You are on page 1of 38

Social Recommender System

By: Ibrahim Sana

15.08.08

Agenda 
Introduction  Background on Collaborative Filtering  Collaborative Filtering Limitation  Using trust in RS  Related works  Research methodology  Evaluation and Results  Conclusion

Introduction 
Recommender system (RS) help users find items

(e.g., news items, movies) that meet their specific needs.  Motivation 

Information overload 

Researches in RS focused on developing methods

and approaches dealing with the Information overload problem.  Main Approaches 
 

Content-Based (Salton, 1989) Collaborative filtering/Social Filtering (Goldberg, 1992 ) hybrid

i w a . Rating prediction: n § p a . Select a subset of the users (neighbors) to use as predictors (recommenders).i  ru ) ! ra  u !1 § u !1 a .Collaborative Filtering (CF)  In the real world we seek advices from our trusted people  CF automate the process of ³word-of-mouth´  General use:    Weight all users with respect to similarity with the active user.u n ( ru w .u .

User-User Collaborative Filtering Active user 3 ? Rating prediction .

 Tends to recommend popular items .CF Limitation  New item problem  Cold start problem  Sparsity (95%-99%)  Controversial user  Easy to attacks  Scalability  Cannot recommend items to someone with unique tastes.

2006)  Molskiing (Massa et al.Solution: using trust relationships  Implicit: Deriving trust score directly from the rating data  Generally based on user prediction accuracy in the past  Explicit: users explicitly ³rate´ other users  FilmTrust (Hendler et al.2005)  Limitation:  Users have on average very few links (trusted sources)  More User¶s effort  Solution  Trust propagation: find unknown user¶s trustworthiness based on the users¶ ³web of trust´ .

1998).Trust inference  Global metrics: computes a single global trust b value for every single user (reputation) 1  Examples:  5 a 2 3 d c 3 PageRank (Page et al.eBuy  Pros:   Based on the whole community opinion Simple to compute Trust is subjective (controversial users)  Cons:  .

Local trust metrics  Local metrics: predicts (different) trust scores that are personalized from the point of view of every single user  Example:   MoleTrust (Massa et al.2006) TidalTrust (Golbeck et al.2005) 1  Pros:  More accurate  Attack resistance  Cons:  Ignoring the ³wisdom of the crowd´  More complicated b ? 5 a 2 d c 3 .

~665K reviews 487K binary trust statement Sparsity=99.Related works(1):Massa et al(2006)  Crawling Epinion.~140K items.  Users can also express their ³Web of Trust´ and their Black list  Dataset:    ~50K users.99135%  Above 50% are cold start users (less than 5 review) .com  users can review items and also assign them numeric ratings in the range 1 to 5.

i  ru ) a .i ! ra  u !1 a .u n ( ru .u §w u !1 .Recommendation method  Using MoleTrust metric Input Estimated trust userXuser Rating MXN Rating predictor output Predicted Ratings MXN n §w pa .

Evaluation and results .

250 rating 350 users with social connection Sparsity=77% . write reviews. and express trust statements in other users based on how much they trust their friends about movies ratings  Rating scale from half start to four start  Trust scale from 1 to 10  Dataset:    500 users.Related works(2):Golbeck et al(2006)  FilmTrust: Online Recommender System  Users can rate films. 11. 100 popular movies.

Recommendation method     Weight ratings by trust value Search recursively for trusted sources Using TidalTrust metric for trust inference Simple Prediction method Alice trust Bob 9 Alice trust Chuck 3 Bob rates the movie ³Jaws´ with 4 stars Chuck rates the movie ³Jaws´ with 2 stars §t i S iS  Example: si im r rsm ! §t si Alice¶s predicted rating for ³Jaws´ is: (9*4+3*2)/9+3=3.5  .

Evaluation and results  Benchmarks: Pure CF and simple average  80% training and 20% testing  Using MAE metric  First analysis. using trust didn¶t appear to be effective Above 50% of the rating were within the range of the mean +/.half star  Trust-based significantly useful only to user who disagree with the average  .

Result .

Limitations  Do not distinguish between various types of social relationships  Researches in marketing and in applied psychology identified different types of social measures impact recipient¶s advice taking  Different types of social relations impact recipient¶s advice taking in different ways .

Dominants Social Measures  Cognitive similarity (Gilly et al. 1998) . 2005)  Competence  Benevolence  Integrity  Social Capital/Reputation (Gilly et al. 1998)  Tie-Strength (Levin & Cross 2004)  Relationship duration  Interaction frequency  Closeness  Trust (Smith et al.

Motivation  Web 2.0 provide opportunity for peoples to interact with each other    Social networks (trust. friendships) Electronic communications (Tie-Strength) Reputation mechanisms (Social Capital) .

Research questions  Can additional relationship information be utilized to enhance recommender system performance?  What types of social relation is most useful? .

Objectives  Identify the difference between similarity based CF and social based CF relations  Explore the contribution of various social  Suggest solution for the cold start problem  Suggest solution for the scalability problem .

Alternative Hypothesis: different social relationships provide similar contribute to the performance of recommender systems H3:Null Hypothesis: different social relationships provide different contribution to the performance of recommender systems.Hypothesis    H1:Null Hypothesis: social relationships don¶t provide any contribution to the performance of recommender systems Alternative Hypothesis: social relationships do contribute to the performance of recommender systems H2:Null Hypothesis: different social relationships provide different contribution to the performance of recommender systems. Alternative Hypothesis: different social relationships provide similar contribute to the performance of recommender systems .

Social dimensions and measurement Social dimension Trust Friendship Interaction Frequency Relationship Duration Social capital Measurement I trust this person I would consider this person a friend How often did you communicate with this person How long have you known this person This person is reputable .

Research Method  Domain: movie recommendation  Subject : 97 4th years student from the IS department (with social relationships)  Tasks:   Provide rating for 160 (popular) items (5 point scale) Select three subject and indicate your social relationships       Some of the relationships we examined Trust Friendship Interaction duration Interaction frequency Reputation .

ovies rating 2-Social net ork building overage ethod .Research method esearch Independent Variables (reco endation ethods) Baseline: ser-based F (Pearson orrelation) ybrids ethod (Si ilarity and Social relations) ( o bination sche es) Social estriction ethod (Pearson orrelation) ependent Variables (Perfor ance) AE Precision and ecall ontrol Variable Subjects Students ith social relationships Tasks 1.

Experiment Environment User Authentication Task1: Movies rating Task2: User's social relationships .

frequency Reputation (Recipient¶s) Sources¶ Qualification s System¶s Source Qualification Component .Research framework Past Ratings Recipient Sources System¶s ReceiverSource Similarity Calculation Systems Prediction Component System¶s Prediction (Recommendation) Recipient-Source similarity Trust. Friendship Interaction duration.

........... u ! ¯ f ± ( Pa ......u ).....if (u  P) ± S ± a ........ S a .............Prediction method 1  Hybrid method     Social relations combined with similarity (Pearson Correlation) Tuning the source¶s weight according to his group Group P: sources similar to the active user Group S: sources belong to the social network of the active user S P P ® a ....if (u  S ) Wa...u ......if (u  P ‰ S ) ±.....Other ise 0 ° ...............u ........u .......

u..if (u  P ‰ S) .Prediction method 2  Social restriction    Social relations used for restriction Consider only sources belong to both groups S and P Using the source¶s similarity S Wa . u ! a ....

Social-based Prediction Prediction item i to user u .

Simulation System Architecture .

82 86 84 0.7405 0.78 MAE Coverage 0.7 0.82 0.72 0.78 0.74 0.8 0.8 0.742 MAE MAE coverage Precision MAE Recall 0.742578 0.76 0.7445 0.7415 0.72 0.76 0.743582 0.74 d=5 d=4 Prediction method d=2 20 d=2 d=1 d=6 70 d=5 d=4 d=3 80 60 50 d=340 30 10 Prediction method ocial tie Prediction method eight pure CF pure CF d=1 0 .7 70 pure CF 0.Results (Hybrid method) oci ocial Weight Im pact ei htin coverage Social weighting ocial eighting 0.68 76 74 0.84 0.744 0.741 72 d=6 100d=690 d=5 d=4 d=3 d=2 d=1 0.743 80 0.74 0.

4 0.6 0.2 0 25 23 21 19 17 15 13 11 9 7 5 3 1 MAE Number of sources MAE-CF MAE-WAA1 .Hybrid method: Cold start users Impact of shared-interest sources 1.8 0.4 1.2 1 0.

746152 0.743428 0.795776 0.798522 0.497685 Improvements .742796 0.741061 0.750318 0.749967 0.18684322 10.794472 0.822165 0.795426 0.731773 0.979698 8.Impact of different social measures Social measures Cognitive similarity Tie-Strength relationship duration interaction frequency Closeness Trust competence benevolence Integrity Social capital AMAE 0.746972 0.744768 0.795328 0.738428 0.748337 0.7585475 9.74938 0.852814 10.741934 0.795456 0.474944 9.797798 0.744079 Precision 0.738412 0.746838 0.797383 Recall 0.752611 0.65368 8.1849155 10.162064 9.796085 0.736044 0.796603 0.74192 9.

7 pureCF pure CF Cover ge MAE .8 0.72 30 0.8 Precision 0.82 80 0.82 70 60 0.78 0.74 20 0.7 0.Result (Social restriction) Socialre triction oci restriction oci re trictin cover ge 0.72 10 0.78 MAE Recall coverage d=6 d=6 d=6 d=5 d=5 d=5 d=4 d=3 d=2 d=4 d=3 d=2 d=4 Prediction method d=3 d=2 Prediction method Prediction method d=1 d=1 d=1 0.76 0.84 0.74 40 0.66 0 pure CF 0.68 0.76 50 0.

M 0. 0 Number of sources . C .Social restriction: cold start users Shared i erest sources i act .

Conclusion  Social relationships is effective in alleviating CF weaknesses:    Cold start problem (Social weighting and social restriction) Scalability problem (Social restriction) Spammers attacks (Social weighting and social restriction) .

Tiella. Italy..: Moleskiing: A Trust-Aware Decentralized Recommender System. 5±53(2004) Massa. Proceedings of the Fourth International Conference on Trust Management. J. 492±508 (2004)1060 C.:Raghavan. and A. Tomkins. ACM Transactions on Information Systems 22..: Social Information Filtering: Algorithms for Automating ¶Word of Mouth¶. of the Thirteenth International World Wide Web Conference. Agia Napa. Kumar. P. Chen Avesani.. Maes. pp. Guha.. P.. Terveen.References  Shardanand. Cyprus.10±217 (1995) Herlocker. pp.A. May 2006.In: Proceedings of the International Conference on Cooperative Information Systems (CoopIS).. J. P. Riedl. R. J. U. L. Pisa. P.: Evaluating Collaborative Filtering Recommender Systems. Massa.: Trust-Aware Collaborative Filtering for Recommender Systems. Hwang and Y.-S. MAY 2004. Konstan. Galway. Propagation of trust and distrust. In: Proceedings of the First Workshop on Friend of a FriendSocial Networking and the Semantic Web.      . P. R. Ireland (2004) Golbeck. R.. In: Proceedings of Human Factors in Computing Systems.-P. P. J: Generating Predictive Movie Recommendations from Trust in Social Networks. In Proc. Avesani.