Exploiting self-reported social networks for routing in delay tolerant networks Greg Bigwood (Devan Rehunathan, Martin Bateman, Tristan Henderson, Saleem Bhatti) School of Computer Science University of St Andrews gjb@cs.st-andrews.ac.uk http://www.cs.st-andrews.ac.uk/~gjb/ Hi, I’m Greg Bigwood – I’m a first year PHD student at the university of st andrews – my supervisor is Tristan Henderson. Today I’m going to present some work we did on exploring how we can use self-reported social networks in routing within ubiquitous computing environments.

Greg Bigwood - University of St Andrews Social Networks Delay Tolerant Networks: episodic connectivity Self-Reported Social Networks: declared social contacts Detected Social Networks: built from encounter trace Understanding how human contacts affect DTNs: Can we use social networks to improve DTN routing? How does a self-reported social network differ from a detected one? In what ways can social networks give us indicators of how a DTN will behave? will social networks help us to effectively find paths through the network that we can use for message forwarding? what are the differences between the social networks individuals define for themselves and the social networks that we detect from human interactions? can we use information from the analysis of the social networks of participants in order to make predictions about how the associated DTN will behave? Eg if a node goes down Greg Bigwood - University of St Andrews

Obtaining Social Networks Detected Social Network (DSN): Mobility tracking experiment 27 participants carrying Tmote invent devices 79 days of encounter tracking Upload via ZigBee basestations DSN ties are encounters Self-Reported Social Network (SRSN): Facebook SRSN ties are Facebook “friends” In order to answere these questions we needed to have both detected social networks and self reported social networks. Devised a mobility tracking experiment in order to obtain the detected socail network We got 27 participants, - mostly undergradautes but we had 3 members of staff and 3 post graduate students. The device can manually upload via ZigBee basestations which were connected to NSLU2 devices, which pushed the data to a central server. We collected records for 79 days. Greg Bigwood - University of St Andrews

Greg Bigwood - University of St Andrews Network Topology Nodes are more reachable in the DSN SRSN (Facebook friends) DSN (Encounters) Here we see the topology of the networks. A tie on the srsn graph indcates a facebook friend link. A tie in the DSN indicates that the participants had an encounter during the data tracking experiment. Nodes map to one another – distances are meaningless. Srsn has different groups – the obvious division was that these nodes were the postgradatuates and staff, whilst the big group was the undergrads. Nodes are more reachable in the DSN – it it more likely messages would reach their destination in the DSN because there area greater number of paths. In the DSN we see that on the whole nodes had more ties in the DSN and therefore you would expect more paths to be available for routing. Extra stuff:- Greg Bigwood - University of St Andrews

Social Network Comparison- Role Equivalence Roles are more clearly defined in SRSN So here we look for roles in the network – a role can be thought of as the way nodes affect information flow. So we could be looking for managers floor workers etc. Explain the role of the top left – is outlying group– and without node 1 it couldn’t communicate. And central nodes overlap with outlying centrals and control their information flow. the blockmodel shows if nodes have ties in ways to each other. We look at the ties of nodes – if they have a suitable leverl of overlap a black block is placed in this matrix. The axis are ordered to show the best grouping of nodes into roles. More well defined roles in SRSN, more sparse cluster in DSN. Whereas in the DSN the roles are harder to detect. They have more been order by number of ties to the center of the network So we would expect that the DSN has more paths available and will result in a higher delivery ratio if used for routing messages in a DTN. Greg Bigwood - University of St Andrews

Social Network Comparison- Role Equivalence Roles are more clearly defined in SRSN So here we look for roles in the network – a role can be thought of as the way nodes affect information flow. So we could be looking for managers floor workers etc. Explain the role of the top left – is outlying group– and without node 1 it couldn’t communicate. And central nodes overlap with outlying centrals and control their information flow. the blockmodel shows if nodes have ties in ways to each other. We look at the ties of nodes – if they have a suitable leverl of overlap a black block is placed in this matrix. The axis are ordered to show the best grouping of nodes into roles. More well defined roles in SRSN, more sparse cluster in DSN. Whereas in the DSN the roles are harder to detect. They have more been order by number of ties to the center of the network So we would expect that the DSN has more paths available and will result in a higher delivery ratio if used for routing messages in a DTN. Greg Bigwood - University of St Andrews

Social Network Comparison- Role Equivalence Roles are more clearly defined in SRSN So here we look for roles in the network – a role can be thought of as the way nodes affect information flow. So we could be looking for managers floor workers etc. Explain the role of the top left – is outlying group– and without node 1 it couldn’t communicate. And central nodes overlap with outlying centrals and control their information flow. the blockmodel shows if nodes have ties in ways to each other. We look at the ties of nodes – if they have a suitable leverl of overlap a black block is placed in this matrix. The axis are ordered to show the best grouping of nodes into roles. More well defined roles in SRSN, more sparse cluster in DSN. Whereas in the DSN the roles are harder to detect. They have more been order by number of ties to the center of the network So we would expect that the DSN has more paths available and will result in a higher delivery ratio if used for routing messages in a DTN. Greg Bigwood - University of St Andrews

Social Network Comparison- Role Equivalence Roles are more clearly defined in SRSN So here we look for roles in the network – a role can be thought of as the way nodes affect information flow. So we could be looking for managers floor workers etc. Explain the role of the top left – is outlying group– and without node 1 it couldn’t communicate. And central nodes overlap with outlying centrals and control their information flow. the blockmodel shows if nodes have ties in ways to each other. We look at the ties of nodes – if they have a suitable leverl of overlap a black block is placed in this matrix. The axis are ordered to show the best grouping of nodes into roles. More well defined roles in SRSN, more sparse cluster in DSN. Whereas in the DSN the roles are harder to detect. They have more been order by number of ties to the center of the network So we would expect that the DSN has more paths available and will result in a higher delivery ratio if used for routing messages in a DTN. Greg Bigwood - University of St Andrews

Social Network Comparison- Role Equivalence Roles are more clearly defined in SRSN So here we look for roles in the network – a role can be thought of as the way nodes affect information flow. So we could be looking for managers floor workers etc. Explain the role of the top left – is outlying group– and without node 1 it couldn’t communicate. And central nodes overlap with outlying centrals and control their information flow. the blockmodel shows if nodes have ties in ways to each other. We look at the ties of nodes – if they have a suitable leverl of overlap a black block is placed in this matrix. The axis are ordered to show the best grouping of nodes into roles. More well defined roles in SRSN, more sparse cluster in DSN. Whereas in the DSN the roles are harder to detect. They have more been order by number of ties to the center of the network So we would expect that the DSN has more paths available and will result in a higher delivery ratio if used for routing messages in a DTN. Greg Bigwood - University of St Andrews

Social Network Comparison- Role Equivalence Roles are more clearly defined in SRSN So here we look for roles in the network – a role can be thought of as the way nodes affect information flow. So we could be looking for managers floor workers etc. Explain the role of the top left – is outlying group– and without node 1 it couldn’t communicate. And central nodes overlap with outlying centrals and control their information flow. the blockmodel shows if nodes have ties in ways to each other. We look at the ties of nodes – if they have a suitable leverl of overlap a black block is placed in this matrix. The axis are ordered to show the best grouping of nodes into roles. More well defined roles in SRSN, more sparse cluster in DSN. Whereas in the DSN the roles are harder to detect. They have more been order by number of ties to the center of the network So we would expect that the DSN has more paths available and will result in a higher delivery ratio if used for routing messages in a DTN. Greg Bigwood - University of St Andrews

Greg Bigwood - University of St Andrews Experiments Hypotheses Configuration More messages to be delivered in DSN Higher delivery ratio (messages delivered/ messages sent) Less message duplication in SRSN Lower delivery cost (medium accesses/ messages sent) Simulate message passing application Source passes SN with message 20 messages per day over whole 79 day trace 100 runs Analyse (against TTL): Delivery ratio Delivery cost Greg Bigwood - University of St Andrews

DSN has only a slightly higher delivery ratio DTN Delivery Ratio DSN has only a slightly higher delivery ratio So we went on to conduct some simulations of a DTN message passingapplication:- Messages were randomly exponentially distirbuted from sender to random destinations at random points during the day. We simulated 20 messages a day, for 79 days. At various ttl’s The scheme worked where a message was sent with the source’s social network and intermediate nodes would forward to those nodes if they encoutnered them. (should I mention infinite buffers?) So we can see from the results here that the DSN did not give an overall substantially greater delivery ratio – at most it was 5-6% Extra stuff: Only 250,000 connections – etc mention problems with this here but ratio was low cause they didn’t encounter each other very much – is different than conference envirnments where people encounter each other all the time. (p< 0.01) Mean diff in range: 0.023% Greg Bigwood - University of St Andrews

SRSN has a much lower delivery cost DTN Delivery Cost SRSN has a much lower delivery cost The next thing we analysed was the delivery cost:- defiend as number of times a messge was attempted at being forwarded divided by the number of messages sent. We can see the the cost for using the SRSN was much lower than the DSN. – there wer more paths available in the DSN and this is why the cost is higher (3 times?). So we can see that it provides an alternative and doesn’t have any learning time – plus the SN data is available. Extra get number for the source vs dynamic routing here (p< 0.01) Mean diff in range: 46.9 Greg Bigwood - University of St Andrews

Greg Bigwood - University of St Andrews Future work How can we bootstrap DTN networks using a mixture of SRSNs and DSNs? How can we exploit role equivalence for DTN routing? How can applications use social network information? What additional experiments do we need to conduct? Explain the bullet points:-cause dsn gives better dliv ration how can we combine htem to get efficient system Havent really used role equivalance at the moment – how can you identify role and use them explicitly for sending How do we pass this info up to the application We only did small experiment is anyone doing similar stuff? Greg Bigwood - University of St Andrews

Greg Bigwood - University of St Andrews Summary Social networking analysis techniques can provide us with insights into DTN performance. Self-reported and detected social networks are different in terms of structure and role equivalence. SRSNs and DSNs have similar performance when used for DTNs, but SRSNs have a much lower cost. Thank you Greg Bigwood - University of St Andrews

Greg Bigwood - University of St Andrews

Greg Bigwood - University of St Andrews Architecture Greg Bigwood - University of St Andrews

Do social networks differ? Structural equivalence Clusters are more clearly defined in the SRSN In order to compare the social networks we performed some social networking analysis on the SRSN and DSN. Here we are looking to see how similar nodes structure is to one another. We look at the euclidean distance of nodes – if nodes I and J have the same ties to the same nodes, then their distance is 0. we compute the distance between all nodes in the network and plot the dendrgrams as seen here. These dendrograms show that the clustering of nodes where they are grouped bottom and and each cluster is defined by having no nodes with a closer euclidean distance that the ones already in the cluster. Nodes on the same branch are considered to have shorted distances between them and are described as being clustered. By looking at the dendrograms we see the clusters on the SRSN but it is harder to identify clusters in the DSN. Greg Bigwood - University of St Andrews

Greg Bigwood - University of St Andrews Delivery Delay (p< 0.01) Mean diff in range: 25 days Greg Bigwood - University of St Andrews

Social network analysis In order to understand if the detected and self-reported are similar. Structural equivalence Nodes with identical ties are structurally equivalent. Role equivalence Nodes are role equivalent if the ways in which they relate to the other nodes is the same. Greg Bigwood - University of St Andrews