) and the Dartmouth College (Dartmouth). The details of
the data sets are listed in Table I.
We choose to use WLAN traces as they are the largest
user behavioral data sets available. The information available
from these anonymized traces contains many aspects of the
network usage (e.g., time-location information of the users
by tracking the association and disassociation events with
the access points, amount of traffic sent/received, etc.). The
richness in user behavioral data poses a challenge in repre-
senting the user behavior in a meaningful way, such that the
representation not only reveals an intrinsic, stable behavioral
profile of a user, but the identified behavioral profile also
leads to practical applications. We show in this paper that the
location visiting preferences (which is only a subset of the user
behavioral data) is a stable attribute for both individual users
and the relationship between users. This property will prove
quite valuable to the design of efficient message dissemination
schemes, which we empirically validate using the above traces.

In this section we introduce our analysis of user behavioral
patterns and its significance on the service design. While
previous works on user classification based on long-term
behavioral trend [4], [20], [19] are useful and in line with
our goal, the stability of such classification over time has
not been studied systematically. In particular, the short-term
behavior of a user may deviate significantly from the norm,
and the stability of user behavioral profiles is a decisive
factor for whether it can be leveraged to represent the user's
future behavior. In this section we investigate the following
questions: (1) How long of behavioral history do we need to
classify a user? and (2) How much does the behavior of a
given user and its relationship with other users change with
respect to time?
We consider the effect of the amount of past history (of user
behavior) on its behavioral profiles. Each user uses the location
visiting preference vectors in the past d days to summarize
the behavior in the most recent history the user retains d
location visiting preference vectors for these days, organize




T, T2

Fig. 2. Illustration: consider the trailing d days of behavioral profile at time
points that are T days apart.

- s 09

M ~

0 10 20
Time gap (T)

30 40

Fig. 3. Similarity metrics for the same user at time gap T apart.

them in a matrix, and use singular value decomposition to
obtain the behavioral profile, as described in section III-A.
We seek to understand how d influences the representation
and similarity calculations. More specifically, we look into two
important aspects: (1) Whether the representation of a given
user is stable across time, and (2) whether the relationships
between user pairs remain stable as time evolves.
We first consider the stability of the representation of a given
user. Considering two points in time that are T days apart,
we obtain the behavioral profiles for the same user at both
end points, using the logs of the trailing d days ending at
those end points, as illustrated in Fig. 2. Then we use the
similarity metric defined in Eq. (3) to compare how stable a
user's behavioral profile is to one's former self after T days
has elapsed. The average results with various values of the
time gap, T, and considered behavioral history d are shown
in Fig. 3. We notice that, even if we collect a short history of
user behavior (say d 3), the representation is similar to the
behavior of the user for a long time into the future. When we
consider T 35 days apart, the behavioral profiles from the
same user still show high similarity, at about 0.6. The amount
of history used does not influence the result too much when
the considered T is large enough to avoid overlaps in the used
behavioral history (i.e., when T > d). We conclude that on
university campuses, the behavioral profile for a given user is

0 10 20
Time gap (T)

30 40

Fig. 4. Correlation coefficient of the similarity metrics between the same
user pair at time gap T apart.

Dart-3days -D rt-y Dart-1Odays
_-USC 3d USC5d -USClOd

stable, i.e., it remains highly similar for the same user across
time. One interesting note is that, when the behavioral profile
includes only part of a week (d < 7), the similarity of the user
to its former self shows a weekly pattern (i.e., when T is an
integer multiple of seven, the similarity peaks), especially in
Second, we try to quantify how the behavioral similarity
between the same pair of users varies with time. For this part,
we use Eq. (3) to calculate the similarity between two users, A
and B, at two points in time, SimT, (A, B) and SimT, (A, B),
where T1 and T2 are T days apart. We perform this calculation
to all user pairs, and then calculate the correlation coefficient
of the similarity metrics obtained after a T-day interval, as

S A,B(X X)(Y Y)
NSxSy (3)
where X =SimT (A, B) and Y = SimT,(A, B), and the
notations X and Sx denote the average and standard deviation
of X, respectively. N is the total number of user pairs. The
correlation coefficient quantifies how stable the relationship
between user pairs is. We repeat the calculation for all pairs
of users with various d and T values to arrive at Fig. 4. We
observe that the similarity metrics between user pairs correlate
reasonably well if the considered time periods are not far
apart. For T smaller than one week, the correlation coefficient
is above 0.62. This indicates, once the similarity between a
pair of user is obtained, it remains a reasonable predictor for
their mutual relationship for some time period into the future.
Although the reliability of the stale similarity data decreases
with respect to time, the current similarity of a user pair
remains moderately correlated to their future similarity, in the
time range up to several weeks. The correlation is above 0.4
for up to five weeks.
The investigation establishes that the user behavioral
profile is a stable feature to represent the users the
representation of an individual user and the relationship
between users are well correlated with the past history
for the near future. Thus we map the behavioral profile to a
virtual behavioral space [3], in which each user's behavior is
quantified as a high dimensional point2. The mutual similarity
metric between users is a function of their respective positions
in this space. In this paper, when we say two users are similar,
it means they are close in the behavioral space (i.e., the
distance between the two users is small). We also use the
term neighborhood of a node to refer to the other nodes that
are similar to this particular node in the behavioral space.

Profiling users based on stable behaviors is a fundamental
step to understand human behavior. Motivated by the stability
of user behavioral profiles, we introduce a behavior-driven
communication paradigm where we use user behavioral pro-
files, instead of network IDs, to represent users. We envision
that such a radical approach has several benefits.

2The dimension of the behavioral space is the same as the mobility
preference vector representation, typically in the order of a hundred for these
two campuses.

7:- -, 7 D

- D-lOdas -D rt5days -D rt3days

First, it enables behavior-aware message delivery in the
network without mapping attributes to network IDs. As each
user maintains its behavioral profile, it is now possible to
deliver announcements about a sports event on campus towards
sports enthusiasts (e.g., people who visit the gym often) or
advertise a performance at the school auditorium to the regular
attendees of such events.
Second, it facilitates the discovery of nodes with certain
behavior patterns. Consider, for example, in the message
ferry [11] architecture where nodes with high mobility move
messages across the network to facilitate the communication
between otherwise disconnected nodes. One can choose a
target profile that reflects a mobility profile and thus eliminate
the need of knowing the identity of the ferry beforehand or
enforcing this mobility pattern on a controlled node a typical
user who happens to have the desired mobility pattern can be
discovered and serve as a ferry.
Our behavior-driven communication paradigm is applica-
ble to several architectures. In the centralized server-based
architecture, user profiles could be collected and stored at
a data repository, and mined for user classification, abnor-
mality detection, or targeted advertisements. In the cellular
networks, the low-bandwidth channel between the users and
the infrastructure can be leveraged to exchange behavioral
profiles and match users. In this paper, however, we consider
a decentralized infrastructure-less networks, and focus on
how stable behavioral profiles are used for better message
dissemination. We name this scheme as CSI, since it is a
Communication scheme based on the Stable, Implicit structure
in human networks.

In this section, we first present our premises and design
requirements for the CSI schemes. We then discuss the design
of the CSI schemes based on in-depth understanding of the
relationship between similar behavioral profiles and encounter

A. Assumptions and Design Requirements
We assume that each node profiles its own behavioral
pattern by keeping track of the visiting durations of different
locations and summarizing the behavioral profile using the
technique discussed in III-A. This is an individual effort by
each node involving no inter-node interactions. This can be
done by the nodes over-hearing the beacon signals from the
fixed access points in the environment to find out its current
location. Note that, the use of these beacon signals is only for
the node to profile its own behavior they are not used to help
the communication in our protocols (we will re-visit detailed
points of this assumption in section VIII). Also, for the ease of
understanding, we assume in this section that nodes are willing
to send its behavioral profiles to other nodes when needed. A
privacy-preserving option that eliminates this operation is also
discussed in section VIII.
The goal of our CSI scheme is to reach a group of nodes
matching with the target profile specified by the sender, under
the following performance requirements: (1) The protocol

should be scalable, in particular not being dependent on a
centralized directory to map target profiles to user identities.
(2) It should work in an efficient manner and avoid transmis-
sion and storage overhead when possible. Also, it should avoid
control message exchanges in the absence of data traffic. (3)
The syntax of the target profile should be flexible, allowing the
target profile to be not in the same context as the behavioral
profiles we use to represent the users. Also the operation of the
protocol should be flexible to allow tradeoff between various
performance metrics. And finally, (4) the design should be
robust and help in protecting user privacy.
We design two modes of operation for the CSI scheme
under the above requirements. When the target profile is in
the same context as the behavioral profile (in our example,
since the behavioral profile is a summary of user mobility, this
corresponds to the scenario when the target profile describes
users that move in a particular way), the CSI:Target mode
(CSI:T) should be used. When the target profile is irrelevant to
the behavioral profile (e.g., when I want to send to everyone
interested in movies on campus), the CSI:D mode should
be used instead. Although it seems that the applicability of
CSI:T is limited, we note that the behavioral profile (in terms
mobility) can sometimes be used to infer other social aspects
of the users, such as affiliations or even interests (e.g., people
who visit the gym often should like sports in general). Such
inferences expand the scenarios in which CSI:T can be used.
When this is not possible, CSI:Dissemination mode (CSI:D)
provides a more generic option.
The major challenge involved in the design process is
that each node is only aware of the behavioral profile of
itself. Furthermore, we require no persistent control message
exchanges for the nodes to "learn" the structure of the network
proactively when they have no message to send. Nodes only
compare their behavioral profiles when they are involved in
message dissemination. Based on this very limited knowledge
about the behavioral space, a node must predict how useful a
given encounter opportunity is in terms of achieving the fore-
mentioned requirements. Since encounter events may occur
sporadically in sparse, opportunistic networks, the nodes must
make this decision for each encounter event independent of
other encounter events (that may occur long before or after
the current one under consideration). Such a heuristic must
rely on the understanding of the relationship between nodal
behavioral profiles and encounters, which we discuss the next.

B. Relationship between Behavioral Profiles and Encounters
We now analyze the relationship between user behavioral
profiles and a key event for user-to-user communication in an
infrastructure-less network encounters. Encounters in mobile
networks refer to events when users are within the radio range
of each other and direct communication between the involved
devices is possible. In this paper, based on the WLAN traces,
we assume that when two users visit the same location during
overlapped time intervals, they encounter with each other.
While it seems intuitive that users visiting similar locations
should encounter with each other with higher probability, this
is not obvious on university campuses. Students and faculty


r 01

M obllty lmllarlty
(c) Similarity of encountered node sets.

Fig. 5. Relationship between the similarity in behavioral pattern and other quantities.

have their own schedules, and they may rarely encounter due
to the difference in their schedules although they might be in
the same building at different times. Hence we investigate the
relationship between behavioral profiles and encounter events,
first as a sanity check of our intuition, and more ilip, i i.miil\
to understand the relationship between the behavioral patterns
and various aspects of the encounter events (e.g., the encounter
probabilities, encounter durations, etc.). This helps reveal the
implicit structure existing in mobile human networks, which
is the key to the design of the CSI schemes in the following
We classify all node pairs into different bins of behavioral
similarity metric (as defined in Eq. (3)), and obtain various
characteristics of encounter events as a function of the pair-
wise behavioral similarity. In Fig. 5 (a), we show the aggregate
encounter time duration between an average pair of nodes
given the behavioral similarity. In Fig. 5 (b), we show the
probability for a given node pair to encounter with each other,
given their similarity. Combining these two graphs, we see
that if two users are similar in behavioral profiles, they
are much more likely to encounter, and the total time they
encounter with each other is much longer an indication
that nodes with similar behavioral profiles indeed are more
likely to have better opportunities to communicate. When
two users are similar enough (with behavioral similarity larger
than 0.3), they are almost guaranteed to encounter at some
point (with probability above 0.9). However, we note that
some "random" encounter events happen between dissimilar
users. For users with very low (almost zero) similarity, the
probability for them to encounter is not zero, although such
encounter events are much less reliable (i.e., they occur with
much shorter durations, see Fig. 5 (a)).
In Fig. 5 (c) we further compare the behavioral similarity
of node A and B versus the sets of nodes A and B encounter.
We denote the set of nodes A encounters with as E(A).
The similarity of the two sets of nodes is quantified by
IE(A) n E(B)I/IE(A) U E(B)1, where | I is the cardinality
of the set. This graph shows, as two nodes are increasingly
similar, there is larger intersection of nodes they encounter.
When an unlikely encounter event between dissimilar nodes
occurs, it helps both nodes to gain access to a very different
set of nodes, which they are unlikely to encounter directly.
The above findings relate to the SmallWorld encounter
patterns between mobile users [14]. The key features of
SmallWorld networks [7] are high clustering coefficient and
low average path length. In the human networks we analyze

in this section, people with similar behavior form "cliques".
The "random" encounter events between dissimilar nodes
build short-cuts between these cliques to shorten the distances
between any two nodes. We leverage these properties in the
protocol design.

C. CSI:Target Mode
In the CSI:target mode (CSI:T), the sender specifies the
target profile (TP) for the recipients which must have the same
format and semantics as that of the user behavioral profile,
i.e., in our case the TP is a summarized mobility preference
vector (i.e., the percentage of times the target node(s) visit
various locations). For example, we could reach people who
like sports by sending messages to those who visit the gym
regularly. This criteria could be set up by specifying the TP
as a vector with only one 1 corresponding to the gym location
(hence only time spent at this location is considered). If a given
user A has Sim(BP(A),TP) > thsm,, i.e., its behavioral
profile, BP(A), is more similar to TP than a sender specified
threshold, we say node A belongs to the group of intended
receivers. This threshold is set by the sender according to
the desired degree of similarity to the TP. The TP and
the threshold, th,i,, are included in the message header to
describe the intended receivers of the message.
We first discuss the intuition behind the design of the CSI:T
mode using Fig. 6 as an illustration. As per section VI-B, to
deliver messages to receivers defined by a given TP, one way is
to gradually move the message towards nodes with increasing
similarity to the TP via encounters, in the hope that such
transmissions will improve the probability of encountering the
intended receivers. Finally, when the message reaches a node
close to the TP (in the behavioral space), most nodes encounter
frequently with this node are also similar to TP. Hence, the
message should be spread to other nodes in the neighborhood
(in the behavioral space) of the node.
Consider the pseudo-code in Algorithm 1. There are two
phases in the operation, the gradient ascend phase and the
group spread phase. (1) Starting from the sender, if node A
currently holding the message is not an intended receiver (i.e.,
Sim(BP(A), TP) < thsi), it works in the gradient ascend
phase, otherwise it works in the group spread phase. (2) In
the gradient ascend phase, for each encountered node, the
current message holder asks the behavioral profile of the other
node, and if the other node is more similar to the TP in the
behavioral space, the responsibility of forwarding the message
is passed to this node. One can imagine that these similarities

0 0 2 0 4 0 1 0 8
M -1 ty _, I nty
(a) Total encounter duration.

(b) Encounter ` '''

Fig. 6. Illustration of the CSI:T scheme in the high dimension behavioral
space. One copy of the message follows increasing similarity gradient to reach
the neighborhood of the target profile, then triggers group spread.

form an inherent gradient for the message to follow and reach
the close neighborhood of the TP in the behavioral space,
hence the name gradient ascend phase. Note that, up to this
point, there is only one copy of the message in the network -
these intermediate nodes who are not similar to the TP only
forward the message once. (3) When the message reaches a
node with similarity larger than thsim to the TP, the group
spread phase starts. This intended receiver holds on to the
message, and requests the behavioral profiles from nodes it
encounters. If they are also intended receivers, copies of the
messages will be delivered to them. All intended receivers,
after getting the message, continue to work in the group
spread phase. Although multiple copies of the message are
generated in the group spread phase, it is triggered only when
the message is close to the TP, thus most of the encounter
events and inquiries will occur among the intended receivers,
reducing unnecessary overhead.

/* BP(A): Behavioral profile of node A */
if node A has the message then
if Sim(BP(A),TP) > thsim then
SInitiate Groupspread();
L Initiate Gradient_ascend();
while the message is not sent do
foreach node E encountered do
Get BP(E) from E;
if Sim(BP(E), TP) > Sim(BP(A), TP) then
L Send message to E;

Group spread(){
foreach node E encountered do
Get BP(E) from E;
if Sim(BP(E), TP) > thsim then
L Send message to E;

Algorithm 1: Algorithm for the CSI:T mode

2 Groupspread: -- "'--"
Starting from the first | *
node with similarity > / P I
th,, all nodes within
the thd,-neighborhood C- "
receive copies of the /
message Sim(BP(C TP) -0 84
Sim(BP(B), TP) -0 B
im(BP(A), TP) 0 25
S 1 Gradient ascend:
Sim(BP(S), TP) 001 A message is sent to
nodes with increasing
similarities to TP

aA A
s As
The "interest space" The "behavioral space"

Fig. 7. Illustrations of the CSI:D scheme. Left chart: The goal is to send a
message to a group of nodes with a similar characteristic in the interest space
(white nodes in the circle). Right chart: However, they may not be similar
to each other in the behavioral space (nodes with the same legend represent
similar nodes in the behavioral space).

D. CSI: Dissemination Mode
In the CSI:Dissemination mode (CSI:D), there does not exist
a direct relationship between the target profiles of the recip-
ients and their measured behavioral profiles. One particular
example is to reach people who like movies on campus. If
there is no movie theaters on campus, the measured behavioral
profiles (i.e., mobility preference) cannot be used to infer such
an interest. This situation is illustrated in Fig. 7. It appears
there is little insight provided by the similarities between the
nodal behavioral profiles to guide message propagation, as
the intended receivers in this case may be scattered in the
behavioral space, and the relationship between the target pro-
file and the behavioral profile cannot be quantified. Although
it is always possible to reach most users through epidemic
routing, this leads to high overhead, and requires all nodes
in the network to keep a copy of the message. The objective
of CSI:D mode is to reduce the numbers of message copies
transmitted and stored in the network, yet make it possible
for most nodes to get a copy quickly, if they belong to the
intended receivers.
We again first discuss the intuition behind the design of the
CSI:D mode in this paragraph, using Fig. 8 as an illustration.
From section VI-B, since the nodes with high similarity
in their behavioral profiles are almost guaranteed to
encounter, there is really no need for each of them to
keep a copy and disseminate the message. Electing a few
message holders within a single group of similar nodes
would suffice. This intuition leads to the construction of
our message dissemination strategy for the CSI:D. We aim
to have only one message holder among the nodes who are
similar in their behavioral profiles (or equivalently, pick only
one message holder within a neighborhood in the behavioral
space. In Fig. 7, this corresponds to having only one message
holder from each group of nodes with the same legend). We
add the messages holders carefully to avoid overlaps in the
encountered nodes among message holders. As suggested by
Fig. 5 (c), we should select nodes that are very dissimilar in
their behavioral profiles to achieve low overlaps. Recall that
dissimilar node pairs still encounter with non-zero probability,
our design philosophy is to leverage these "random" encounter
events as short-cuts to navigate through the behavioral space
efficiently, hopping across the space to reach dissimilar nodes
with relatively few message transmissions. Such a design
philosophy is also related to the SmallWorld human network
structure a message will be received by an intended receiver
shortly once it has reached someone in the receiver's "clique".

Fig. 8. Illustration of the CSI:D scheme. The idea is to select the message
holders in a non-overlapping fashion to cover the entire behavioral space.

Consider the pseudo-code in Algorithm 1. (1) The sender
itself starts as the first message holder in the network. (2) Each
message holder tries to strategically add additional message
holders in the network. When it encounters with other nodes,
it asks for the behavioral profile of the other node to be
considered as a potential additional message holder. Each
message holder keeps a list of the behavioral profiles of
all known message holders3, and the new node has to be
dissimilar (with the similarity metric lower than a threshold,
thfwd) to all known holders to be added as a new message
holder and keep another full copy of the message. (3) If, on
the other hand, this node is similar to the message holder
(i.e., within similarity threshold thnbr), it uses a single bit to
remember that there is a message holder in its neighborhood
and propagates this information to similar nodes. This bit
is used to prevent excessive message holders in the same
neighborhood, even if some nodes have not encountered with
the message holders directly. (4) When holders encounter, they
update each other with the behavioral profiles of the known
holders list, to gain a better view of the situation of message
spreading. (5) If two similar holders encounter, one of them
should cease to be a holder to reduce duplicated efforts.
Each message holder is responsible for disseminating the
actual message to the intended receivers. The message holders
sends the TP specified by the sender in the message to the
encountered nodes. If the encountered node is an intended
receiver, the full message will be transferred.

In this section, we perform extensive simulations with the
CSI schemes, based on the derived encounters between users
from the two empirical traces. We compare the performances
of our proposal to oracle-based forwarding decisions to show
that our performance is close to the optimum (in terms of the
delivery success rate and the overhead), and does not fall much
behind in delay. We also compare CSI to epidemic routing [5]
and variants of random walk4. In all the simulation cases, we
split the traces into two halves, use the first half to obtain the
behavioral profiles for all users, and then use the second half
of the trace to evaluate the success of our proposed schemes.

3Note this list does not necessarily contain all holders in the network.
Message holders that are added by a particular message holder are not known
to other holders until they meet and sync the lists.
4The CSI could not be directly compared with existing routing schemes
(e.g., [17], [3], [6], [10]) in DTN as most of them have a different routing
objective: reaching a particular network ID.

/* BP(A): Behavioral profile of node A */
/* Hi(A): The i-th known holder of node A */
/* holder_in_group(A): If A knows there is a message
holder in its neighborhood */
if node A is a message holder then
foreach node E encountered do
Get BP(E);
if E is not a holder then
if Sim(BP(E), BP(Hi(A))) < thfdVi and
holder_in_group(E) false then
Elect E as an holder;
Add BP(E) to holder list;
Send the message;
Send BP(H,(A)),Vi;
else if Sim(BP(E), BP(Hi(A))) > thnbr
for any i then
L Let E set holder_in_group(E) = true;
if Sim(BP(E), BP(A)) > thnbr then
SA ceases to be a holder;
L Sync holder lists between node A and E;

else if holder_in_group(A) true then
foreach node E encountered do
Get BP(E);
if Sim(BP(A), BP(E)) > thnb then
L Let E set holder_in_group(E) = true;

Algorithm 2: Algorithm for CSI:D mode.

A. CSI:Target Mode
1) Simulation Setup: In the scenario of CSI:T mode, the
sender specifies the TP and a threshold of similarity ths,,. If
a node shows a similarity metric higher than th,i, to the TP,
it is an intended receiver. In our evaluation, we use the top-
10 dominant behavioral profile5 (i.e., the behavioral profiles
with the most number of people following it, typically in the
order of hundreds) in our traces as the TP, and for each TP we
randomly pick 100 users as the senders generating messages
targeting at the TP. We use the threshold ths,, = 0.8 as the
transition point between the gradient ascend phase and the
group spread phase.
We compare our CSI:T scheme with several other protocols
discussed below. The epidemic routing [5] is a message
dissemination scheme with simplistic decision rules: all nodes
in the network send copies of messages to all the encountered
nodes who have not received the message yet. The random
walk (RW) protocol generates several copies of the message
from the sender, and each copy is transferred among the nodes
in a random fashion, until the hop count reaches a pre-set
TTL value. Group spread only is a simplified version of
our protocol. It uses only the group spread phase, i.e., the
original sender holds on to the message until it encounters

5We have also experimented with other target profiles, such as rarely
visited locations on campuses or profiles that contain a combination of several
locations, and the results are similar to those presented in this section.

with someone who is more similar than thstm to the TP and
starts the group spread phase directly from there.
We also consider two protocols that require global knowl-
edge of the future. The optimal protocol sends copies of the
message only to the nodes which lead to the fastest delivery to
the targeted receivers, and no one else. This is the oracle-based
optimal protocol achievable if one has perfect knowledge of
the future, and serves as the upper bound for performance. The
optimal single-forwarding-path is the oracle-based protocol to
find the fastest path to deliver the message to the neighborhood
of the TP Using the knowledge of the future, it identifies the
path that leads to the earliest message delivery to one of the
intended receivers. Once a copy of the message is delivered to
the thsim-neighborhood to the TP, it follows the same group
spread phase as in CSI:T. This is the optimal performance
(upper bound) for the family of protocols delivering one copy
of message to the neighborhood of the target profile, if one
chooses a good (shortest delay) path note that this shortest-
delay path may not always follow an increasing gradient of
similarities to the TP.
We compare these message dissemination schemes with
respect to three important performance metrics: delivery ratio,
average delay, and transmission overhead. The delivery ratio
is defined as the percentage of the intended receivers (those
with similarity greater than thsm to the TP) actually received
the message. We account for the transmission overhead as the
total number of messages sent in the process of delivery. See
more discussions on the additional overhead of exchanging the
behavioral profiles later in section VIII-A.
2) Simulation Results: We show the normalized perfor-
mance metrics with respect to that of epidemic routing (the
relative performance for each protocol assuming epidemic
routing is 1.0) and its .'. confidence intervals in Fig. 9. We
observe that epidemic routing leads to the highest overhead
while its aggressiveness also results in the highest possible
delivery ratio and the lowest possible delay. The random walks
do not work well regardless the number of copies and the value
of TTL, as they use no information to guide the propagation
of the message towards the right direction. Our CSI:T protocol
leads to a success rate close to the epidemic routing (0.96 for
USC, 0.94 for Dartmouth) with very small overhead (0.02 for
USC, 0.018 for Dartmouth). For the simplified version, group
spread only, the delay is longer and the success rate is lower
than our protocol. We will further investigate this phenomenon
When comparing CSI:T with the protocols with future
knowledge, we see that there is really not much room for
improvement in terms of the success rate and the overhead.
Our gradient ascend approach in CSI:T is similar to what is
achievable even one has the knowledge of the future in these
two aspects. Specifically, CSI:T has more than I' of delivery
rate and uses less than I' overhead of the optimal strategy.
The delay, on the other hand, has some room for improvement.
Our gradient ascend phase generates only one copy of message
from the sender and it moves towards the TP following strictly
ascending similarity. Comparing with the best IlI.i scI I path to
the TP used in the optimal single-forwarding-path, our CSI:T
has 1.40 and 1.47 times more delay, for USC and Dartmouth,

Dellvery ratio ODeIl


Epidemic routig
Group spread only
RW TTL 500 copy 1
RW TTL 500 copy 3 +<
RW TTL 10 copy 150
0 05 1 1 5 2
(a) USC.
Delveryrato Delay Overhead
Epidemic routmg
Group spread only
Optimal I path
RW TTL 500 copy 3
RW TTL 5 copy 300
0 1 2 3 4 5 6 7
(b) Dartmouth.

Fig. 9. Performance comparison of CSI:T to other protocols.

respectively. If we compare with the optimal strategy, where
multiple copies are generated whenever it helps to improve
the delay, the difference is even larger. This calls for a further
investigation of selecting good path(s) from the sender to the
TP, which we leave out for future work.
We take a closer look at the performance metrics by splitting
the simulation cases into categories, depending on the original
similarity metric between the sender's behavioral profile and
the TP, Sim(BP(S),TP). By the split statistics shown in
Fig. 10, we see why the gradient ascend phase is needed
to improve the success rate and reduce the delay. When we
use only the group spread phase, and the sender is dissimilar
from the TP, it takes a longer time before any encounter
event happens directly between the sender and anyone in the
neighborhood of the TP, if it happens at all hence the delay
is longer, and the success rate is lower.
Comparing the differences between two versions of random
walks, few long threads and many short threads, reveals an
interesting difference. The concept that leads to the difference
is illustrated in Fig. 11. Many short threads are better if the
sender is close to the TP, in terms of both delivery ratio and
delay, as the sender generates a lot of threads to "occupy"
the neighborhood since the threads are short, and similar
users encounter more frequently, they are likely to stay in the
neighborhood. Contrarily, if the sender is far away from the
TP, long random walk threads provide a legitimate chance of
moving close to the TP, while short threads provide less hope.

B. CSI:Dissemination Mode
1) Simulation Setup: In the scenario of CSI:D mode, the
target profile specified by the sender cannot help to determine

- slni0.0001
I 0.0001 sim<0.001
] 0.01 SO.1 sim

0.4 -
0.2 -


Epidemic Groupspread CS1 FewlongRW Many hort
r ) Ding only RW
(a) Delivery ratio.

Epidemic Groupspread CSI Few long Manyshort
routing only RW RW
(b) Average delay.

Fig. 10. Split performance metrics by the similarity between the sender and
the target profile (USC).

Single long RW Multiple short RW Single long RW Multiple short RW

Sender is similar to TP Sender is dissimilar from TP

Fig. 11. Illustrations for the comparison between one long random walk and
many short random walks.

to where the message should be sent in the behavioral space.
Hence, the strategy seeks to keep one copy in every neigh-
borhood in the behavioral space. In our evaluation, we start
from 1000 randomly selected users as the senders. Since the
target profile of the intended receivers can be orthogonal to
the behavioral profile, we create the scenario for evaluation
by randomly selecting 500 nodes as the intended receivers
for each sender, and consider the average performances. We
vary the two thresholds, thfwd and thnbr in our CSI:D mode
scheme proposed in VI-D, to adjust the aggressiveness of the
forwarding scheme. Setting low values for both thresholds
leads to less aggressive operations and inferior performances.
At the same time is also leads to lower overheads, as the mes-
sages are copied to fewer message holders, and the existence
of a message holder prevents nodes in a larger neighborhood
from becoming another message holder.
We compare various parameter settings of our CSI:D mode
with two baseline protocols, the epidemic routing and the
random walk. The epidemic routing works the same way as
before, serving as the baseline for comparison. In the random
walks, the visited nodes along the walks become message
holders and they will later disseminate the messages further
when encountering with the intended receivers. The optimal
protocol again assumes global view of the network and the
knowledge of the future. Every node in the network knows

who the intended receivers are, and sends the messages to
other nodes only if they lead to the fastest delivery to the
message to one of the receivers.
The performance metrics we consider are delivery ratio, av-
erage delay, transmission overhead, and, in addition, storage
overhead. Here the transmission overhead refers to the total
number of transmissions to reach the message holders and
the intended receivers. The storage overhead is the number
of eventual message holders that remains in the network after
our scheme is stabilized (recall that some message holders
may decide to cease performing the task if another message
holder is found with similar behavioral pattern in CSI:D). This
is the overall amount of storage space invested by the nodes
collectively to deliver the message6. In the epidemic routing
and the optimal protocol, all nodes that receive the message
hold on to the message for future transmissions (there is no
distinction between the message holder and a regular node),
hence the transmission overhead and the storage overhead are
the same.
2) Simulation Results: In Fig. 12 we show the average
result of the 1000 simulation cases with the I ".'. confidence
interval. We use the legend CSI:D-thfwd-thnbr for our CSI:D
scheme. Comparing with the epidemic routing, our protocol
saves a lot of transmission and storage overhead. It is possible
to use only about 7._" strategically chosen nodes as the
message holder and reach the intended receivers with little
extra delay (about :_". more), when thfwd 0.3 and
thb, = 0.7. Notice that the storage overhead of the CSI:D
scheme is even lower than the optimal protocol (less than .1 .
with the objective of minimizing the delay. If one desires
further reduction in the overhead, setting lower threshold
values provide a way to trade performance for overhead, e.g.,
setting thfwd = 0.1 and thnbr 0.6 cuts the storage overhead
to about :'. of the epidemic routing. The delay of the CSI:D
is not much more than the epidemic routing or the optimal, at
around .'' to :_" more when thfwd 0.3 and thnb, 0.7.
For the random walks, we have configured the TTL values
for them to have similar overhead with the CSI:D (i.e.,
compare RW TTL=350 with CSI:D-0.7-0.3 and RW TTL=150
with CSI:D-0.6-0.1). We notice that although the delivery
rate of the random walk is also pretty good (1.".'. to I1'.
inferior to the corresponding CSI:D), thanks to the non-zero
encounter probability between dissimilar nodes, its delay is
much longer than the corresponding CSI:D (between ".i'.
to 1 I-' more). This is because the random walk does not
leverage the implicit structure of the human network to select
the message holders wisely, as the CSI:D does. The random
walk leaves copies within the same neighborhood of the
original sender with higher probability, as similar nodes are
more likely to encounter (i.e., the random walk will not "leave
the neighborhood" in a small number of hops). Hence, there
exists ,igiiik.lmi overlap between the nodes encountered by
the selected message holders, and the other nodes that are
dissimilar to these holders have to wait for a long time before

6Typically, only about a couple dozens of message holders drop the message
in the simulation cases. Even if we have accounted for the temporarily invested
storage, it adds less than 1% additional storage overhead.

Epidemic routing
CSI:D-0 7-0 3
CSI:D-0 6-0 1
CSI:D-0 5-0 05
RW TTL-350
RW TTL=150

D delivery ratio
SStorage overhead
MI Tx overhead

0 05 1 15
(a) USC.

Epidemic routing
CSI:D-0 7-0 3
CSI:D-0 6-0 1
CSI:D-0 5-0 05
RW TTL-350
RW TTL-150

(b) Dartmouth.

Fig. 12. Performance comparison of CSI:D to other protocols.

some "random" encounter events occur to receive the message,
resulting in the longer delay.

A. Additional Overhead
In addition to the message transmission and storage, in our
proposed CSI schemes, due to the need for exchanging and
maintaining the behavioral profiles, there are some additional
overhead. We discuss them in details in this section.
Overhead for exchanging the behavioral profiles We iden-
tify some additional components to the actual message trans-
missions when the encounter events between mobile nodes are
leveraged for message dissemination. Some of the components
are common to any message dissemination schemes, and the
others are unique to our CSI schemes.
The common overhead for all the DTN message dissem-
ination schemes considered include the beacon signals
for nodes to discover each other when they encounter,
and the exchange of a list of "messages I have seen" to
avoid a given node receiving duplicated messages from
different nodes. This type of overhead is a function of the
encounter patterns itself and is independent of the actual
protocol used. We ignore these common factors in our
Exchanging the behavioral profiles for the evaluation
of mutual similarity is an additional component that
exists only in our behavior-aware protocol. These profiles
are a handful of vectors associated with its weights.
For most of the users, empirically, five to seven eigen-
behavior vectors capture more than 1' of the power in
their association matrices [4]. This is a small constant

overhead we pay for each encounter when one of the
nodes has some message to send. If the message size
is much larger than the overhead, which is usually the
case as messages are transferred in a bigger unit (i.e., a
"bundle") in DTNs, it is worthwhile to pay this overhead
to gain the reduction of transmission counts as we see in
section VII. Furthermore, with CSI, if there is no message
to send, there is no need to exchange the behavioral
profile. Thus, comparing with the protocols that require
proactive, persistent exchanges of control messages when
nodes encounter (e.g., ProPHET [17] requires the ex-
change of encounter probability vectors), qualitatively,
the CSI schemes have lower overhead, especially when
the volume of traffic is low in the network.
The actual message size has to be augmented with the
TP as well. This is a constant overhead, and it can be
reduced if the target vector is 'uiP.,i (e.g., if the TP
considers only the visits to the gym exclusively, there
is only one 1 in the vector. Instead of adding a vector
(0, ..., 0, 1, 0,....) in the header, the vector can be encoded
(i.e., by specifying (gym, 1)) to save space.).
In the CSI:D mode, the message holders have to exchange
the list of behavioral profiles of known holders. This
happens only between a small subset (less than .' .) of
the nodes, and the exchange is necessary only when there
is a difference in the lists. To further alleviate this, the
two nodes can compare their known holder lists using a
hash value, and exchange only the difference.
Overhead for maintaining the behavioral profiles In order
to maintain the behavioral profile, the nodes have to keep
track of its visiting time to various locations. Note this does
not require a node be aware of all possible locations in the
environment it has to keep track of only the ones it has
been to. When two nodes exchange the behavioral profiles,
each entry in the behavioral profile contains only a subset of
locations with annotations for these locations (e.g., Node A
specifies (library, gym) = (0.8, 0.2) while node B specifies (li-
brary, computer lab) = (0.4, 0.6)). The nodes will take a union
of the location sets when comparing their similarities (e.g.,
in the previous example, when node A sends the behavioral
profile to B, B will convert the profiles to BP(A): (library,
gym, computer lab) = (0.8, 0.2, 0) and BP(B): (library, gym,
computer lab) = (0.4, 0, 0.6) before comparing). The required
storage on each node is minimal, as we show about three to
five days of summarized mobility preference is sufficient to
establish a stable behavioral profile for the user in section IV.
In addition, if the beacon signals from locations are not
available, it is possible to use the mutual encounter vectors as
the behavioral descriptors for the nodes nodes who move
similarly should have similar encounter sets. In this sense, we
could replace the representation to be totally independent of
the infrastructure.

B. Privacy Issues
While the behavior-aware message dissemination schemes
achieve good performance with ilgnilik.ill overhead reduction,
it also raises user privacy concerns. In some cases, individuals

may not want to reveal their own behavior. We discuss privacy-
preserving options with our CSI scheme below.
First we emphasize that the original design of CSI presented
in section VI inherently possesses a privacy-preserving feature:
we only use a small subset of user behavior (specifically, the
mobility preference) in the behavioral profile, and with the
singular value decomposition, we reveal only the summarized
trend, not detailed location visiting events for the user. In
addition, the behavioral profiles are exchanged only between
nodes, not stored in any public directory, and it limits only to
when a given node is involved in message dissemination.
We can further reduce the behavioral profile exchanges
in the CSI scheme, and hence help to preserve privacy as
follows. For the CSI:T mode, when nodes encounter, instead
of exchanging their behavioral profile, the node with a message
to send would first send to the other node the TP of the
message and its similarity score to the TP. The other node
silently calculates its similarity to the TP and decides whether
to request for the actual message. This completely removes
the need for behavioral profile exchanges in CSI:T mode.
For the CSI:D mode, when a message holder looks for
potential new holders, instead of asking other nodes to send
the behavioral profile, the message holder sends the list of
known holder's behavioral profiles to the other node. Since
this list contains only the behavioral profiles of the known
holders, not their identities, dissemination of such lists in
the network does not pose a threat to the privacy of the
message holders. Furthermore, when there are multiple holders
in the list, the other node is not able to tell which behavioral
profile corresponds to the holder who sends out the list. If the
other node decides to become a message holder, its behavioral
profile has to be added to the list of known holders. Instead of
immediately sending the behavioral profile of the new holder
to the old holder, which poses an opportunity for the old
holder to link the identity and the behavioral profile of the
new holder, the new holder only adds its behavioral profile to
its own known holder list, and delays the dissemination for a
later holder profile list exchange.
Finally, as a last resort, privacy-minded individuals can
always opt-out of the service, and we expect this would not
impact the performance severely, as it has been shown that
the encounter pattern between nodes in mobile networks is
rich enough to sustain up to I"'. of nodes opting out before
observing a performance degradation [14].


In this paper, we propose a paradigm to represent, summa-
rize and manipulate behavioral profiles and use such profiles
as targets for the communication. We have presented a novel
service of message dissemination in infrastructure-less mobile
human networks based on the behavioral profiles of the users.
The CSI schemes meet the design goals outlined in section VI-
A with respect to efficiency, flexibility and privacy preserving
properties. The CSI schemes perform closely to the delay-
optimal protocols (with I' or more success rate, less than
. of overhead, and the delay is inferior by il1' or less). In
addition, we also observe that human behavior as observed in

the large scale empirical traces is quite robust and only a few
days' worth of data is adequate to summarize and leverage for
message dissemination, which is quite surprising.
We are working toward an implementation of the CSI
schemes based on mobile devices and consider a real-world
evaluation. One key issue is to adapt our algorithm in a more
privacy-preserving fashion which is also resistant to spam
(e.g., include a reputation system). We are also considering
different applications of behavioral profiles, including targeted
advertising via our CSI schemes.

[1] MobiLib: Community-wide Library of Mobility and Wireless Networks
[2] CRAWDAD: A Community Resource for Archiving Wireless Data At
[3] J. Leguay, T. Friedman, and V. Conan, "Evaluating Mobility Pattern Space
Routing for DTNs," in Proceedings of IEEE INFOCOM, April, 2006.
[4] W. Hsu, D. Dutta, and A. Helmy, "Extended Abstract: Mining Behavioral
Groups in Large Wireless LANs" In Proceedings of ACM MOBICOM,
Sep. 2007.
[5] A. Vahdat and D. Becker, "Epidemic Routing for Partially Connected
Ad Hoc Networks," Technical Report CS-200006, Duke University, April
[6] E. Daly and M. Haahr, "Social Network Analysis for Routing in Discon-
nected Delay-Tolerant MANETs," In Proceedings of ACM MOBIHOC,
Sep. 2007.
[7] D. J. Watts and S. H. Strogatz. "Collective Dynamics of 'Small-World'
Networks," Nature, vol. 393, pp. 440-442, 1998.
[8] P. Costa, C. Mascolo, M. Musolesi, and G. Picco, "Socially-aware Rout-
ing for Publish-Subscribe in Delay-tolerant Mobile Ad Hoc Networks,"
to appear in IEEE Journal on Selected Area of Communications.
[9] A. Miklas, K. Gollu, K. Chan, S. Saroiu, K. Gummadi, and E. Lara,
"Exploiting Social Interactions in Mobile I.., in Proceedings of
9th International Conference on Ubiquitous Computing, Sep. 2007.
[10] M. Thomas, A. Gupta, and S. Keshav, "Group Based Routing in
Disconnected Ad Hoc Networks", in Proceedings of 13th Annual IEEE
International Conference on High Performance Computing, Dec. 2006.
[11] W. Zhao, M. Ammar, and E. Zegura, "A Message Ferrying Approach
for Data Delivery in Sparse Mobile Ad Hoc Networks," in Proceedings
of ACM Mobihoc 2004, May 2004.
[12] W. Hsu and A. Helmy, MobiLib USC WLAN trace data set. Downloaded
[13] D. Kotz, T. Henderson and I. Abyzov, CRAWDAD data set dart-
mouth/campus/ movement/01_04 (v. 2005-03-08). Downloaded from
[14] W. Hsu and A. Helmy, "On Nodal Encounter Patterns in Wireless
LAN Traces," the Second International Workshop On Wireless Network
Measurement (WiNMee 2006), April 2006.
[15] W. Hsu, D. Dutta, and A. Helmy, "Profile-Cast: Behavior-Aware Mobile
Networking," in Proceedings of IEEE WCNC, Las Vegas, NV, Mar. 2008.
[16] N. Eagle and A. Pentland, k.. iii. mining: sensing complex social
Sin Journal of Personal and Ubiquitous Computing, vol.10, no.
4, May 2006.
[17] A. Lindgren, A. Doria, and O.Schelen, "Probabilistic Routing in Inter-
mittently Connected Networks," Lecture Notes in Computer Science, vol.
3126, pp. 239-254, Sep. 2004.
[18] M. Motani, V. Srinivasan, and P. \, ..ii ii "PeopleNet: Engineering
A Wireless Virtual Social Network." in Proceedings of MOBICOM 2005,
Sep. 2005.
[19] M. Kim and D. Kotz, "Periodic properties of user mobility and access-
point ', ..'I Journal of Personal and Ubiquitous Computing, 11(6),
Aug. 2007.
[20] J. Ghosh, M. J. Beal, H. Q. Ngo, and C. Qiao, "On Profiling Mobility
and Predicting Locations of Wireless Users," in Proceedings of ACM
REALMAN, May 2006.
[21] R. Horn and C. Johnson, Matrix Analysis, Cambridge University Press,
published 1990.

University of Florida Home Page
© 2004 - 2010 University of Florida George A. Smathers Libraries.
All rights reserved.

Acceptable Use, Copyright, and Disclaimer Statement
Last updated October 10, 2010 - - mvs