SOFIA: Social Filtering for Niche Markets

Transcription

1 Social Filtering for Niche Markets Matteo Dell'Amico Licia Capra University College London UCL MobiSys Seminar 9 October 2007 : Social Filtering for Niche Markets

2 Outline 1 Social Filtering Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern 2 From HITS to 3 Datasets Hidden Judgements Sybil Attacks : Social Filtering for Niche Markets

3 Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern The Long Tail Chris Anderson, 2006 Digital distribution: millions of dierent products are available to consumers. An enormous market for niche content is appearing. Users need help to nd interesting content. Filters are essential to connect supply and demand. Our Problem Creating an ecient and robust lter. : Social Filtering for Niche Markets

4 Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern The Long Tail Chris Anderson, 2006 Digital distribution: millions of dierent products are available to consumers. An enormous market for niche content is appearing. Users need help to nd interesting content. Filters are essential to connect supply and demand. Our Problem Creating an ecient and robust lter. : Social Filtering for Niche Markets

5 Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern Outline 1 Social Filtering Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern 2 From HITS to 3 Datasets Hidden Judgements Sybil Attacks : Social Filtering for Niche Markets

6 Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern Collaborative Filtering Which items might I like? Let's look at what similar users did. Similarity in reviews, behaviour... They are competent: they express (subjective!) judgements we agree with. : Social Filtering for Niche Markets

7 Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern Propagating Trust: Competence Alice expressed judgement X (I like eating at SOAS). Bob agrees with Alice on X, therefore Alice ranks Bob as a competent evaluator. Bob also expressed judgement Y (They make good burgers at ULU). Alice decides to trust Bob's advice and tries ULU. : Social Filtering for Niche Markets

8 Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern Sybil Attack Also known as... How to trick Alice? Create lots of false users (Sybils) that copy Alice's judgements. All Sybils vote for a malicious judgement they want to increase the ranking of. Since the Sybils look competent to Alice, she will trust them. Prole injection, shilling (CF), web spam (webpage ranking). In Social Filtering, Alice leverages on her social ties to isolate Sybils. : Social Filtering for Niche Markets

9 Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern Sybil Attack Also known as... How to trick Alice? Create lots of false users (Sybils) that copy Alice's judgements. All Sybils vote for a malicious judgement they want to increase the ranking of. Since the Sybils look competent to Alice, she will trust them. Prole injection, shilling (CF), web spam (webpage ranking). In Social Filtering, Alice leverages on her social ties to isolate Sybils. : Social Filtering for Niche Markets

10 Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern Outline 1 Social Filtering Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern 2 From HITS to 3 Datasets Hidden Judgements Sybil Attacks : Social Filtering for Niche Markets

11 Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern Propagating Trust: Intent Web of Trust: a social network where A links to B if A trusts B to behave honestly. Created explicitely by users (e.g., Facebook) or automatically (e.g., logs). Trust Transitivity: I trust the friends of my friends. Alice thinks Bob is honest. Bob recommends Charlie to Alice. Since Alice trusts Bob, she decides to trusts Charlie as well. Iteratively, Alice derives trust for Dave. : Social Filtering for Niche Markets

12 Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern Isolating Sybils There is no way to recognize legitimate users only by looking at their judgements. It is costly for the attacker to convince honest users to trust it. A small number of honest users are connected to the Sybil network via attack edges (Yu et al., ACM SIGCOMM '06). We can isolate Sybils if we limit the amount of trust propagated through the attack edges. : Social Filtering for Niche Markets

13 Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern Discussing Trust Transitivity Pro Users get trusted if they behave honestly. If reciprocative behaviour is adopted, the rational choice for selsh users is to behave honestly (Feldman et al., ACM EC '04). Sybils can get isolated. Con Trust transitivity does not take into account the tastes of the users. This is a big problem in niches, where subjectivity is extreme. : Social Filtering for Niche Markets

14 Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern Discussing Trust Transitivity Pro Users get trusted if they behave honestly. If reciprocative behaviour is adopted, the rational choice for selsh users is to behave honestly (Feldman et al., ACM EC '04). Sybils can get isolated. Con Trust transitivity does not take into account the tastes of the users. This is a big problem in niches, where subjectivity is extreme. : Social Filtering for Niche Markets

15 Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern Outline 1 Social Filtering Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern 2 From HITS to 3 Datasets Hidden Judgements Sybil Attacks : Social Filtering for Niche Markets

16 Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern Propagating Trust: Social Filtering We trust users who are both willing and able to give good judgements. Alice trusts Dave's intent because a path in the web of trust connects her to him. She trusts his competence because they agree on X. Since Dave is honest and competent, Alice trusts his judgement Y. : Social Filtering for Niche Markets

17 From HITS to Outline 1 Social Filtering Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern 2 From HITS to 3 Datasets Hidden Judgements Sybil Attacks : Social Filtering for Niche Markets

18 From HITS to PageRank Google's algorithm to rank the importance of web pages. Intuitive consideration: an authoritative page is linked by many authoritative pages. A random surfer following links at random is more likely to stumble in more important pages. With 1 α probability of stopping at each step, PageRank computes the probability that the random surfer stops at any given page. In Webs of Trust The same principle applies: reputable users are recommended by other reputable users. We swap the WWW graph with the social network. : Social Filtering for Niche Markets

19 From HITS to PageRank Google's algorithm to rank the importance of web pages. Intuitive consideration: an authoritative page is linked by many authoritative pages. A random surfer following links at random is more likely to stumble in more important pages. With 1 α probability of stopping at each step, PageRank computes the probability that the random surfer stops at any given page. In Webs of Trust The same principle applies: reputable users are recommended by other reputable users. We swap the WWW graph with the social network. : Social Filtering for Niche Markets

20 From HITS to Personalized PageRank PageRank does not take into account subjectivity, which is essential to isolate Sybil nodes. We force the random walk to start in the evaluating node: this assures that the walk starts at a honest node. The trust obtained by Sybil nodes is limited by the probability of following an attack edge. : Social Filtering for Niche Markets

21 From HITS to Personalized PageRank - The α Parameter (1) α: probability that our random walk continues at each step. Low α implies shorter paths. Pro: Con: Fast convergence Close social ties may have related tastes (i.e., my friends listen to similar music) We don't trust honest users because they're socially far away. : Social Filtering for Niche Markets

22 From HITS to Personalized PageRank - The α Parameter (2) High α implies longer paths: Pro: Con: We have more information about nodes. Attack edges are more likely to be traversed: lower attack resilience. : Social Filtering for Niche Markets

23 From HITS to Outline 1 Social Filtering Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern 2 From HITS to 3 Datasets Hidden Judgements Sybil Attacks : Social Filtering for Niche Markets

24 From HITS to HITS: the Idea Jon Kleinberg, JACM 1999 Web pages are seen as hubs and authorities: authorities are the authoritative pages; hubs are pages that link to authorities. Good hubs point to good authorities; good authorities are pointed by good hubs. In our case Users instead of hubs; judgements instead of authorities. : Social Filtering for Niche Markets

25 From HITS to HITS: the Algorithm We have a bipartite graph with hubs/users (circles) and authorities/judgements (squares). All hubs start with the same weight. Iteratively, until convergence: Weights on authorities are the sum of weights on all hubs that link them; Weights on hubs become the sum of weights on authorities they link; Weights on hubs get renormalized. : Social Filtering for Niche Markets

26 From HITS to HITS: Example (1) Initialization Weights on hubs get initialized. : Social Filtering for Niche Markets

27 From HITS to HITS: Example (2) Forward step Weigths on authorities are the sum of hubs who link them. : Social Filtering for Niche Markets

28 From HITS to HITS: Example (3) Backward step Weigths on hubs are the sum of linked authorities. : Social Filtering for Niche Markets

29 From HITS to HITS: Example (4) Normalization Weigths on hubs get renormalized. Back to the Forward Step. : Social Filtering for Niche Markets

30 From HITS to Outline 1 Social Filtering Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern 2 From HITS to 3 Datasets Hidden Judgements Sybil Attacks : Social Filtering for Niche Markets

31 From HITS to Change 1: Tightly Knit Communities SALSA: Lempel and Moran, 2001 HITS rewards disproportionately communities where users and judgements are highly correlated. In the graph on the left, the ranking of nodes in the less dense blue community goes to 0. Fix: perform a random walk on the judgement graph and compute the equilibrium distribution. Side eect: niche judgements are rewarded, since their weight is redistributed to less nodes. : Social Filtering for Niche Markets

32 From HITS to Change 2: Subjective Ranking Problem The results of HITS are independent from tastes of the evaluating node. It is essential to have personalized results. Fix Same approach as in PageRank: we start the random walk from the evaluating node. To reward shorter paths, we stop at each iteration with probability 1 β. Low β implies higher subjectivity and faster convergence. High β favours longer paths of trust propagation. : Social Filtering for Niche Markets

33 From HITS to Change 2: Subjective Ranking Problem The results of HITS are independent from tastes of the evaluating node. It is essential to have personalized results. Fix Same approach as in PageRank: we start the random walk from the evaluating node. To reward shorter paths, we stop at each iteration with probability 1 β. Low β implies higher subjectivity and faster convergence. High β favours longer paths of trust propagation. : Social Filtering for Niche Markets

34 From HITS to Change 3: Take Intent into Account Problem Fix As said before, we don't want to trust dishonest nodes. Culprit for HITS: backwards step. The fact that a user expressed a judgement does not insure they are well intentioned. 1 Compute intent ranking using Personalized PageRank. 2 Redistribute trust to users proportionally to their intent ranking. : Social Filtering for Niche Markets

35 From HITS to Change 3: Take Intent into Account Problem Fix As said before, we don't want to trust dishonest nodes. Culprit for HITS: backwards step. The fact that a user expressed a judgement does not insure they are well intentioned. 1 Compute intent ranking using Personalized PageRank. 2 Redistribute trust to users proportionally to their intent ranking. : Social Filtering for Niche Markets

36 From HITS to in Synthesis : SOcial FIltering Algorithm HITS-like trust propagating algorithm. 3 key modications: 1 Random walk trust propagation as proposed in SALSA 2 The starting point is the evaluating node; the random walk continues at each step with probability β. 3 In the backward step, trust is redistributed from judgements to users according to their intent ranking computed using Personalized PageRank on the web of trust. : Social Filtering for Niche Markets

37 Datasets Hidden Judgements Sybil Attacks Outline 1 Social Filtering Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern 2 From HITS to 3 Datasets Hidden Judgements Sybil Attacks : Social Filtering for Niche Markets

38 Datasets Hidden Judgements Sybil Attacks CiteSeer Large dataset of scientic collaborations. Social network: co-authorship data. Authors A and B are connected if they wrote papers together. Judgements: citations. Graph data If X cites Y, the implicit judgement is Y is relevant to X's topic. A highly clustered subset of the whole graph. 10,000 authors. 182,675 papers. : Social Filtering for Niche Markets

39 Datasets Hidden Judgements Sybil Attacks Last.fm Social networking website devoted to music. Social network: friend lists. Same as Facebook, MySpace,... Judgements: most listened artists chart for each user. Implicit judgement: I like to listen to songs by X. Graph data A BFS crawl of 10,000 users. 51,654 dierent artists. : Social Filtering for Niche Markets

40 Datasets Hidden Judgements Sybil Attacks Outline 1 Social Filtering Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern 2 From HITS to 3 Datasets Hidden Judgements Sybil Attacks : Social Filtering for Niche Markets

41 Datasets Hidden Judgements Sybil Attacks Hidden Judgements How to evaluate the accuracy of 's ranking on judgements? We want to rank highly the judgements that a user would approve. We hide a random judgement and execute. If the algorithm performs well, the hidden judgement will have a high ranking. In Citeseer, we try to guess a missing citation from a paper. In Last.fm, we try to nd the missing artist in a chart. : Social Filtering for Niche Markets

42 Datasets Hidden Judgements Sybil Attacks Hidden Judgements - Citeseer no intent ranking Personalized PageRank Ratio Rank Medians: 4 (), 12 ( - no intent ranking), 30 (Personalized PageRank). : Social Filtering for Niche Markets

43 Datasets Hidden Judgements Sybil Attacks Hidden Judgements - Last.fm no intent ranking Personalized PageRank Ratio Rank Medians: 174 (), 157 ( - no intent ranking), 344 (Personalized PageRank). : Social Filtering for Niche Markets

44 Datasets Hidden Judgements Sybil Attacks Outline 1 Social Filtering Competence: Taste Similarity Intent: Trust Transitivity The Social Filtering Pattern 2 From HITS to 3 Datasets Hidden Judgements Sybil Attacks : Social Filtering for Niche Markets

45 Datasets Hidden Judgements Sybil Attacks Sybil Attack We simulated an attack trying to inate the rating of a malicious judgement X on a victim node A. A coalition of 100 Sybil nodes is created. All Sybils copy A's judgements, then add a link to X. We study how the ranking of X changes before and after the attack, on the victim node A and on other nodes. : Social Filtering for Niche Markets

46 Datasets Hidden Judgements Sybil Attacks Sybil Attack - Last.fm (1) Attack Percentiles Algorithm edges Role Any no attack 12,914 25,827 38,741 - no intent victim other 348 1,185 3, ,730 20,493 33,322 Pers. PageRank 10 4,759 8,757 13, ,092 2,012 3,101 1 victim 3,406 11,182 31,765 other 9,599 19,186 33, victim 469 1,311 2,815 other 4,612 8,779 14, victim other 1,040 2,649 5,571 : Social Filtering for Niche Markets

47 Datasets Hidden Judgements Sybil Attacks Sybil Attack - Last.fm (2) Attack Percentiles Algorithm edges Role (α = 0.9) 100 (α = 0.5) 100 victim other 1,040 2,649 5,571 victim other 1,578 3,106 5,128 Tradeo between accuracy and attack resilience. : Social Filtering for Niche Markets

48 Datasets Hidden Judgements Sybil Attacks Conclusions Social Filtering Integrating information about social networks and subjective preferences we obtain recommendations that are: Accurate (due mainly to preferences) Attack resilient (thanks to social networks). Incorporating social network may increase accuracy. A particular implementation of Social Filtering. Future Work P2P/mobile decentralised implementation Other social ltering algorithms? : Social Filtering for Niche Markets