1 Intelligent E-Commerce Systems Laboratory,

1. DISTRIBUTED COLLABORATIVE FILTERING FOR ROBUST RECOMMENDATION AGAINST SHILLING ATTACKSAE-TTIE JI1, CHEOL YEON1, HEUNG-NAM KIM1, AND GEUN-SIK JO2 1 Intelligent E-Commerce Systems Laboratory, Department of Computer Science & Information Engineering, Inha University {aerry13, entireboy, nami}@eslab.inha.ac.kr 2 School of Computer Science & Engineering, Inha University, 253 Yonghyun-dong, Incheon, Korea 402-751 gsjo@inha.ac.kr

2. INTRODUCTION & BACKGROUNDS • A Robustness Analysis of Collaborative Filtering • Userprofiles made by anonymous unauthenticated users  Vulnerability to Profile Injection Attacks • PocketLens - Distributed Personal Recommender  It can partially improve the effects of PIA from system providers. • Trust in Recommender Systems • But, it is still not safe from anonymous attackers! • “Trust” in Recommender systems • Automated attack detection schemes and robustness of recommendation algorithms. • Correlation between trust and user similarity

3. TCFMA ARCHITECTURETRUST-BASED COLLABORATIVE FILTERING WITH MOBILE AGENTS • Credibility of recommendations • To achieve robustness against shilling attacks Distributed Personal Recommender Web of Trust • Trust Propagation • To overcome sparseness of webs of trust  The Advogato trust metric • Scalability • To raisethe efficiency of distributed computing  Mobile Agent Framework

4. Architecture Fig. 1. Overview of trust-based collaborative filtering with mobile agents

5. THE MEANING OF NOTATIONS Table 1. The meaning of notations

6. TRUST-BASED USER SELECTION • AGENTPo finds the migration path {PATHPo} that includes users trusted by PO for a mobile agent AGENTMPo. • The neighbors of target user PO are chosen from the users included in {PATHPo}. • PO’s personal agent AGENTPo creates a mobile agent, AGENTMPo, to find neighbors and build a similarity model based on them incrementally. • AGENTMPo traces the path recursively until no users exist in {PATHPo}∩{TRUSTPc}. • AGENTMPo is disposed of from the last node after visiting all users in {PATHPo}.

7. Trust-based User Selection • The Advogato maximum flow algorithm • Discover which users are trusted by credible members of an online community and which are not. • The bottleneck property • “the total trust quantity accorded to an s → t edge is not significantly affected by changes to the successors of t” • The minimum number of profiles that make the attack succeed is not included in the process of collaborative filtering.

8. Incremental Model Building • AGENTMPO identifies IOi and IPj that are {ITEMSPO}∩{ITEMSPC} and {ITEMSPC} - {ITEMSPO} respectively, by communicating with a neighbor agent AGENTPC. • For each pair (IOi, IPj), AGENTMPO calculates values and sends the values to its own user agent AGENTPO. (cosine and adjusted cosine similarity)

9. Incremental Model Building • AGENTPO adds up these values incrementally until AGENTMPO sends values of all users in {PATHPO} except for those which don’t have IOi. • AGENTPO calculates the similarity of item pair (IOi, IPj).

10. AGENTS’ TASKS IN EACH CASE Fig. 2. Agents’ tasks in each case

11. RECOMMENDATIONS & FEEDBACK • Predictions • Feedback Fig. 3. Recommendations and propagation user’s feedback

12. DATASETS & EVALUATION METRICS • Datasets • Crawling through epinions.com in May 2006 • http://www.epinions.com • Numeric rating of item is in the range of 1 to 5 • Web of Trust among users • Users who had rated at least 5 item • Users who had expressed trust opinion to at least 25 users • Items that had been rated by at least 10 users Table 1. Dataset for Experiment

13. DATASETS & EVALUATION METRICS • Evaluation Metrics • Mean absolute error (MAE) • Absolute Prediction Shift (APS)

14. Performance Evaluation • Prototype system implemented using IBM aglet Software with JDK 1.4.2 • Benchmark system to compare the performance • Random model building (in PocketLens) • Miller, B., Konstan, J., Terveen, L., Riedl, J.: PocketLens: Towards a Personal Recommender System. In ACM Transactions on Information Systems 22 (2004) 437-476

15. Performance Evaluation • Overall Performance of Prediction Quality • TCFMA + cosine-based scheme showed better prediction quality than the other two methods. • Even a small number of users can result in a relatively better model with our proposed methods Table 2. Overall Performance of Prediction Quality

16. PerformanceEvaluation • Positive Effect of Trust for Prediction • Datasets with users who have more than x trusted users. • The more trust opinions are included in each user, the better the prediction quality obtained. • Direct trust opinions have a positive influence on prediction quality. Table 3. Sensitivity of trust on MAE (neighbor peer size = 50)

17. Performance evaluation • Robustness of the shilling problem • The set of manipulated users including arbitrary 50 ratings were inserted into the training dataset. Fig. 4. Comparison of robustness on manipulated users

18. Performance evaluation • Efficiency of similarity model building • The time required for model building • The number of neighbors required for model building • The proposed method is far superior with respect to the effectiveness of similarity model building. Table 4. Comparison of required time and accessed users (neighbor user size = 50)

19. Conclusion • We proposed a novel TCFMA architecture to solve the problems that can occur in online CF recommender systems related to an improper use of personal information and a profile injection attack. • We obtained very good robustness from malicious attackswithout any degradation of prediction quality, compared to general peer-to-peer CF recommender systems. • We also achieved efficient distributed computing for building item-item similarity models by adding useful functionalities of mobile agents.

20. FUTURE WORK • Trust Decay • The trust relationship becomes weaker as it forwards to its successors. • It is essential to take this phenomenon into consideration for applying trust propagation algorithms to real-world applications. • Attack Detection • Automated attack detection algorithms based on diverse types of attack models can lead to more robust recommendation algorithms.

21. !!!!THANKYOU!!!!

22. Trust Graph Conversion - Advogato Advogato graph transform functiontransform ( G = (V, E, CV)) { set E′ 0, V′ 0; for all x ∈ V do add node x+ to V′ ; add node x- to V′ ; if CV (x) >= 1 then add edge (x-, x+) to E′; set CE′ (x-, x+)  CV (x) -1; for alledge (x, y) ∈ E do add edge (x+, y-) to E′; set CE′ (x+, y-)  ∞; end do add edge (x, supersink) to E′; set CE′ (x-, supersink)  1; end if end do return G′ =(V′, E′, CE′ ); }

23. Capacity assignment

24. Converted graph

25. Trust Propagation & Finding Migration Path Ford-fulkerson maxflow algorithm functionmaxflow (G′, seed, supersink) { for eachedge (x, y) ∈ E′ in G′ do F (x, y)  0; F (y, x)  0; end do while there exists a path P from seed to supersink in the residual Network G′Fdo CF(P) min {CF (x, y) : (x, y) in P}; for eachedge (x, y) in P do F (x, y)  F (x, y) + CF (P); F (y, x)  -F (x, y) end do end while }

26. Examples

27. 4*3+3*1+1*2+4*3 32+12+22+32 42+32+12+42 0.9134 Examples Mi: Model owner’s items Ni: Neighbor’s items Ni Mi Ni Mi

28. 4*3+3*1+1*2+4*3+3*1 32+12+22+32+12 42+32+12+42+32 0.9146 Examples Ni Mi Ni Mi Mi: Model owner’s items Ni: Neighbor’s items

29. 4*3+3*1+1*2+4*3+3*1+1*5 32+12+22+32+12+52 42+32+12+42+32+12 0.7329 Examples Ni Mi Ni Mi