Capturing Interactions in Face-To-Face Social Networks

Francesco Ficarola and Andrea Vitaletti

Department of Computer, Control, and Management Engineering,

Sapienza University of Rome, Via Ariosto 25, 00185 Rome, Italy

Keywords:

Mobile, Interactions, Social Networks, RFID, Proximity, Face-To-Face.

Abstract:

Online social networks, formed by cyber interactions between users, are nowadays explored in a number of

papers. In this work, we present our experimental activity on Face-To-Face (F2F) social networks tracing

physical interactions of humans in real-world scenarios. We brieﬂy present the technologies to observe F2F

social networks focusing on the SocioPatterns platform that we have employed in our real-world experiments.

Motivated by the requirements of heterogeneous applications, we discuss how to tune the SocioPatterns col-

lection protocol parameters in order to better capture fast interactions between users; carried out experiments

conﬁrm the effectiveness of such tuning.

1 INTRODUCTION

The participation in online social networking is con-

tinuously growing and more and more people use

websites such as Twitter, Facebook or Google+. This

new online phenomenon has led to a greater atten-

tion of the research community to social networks,

also motivated by relevant application scenarios, such

as reputation management (Yu et al., 2010), recom-

mendation systems (Adomavicius and Tuzhilin, 2005;

Walter et al., 2008) or information sharing platforms

such as Quora. However, a new trend involving our

“real life” relationships is becoming more and more

contemporary due to the continuous development of

new devices capable to track physical interactions. In

this paper we are interested in Face-To-Face (F2F)

social networks, namely networks made of evolving

graphs in which nodes are humans and edges between

nodes dynamically appear whenever a F2F interaction

between humans takes place. Over the last decade,

some papers such as (Hui et al., 2005), have focused

on tracking physical proximities, however the size

of those experiments is relatively small and some of

them were deployed employing unsuitable technolo-

gies; so far, the main reason behind the lack of anal-

ysis of real-world interactions is mainly due to hard-

ware limitations (e.g., bluetooth is too inaccurate to

obtain an efﬁcient face-to-face measure between peo-

ple). One of the ﬁrst attempt to effectively track face-

to-face interactions has been made by the SocioPat-

terns researchers (soc, nd; Barrat et al., 2008) using

the Radio Frequency IDentiﬁcation (RFID) technol-

ogy. We believe SocioPatterns tags are among the

most promising devices to track F2F interactions in

real-world scenarios. For this reason, our experi-

ments have been conducted using such devices with

the main purpose of tuning the multiple access con-

trol (MAC) (Wu and Pan, 2008) protocol programmed

into the tags to better suit heterogeneous application

scenarios. Notice that, standard MAC protocols for

Wireless Sensor Networks (WSNs) are not suitable in

real-world social scenarios where the dynamics of in-

teractions are extremely fast and unpredictable when

compared to the WSNs ones.

Contribution. The ability of capturing F2F interac-

tions opens the possibility of quantitatively exploring

the way in which humans interact in the real world.

In this paper, we compare

the SOCIOMAC protocol

proposed by the SocioPatterns initiative and already

used in monitoring activities, with the new PROXMAC

protocol. Our experimental results show that a proper

tuning of the protocol parameters can better support

the rich heterogeneity of F2F real-world scenarios and

guarantees an effective trade-off between interaction

resolution and network life-time.

Roadmap. In the next section we introduce more for-

mally F2F social networks, presenting some of their

applications and the technologies employed to track

the interactions. In section 3 we discuss the issues

in tracking F2F interactions in heterogeneous applica-

This work has been possible thanks to the cooperation

with the SocioPatterns collaboration (soc, nd).

613

Ficarola F. and Vitaletti A..

Capturing Interactions in Face-To-Face Social Networks.

DOI: 10.5220/0005432206130620

In Proceedings of the 11th International Conference on Web Information Systems and Technologies (WEBIST-2015), pages 613-620

ISBN: 978-989-758-106-9

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

tion scenarios and the corresponding MAC protocols.

In section 4 we show the results of the experiments

conducted on F2F social networks and on a simula-

tion environment to test the performance of the pro-

posed MAC protocols. Conclusions are discussed in

section 5.

2 F2F SOCIAL NETWORKS

A Face-To-Face (F2F) social network is a dynamic

network made by linking nodes (i.e., humans) that in-

teract at short range (e.g., 1-1.5 meter of distance) for

a sufﬁcient amount of time to potentially exchange

meaningful information. More formally, a F2F social

networks is a dynamic evolving network G made of a

ﬁnite sequence G

, G

, . . . , G

of t static networks

over the same vertex set V and a variable edge set E

(Avin et al., 2008); a link connecting nodes hu, vi ∈ E

iff at least a F2F interaction between u and v occurred

in the interval in time ∆ between G

and G

i+i

. We call

∆ the resolution of the network.

Technologies. At present, there are a limited num-

ber of solutions able to track interactions between in-

dividuals in a distributed way that include sensors and

wireless technologies, such as Bluetooth, WiFi and

RFID. One of the ﬁrst experiments to collect infor-

mation from a real group of people was made in (Hui

et al., 2005), where 54 individuals participating in a

conference were given Intel Imote devices, equipped

with a Bluetooth radio to discover nodes in proxim-

ity. However, Bluetooth does not allow a ﬁne-grained

recording of social interactions because of two rea-

sons: 1) the discovery process to identify potential

F2F neighbors is slow and 2) since the radio range is

≈ 5-10 meters, it is possible to record as “social in-

teraction” the simple fact of being in the same room,

even if the users are not interacting in any way.

Jayagopi et al. (Jayagopi et al., 2010) deployed

some experiments involving 24 groups of 4 members

each. Every participant was asked to wear a sociomet-

ric badges capable of recognizing speech activity and

line-of-sight presence. However, all the experiments

performed in (Jayagopi et al., 2010) are based on rel-

atively small groups in which interactions are limited

inside the group, while we are interested in a technol-

ogy to study the dynamics of larger groups in which

members are free to move.

SocioPatterns is an interdisciplinary research col-

laboration that supports the development of the So-

cioPatterns Sensing Platform, an infrastructure in-

cluding new experimental RFID sensors that can be

worn by humans in order to track their mobility and

F2F interactions in real-world scenarios. The So-

cioPatterns platform is made of two main entities: ac-

tive RFID tags and RFID readers. When two per-

sons are in proximity within 1-1.5 meter (see Fig-

ure 1), their tags exchange proximity packets con-

taining their IDs (process 1) so that each tag is able

to know who is talking to. Eventually, tags send the

received proximity packets to close-by readers (pro-

cess 2), which in turn will forward those messages

to a central server running a UDP logger (process

3). Proximity ranges can be controlled via ﬁrmware

by setting the transceiver’s transmission at 4 differ-

ent levels of powers: 0, −6, −12, −18 dBm. Low-

power transmissions entail lower ranges of proximity.

The SocioPatterns platform has proved to work well

in a number of F2F application scenarios that are dis-

cussed in the next section. For this reason the experi-

ments presented in this paper have been implemented

using such a platform.

Figure 1: The communication process.

Applications. SocioPatterns performed several in-

stallations in different social contexts analyzing the

obtained dynamics, such as the study of relationships

between attendees in conferences (Barrat et al., 2010;

der Broeck et al., 2010) or the spread of disease in

hospitals and schools (Isella et al., 2011; Vanhems

et al., 2013; Stehl

e et al., 2011). One of the most

important work is certainly Live Social Semantics

(LSS) (Barrat et al., 2010), which is an applications

that relates virtual and real interactions among indi-

viduals during conferences. Furthermore, since sens-

ing human behaviors in real-world scenarios opens

new frontiers in ubiquitous areas, SocioPatterns have

started studying and analyzing characteristics in this

kind of social networks (Barrat et al., 2013). Anal-

ogous experiments to the SocioPatterns’ ones were

deployed by Chin et al. (Chin et al., 2012) giving

each person an active RFID badge during the course

of a conference. They were interested in realizing

a system able to ﬁnd and connect people to each

other. Analyzing such an experiment, they discov-

ered that more proximity interactions result in an in-

creased probability for a person to add another as a

social connection. Finally, Becchetti et al. deployed

WEBIST2015-11thInternationalConferenceonWebInformationSystemsandTechnologies

614

two large-scale social experiments (Becchetti et al.,

2012), one at the Department of Computer Engineer-

ing in Rome and another one at the MACRO museum

during the opening of the NEON exhibition. They

collected data coming from active RFID tags worn by

around 120 volunteers normally moving and interact-

ing in the monitored areas.

3 CAPTURE F2F INTERACTIONS

A number of protocols nowadays exists for wireless

sensor networks applications (Demirkol et al., 2006),

opportunistic networks contexts (Conti et al., 2010)

or mobile ad hoc networks (Taneja and Kush, 2010).

Nevertheless, most of those protocols are not well

suited to track fast-changing F2F networks where in-

teractions are very dynamic and potentially might

change the structure of the network at every time-step.

We stress that the main purpose of our experiments is

to accurately track F2F interactions rather than to al-

low users to exchange messages. In this perspective,

the proposed MAC protocols are primarily designed

to quickly exchange small packets to build a neigh-

borhood map rather than to exchange relatively big

packets.

In this section we introduce SOCIOMAC and

PROXMAC, two MAC protocols speciﬁcally designed

to track F2F interactions, and we discuss how the

proper tuning of the protocol parameters can allow

us to effectively meet the needs of heterogeneous F2F

tracking scenarios. In particular, depending on the ex-

pected dynamics of the underlying social graph (e.g.,

the interaction time), the protocols have to strike a

balance between accuracy and network life-time.

SocioPatterns tags were initially provided with

SOCIOMAC (soc, 2014) that has been designed for ex-

periments characterized by long-lasting F2F interac-

tions. SOCIOMAC allows to identify nodes in the prox-

imity and to deliver the position and the list of en-

countered nodes to the readers. Communication is en-

crypted to support user privacy and packets are broad-

cast to the readers in a best-effort fashion with no ac-

knowledgment mechanism to minimize the overhead

of the communication. The three types of packets

used by the protocol are: contact (or proximity), bea-

con (or sighting), report. Contact packets are used

to identify close-by tags. This information, as well as

the position of the node, is then delivered to the reader

using the report and beacon packets, respectively.

Figure 2 illustrates an entire phase of the default

SOCIOMAC protocol. Each phase, composed of 8 cy-

cles, represents the main policy of the protocol. Every

cycle starts with a random sleeping period S, between

Figure 2: The SOCIOMAC protocol.

50 and 91 ms, to avoid periodic collisions, then it is

concluded alternatively by the transmission of con-

tact packets TX(C) or by the transmission of beacon

packets TX(B). Before performing a TX(C), tags go

in receiving mode RX for 5 ms so as to receive con-

tact packets from possible neighbors. A phase is con-

cluded with the 8th cycle, when a report packet TX(R)

is sent to the reader. It can be computed that the aver-

age duty cycle of SOCIOMAC is less than 10% making

this protocol suitable for long-lasting deployments;

however the probability P

succ

of exchanging a con-

tact packet between two neighbors in a phase “only”

ranges between 10% and 17%.

According to the analysis and evaluation given in

(Cattuto et al., 2010), a resolution of ∆ = 20 seconds

is considered sufﬁcient to accurately track F2F inter-

actions and to consequently assess proximity. More

speciﬁcally, a contact between two individuals is es-

tablished with a probability in excess of 99%, namely

there is a link in G

, as long as at least one of them

receives a contact packet from the other every 20 sec-

onds. However, such a time scale could be not sufﬁ-

cient in some cases. In particular, we consider a sce-

nario to investigate the so called “wisdom of crowds”

(Surowiecki, 2004), where a group of people is asked

to answer a set of questions in two separate phases. In

the ﬁrst phase participants have to guess the answer

without having any kind of relationship with others,

while in the second phase they are allowed to interact

and exchange their opinions. Since users are already

aware of the questions asked in the ﬁrst phase, in the

second round fast interactions, such as “What did you

answer the ﬁrst question?”, are likely to happen. In

this case, a higher resolution (i.e., a lower time scale,

such as 5 seconds) may be required in order to cap-

ture all those kinds of interactions, consequently it is

crucial a ﬁne-tuning of the SOCIOMAC parameters to

increase P

succ

Figure 3: The PROXMAC protocol.

CapturingInteractionsinFace-To-FaceSocialNetworks

615

Since in short-lasting deployments energy con-

straints are less demanding and interactions tend to

be very fast and dense, at the cost of a higher duty

cycle, the ﬁrst tuning that had to be made was the in-

crease of the receiving interval so as to collect with

a higher probability incoming packets from fast inter-

actions. We recall that the SOCIOMAC protocol sets

the RX interval to 5 ms; because the transmission of a

packet approximately lasts the same order of magni-

tude, the protocol is not always able to capture packets

in their ﬁrst exchanges. For that reason, we decided

of redesigning the main policy of the protocol, named

PROXMAC and shown in Figure 3, so that a ﬁne-tuning

of the parameters would have been achievable. In the

next section we will show how P

succ

increases, mak-

ing this protocol more suitable for detecting fast dy-

namics. Values of the sleeping period and the receiv-

ing interval can be conﬁgured according to the sce-

nario which is going to be deployed.

A sketch of the phase is shown in Figure 3. Notice

that, in this new policy, the report packet can be sent

after every receiving operation only if the tag has pre-

viously received some contact packets, otherwise this

transmission is skipped. This feature guarantees com-

munications broadcasting only useful packets and re-

ducing contingent collisions towards the readers. Fur-

thermore, to better discriminate F2F links, TX(C) are

transmitted with signal strengths 0 and 1, namely −18

and −12 dBm, while the default SOCIOMAC proto-

col uses values of 1 and 2, making it more prone to

false positives. Finally, a tag running PROXMAC sends

a beacon packet only every t phases, so as to avoid to

incessantly ﬂood the channel. The t parameter is usu-

ally conﬁgured to receive at least one beacon packet

at each time-step. That ensures a continuous moni-

toring of the tag inside the network. Both the beacon

and report packets are sent using a signal strength of 3

in order to ensure with a good probability the collec-

tion of the packet by close readers. Some other minor

optimizations of the protocol have been implemented

but will not be discussed in this paper for the sake of

space.

Of course, depending on the chosen parameters,

we can experience with different duty cycles and

probabilities P

succ

of successfully receiving a packet.

As sample of setting, also used in the experiments dis-

cussed in the next section, we pick out a sleeping pe-

riod ranging in [20, 30) ms and a receiving interval

in a range between 30 and 39 ms. Such values, as

well as other settings, were tested in our labs and, af-

ter a long time of measurements, we found that such

a conﬁguration was a good compromise in terms of

performance for our purposes and social experiments.

Indeed, we were interested in deploying social exper-

iments where a high number of iterations of the main

phase in a time-step was required to capture fast in-

teractions as much as possible in dense real F2F so-

cial networks. Other values of S and RX are clearly

available and allowed; they only depend on the pur-

pose of the deployment. However, you should take

into account the fact that the minimum value of the

receiving interval should always be equal or greater

than the maximum value of the sleeping period, so

that the PROXMAC policy is still guaranteed. Indeed, if

that condition is not fulﬁlled, then it may happen that

a whole receiving interval completely falls in the mid-

dle of the sleeping period, thus compromising P

succ

As we will better see in the next section, using

those parameters, PROXMAC can almost always en-

sure the reception of at least one proximity packet

every 2 or 3 time-steps if two tags are facing each

other at 1-1.5 meter. Therefore, people having very

fast interaction can be now tracked and logged. Un-

like SOCIOMAC where the recommended resolution

is ∆ = 20 seconds and the maximum signal strength

for collecting contacts is 2, PROXMAC can exploit a

more ﬁne-grained ∆ and better manage false positives

thanks to a reduced signal strength. An improvement

of the performance even depends on the new length of

the phase being very short, so that the protocol can it-

erate the main cycle several times within a time-step,

which is the most ﬁne-grained resolution.

4 EXPERIMENTS AND

EVALUATION

In this section we evaluate the SOCIOMAC and

PROXMAC’s performance under different conditions in

real-world testbeds.

4.1 Preliminary Experiments

The ﬁrst experiment measures the number of contact

packets exchanged in 10 minutes between two F2F

tags 1 meter apart and then collected by a reader.

This kind of setting is used to simulate an interac-

tion as happens in real-world scenarios. Of course, in

real-world testbeds, communications are more chal-

lenging with respect to this preliminary experiment

because of overlapping transmissions and dense net-

works. However, this ﬁrst experiment allows us to ﬁg-

ure out how protocols perform in an ideal setting, so

that we can understand when to use determined set-

tings in real-world social networks in which perfor-

mance measures are too difﬁcult to explore and an-

alyze in detail. The plots depicted in Figure 4 show

the results of the experiments. The x-axis represents

WEBIST2015-11thInternationalConferenceonWebInformationSystemsandTechnologies

616

the time of the testbed, while the y-axis the total num-

ber of distinct proximity packets collected by both the

protocols. Notice that, the slope of both the plots is

less than 45 degrees. This means that for both the

protocols, a resolution of 1 second (i.e., the minimum

time-step in real-world scenarios) cannot be achieved.

However, as expected, PROXMAC provides better per-

formance in terms of proximity collection. In this

simple setting it is able to support an average reso-

lution of ∆ = 1.8 seconds, while SOCIOMAC supports

a resolution of ∆ = 2.9 seconds.

Figure 4: All the packets exchanged over time.

The second experiment is shown in Figure 5 and

is used to evaluate the number of false positives re-

ported by both the protocols, namely all those con-

tact packets exchanged by two tags not facing each

other. In this experiment two couples of nodes are

placed at a distance of 2.5 meters. As in the previ-

ous experiment, the two tags in a couple are 1 meter

apart. In principle, a tag should detect only the other

one in the couple (i.e., a F2F interaction). However,

due to the proximity of the two couples, false posi-

tives may occur. Of course, similar behaviors occur

in real-world scenarios whenever there are high net-

work densities. For instance, it can happen that two

independent groups of people at close range wrongly

exchange contact packets.

Figure 5: Correct contacts and false positives.

As shown in Figure 5(b), most of the PROXMAC

contact packets are correctly exchanged (blue arcs)

between the tags in the couple. On the contrary, that is

not the same for SOCIOMAC (see Figure 5(a)), where a

number of false positive occurs (red arcs). The thick-

ness of the edges is proportional to the number of con-

tact packets exchanged between the tags. A ticker

edge indicates a longer interaction, while a thin arc

shows a short-lasting interaction. This feature helps

to visualize what kinds of contacts were common. In

the SOCIOMAC case, the formed network is the re-

sult of 5 edges, including two regular interactions and

three false positives. The two more thin false posi-

tives could be excluded using a ﬁlter that analyzes the

structure of the network enforcing a certain thresh-

old, but the third arc has a thickness pretty similar

to one of the two regular. Therefore, there is a sig-

niﬁcant problem in selecting which arc to keep and

which arc to discard. The main cause of this issue

in SOCIOMAC depends on the signal strength used by

the protocol and the corresponding policy. As already

described in section 3, SOCIOMAC allows tags to ex-

change contact packets using three level of power, in-

cluding the signal strength 2. This brings to a wider

range of communication catching farther tags which

cannot be considered face-to-face. This issue is rather

reduced using PROXMAC. Indeed, the only false pos-

itive registered during the experiment was the thin

edge between the nodes 1014 and 1244. However,

when the thickness of false positives is quite thin with

respect to other regular edges, then that arc can be

easily rejected using some techniques of ﬁltering in

post-processing.

4.2 Real-World Social Experiments

WSDM 2013 Conference. On February 2013 we

deployed a real-world social experiment in Rome

during the WSDM Conference, where 69 attendees

agreed to wear our tags running the SOCIOMAC proto-

col. The experiment lasted more than 1 hour, where

50 minutes were allocated for the social interaction,

thus collecting data from a large area, including sev-

eral rooms, the corridor and common spaces. The

main purpose of the experiment was to collect data

to study how F2F interactions can possibly inﬂuence

the wisdom of a group of people, which is usually

called the “wisdom of crowds” phenomenon (Lorenz

et al., 2011). People were not constrained to talk with

a restricted group of other participants, but they were

allowed to interact with anyone they wanted.

At the end of the experiment we collected more

than 23000 single proximity packets. Figure 6 shows

the social network built aggregating all graphs col-

lected over time. Each of them was built with a reso-

lution of ∆ = 3 seconds and an edge threshold of θ = 5

proximity contacts. This means that an edge was built

only if at least 5 proximity packets were collected,

each of which had to be captured within a 3-second

CapturingInteractionsinFace-To-FaceSocialNetworks

617

time interval. Such a resolution was chosen accord-

ing to the results obtained in section 4.1, while that θ

empirically gives a good chance to record a real inter-

action. Indeed, assuming to choose a lower θ, an edge

may be established by the simple fact that two users

can bump into each other just for a moment. Darker

and bigger nodes of the graph represent people that in

the whole experiment accumulated the bigger number

of F2F interactions, while the thickness of an edge is

proportional to the number of interactions observed

between two nodes over the whole time.

Figure 6: The social graph of the network formed during

the experiment in Priverno.

The graph is made of 69 nodes and 133 undirected

edges. The average degree is 3.855, while the net-

work diameter and the average path length are 7 and

3.222, respectively. Only 6 nodes are isolated, while

the maximum degree is 13. Finally, the average den-

sity of the graphs in the evolving network is 0.003.

Priverno’s Country Fair. Similarly to the previous

experiment and with the same modalities, on May

2014 we deployed another real-world social experi-

ment in Priverno (LT), Italy, during a country fair,

recruiting 60 volunteers. As usual, each participant

wore a tag, but this time running the PROXMAC proto-

col. They were free to move within a large monitored

area (around 10 x 15 meters) in a green park. The to-

tal experiment lasted around 30 minutes to study the

same social phenomenon of WSDM. The interaction

part of the experiment lasted around 10 minutes, time

in which we collected more than 11000 single prox-

imity packets. The graph in Figure 7 depicts the ag-

gregated result of the social interaction. Since the pre-

liminary experiments showed that PROXMAC was able

to collect almost twice the number of proximity pack-

ets than SOCIOMAC, in this experiment we leave a res-

olution of ∆ = 3 seconds, but we set an edge threshold

of θ = 10. The same resolution allowed us to compare

this experiment to the previous one, while a double

Figure 7: The social graph of the network formed during

the WSDM experiment.

edge threshold tested the performance of PROXMAC

measured in the ideal experiments.

The graph formed of 60 nodes and 128 undirected

edges has an average degree of 4.267, a network di-

ameter and the average path length of 9 and 3.713,

respectively. 7 users result isolated, while the maxi-

mum degree is 10. Finally, the average density of the

graphs is 0.012. Despite this network is quite similar

in terms of size with respect to the WSDM’s one, we

can notice how the Priverno’s density is an order of

magnitude greater than the WSDM’s one. This fact,

besides a likely higher participation, is also due to

PROXMAC which could collect many more proximity

packets. Clearly, these small real-world social net-

works are only a starting point to test such collection

protocols. For this reason, in the next section, we sim-

ulated the behavior and the performance of both the

protocols in larger graphs and higher densities.

4.3 Simulation on Larger and Denser

Graphs

A simulation on graphs having many more nodes

and higher densities allows us to analyze what kind

of situation we may encounter in future and bigger

real-world scenarios. First of all, we remind that

SOCIOMAC sends report packets to the readers only

after 8 cycles, while PROXMAC can report after ev-

ery receiving operation if new contact packets have

been received. Due to this policy, in dense networks

SOCIOMAC can more easily experience a buffer over-

ﬂow (tags have room for only 4 packets) and conse-

quently discard new incoming contacts. We set up a

simulation for different graph densities and sizes for

a total of 1200 time-steps. For each step, we gener-

ate a random graph based on the Erdos-Renyi model

(Erds and Rnyi, 1960), with n = {60, 100, 200, 500}

and p = {0.01, 0.02, 0.05, 0.1}, where n is the num-

WEBIST2015-11thInternationalConferenceonWebInformationSystemsandTechnologies

618

Figure 8: The percentage of lost packets for different graph densities and number of nodes.

ber of nodes and p is the probability of including an

edge in the graph. Notice that for a sufﬁcient large

amount of time-steps, the probability p approximates

the average density of the whole evolving network.

Figure 8 shows the results of our simulation.

The equivalent case to the real-world experiment

deployed in Priverno is depicted in Figure 8(a), where

the number of nodes is 60 and the density is 0.01.

PROXMAC turned out to be very efﬁcient in this case,

where the percentage of lost packets is around 7%.

Vice versa, SOCIOMAC was not able to collect almost

the 40% of the total possible packets. A similar be-

havior can be observed also in the other cases con-

sidered in the simulation where PROXMAC is able to

better support denser graphs. Indeed, we recall that

PROXMAC has a possible forwarding action after ev-

ery receiving operation, while SOCIOMAC forwards

its contact packets only every 8 cycles. Also, the

whole phase lasts less than the SOCIOMAC’s one. This

brings to a faster actions cycle in PROXMAC, which

is able to execute each operation several times in ev-

ery time-step. Still considering the plot with n = 60,

PROXMAC has a similar percentage of lost packets for

higher densities as well, while the SOCIOMAC’s trend

slightly increases. Their behavior changes little for

the graph having 100 nodes, but increases in graphs

with n = 200 and n = 500. However, for a density

of 0.01, PROXMAC was able to maintain a very low

level of loss. On the contrary, SOCIOMAC starts with

a level of loss equals to 40% and 60%, respectively.

Then, their trends increase and reach more than 80%

for SOCIOMAC with a density 0.1, while PROXMAC

amounts to 40% in n = 200 and 60% in n = 500. We

recall, however, that very high densities are unlikely

in real-world scenarios, as already observed in our so-

cial experiments.

5 CONCLUSION AND FUTURE

WORK

The ability of capturing F2F interactions opens the

possibility of quantitatively exploring the way in

which humans interact in real-world scenarios. The

heterogeneity of such scenarios, requires a proper tun-

ing of the collection protocols in order to achieve the

most effective trade-off between resolution and net-

work life-time.

In this paper we compared the SOCIOMAC protocol

proposed by the SocioPatterns initiative and already

used in monitoring activities, with the new PROXMAC

protocol, speciﬁcally designed to support the moni-

toring of fast-changing interactions. The experimen-

tal results, performed both in real-world social exper-

iments and in denser and bigger simulated networks,

conﬁrm that PROXMAC can collect twice the number

of proximity packets than SOCIOMAC, thus provid-

ing a better resolution for capturing fast interactions.

Thanks to an optimized protocol policy, PROXMAC de-

livers report packets more efﬁciently than SOCIOMAC,

which suffers in denser networks.

In our vision, the collection of reliable F2F in-

teractions will support a better understanding of

the so-called “wisdom of crowds” phenomenon

(Surowiecki, 2004). Indeed, all the proposed models

in this research area (such as (DeGroot, 1974; Fried-

kin and Johnsen, 1990; Bindel et al., 2011)) rely on

the availability of interaction graphs.

In the future, we plan to extend the real-world

testbeds considering a higher number of individuals

in order to study bigger and denser networks.

CapturingInteractionsinFace-To-FaceSocialNetworks

619

REFERENCES

(2014). The SOCIOMAC patent. https://www.google.com/

patents/US8660490.

(n.d.). SocioPatterns. http://www.sociopatterns.org/.

Adomavicius, G. and Tuzhilin, A. (2005). Toward the

next generation of recommender systems: a survey of

the state-of-the-art and possible extensions. Knowl-

edge and Data Engineering, IEEE Transactions on,

17(6):734–749.

Avin, C., Kouck

y, M., and Lotker, Z. (2008). How to ex-

plore a fast-changing world (cover time of a simple

random walk on evolving graphs). In Automata, Lan-

guages and Programming, pages 121–132. Springer.

Barrat, A., Cattuto, C., Colizza, V., Gesualdo, F., Isella, L.,

Pandolﬁ, E., Pinton, J.-F., Rav, L., Rizzo, C., Romano,

M., Stehl, J., Tozzi, A., and Broeck, W. (2013). Empir-

ical temporal networks of face-to-face human interac-

tions. The European Physical Journal Special Topics,

222(6):1295–1309.

Barrat, A., Cattuto, C., Colizza, V., Pinton, J.-F., den

Broeck, W. V., and Vespignani, A. (2008). High res-

olution dynamical mapping of social interactions with

active rﬁd. CoRR, abs/0811.4170.

Barrat, A., Cattuto, C., Szomszor, M., den Broeck, W. V.,

and Alani, H. (2010). Social dynamics in conferences:

Analysis of data from the live social semantics appli-

cation. In Proceedings of the 9th International Seman-

tic Web Conference (ISWC 2010).

Becchetti, L., Bergamini, L., Ficarola, F., and Vitaletti, A.

(2012). Population protocols on real social networks.

In Proceedings of the 9th ACM Symposium on Per-

formance Evaluation of Wireless Ad Hoc, Sensor, and

Ubiquitous Networks, PE-WASUN ’12, pages 17–24,

New York, NY, USA. ACM.

Bindel, D., Kleinberg, J., and Oren, S. (2011). How bad

is forming your own opinion? In Proceedings of the

2011 IEEE 52Nd Annual Symposium on Foundations

of Computer Science, FOCS ’11, pages 57–66, Wash-

ington, DC, USA. IEEE Computer Society.

Cattuto, C., Van den Broeck, W., Barrat, A., Colizza, V.,

Pinton, J.-F., and Vespignani, A. (2010). Dynamics

of person-to-person interactions from distributed rﬁd

sensor networks. PLoS ONE, 5(7):e11596.

Chin, A., Xu, B., Wang, H., and Wang, X. (2012). Link-

ing people through physical proximity in a confer-

ence. In Proceedings of the 3rd international work-

shop on Modeling social media, MSM ’12, pages 13–

20. ACM.

Conti, M., Giordano, S., May, M., and Passarella, A. (2010).

From opportunistic networks to opportunistic comput-

ing. Communications Magazine, IEEE, 48(9):126–

139.

DeGroot, M. H. (1974). Reaching a consensus. Journal

of the American Statistical Association, 69(345):118–

121.

Demirkol, I., Ersoy, C., and Alagoz, F. (2006). Mac proto-

cols for wireless sensor networks: a survey. Commu-

nications Magazine, IEEE, 44(4):115–121.

der Broeck, W. V., Cattuto, C., Barrat, A., Szomszor, M.,

Correndo, G., and Alani, H. (2010). The live so-

cial semantics application: a platform for integrat-

ing face-to-face presence with on-line social network-

ing. In First International Workshop on Communica-

tion, Collaboration and Social Networking in Perva-

sive Computing Environments (PerCol 2010).

Erds, P. and Rnyi, A. (1960). On the evolution of random

graphs. In Publication of the Mathematical Institute

of the Hungarian Academy of Sciences, pages 17–61.

Friedkin, N. and Johnsen, E. (1990). Social-inﬂuence and

opinions. Journal of Mathematical Sociology, 15(3-

4):193–205.

Hui, P., Chaintreau, A., Scott, J., Gass, R., Crowcroft, J.,

and Diot, C. (2005). Pocket switched networks and the

consequences of human mobility in conference envi-

ronments. In Proceedings of ACM SIGCOMM ﬁrst

workshop on delay tolerant networking and related

topics.

Isella, L., Romano, M., Barrat, A., Cattuto, C., Colizza,

V., Van den Broeck, W., Gesualdo, F., Pandolﬁ, E.,

Rav

a, L., Rizzo, C., and Tozzi, A. E. (2011). Close en-

counters in a pediatric ward: Measuring face-to-face

proximity and mixing patterns with wearable sensors.

PLoS ONE, 6(2):e17144.

Jayagopi, D. B., Kim, T., Pentland, A. S., and Gatica-Perez,

D. (2010). Recognizing conversational context in

group interaction using privacy-sensitive mobile sen-

sors. In Proceedings of the 9th International Con-

ference on Mobile and Ubiquitous Multimedia, MUM

’10, pages 8:1–8:4, New York, NY, USA. ACM.

Lorenz, J., Rauhut, H., Schweitzer, F., and Helbing, D.

(2011). How social inﬂuence can undermine the wis-

dom of crowd effect. Proceedings of the National

Academy of Sciences, 108(22):9020–9025.

Stehl

e, J., Voirin, N., Barrat, A., Cattuto, C., Isella, L., Pin-

ton, J.-F., Quaggiotto, M., Van den Broeck, W., Regis,

C., Lina, B., and Vanhems, P. (2011). High-resolution

measurements of face-to-face contact patterns in a pri-

mary school. PLoS ONE, 6(8):e23176.

Surowiecki, J. (2004). The Wisdom of Crowds. Anchor

Books.

Taneja, S. and Kush, A. (2010). A survey of routing

protocols in mobile ad hoc networks. International

Journal of Innovation, Management and Technology,

1(3):2010–0248.

Vanhems, P., Barrat, A., Cattuto, C., Pinton, J.-F., Khanafer,

N., Rgis, C., Kim, B.-a., Comte, B., and Voirin, N.

(2013). Estimating potential infection transmission

routes in hospital wards using wearable proximity sen-

sors. PLoS ONE, 8(9):e73970.

Walter, F., Battiston, S., and Schweitzer, F. (2008). A model

of a trust-based recommendation system on a social

network. Autonomous Agents and Multi-Agent Sys-

tems, 16(1):57–74.

Wu, H. and Pan, Y. (2008). Medium Access Control in Wire-

less Networks. Nova Science Publishers.

Yu, H., Shen, Z., Miao, C., Leung, C., and Niyato, D.

(2010). A survey of trust and reputation management

systems in wireless communications. Proceedings of

the IEEE, 98(10):1755–1772.

WEBIST2015-11thInternationalConferenceonWebInformationSystemsandTechnologies

620