Improving the efficiency of DTN with social graph

Transcription

1 1 Improving the efficiency of DTN with social graph Jiahua Chang Communication department of Aalto University School of Science & Technology Espoo, Finland Abstract Delay Tolerant Networks (DTN) are networks of self-organizing wireless networks, the connectivity in intermittent between the end-end nodes. The advantage of these networks is low cost, both the infrastructure cost and the maintenance fee. The biggest disadvantage of the networks is low efficiency. In this paper, I summarized two papers From Contacts to Graphs: Pitfalls in Using Complex Network Analysis for DTN routing and Real World Routing Using Virtual World Information and propose my own opinion in the end of this paper. The first paper proposes a simple contact mapping algorithms. The second paper proposes to leverage social graphs from Online Social Networks (OSN) to improving the forwarding efficiency of mobile networks. My own opinion is about how people have motivation to transfer information inside of social graph and how the social graph be used in mobile device. T Index Terms DTN, Social graph, efficiency I. INTRODUCTION he Delay Tolerant Networking (DTN) paradigm has been proposed to support emerging wireless networking applications, where end-to-end connectivity cannot be assumed for technical or economical reasons [3], [4], [5]. Due to the lack of end-to-end paths, traditional routing protocols perform poorly, and numerous opportunistic routing algorithms have been proposed instead [6], [7], [8], [9], [10], and [11]. There, multiple replicas of the same content are often routed in parallel to combat the inherent uncertainty of future communication opportunities between nodes [6]. In order to carefully use the available resources (i.e., limit the number of content copies in the network) and still get a short delay, many protocols attempt to predict which nodes are likely to deliver content or bring it closer to the destination [8]. In most of the networks discussed, node mobility (and resulting communication opportunities) is not entirely random. Instead, weak or strong patterns are present, which a node can attempt to infer and use to predict future contact opportunities. To this end, numerous utility-based routing schemes have been proposed, where various contact properties such as time of last encounter between two nodes [7], frequency of past encounters [8], and mobility profiles [9] are maintained and analyzed to assess the probability of a given node to get closer to the destination. Nowadays, there are more than 6 billion wireless devices (PDAs, mobile phone, laptops) in all over the world. These devices are often carried by humans and communicate when in close proximity. (e.g., PodNet [30], pocket switched networks [5]). So the social interaction will directly affect the communication patterns between the devices. The principal challenge is that mobile devices have relatively small storage. Thus the routing information stored on the device has to be succinct. Secondly, because of privacy considerations, it is not feasible to store social graph information beyond direct friendship links of the owner of the mobile device. Thus, the PSN routes data over humans who are close to each other in real space. We call the graph induced by links formed this way as the proximity graph. The unpredictability of human contacts [23] makes stable routing difficult in the proximity graph. On a small scale, forwarding using community information has been shown to be more effective than other methods [16]. However, inferring community information from the proximity graph on a large scale such as a city is a challenging problem: The proximity graph evolves over time. Observing this over a random time window in the recent past may not reveal all intra-community links. This paper will be organized in the following order: what is social graph (Section II), analysis social graph (section III), real world routing using virtual world information (section IV), and in the section V and VI, I will include my own thinking. II. WHAT IS SOCIAL GRAPH There is theory shows any two person in the world can be connect with maximum seven people. The world is small than we thought, the problem is can we find the correct connection. So social graph is important in connect people and share information with others. The social graph consist three parts: nodes, links and communities. A. Nodes At its core is the idea that as the storage capacities of mobile devices increase, and support for Bluetooth and other short-range data transfer protocols becomes more prevalent, we could use these devices to construct data paths in a store-carry forward fashion: Various intermediate nodes store the data on behalf of the sender and carry it to another contact opportunity where they forward the data to the destination or another node that can take the data closer to the destination. [1] B. Links A social graph offers a natural, compact representation of the

2 2 resulting contact set over time. A graph link could mean that two nodes see each other frequently because they have a social connection (friends), or because they are frequently in the same place without actually knowing each other (familiar strangers); hence, a link is intended to have predictive value for future contacts. Nevertheless, the aggregation of contacts between nodes over time into a static social graph presents an inherent mapping tradeoff, where some information about timing of contacts is lost1. One could create a link if at least one contact has occurred in the past between the two nodes [10], but this would result in an overly dense graph, after a certain network lifetime. On the other hand, a past contact could represent a link only if occurred during a given time window [11]. However, if the sliding time window is too small, the resulting graph might be too sparse. In both cases, meaningful differentiation between nodes using complex network analysis may be rendered impossible. It is thus important to carefully design this mapping in order for links in the graph to maintain their predictive value. [1] C. Communities Nodes could have chance encounters with strangers. Communities are firmed by group of nodes, who has common interest, affiliations, hobbies, political stands etc. The communities are not only mean by actually communities, but also virtual communities. In this work, we show that the social graph on online social networks (OSNs) can be used to efficiently infer routes on a large scale. OSNs such as Facebook, Orkut, Flickr, and LiveJournal are attractive because of their large user base and explicitly declared, stable friendship links. III. ANALYSIS SOCIAL GRAPH When we know about what is social graph, we will analysis the social graph and organize it in a logic way. A. Related work A number of proposed DTN forwarding schemes rely implicitly on assessing the strength of social connections between nodes. Two recently proposed forwarding schemes, SimBet [10] and Bubble Rap [11], use the social structure of the network more explicitly and apply complex network analysis to assess the utility of a node for forwarding. Both are based on the idea that nodes (or rather the users carrying the devices) are clustered to communities of highly connected nodes, and some nodes form bridges between such communities. Further, they assume that this structure will be reflected in the social graph they construct. Although preliminary results for SimBet and Bubble Rap demonstrate promising performance [10], [11], we will show that this performance heavily depends on the way contact aggregation is done. Below, we give a brief description of the two protocols. They will serve as our case studies. SimBet: [10] assesses similarity to detect nodes that are part of the same community, and betweenness centrality to identify bridging nodes, that could carry a message from one community to another. The decision to forward a message depends on the similarity and centrality values of the newly encountered node, relative to the current one: If the former node has a higher similarity with the destination, the message is forwarded to it; otherwise, the message stays with the most central node. The goal is to first use increasingly central nodes to carry the message between communities, and then use similarity to home in to the destination s community. In the original algorithm [10], betweenness and similarity are calculated over a social graph, where there is an edge between two nodes if there has been at least one contact between them at any time in the past. Bubble Rap: [11] uses a similar approach. Again, betweenness centrality is used to find bridging nodes until the content reaches the destination community. Communities here are explicitly identified by a community detection algorithm, instead of implicitly by using similarity. Once in the right community, content is only forwarded to other nodes of that community: a local centrality metric is used to find increasingly better relay nodes within the community. Regarding contact aggregation, it is performed at two points. First, as in SimBet, all contacts in the past are considered. Second, for the two centrality values the time is split into 6h time windows, instead. All contacts in such a 6h window form edges of the graph. [1] B. Contact aggregation For both SimBet and Bubble Rap to function properly, social structures which drive node mobility, such as communities and bridges, must be correctly reflected in the social graph. Here, we argue that this heavily depends on the way this graph is constructed out of observed contacts (i.e., contact aggregation). We illustrate and define the contact aggregation problem and propose simple aggregation mappings. Contact aggregation problem With nodes shaded according to their betweenness centralities, we see in Figure 1 that after a short network lifetime (e.g., after 1 hour) most nodes have the same color since they did not have any contacts yet and thus their betweenness centrality is not defined. After 2 hours, enough contacts have occurred to differentiate many nodes. However, after 72 hours of running time, all nodes have seen each other and the nodes have again the same betweenness centrality. In this case, forwarding decisions gradually degenerate to random, significantly affecting the performance of the two protocols. Fig. 1. Aggregated contacts for the ETH trace at different time instants. Contact aggregation mapping There are different ways to define this indicator function, by time or contacts.

3 3 (1) Growing Time Window: As we saw in Section II, some state-of-the-art algorithms [10] use what we will refer to as growing time window mapping, that is, 1, if u. v Co, n e, u, v (1) 0, otherwise As explained earlier, the problem with this aggregation is that for large n the graph gets fully meshed with all nodes appearing equal with respect to social metrics (e.g., Figure 1.c). (2) Sliding TimeWindow: A generalization of the time window aggregation is to aggregate over a fixed time window instead of the whole lifetime of the network. Denoting the time window length as T, we write 1, if u. v Cn T, n e, u, v (2) 0, otherwise The crucial question here is how large the fixed time window should be for an optimal aggregation. For example, Bubble Rap uses a 6h window for the centrality aggregation. A fixed size time window improves the situation. However, it can have different implications depending on the scenario in hand. In a smaller network, such as in an office or conference, the graph can be fully meshed after few days or hours already, as shown by Figure 1. On the other hand, in a larger network such as a campus or a city-wide scenario, a window of few hours may result in a very sparse graph. (3) Most Recent Contacts: In many scenarios, it is reasonable to assume that very old contacts may not have the same predictive power as more recent ones. Hence, we could aggregate only the En most recent edges to construct the social graph. To achieve this, each edge in the graph is labeled with the last time of appearance in order to know which edge to replace in case of a new contact. Specifically, we assign each pair of nodes fu; vg a timestamp tfu;vg, and maintain a time variable toldest;n that keeps track of the oldest edge in Gn. We can then write our indicator function at time n as 1, iftu, v toldest, n e, u, v (3) 0, otherwise (4) Most Frequent Contacts: Another option is to aggregate only the set of most frequent contacts in En. In many scenarios, a more frequent contact has a higher probability of occurring again soon, and thus reflects a stronger social link. Instead of a timestamp of last appearance, we now maintain, for each pair of nodes, a counter cfu;vg indicating how often a contact was seen in the past. Further, the least frequent contact ID in En and the respective number of times it was seen cleast;n is maintained. Then, the indicator function corresponding to the above mapping is: 1, ifc u, v cleast, n e, u, v (4) 0, otherwise It is important to note that a large number of different and more sophisticated mappings are possible such as, for example, weighted graphs [38]. Our goal here is not to derive an optimal aggregation function, but rather to demonstrate that, even with simple aggregation functions, one can considerably influence the performance of DTN routing. [1] C. Social-based forwarding Routing or forwarding issues in DTN and Mobile Ad Hoc Networks (MANET) are important research topics for researchers. Many MANET and some DTN routing algorithms [17] [18] accomplish forwarding by building and updating routing tables whenever mobility occurs. This approach is considered to be cost ineffective for a human mobile network, since human mobility is often unpredictable, and topology changes can be rapid. For instance, Fig 1 shows the contact occurrence distributions in two small-scale proximity graphs created from subsets of data gathered in the MIT Reality mining [14] and UCSD wireless topology discovery [24] projects. It shows that a random pair of nodes is likely to be connected only very rarely. There is a high probability that an edge will occur fewer than 10 times in the trace, and cannot be easily be learned by distributed route computation algorithms Yet, as shown in [23], these rare edges are important for data delivery. Rather than exchange much control traffic to create unreliable routing structures, we propose to use social information to choose the next-hop relays, which are less volatile than mobility [16]. It was shown using real human mobility traces that by leveraging social information such as centrality [15] and community [21], the proposed forwarding algorithm, Bubble, can significantly improve forwarding efficiency by increasing the fraction of delivered messages at a lower cost. Inferring proximity from social graph Let us denote a social graph from one of the OSN as G = (V;E), where V is the set containing all the nodes in the graph and E is the set of all edges which connect nodes pair (u;v) in the graph. Here a node v 2 V is a user on the OSN, and e 2 E defines the relationship between two users (e.g. friendship). Considering that G consists of users scattered all over the world, some users connected by an edge may be located in different parts of the world and not useful for city-wide mobile computing. In this case, we can first remove all the edges with the geographical locations not in the same target city. By this process, we can create sub graphs for each city i. Our datasets do not have geographical information, so we cannot evaluate the sizes of the geographical subgraphs in this paper. However, Wilson et al. have Facebook regional networks, and found that there are more than 2 million Facebook users in London [25]. This number is large enough for bootstrapping many mobile computing applications in a large city like London. In human society, we can build up hierarchical trees according to the closeness of the relationship between the nodes. For example, we can structure it in such a way that the leaf nodes are the individuals, the first layer consists of family members who represent the closest links, followed by relatives in the second layer, and the top layer will be all the members in the society. The nodes in each layer represent the members in each community at that layer. The closeness of the relationship decreases when we go up the tree. There are many other ways of constructing the hierarchies, such as using affiliation instead of kinship. The lower the layer two nodes belonging to the same community, the closer their relationship and the higher chance that they are good relays of messages for each other. By similar

4 4 means, we can extract a hierarchical structure, Hs from graph G based on the cohesiveness of the connection of the nodes on the graph. Figure 2 illustrates the process of building up the routing structure from graph G. The leaf nodes are the individual nodes on the original graph. Layer 1 is the first layer of the hierarchy, with each node representing the most densely connected community, and vice versa up the tree. between nodes of the same community become a self-loop on the new node. We call the communities detected with a run consisting of these two phases the communities on that particular hierarchical level. The modularity optimization process from the first phase is repeated on the new network. These two phases are repeated iteratively until the global modularity maximum is achieved. [2] IV. REAL WORLD ROUTING USING VIRTUAL WORLD INFORMATION In this Section, an idea will be issued. P. Hui and N.Sastry propose to leverage social graphs from Online Social Networks (OSN) to improve the forwarding efficiency of mobile networks, more particularly Delay Tolerant Networks. [2] A. What is OSN OSN Contain millions of nodes, e.g. facebook, YouTube, Flickr, Livejournal Forwarding using communities We can then define a probability of forwarding in each community layer as Pl, where l is the layer number. Considering an encounter scenario, where device vi meets device vj, and vi has a message m for device vk. If vj and vk are in the same community at layer l, the probability that vj is a good relay for vk is proportional to the probability Pl. For example, we can let P1 = 1. [2] D. Fast community detection In order to know how the real OSN social graph can be fitted to the requirements of mobile devices, we look at three popular OSNs and see what kind of hierarchical structures we can extract from these social graphs. In order to do this, we use a community detection algorithm to cluster the datasets. The algorithm we used in this paper is a fast community detection algorithm based on modularity optimization [13]. Modularity is defined as the difference between this fraction and the fraction of the edges that would be expected to fall within the communities if the edges were assigned randomly but keeping the degrees of the vertices unchanged. The fast algorithm [13] is divided into two phases and repeated iteratively. At the beginning, each node is assigned to a different community, then, for each node i and its neighbour node j, we evaluate the gain of modularity by removing i from its community and placing into the community of node j. We place node i into the neighbour community for which the gain is the maximum. This process is repeated until no more increase in modularity can be achieved by replacing nodes. The second phase of the algorithm involves building a new network whose nodes the communities are found during the first phase. The weights of the links between nodes in the new network are now the sum of the weight of the links between nodes in the corresponding two communities, and the links B. OSN datasets The OSNs we study in this paper includes Flickr, LiveJournal, and YouTube. The data were collected by crawling in late 2006 and 2007 [20]. Of course, these networks have evolved since they were crawled, as would be expected for any real social network. Studying a human social network at a certain point in time can still give us some general insight and knowledge into the human anthropology and sociology. More importantly, we believe that the community characteristics of the data sets captured would be representative of the current states of these OSNs. Table 1 shows huge amount of nods, links and communities in the OSN. C. Hierarchical structure We apply the hierarchical modularity optimization algorithm from Section III on the three data sets from Section IV. The algorithm can handle the whole crawled graphs for Flickr, LiveJournal, and Youtube on a computer with 2GB memory. Previous work has studied proximity communities of several OSNs [22], but the scale is only up to thousands of nodes. Here we study networks of millions of nodes and we believe this can give a better overall picture about the clustering and community properties of these social networks. [2]

5 5 D. Storing communities on mobile devices The routing hints we propose in Section II-B are the community memberships of the person at different levels of hierarchy. Below, we examine whether it is practical to store this information on a mobile device, given the typical communities we find in the previous section. A. Succinctness The mean level-1 community size for the LiveJournal network is 21, which can easily fit into the storage of a mobile device. To store the Bluetooth IDs (6 bytes) of each node in the community on the mobile, we need 126 bytes, which is nothing compared to the size of modern flash memory, but of course each user can have multiple Bluetooth IDs. Even for an extremely large community of 100,000 nodes, we only need 0.6M bytes of storage, again within the reach of a modern mobile phone, which can have a few gigabytes of flash memory. B. Privacy Even if the community is small enough to store on a mobile device, it may not be preferable to store this information on a low-security device such as a mobile phone, due to privacy concerns. A solution to this is to store a one-way hash of the Bluetooth IDs. For instance, with SHA-1, we will only need 160 bits per ID, leading to further compression. [2] V. DISCUSSION & FUTURE DIRECTION A. Improving forwarding efficiency The Bubble algorithm uses both community and centrality information to improve forwarding efficiency [16]. It is possible calculate the centrality value of each node on the social graphs using egocentric approximation [19], and merge together with the hierarchical community information to achieve the same performance as Bubble. In P.Hui, and N. Sastry s paper, the main contribution is not to propose and analyze yet another social-based forwarding algorithm, but to analyze how we can leverage information from OSN social graphs for more stable routing in large-scale PSNs. They do not claim that the algorithm used in this paper to extract hierarchical information is optimal. It is intended merely as a proof-of-concept. There are other methods in the literature that can be used, for example the algorithm by Clauset et al. for extracting hierarchical. [2] B. How people have motivation to transfer information Even the algorithm about social graph works well in theory, there is still one key question: how people have motivation to do that. In my opinion, there are three power may drive the users. First, friendship, sometimes forwarding information takes energy and time, but if you are thinking about helping your friends and making your friends happy, you may have the motivation to do it. Of course, your friends will probably do the same thing for you. Second, common interest and curious, in more situation the users do not know from each others. They will forwarding the information because they have common interest and want to share with people. In this case, the content itself should be attractive. Third, bonus points in virtual world, if you share more information with others or forwarding more data for others, you will have more bonus points or higher rank in the virtual world, then many people are willing to do so. C. How the social graph be used in mobile device The purpose of doing this research is that find the possible direction for DTN. So another key question is how the mobile devices can be connecting with the social graph. As we mentioned in the introduction the principal challenge is that mobile devices have relatively small storage and also the privacy should be protected. Nowadays, the storage can be solved by new technology. The internet file may be an efficiency way for mobile devices connecting with social graph. Many big IT companies such as Nokia, Microsoft, Google and Apple opened personal file for their costumers on the internet. Their costumers can upload and download files and data from internet space and backup their personal information on the internet. The security is better than before, because they save the data on sever instead of personal devices. Mobile device can connect with each other not only with Bluetooth, but also share internet file through wireless network. VI. CONCLUSION To arrange the social resource more efficiency will help improve the usage of DTN. Even there are many problems with the social graph, but still will improve the efficiency of DTN and have a strong potential in the future. REFERENCES [1] T. Hossmann, F. Legendre, and T. Spyropoulos, From Contacts to Graphs: Pitfalls in Using Complex Network Analysis for DTN routing, in ETH Zurich, [2] P.Hui, and N. Sastry, Real World Routing Using Virtual World Information, in Deutsche Telekom Lab and University of Cambridge, [3] P. Juang, H. Oki, Y. Wang, M. Martonosi, L. S. Peh, and D. Rubenstein, Energy-efficient computing for wildlife tracking: design tradeoffs and early experiences with zebranet, in ASPLOS-X, [4] P. Basu and T. D. C. Little, Networked parking spaces: architecture and applications, in Vehicular Technology Conference, [5] P. Hui, A. Chaintreau, J. Scott, R. Gass, J. Crowcroft, and C. Diot, Pocket switched networks and human mobility in conference environments, in WDTN, [6] T. Spyropoulos, K. Psounis, and C. S. Raghavendra, Spray and wait: an efficient routing scheme for intermittently connected mobile networks, in WDTN, [7] H. Dubois-Ferriere, M. Grossglauser, and M. Vetterli, Age matters: efficient route discovery in mobile ad hoc networks using encounter ages, in MobiHoc, [8] A. Lindgren, A. Doria, and O. Schel en, Probabilistic routing in intermittently connected networks, SIGMOBILE Mob. Comput. Commun. Rev., vol. 7, no. 3, pp , July [9] J. Leguay, T. Friedman, and V. Conan, Evaluating mobility pattern space routing for dtns, in INFOCOM, [10] E. M. Daly and M. Haahr, Social network analysis for routing in disconnected delay-tolerant manets, in MobiHoc, [11] P. Hui, J. Crowcroft, and E. Yoneki, Bubble rap: Social-based forwarding in delay tolerant networks, in MobiHoc, May [12] M. E. J. Newman, Analysis of weighted networks, [13] V. D. Blondel, J.-L. Guillaume, R. Lambiotte, and E. Lefebvre. Fast unfolding of communities in large networks. J.STAT.MECH., page P10008, [14] N. Eagle and A. S. Pentland. CRAWDAD data set mit/reality (v ). Downloaded from July [15] L. C. Freeman. A set of measures of centrality based on betweenness. Sociometry, 40:35 41, [16] P. Hui, J. Crowcroft, and E. Yoneki. Bubble rap: Social based forwarding in delay tolerant networks. In MobiHoc 09: Proceedings of the 9th ACM

6 international symposium on Mobile ad hoc networking & computing, May [17] E. P. C. Jones, L. Li, and P. A. S. Ward. Practical routing in delay-tolerant networks. In Proc. WDTN, [18] A. Lindgren, A. Doria, and O. Schelen. Probabilistic routing in intermittently connected networks. In Proc. SAPIR, [19] P. V. Marsden. Egocentric and sociocentric measures of network centrality. Social Networks, 24(4): , October [20] A. Mislove, M. Marcon, K. P. Gummadi, P. Druschel, and B. Bhattacharjee. Measurement and analysis of online social networks. In Proceedings of the 5th ACM/USENIX Internet Measurement Conference (IMC 07), October [21] G. Palla, I. Derenyi, I. Farkas, and T. Vicsek. Uncovering the overlapping community structure of complex networks in nature and society. Nature, 435(7043): , [22] B. Saha and L. Getoor. Group proximity measure for recommending groups in online social networks. In 2nd ACM SIGKDD Workshop on Social Network Mining and Analysis, [23] N. Sastry, K. Sollins, and J. Crowcroft. Delivery properties of human social networks. In Proc. IEEE INFOCOM Miniconference, [24] UCSD. Wireless topology discovery project. edu/wtd/wtd.html, [25] C. Wilson, B. Boe, A. Sala, K. P. Puttaswamy, and B. Y. Zhao. User interactions in social networks and their implications. In EuroSys 09: Proceedings of the fourth ACM european conference on Computer systems, pages , New York, NY, USA, ACM. 6