网络专题选讲 华中科技大学 电子与信息工程系 程文青 chengwq@mail.hust.edu.cn 2013年1月 社会网络应用专题选讲 华中科技大学 电子与信息工程系 互联网技术与工程研究中心 黑晓军 Email: heixj@hust.edu.cn Web: http://itec.hust.edu.cn/~heixj 2013.1 Outline Introduction Case study 《网络专题选讲 》 Traffic transport NetTube: Exploring Social Networks for Peer-to-Peer Short Video Sharing, 2009 Incentive P2P Trading in Social Networks: The Value of Staying Connected, 2010 Recommendation Circle-based Recommendation in Online Social Networks, 2012 -3- Internet Topology Introduction 我们生活在一个关系的社会 5 社会网络应用 6 Friend network in Facebook 7 Co-authorship network 8 Co-authorship in network science 9 Ingredient networks 10 911事件——犯罪网络 11 Social Networking(General public) Social Networking(General public) Social Networking(Academia) 专著 专著 研究现状 主要研究机构 国外:MIT、Stanford、Maryland、USC、HP、Michigan 国内:IBM中国研究院、微软亚洲研究院、中科院、中国传媒大学 、清华大学、南京大学 近年来社会网络成为国内外研究热点 美国国家科学基金会(NSF)将社会计算研究领域提供专项资金(2010) 美国计算机协会ACM,Workshop on Social Network Mining and Analysis(2007~2012) WWW会议成立“Social Network and Web2.0 Track”论坛(2009) SIGCOMM,ACM SIGCOMM Workshop on Online Social Networks(2009~2012) EuroSys, Workshop on Social Network System(2009~2012) 互联网测量会议(IMC),海量数据仓库国际会议(VLDB),信息与知识管理 (CIKM)大量关于社交网络文章 全国网络科学论坛(2004~2012),全国复杂网络会议(2005~2011) 17 NetTube: Exploring Social Networks for Peer-to-Peer Short Video Sharing IEEE INFOCOM, 2009 Xu Cheng and Jiangchuan Liu School of Computing Science Simon Fraser University British Columbia, Canada October 2009 Background (1) Social Networked Media Sharing – new killer Internet application Since 2005 Rich user-generated content (UGC) sharing Social networks Among users Among videos Changing the popular culture Background (2) YouTube – a representative Popular Market share of around 43% More than six billion videos viewed in January 2009 Consumed as much bandwidth as the entire Internet in 2000 3rd visits among all Internet sites (after Google and Yahoo) Fast growing 20% growth rate per month 15 hours of new videos are uploaded every minute Motivation (1) The YouTube Crisis – all other sites’ challenge Severely hindered by client/server architecture Bandwidth costs Consumed as much bandwidth as the entire Internet in 2000 $1 million a day for server bandwidth! Sold to Google for $1.65 billion in Nov. 2006 Performance and scalability “Slow” among the surveyed sites by Alexa.com Motivation (2) Peer-to-peer (P2P) – alternative to Client/Server New generation of communication paradigm Scale well with larger user base Each peer contributes its bandwidth to serve others More users, more resources contributed Success already seen in BitTorrent, eMule, eDonkey (file sharing) Video broadcasting … P2P架构 没有永远在线的服务器 任意主机可以同另一个主 机进行通信 peer-peer 节点可以间歇性的连入系 统,IP地址可能会变化 23 对等网络流媒体系统 两大设计空间 如何形成重叠网络? 如何传输内容? 现有体系结构 树状拓扑 + 推式内容传输 ESM, Yoid, CoopNet, SplitStream, Bullet, Chunkspread … 网状拓扑 + 拉式内容传输 《网络专题选讲》 -24- 网状-拉式对等网络流媒体系统 这类系统非常类似于BitTorrent 《网络专题选讲》 -25- 节点软件结构 双缓存 《网络专题选讲》 积极下载 vs 保守下载 处理丢失的数据块 缓存控制 -26- 缓存映像(buffer map) 缓存映像反映了节点缓存所拥有的数据块信息 此映像可以被用来评估用户的播放质量 《网络专题选讲》 -27- 对等网络中的问题 内容组织和搜索 内容传输 信誉、激励及安全相关问题 28 CoolStreaming The first practical large-scale P2P NetTV Origin of data-driven mesh design With many follow-ups: PPLive, PPStream, UUSee … X. Zhang, J. Liu, B. Li, and T.-S. P. Yum, CoolStreaming/DONet: A Datadriven Overlay Network for Live Media Streaming, IEEE INFOCOM'05, March 2005. >800 citations J. Liu, S. G. Rao, B. Li, and H. Zhang, Opportunities and Challenges of Peerto-Peer Internet Video Broadcast, Proceedings of the IEEE, Vol. 96, No. 1, pp. 11-24, January 2008. Data-Driven Mesh Core operations Easy to implement no need to construct and maintain a complex global structure Efficient Every node periodically exchanges data availability information with a set of partners Then retrieves unavailable data from one or more partners, or supplies available data to partners data forwarding is dynamically determined according to data availability Robust and resilient adaptive and quick switching among multi-suppliers Challenges and Opportunities (1) Challenges – Drastically different statistics 1.5 year measurement of 5 million videos http://netsg.cs.sfu.ca/youtubedata/ Short video clips – stability 99.6% are less than 700 seconds “I don’t want to wait for 30 seconds for a two-minute video!” Searching for sources High churn rate: join/leave system Huge number of videos – scalability Highly skewed Inefficient for unpopular videos Very few users watch the same one Challenges and Opportunities (2) Opportunities – Social networks No longer independent – videos have related videos Small-world – strong clustering Important role NetTube Design (1) Bi-layer overlay network Lower-layer – per-video Download and uploading Peers stay in previous overlays as sources Larger and more stable Upper-layer – social network Connected by the same peers in different lower-layer overlays Conceptual relation for searching Social network brings similar peers closer Clustering Efficient searching NetTube Design (2) Bloom filter based indexing An efficient approach to keep track of peers’ cached videos Bloom filter An m-bit array using k hash functions Space-efficient Scalable indexing table Fast searching Table size is scalable with the number of videos Search locally and search in the upper-layer overlay Social network clustering the similar video NetTube Design (3) Transmission scheduling: From which partner to fetch which data segment ? Constraints Data availability Playback deadline Heterogeneous partner bandwidth Rarest-first (BitTorrent’s) doesn’t work ! NetTube Design (3) Variation of Parallel machine scheduling NP-hard Conventional Heuristics Message exchanged Window-based buffer map (BM): Data availability Segment request (piggyback by BM) Less suppliers first Multi-supplier: Highest bandwidth within deadline first NetTube Design (3) Short video ? CODAS: Collaborative Delay-Aware Scheduling NetTube Design (4) Social network assisted pre-fetching Most peers finish downloading before playback ends - free time available (about 80 seconds on average) Using free time to reduce startup delay Prefix pre-fetching Multiple pre-fetching Avoid wasting bandwidth and space Enable multiple pre-fetching Accuracy increases greatly Accuracy increases as watch more videos Pre-fetching among neighbors Easy to implement Social network helps improve efficiency Performance Evaluation (1) Simulation Configuration Based on about 7,000 crawled videos Scale to more than 10,000 heterogeneous clients Compare with PA-VoD (MSN Video) Bandwidth reduction Save significantly more More scalable Performance Evaluation (2) Simulation Impact of social network Find more sources: more than 95% within 2 hops Greatly increase pre-fetching accuracy Performance Evaluation (3) PlanetLab experiment Configuration Maximum 235 PlanetLab nodes Experiment results Server bandwidth reduction: more than 40% Startup delay: average 2.2 s Playback continuity Summary Contribution Techniques First social network assisted P2P system for short video sharing IWQoS’08, INFOCOM’09, IEEE Transactions on Multimedia Bi-layer overlay network Bloom filter based indexing Social network assisted pre-fetching Collaborative delay-aware scheduling Evaluation results Greatly reduce server bandwidth Much lower maintenance cost: $1 million → $60 K Inherently scalable – P2P Greatly reduce playback delay Satisfying startup delay Continuous playback P2P Trading in Social Networks: The Value of Staying Connected IEEE INFOCOM, 2010 Zhengye Liu, Hao Hu, Yong Liu, Keith Ross, Yao Wang, and Markus Mobius Polytechnic Institute of NYU Dept. of Economics, Harvard Unviversity 43 Outline Background: P2P Incentive Networked Asynchronous Bilateral Trading (NABT) NABT Efficiency Theory NABT Simulations Conclusions 44 P2P Apps: BitTorrent 45 P2P Apps: Skype 46 Peer-Assisted Video Streaming Large scale deployments on Internet Leading P2P Video Companies 47 thousands of live/on-demand channels millions of world-wide users daily CoolStreaming PPStream PPLive Sopcast UUSee 1 6 2 5 3 4 Major P2P Issues Traffic localization Security 48 P4P Attacks on Attacks from Lack of uniform API Incentives for peers to contribute resources Partially Successful P2P Incentive BitTorrent is popular 50+ client implementations Dozen public trackers 5-10 million users Why BitTorrent? P2P design Incentives Resources First generation P2P applications: Gnutella 49 + 70% of users are free-riders Second generation P2P applications: BitTorrent The BitTorrent Incentive To get files faster… contribute more bandwidth Implementation of incentive: 50 The rich play/trade with the rich BitTorrent: Tit-for-tat (0) Everyone nominally has four trading partners (1) Alice tries sending to Bob. Is he rich? (2) Alice becomes one of Bob’s top-four providers; Bob reciprocates. (3) Bob becomes one of Alice’s top-four providers. Tit-for-Tat: Live P2P Video To get better video quality… contribute more bandwidth Layer 3 LC31 LC32 LC33 LC34 Layer 2 LC21 LC22 LC23 LC24 Layer 1 LC11 LC12 LC13 LC14 “LayerP2P: Using Layered Video Chunks in P2P Live Streaming”, Z. Liu, Y. Shen, K.W. Ross, S. Panwar, Y. Wang, IEEE Transactions on Multimedia, November 2009. 52 “Substream Trading: Towards an Open P2P Live Streaming System”, Z. Liu, Y. Shen, K.W. Ross, S. Panwar, Y. Wang, Inter Conf on Network Protocols (ICNP), October 2008 Limitations of Tit-for-Tat Tit-for-Tat is synchronous trading Tit-for-Tat == Barter (物物交换) in primitive economy Barter is highly inefficient 53 Alice and Bob can trade if and only if they simultaneously have data for each other in a short time period fails if lack of “double coincidence of wants” failure example: Tit-for-tat does not provide incentive for seeding Currency-based Trading Currency improves trading efficiency in modern economy Asynchronous trading regulated by money 54 users accumulate for providing services and later spend for acquiring services Major Issues/Solutions Solutions: Cheating Counterfeit Dispute Resolution 55 Banking System Market Regulation Trading Policy Court System Law Enforcement …… Global Currency in P2P? Peers trade with each other using digital cash Heavyweight coordination infrastructure needed 56 earn cash by contributing resources to provide services to other peers, pay cash to consume services provided by other peer. banking/regulation/court/enforcement hard to justify for P2P trading goods carrying low value. Only limited research attempts, no large-scale deployment Desirable P2P Incentive Mechanism High Trading Efficiency Cheating-proof isolate and punish cheaters prohibit collusions Low-degree of Coordination 57 trade asynchronously trade with many peers trade diverse set of goods/services light-weight and distributed protocols low management cost Outline Background: P2P Incentive Networked Asynchronous Bilateral Trading (NABT) NABT Efficiency Theory NABT Simulations Conclusions 58 Alternative Trading Systems in Social Networks Asynchronous Trading Networked Trading 59 exploit trust between friends allow debt: providing a service without immediate payment exploit trust in network of friends trade with indirect friends “Trust and social collateral”. Dean Karlan, Markus Mobius, Tanya Rosenblat, and Adam Szeidl. Quarterly Journal of Economics, 2008. Friendship as Trading Collateral (抵押) Resolve cheating/disputes: Terminate friendship! 60 P2P Trading in Social Networks Networked Asynchronous Bilateral Trading (NABT) 61 Social network: peers belong to an underlying social network Pair-wise credit: friends maintain pair-wise credits Asynchronous trading: peers can use their credits anytime they want Credit limit: each peer sets a credit limit for each of its friends Networked trading: peer trades with a remote peer by transferring credits through a chain of friends links. Async Trading Between Direct Friends A pair of friends maintain local credit balance bij = amount of credits b + ∆ that j owes i bij=-bji Alice update balance upon services AB Control risk of defaulting 62 bBA - ∆ Cij = credit limit for j set by i - Cji ≤ bij ≤ Cij incentivizes users Bob Networked Trading via Intermediaries To access service on a remote peer 1. 2. 3. 4. find a path of friend links in social network arrange a series of credit transfers along path intermediaries update credit balances with upstream and downstream friends, and break even remote peer provides b +∆ b -∆ , bBA-∆, b +∆ requested service CB AB BC Alice Bob 63 Charlie NABT Issues NABT is decentralized, and effective for resolving disputes. But Is NABT efficient? How to set credit limits Cij ? Can users free-ride in NABT? 64 Outline Background: P2P Incentive Networked Asynchronous Bilateral Trading (NABT) NABT Efficiency Theory NABT Simulations Conclusions 65 NABT Efficiency Single trade can be exercised if and only if a credit transfer can be arranged Multiple trades coupled through the underlying social network 66 subject to social network connectivity obey credit limit on each social link later trades work with credit balance resulted from earlier trades concurrent trades compete for credit transfer NABT Efficiency Model Given: underlying social network: credit limits as link weights: service demand matrix: : cost charged by user k to serve user l. Find credit transfer flows for all demands resulting credit balance bounded by credit limits 67 : credit flow for demand d on social link <i,j> credit flow conservation on intermediaries NABT Credit Flow Routing Similar to classical network flow problem, but: credit balance on link can be negative credit flows in opposite directions cancel Example: Circular Service Demands: A wants a file on B, B wants a file on C, and C wants a file on A B bAC=1 A bBA=1 B C bCB=1 credit routing scheme 1 68 bAC=0 A bBA=0 C bCB=0 credit routing scheme 2 Balanced Demand For each user k, total service he provides (regardless of receivers) equals total service received (regardless of providers) Theorem 1: Any balanced demand can be executed as long as users involved in the demand sets are connected. NABT is as efficient as global currency 69 networked Tit-for-Tat: peers play tit-for-tat with whole network instead of another peer Unbalanced Demand For at least one user, service contribution does not equal to service consumption. net-service contribution: service sources: service sinks: aggregate net-service imbalance 70 Extended Social Network augment social network with a virtual source, a virtual sink, virtual links Example aggregate net-service imbalance 71 Efficiency with Unbalanced Demand Theorem: An unbalanced demand is executable iff the min-cut between the source s+ and sink s- in extended social network is greater than or equal to the aggregate net-service imbalance. What matters: What does not matter: 72 underlying social network topology credit limits on social links service imbalance between a user and whole network service imbalance between individual pairs of users Dynamic Payment Routing 73 Time is slotted Demands are now sequential H(1), H(2),… Suppose we succeed at executing H(1),…, H(k-1). Theorem: To successfully execute H(k), we do not have to worry about how we executed H(1),…,H(k-1). Outline Background: P2P Incentive Networked Asynchronous Bilateral Trading (NABT) NABT Efficiency Theory NABT Simulations Conclusions 74 Preliminary NABT Protocol Design On-demand credit flow routing Dynamic credit-limit setting 75 locate service providers send out credit-transfer request through controlled flooding request propagates along friends links with enough credit space When request hits one providers, it sends back reply through reverse path to establish credit transfer on intermediaries. Complete credit transfer and service increase credit-limit linearly after each fulfilled transaction decrease credit-limit multiplicatively after each unfulfilled/disputed transaction Simulation Study Trading with global currency (GCT): Synchronous Trading (ST): Two peers can trade if and only if they can supply files to each other simultaneously If peer i downloads a file from peer j, peer j will download a file from peer i. Two-hop NABT: 76 Global currency and a centralized bank Each peer has Bi initial credits and each file costs one credit If peer i downloads a file from peer j, peer i pays 1 credit to peer j Peers are connected in an underlying social network A requesting peer requests files from its friends (one-hop friends) and the friends of its friends (two-hop friends) If peer i downloads a file from peer j within two hops, peer i passes 1 credit to peer j Simulation Setup Peer profile File profile 77 Social network with a topology collected from MySpace Totally 10,000 peers Peer upload bandwidth 37% Ethernet users (1.2Mbps) + 63% residential users (400 kbps) Willingness for sharing 10% content-rich peers (1,000 files) + 90% content-scarce peers (50 files) Online and offline Markov ON-OFF process (On time = Off time = 12 hours) Totally 10,000 different files Files are small and have the same size of 3MB File popularity follows a Zipf distribution Trading Efficiency Request success ratio: The ratio of fulfilled requests to the total number of requests 78 CDF of request success ratio Importance of Trading Intermediaries 79 CDF of request success ratio for the systems with and without intermediaries Service Differentiation of NABT 80 Relation between request success ratio and upload contribution (in terms of number of uploaded files) Conclusion NABT -- a new P2P trading paradigm over social networks NABT is efficient 81 exploits trust between friends, and friends network trade asynchronously, and over network, light-weight, distributed almost as efficient as global currencies support networked tit-for-tat topology and credit limits matters memoryless Open Research Issues incentives for intermediaries isolate and punish cheators dynamic credit-limit setting heterogeneous NABT market 82 diverse set of services exchange ratio between pair-wise credits deal-making …… Take Away Messages Asynchronous incentives are critical for taking P2P to the next level Async incentives require money The future of P2P may lie in social networks 83 Circle-based Recommendation in Online Social Networks ACM KDD 2012 Xiwang Yang, Harald Steck*, and Yong Liu Polytechnic Institute of NYU * Bell Labs/Netflix 84 Outline Background & Motivation Circle-based RS 85 Trust Circle Inference Trust Value Assignment Model Training Evaluation Conclusion & Future work Social Recommenders Everywhere 86 Collaborative Filtering (CF) Most Used and Well Known Approach for Recommendation Finds Users with Similar Interests to the target User Aggregating their opinions to make a recommendation. 87 User Based Collaborative Filtering Target Customer Aggregator u w u ru , i Prediction 88 u wu Item-based Collaborative Filtering 89 Item-Item Collaborative Filtering Aggregator 90 i w i ru , i i Prediction wi Matrix Factorization (BaseMF) [NIPS08] Introduced by R. Salakhutdinov and A. Mnih Probabilistic matrix factorization. In NIPS 2008 Model based approach Latent features for users QR u0 d Latent features for items i d PR0 91 P and Q have normal priors Matrix Factorization (BaseMF) Prediction Model T Rˆ rm Q P P (R | P, Q , ) 2 R [N ( R u ,i 2 | Rˆ u , i ), R ] R I u ,i all u all i [N ( R | rm Q u Pi ), ] T u ,i 2 R R I u ,i all u all i Objective Function 1 2 92 ( u , i ) obs . ( R u ,i 2 2 2 ˆ R u , i ) (|| P || F || Q || F ) 2 P and Q have normal priors Related Work-Social Recommender Social Recommendation (SoRec) Model Social Trust Ensemble (STE) Model 93 SIGIR’09 User’s rating influenced by social friends SocialMF Model CIKM’08 Factorizing social trust matrix together with user rating matrix RecSys’10 User’s latent feature (taste) influenced by social friends Handle trust propagation in social network Using whole trust network for item rating prediction SocialMF [RecSys2010] Social Influence behavior of a user u is affected by his direct neighbors Fu. Latent factor of a user depend on his neighbors. * S u ,v is the normalized trust value. Prediction Model: Objective: 1 2 2 ( R u , i Rˆ u , i ) ( u , i ) obs . * * T ( Q u S u , v Q v )( Q u S u , v Q v ) 2 all u v v 2 94 (|| P || F || Q || F ) 2 2 Proposed Improvements for Current Social Recommender Social networks include multiple circles A more refined social trust information—richer information Incorporate circle information in Social Recommender Use trust circles specific to an item category when predict rating in this category 95 e.g. Trust Circle of “Music”, Trust Circle of “Cars”, etc Proposed Improvements for Current Social Recommender Existing circles (Google+, Facebook) not corresponding to an item category 96 Proposed Improvements for Current Social Recommender 97 In existing multi-category rating datasets, no circle information User trusts different subsets of friends in different domains (Cars, Music…) User trusts different friends differently, related to friend’s expertise value Should use trust circle specific to item category Outline Background & Motivation Circle-based RS 98 Trust Circle Inference Trust Value Assignment Model Training Evaluation Conclusion & Future work Trust Circle Inference User v is in inferred circle c of u iff u trust v in original social network and both of them have rating in category c Original Social Network Inferred circle for category C1 99 Inferred circle for category C2 Inferred circle for category C3 Outline Background & Motivation Circle-based RS 100 Trust Circle Inference Trust Value Assignment Model Training Evaluation Conclusion & Future work Trust Value Assignment 101 CircleCon1: Equal Trust Trust Value Assignment CircleCon2: Expertise-based Trust 102 assign a higher trust value or weight to the friends that are experts in the circle / category. CircleCon2: Expertise-based Trust Variant a: 103 Expertise based on number of ratings in a circle CircleCon2: Expertise-based Trust Variant b: (c) Ev N v v (c) (c) Dw records the proportions of ratings user w assigned in all categories. It reflects the interest distribution of w cross all categories 104 CircleCon3: Trust Splitting Original trust link 105 trust link in c1 trust link in c2 Trust due to followee’s rating in one category Likelihood u2 trusts u1 in C1, C2 ? Infer likelihood proportional on u2’s number of ratings in C1 and C2. Assign trust value in a category proportional to the likelihood u2 trusts u1 in a category CircleCon3: Trust Splitting N 106 c1 u1 9, N c2 u1 1 S u 2 1, u1 0.9, S u 2 2, u1 0.1 (c ) (c ) Normalize across followees Outline Background & Motivation Circle-based RS 107 Trust Circle Inference Trust Value Assignment Model Training Evaluation Conclusion & Future work Model Training Training with ratings from each category Predict user’s rating in category c Input rating: rating in category c Input social network: Circle c (c) (c) (c) ( c )T Rˆ u , i rm Q u Pi L (c ) 1 2 (R ,Q (R (c ) (c) u ,i ,P (c ) ,S ) ( c )* is social information weight (c) 2 Rˆ u , i ) ( u ,i ) o b s . 2 (c ) 2 (c) (Qu u a ll S ( c )* u ,v (c) Qv v (c) )( Q u S ( c )* u ,v v T ) (|| P (c) || 2 F || Q (c) 2 F || ) i (c) 0 d u0 d ,Q R P R i0( c ) is the number of items in category c (c) Solved by gradient descent 108 (c) Qv (c) Model Training 109 Training with ratings from each category Model Training Training with ratings for all categories Predict user’s rating in category c Input rating: rating from all categories Input social network: Circle c L (c ) 1 2 (R,Q (c ) 110 (c ) ( c )* ,S ) 2 ( R u , i Rˆ u , i ) ( u ,i ) o b s . (c) ( Q u 2 a ll u (|| P (c) 2 P ,P (c) 2 ,Q (c) v (c) (c) )( Q u v || F || Q i d R0 ( c )* S u ,v Q v (c) 2 || F ) R u0 d ( c )* (c) S u ,v Q v T ) Outline Background & Motivation Circle-based RS 111 Trust Circle Inference Trust Value Assignment Model Training Evaluation Conclusion & Future work Epinions Data 112 Performance Metrics RM SE ( u , i ) R test 2 ( R u , i Rˆ u , i ) | R test | M AE ( u , i ) R test | R u , i Rˆ u , i | | R test | 113 Training with per-category ratings 114 Training with per-category ratings L (c ) (R (c ) ,Q (c ) ,P (c ) ,S ( c )* ) (c) (c) 2 (c) 2 (c) 2 ( R u , i Rˆ u , i ) (|| P || F || Q || F ) 2 ( u ,i ) o b s . 2 1 2 115 (c) (Qu u a ll v ( c )* (c) S u ,v Q v (c) )( Q u v ( c )* (c) S u ,v Q v T ) Training with ratings from all categories 116 CircleCon3 of training with per-category rating Training with ratings from all categories 117 Training with ratings from all categories 118 Summary Propose a novel Circle-based Social Recommendation framework Split original social network to different circles, one circle corresponding to one item category User trusts different subsets of friends in different domains(Cars, Music…) User trusts different friends differently, based on friend’s expertise Outperforms the state-of-the-art social collaborative filtering algorithms Show the promising future of circle-construction techniques in Social Recommender 119 小结 Social networking has been changing the way which people communicate! 120 Reading List Lada Adamic, Social Network Analysis, https://class.coursera.org/sna-2012001/wiki/view?page=syllabus World By David Easley and Jon Kleinberg, Networks, Crowds, and Markets Reasoning About a Highly Connected, Cambridge University Press, 2010 http://www.cs.cornell.edu/home/kleinber/networks-book/ Xu Cheng and Jiangchuan Liu, "NetTube: Exploring Social Networks for Peer-to-Peer Short Video Sharing", IEEE INFOCOM, 2009. Zhengye Liu, Hao Hu, Yong Liu, Keith Ross, Yao Wang, and Markus Mobius, “P2P Trading in Social Networks: The Value of Staying Connected”, in the Proceedings of IEEE Conference on Computer and Communications IEEE INFOCOM, 2010 Xiwang Yang, Harald Steck and Yong Liu, “Circle-based Recommendation in Online Social Networks ”, in the Proceedings of ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2012), Long Paper, August, 2012 -121-