Social-Aware Stateless Forwarding
in Pocket Switched Networks
Alessandro Mei
Computer Science Department
Sapienza University, Rome, Italy
Giacomo Morabito
DIIT
University of Catania, Italy
Paolo Santi
IIT
CNR, Pisa, Italy
Julinda Stefa
Computer Science Department
Sapienza University, Rome, Italy
Abstract—In this paper we describe SANE, the first forwarding
mechanism that combines the advantages of both social-aware
and stateless approaches in pocket switched network routing.
SANE is based on the observation—that we validate on real-
world traces—that individuals with similar interests tend to meet
more often. In our approach, individuals (network members) are
characterized by their interest profile, a compact representation
of their interests. Through extensive experiments, we show the
superiority of social-aware, stateless forwarding over existing
stateful, social-aware and stateless, social-oblivious forwarding.
An important byproduct of our interest-based approach is that
it easily enables innovative routing primitives, such as interest-
casting. An interest-casting protocol is also described, and exten-
sively evaluated through experiments based on both real-world
and synthetic mobility traces.
I. INTRODUCTION
The vision of a near future in which a multitude of hand-
held devices establish direct wireless communication links in
an opportunistic fashion has recently attracted the attention of
the research community. Indeed, powerful hand-held devices
are becoming increasingly popular and that these devices are
typically endowed with wireless technologies allowing direct
communication between them (e.g., Bluetooth). The above
vision has motivated researchers to focus on a specific type of
delay tolerant network, called opportunistic or pocket switched
network [8], in which nodes are individuals carrying hand-held
devices and links appear and disappear as these individuals
move and get in physical proximity. In particular, several
authors have tried to use the social ties between individuals
to optimize the performance of unicast communications [3],
[6], [10], as well as multicasting [5], and publish-subscribe
mechanisms [1], [2], [9].
Social-aware routing protocols have been shown to have
superior performance to social-oblivious routing protocols
such as, e.g., BinarySW [14]. This performance improvement
comes at the expense of storing a significant amount of state
information (e.g., history of past encounters, portion of the
“social network” graph, etc.) at the local memory of the
nodes. In other words, a common feature of the social-aware
routing approaches introduced so far is that they heavily build
upon a notion of state. Our goal in this paper is to design a
Alessandro Mei is supported by a Marie Curie Outgoing International
Fellowship funded by the European Union Seventh Framework Programme
(FP7/2007-2013) under grant agreement n. 253461.
The work of Paolo Santi has been partially supported by the Italian MIUR,
program PRIN, Project COGENT.
suite of social-aware, stateless routing protocols that combine
the advantages of social-aware forwarding with the negligible
extra storage requirements typical of stateless protocols.
We start by giving a model for representing user interests
and their similarity. In our approach, the collection of interest
are represented in an m-dimensional interest space, and the
individuals of the network are characterized by their interest
profile, an m-dimensional vector corresponding to a point
in the interest space. The forwarding strategy is then driven
by a measure of similarity between interest profiles that,
at least indirectly, expresses strength of social ties between
the corresponding individuals. Through extensive experiments
based on both real-world and synthetic mobility traces, we
show the superiority of our proposed social-aware, stateless
routing approach over existing stateful, social-aware as well
as stateless, social-oblivious routing approaches.
II. RELATED WORK AND CONTRIBUTION
The idea of exploiting information regarding social ties
between network nodes in PSNs is not new. For instance, in [3]
the authors use the notions of “ego-centric betweenness” and
“social similarity” to improve end-to-end routing performance.
In [6], the authors propose to use a social “centrality” metric
to achieve the same purpose. In [10], the authors use a “social
similarity” metric locally computed from the history of past
encounters to route messages within the network. Recently,
a social-based approach based on a notion of “ego-centric
betwenness” has been proposed also to optimize multicast
performance [5].
The above protocols have shown how the social structure
of a PSN can be successfully exploited to improve traditional,
social oblivious approaches. However, existing social-aware
approaches heavily build upon the ability of storing a large
amount of information at the nodes (typically, to keep trace of
past encounters), i.e., their are stateful approaches. This fact
has important implications for what concerns i) scalability and
ii) effects of memory size on routing performance. As for
i), we observe that relying on a rich state (in some cases,
O(n
2
) storage capacity is required at the nodes, where n is
the number of network nodes) might impose severe limits to
the ability of these approaches to scale up to networks of
even medium size. As for ii), to the best of our knowledge,
the effect of limited memory size on social-aware routing
performance has not been investigated so far.
III. INTEREST SPACE AND PROFILES
We assume each individual in the network can be repre-
sented through her interest profile, a compact representation of
her interests within the interest space. We represent the interest
space as an m-dimensional unit cube C = [0, 1]
m
, where m
is the total number of interests in the network. Interests are
intended in a very broad sense, they might represent degree
of interest in a certain topic (e.g., cinema, literature, etc.), the
fact that an individual belongs to a certain physical or virtual
community (e.g., living in a certain neighborhood, member of
a Facebook interest group, etc. ), and so on.
Given the above definition of interest space, it is natural
to represent the interest profile of an individual A with an m-
dimensional vector reporting, for each possible interest dimen-
sion, A’s degree of interest in the particular topic/community
(either a real number or a binary value). To express similarity
between individual interests, and thus quantitatively measure
“homophily”—degree of interest similarity [11]—we use the
well-known cosine similarity metric [4]:
Definition 1: Given two m-dimensional vectors A and B,
the cosine similarity metric, denoted Θ(A, B), is defined as
follows:
Θ(A, B) = cos(∠AB) =
A · B
k A kk B k
,
where k X k represent the length of vector X.
Note that, given the definition of interest space, 0 ≤
Θ(A, B) ≤ 1 in our model, with higher values of Θ(A, B)
corresponding to a higher “homophily” degree.
Our stateless protocols are based on a simple and natural
observation from everyday life: Our movements are guided by
our interests. To validate this intuition, we use traces collected
during an experiment done with real Bluetooth communicating
devices distributed to participants of the Infocom 2006 con-
ference [6], [7]. This data trace contains not only contact logs,
also does it report information on participants’ nationality,
residence, languages spoken, affiliation, scientific interests,
etc.. From this information we generate interests profile as
defined above. In the process, we discard participants that have
not declared any interest, in order to remove erroneous profiles.
This way, the number of participants reduces to 61 nodes.
To support our intuition, we first calculate the cosine sim-
ilarity between the interest profiles for every pair of partici-
pants. Then, we compute the Pearson correlation index among
this value and the total meeting duration/meeting frequency
among every couple. We then compute the correlation coeffi-
cients among profile similarities and meeting duration/meeting
frequency, only for pairs of individuals who spend, on the
average, more than a certain amount of time together. This way
the effect of casual short meetings is attenuated. The results are
presented in Table I. As can be seen, when we focus on longer
meetings, the correlation of meeting frequency and similarity
of interest profiles is considerably high, reaching 0.67. These
results support the conclusion that our intuition is sound and
that it can be used as the basic mechanism of social-aware,
stateless forwarding protocols.
AVG meet time C
d
C
f
Nodes
> 0 (min) .28 .08 61
> 5 (min) .55 .57 53
> 10 (min) .67 .67 26
TABLE I
CORRELATION BETWEEN INTERESTS PROFILES AND PARTICIPANTS’
ENCOUNTERS. C
d
AND C
f
INDICATE THE PEARSON CORRELATION
COEFFICIENT BETWEEN PARTICIPANTS’ COUPLES PROFILES AND
RESPECTIVELY TOTAL MEETING DURATIONS AND MEETING RATES.
IV. SOCIAL AWARE NETWORKING (SANE)
In this section, we describe Social Aware NEtworking
(SANE), a protocol suite that enables the efficient delivery of
information to relevant destinations in PSNs. SANE supports
a novel communication service, that we call interest-cast (see
Section IV-B), besides the traditional unicast.
We assume that each node can be a forwarder and therefore
maintains a buffer of messages that must be relayed to the
respective destinations. Each message M has a header that
contains a target interest profile that we call message relevance
profile, an integer value N
replicas
representing the number of
replicas of the message that the node is allowed to forward to
other relays, and a time-to-live value TTL. In PSNs nodes can
exchange information as a communication opportunity arises.
Accordingly, SANE procedures are triggered each time a node,
say A, enters within the radio coverage of another node, say
B. Initially, nodes exchange their interest profile (IP), then
each node start scanning its buffer for messages to relay.
A. Unicast
In the unicast case the message relevance profile is set
equal to the interest profile of the destination. According to
our interest-based approach, a message M should preferably
be forwarded to individuals whose interest profile closely
resembles the one of the destination. More specifically, as
in [14], we assume that in order to keep the communication
overhead under control, the same message can be relayed at
most for N
∗
replicas
times. Message relaying obey the following
rules: Message M should be relayed to a node B if an only
if both the two following conditions hold:
– the current value of N
replicas
is higher than 1.
– the cosine similarity metric between the relevance of
message M , denoted as R(M ), and the IP of B, denoted
IP (B), is higher than a given threshold ρ that we call
relaying threshold.
In case, the value of N
replicas
in the message header is halved
and a copy of the message is sent to B. Obviously the message
is transmitted to node B regardless of the value of N
replicas
if
B the destination of the message. In this case, node A removes
the message from the buffer after this is relayed to B.
Note that, as the threshold ρ decreases, the forwarding
strategy becomes more aggressive. This results in the decrease
of the delivery delay, and in the increase of both the delivery
success probability and the communication overhead (cost)
incurred for the delivery of the message M, that we denote
as c(M). Observe that the cost c(M) is proportional to the
number of copies of the message M spread in the network. A
few extreme cases can be considered:
• N
∗
replicas
= ∞: in this case there is no bound on
the number of copies of the message circulating in the
network. We call the resulting version of our protocol
suite epidemic unicast SANE, and we denote it with UN-
SANE EP. The SANE version corresponding to the case
N
∗
replicas
< ∞ is instead called spray & wait unicast
SANE and denoted UN-SANE SW.
• ρ = 0: in this case, the relay threshold is not used, and the
proposed forwarding strategy becomes the same as Bina-
rySW [14]. Furthermore, if N
∗
replicas
is set equal to ∞
then our protocol behaves like epidemic forwarding [15],
which is the policy achieving the lowest delivery delay
(but also the highest cost).
• ρ = 1: in this case, only direct message delivery from
source to destination is possible: Message delivery cost
is minimized, but message delivery delay is very high.
In Section V, we will show the impact of the threshold ρ on
the performance of the forwarding strategy through numerical
examples.
B. Interest-cast
PSNs can create innovative services. Interest-cast is an ex-
ample of such services in which a user wants to communicate
a certain information (for instance, a movie at a local theater
about opera composer Puccini) to the maximum possible
number of interested users, within a certain time (e.g., the
time of the last movie show). Interested users might have an
interest in opera, or cinema, or both, and may be located in the
“neighborhood” of the theater, so to be able to reach the theater
if interested. This type of communication paradigm matches
very well with the localized nature of PSN communications:
the information is spread relatively fast in the neighborhood of
the sender, while it takes longer to propagate to remote areas
(which are typically less interested in the information).
Assume individual C wants to send a message M to all the
individuals within the network that are potentially interested.
First, C sets R(M), the message relevance profile of M, which
is done by setting a “relevance” value in [0, 1] for each of the
m interest dimensions. Whether a message M is relevant for
a certain individual B is determined using a certain relevance
metric. As we already explained, in this paper we use the
well-know cosine similarity metric [4] to determine whether
message M is relevant for individual B.
Note that, since both individuals’ interests and message rele-
vance profiles take values in the same m-dimensional interest
space, we have that, for any individual B and message M,
the angle between IP (B) and R(M ) is in [0, π/2], implying
that Θ(B, M ) is indeed in [0, 1]. In this paper, we use the
following simple rule to determine whether message M is
relevant to individual B: The message is relevant if and only
if Θ(IP (B), R(M)) ≥ α, where α is a suitably chosen
relevance threshold.
V. EXPERIMENTAL SETUP AND RESULTS
Here we present experimental results on the performance
of SANE (the interest-cast version) and UN-SANE (the uni-
cast version) compared to that of well known opportunistic
forwarding protocols. For the evaluation we use both real-
world traces (Infocom 06) and synthetic ones obtained with
the SWIM mobility model [12]. We use also synthetic mobility
traces to evaluate protocol performance because of the limited
real-world traces enriched with user profiles, which does
not allow evaluating performance under different conditions
for what concern, e.g., the degree of correlation between
individual meeting rates and similarity of their profiles. Such
different mobility scenarios can instead be easily realized in
SWIM by properly tuning the model parameters.
A. SANE with limited buffer in Infocom 06
To validate the protocols on the Infocom 06 trace we
average the results of the following experiment, repeated
100 times: We generate a message with a uniform traffic
pattern (source-destination chosen uniformly at random), and
set message’s relevance profile to the destination’s interest
profile. Then, we let the message to be forwarded in the
network according to the different forwarding schemes. We
have evaluated all the protocols with limited buffer size of
the nodes, for different limits. Here, for the sake of space, we
present only the results where the limit is 40 packets per node.
As already discussed, the correlation between node interest
profiles and their meeting frequencies may be low (see first
row of Table I) without filtering out short meetings; on the
other hand, filtering out short meetings to increase correlation
would considerably reduce the size of the data set, making
simulation results scarcely significant. In view of this, we
have decided to keep the user population as large as possible
(61 users, with a 0.08 meeting frequency correlation); thus,
reasonably low values for the relay and relevance thresholds
ρ and α should be chosen (ρ = .25 and α = .45 in our case)
for both unicast and inter-cast.
1) Unicast: We compare the unicast version of SANE (UN-
SANE) to well known stateless forwarding protocols such as
BinarySW [14] and Epidemic [15], and to a state-of-the-art
of social-aware forwarding protocol, BUBBLE [6]. In imple-
menting BUBBLE, we took care of putting the protocol in the
best possible conditions, i.e., complete knowledge of the social
graph and of the local/global ranking metrics. We consider
both the SW and the uncontrolled version of UN-SANE in
our experiments, denoted UN-SANE SW and UN-SANE EP,
respectively. Being the network considered of only 61 nodes,
parameter N
∗
replicas
(number of message copies) of BinarySW
and UN-SANE SW is set to 4. The experiments are repeated
for various values of the TTL’s, and in each case we measure
the average delay (average delivery time for successfully
delivered messages), the cost (average number of message
copies in the network per delivered message, computed only
for successfully delivered messages), and success percentage.
As you can see in Figure 1, both versions of UN-SANE
provide significantly higher success percentage than that of
all the competing protocols, at a lower cost and at a similar
or smaller delay. Quite interestingly, buffer limit yields the
surprising result that UN-SANE EP has higher success rate
than Epidemic itself.
2) Interest-cast: Here, we show results related to the two
interest-cast versions of our protocol: SANE SW, and SANE
EP. Since there is no immediate way of extending BUBBLE
into an interest-cast protocol, we compare SANE protocols
only to Epidemic and BinarySW, whose interest-cast versions
are straightforward (simply delivers a copy of the message
to all relevant destinations). The way we generate messages
and the input tuning parameters of BinarySW and SANE
SW are the same as in unicast. The results are shown in
Figure 2. In this case, coverage refers to the percentage of
relevant destinations holding a copy of the message when
the TTL expires. As seen from the figures, SANE protocols
perform very well, providing comparable coverage of relevant
destinations to that of Epidemic (for TTLs values large than
30 min), but with a much reduced cost (as much as 10-fold
cost reduction with respect to Epidemic, in case of SANE SW).
The benefits of social-aware forwarding are evident comparing
the relative performance of BinarySW and SANE SW: with
a comparable cost, SANE SW provides higher coverage and
lower delay than BinarySW.
B. SANE with Synthetic Traces
The synthetic traces we use for evaluation have been ob-
tained from the SWIM mobility model [12], [13]. In SWIM,
nodes are assigned a home point in the network area, assumed
to be a square. Each time a node has to choose its next desti-
nation, it tradeoffs distance from its home point and popularity
of the possible destinations. Thus, nodes with relatively close
home points (neighbors) tend to go to the same locations and
get in contact more often. In order to run SANE on SWIM’s
traces, we do the following setup: First, we generate a the
network, and a given number of network nodes. For each node,
a 4-dimensional interest profile vector is randomly generated,
with entries chosen independently and uniformly at random
in [0, 1]. Each profile vector is then normalized to 1—this
way, we make sure that no node has very low interests or no
interests at all.
In SWIM, neighbors tend to have a higher meeting rate.
The amount of correlation between vicinity of home points
and meeting rate in SWIM is controlled by a parameter
that here we denote as η (see [12], [13] for details): The
higher this parameter, the higher the correlation. Therefore, we
can generate SWIM mobility traces controlling the resulting
correlation between node profile similarity and their meeting
frequency by tuning SWIM’s η parameter. Due to space
limitation, in the following we will only show results for a
SWIM simulation with 200 nodes scattered in a square area of
500m×500m and with η set in such a way that the correlation
between interest profile similarity and pairwise meeting rates
is about .7.
Unfortunately, due to lack of space, here we do not present
SWIM-based comparison results of SANE with the afore-
mentioned well-known forwarding based protocols. Still we
want to stress that due to the high correlation between node-
profiles and pairwise meeting rates the advantage of the SANE
protocols over the competitors becomes even more evident
than in Infocom 06 simulations. Figures 3 and 4 we show the
success rate, average cost and average delay per received copy
when the relevance threshold is α = .95, versus the value of
the relaying threshold ρ. As expected, the communication cost
increases as the value of ρ decreases.
VI. CONCLUSIONS
In this paper, we have first validated the intuition that
individuals with similar interests tend to meet more often
than individuals with diverse interests, and then used this
intuition to design the first social-aware, stateless forwarding
mechanism for opportunistic networks, called SANE. A nice
feature of the SANE forwarding approach is that it can
be used not only for traditional unicast communication, but
also for realizing innovative networking services for PSNs,
such as interest-casting. When collectively considered, the
experimental results clearly show the superiority of SANE
protocols over both social oblivious, stateless and social-aware,
stateful approaches. Quite astonishingly, SANE provides bet-
ter performance than competitors even when the degree of
correlation between interest profile similarity and pairwise
meeting rates is modest, as in the Infocom 06 scenario. If
this correlation is higher, as it might be expected in practical
situations, we expect that the advantages of SANE protocols
over competitors become substantial.
REFERENCES
[1] C. Boldrini, M. Conti, A. Passarella, “ContentPlace: Social-Aware Data
Dissemination in Opportunistic Networks”, ACM MSWiM 2008.
[2] P. Costa, C. Mascolo, M. Musolesi, G.P. Picco, “Socially-Aware Routing
for Publish-Subscribe in Delay-Tolerant Mobile Ad Hoc Networks”,
IEEE JSAC, Vol. 26(5), 2008.
[3] E. Daly, M. Haahr, “Social Network Analysis for Routing in Discon-
nected Delay-Tolerant MANETs”, ACM MobiHoc 2007.
[4] M.M. Deza, E. Deza, Encyclopedia of Distances, Springer, Berlin, 2009.
[5] W. Gao, Q. Li, B. Zhao, G. Cao, “Multicasting in Delay Tolerant
Networks: A Social Network Perspective”, ACM MobiHoc 2009.
[6] P. Hui, J. Crowcroft, E. Yoneki, “BUBBLE Rap: Social-based Forward-
ing in Delay Tolerant Networks”, ACM MobiHoc 2008.
[7] P. Hui, E. Yoneki, S-Y. Chan, J. Crowcroft, “Distributed Community
Detection in Delay Tolerant Networks”, ACM MobiArch 2007.
[8] P. Hui, A. Chaintreau, J. Scott, R. Gass, J. Crowcroft, C. Diot, “Pocket-
Switched Networks and Human Mobility in Conference Environments”,
ACM WDTN 2005.
[9] S. Ioannidis, A. Chaintreau, L. Massoulie, “Optimal and Scalable
Distribution of Content Updates over a Mobile Social Networks”, IEEE
Infocom 2009.
[10] F. Li, J. Wu, “LocalCom: A Community-Based Epidemic Forwarding
Scheme in Disruption-tolerant Networks”, IEEE Secon 2009.
[11] M. McPherson, “Birds of a feather: Homophily in Social Networks”,
Annual Review of Sociology, vol. 27, n. 1, pp. 415–444, 2001.
[12] A. Mei, J. Stefa, “SWIM: A Simple Model to Generate Small Mobile
Worlds”, IEEE Infocom 2009.
[13] S. Kosta, A. Mei, J. Stefa, “Small World in Motion (SWIM): Modeling
Communities in Ad-Hoc Mobile Networking”, IEEE Secon 2010.
[14] T. Spyropoulos, K. Psounis, C. S. Raghavendra, “Efficient Routing
in Intermittently Connected Mobile Networks: The Multi-copy Case”,
IEEE Trans. on Networking, vol. 16(1), 2008.
[15] A. Vahdat, D. Becker, “Epidemic Routing for Partially Connected Ad
Hoc Networks”, TR CS-200006, Duke Univ., 2000.
(a) Success Percentage (b) Average Cost (c) Average Delay
Fig. 1. Performance of unicast protocols on Infocom 06 traces. Buffer limit set to 40 messages.
(a) Success Percentage (b) Average Cost (c) Average Delay
Fig. 2. Performance of multicast protocols on Infocom 06 traces. Buffer limit set to 40 messages.
(a) Coverage (b) Average Cost (c) Average Delay
Fig. 3. SANE in dependence of the relay threshold ρ.
(a) Succes Percentage (b) Average Cost (c) Average Delay
Fig. 4. UN-SANE in dependence of the relay threshold ρ.
