2. neal lathia (2nd yr phd): stephen hailes & licia capra department of computer science, university college london intern @ telefonica i+d: xavier amatriain
3.
4.
5. source: O. Celma & P. Lamere “Music Recommendation Tutorial” ISMIR 2007 recommender systems
6.
7. … a method to classify content correctly data predicted ratings intelligent process our focus: k-nearest neighbours (kNN)
8. ? kNN collaborative filtering 1. find neighbours 2. make predictions 3. recommend source: N. Lathia “Computing Recommendations with Collaborative Filtering” To appear, Collaborative and Social Information Retrieval and Access: Techniques for Improved User Modeling
9. e-commerce machine learning computer supported collaborative work human computer interaction mobile systems m-commerce agents artificial communities distributed systems trust management reputation systems social networks who does the research?
10. e-commerce machine learning computer supported collaborative work human computer interaction mobile systems m-commerce agents artificial communities distributed systems trust management reputation systems social networks example?
11. distributed content dissemination via trustworthy peers filter out the spammers source: D. Quercia, S. Hailes, L. Capra “Lightweight Distributed Trust Propagation” ICDM 2007
13. bootstrap cooperation with privacy source: N. Lathia, S. Hailes, L. Capra “Private Distributed Collaborative Filtering Using Estimated Concordance Measures” RecSys 2007, Minn, USA method: estimate transitive similarity evaluation: netflix dataset & simulated profiles
14. effect on simulated profiles High Estimation Error high correlation estimation error, but prediction accuracy remains?
15. similarity values depend on the method used: there is no agreement between measures [2] [3] [1] [5] [3] [4] [1] [3] [2] [3] my profile your profile pearson -0.50 weighted- pearson -0.05 cosine angle 0.76 co-rated proportion 1.00 concordance -0.06 bad near zero good very good near zero
16. … the pearson distribution intelligent process
17. … the modified pearson distributions weighted-PCC, constrained-PCC
18. … and other measures intelligent process somers’ d, co-rated, cosine angle
19. java.util.Random r = new java.util.Random() for all neighbours i { similarity(i) = (r.nextDouble()*2.0)-1.0); } … what if..? source: N. Lathia, S. Hailes, L. Capra “The Effect of Correlation Coefficients on Communities of Recommenders” ACM SAC TRECK 2008, Fortaleza, Brazil.
22. a) our error measures are not good enough? b) is there something wrong with the dataset?
23. a) our error measures are not good enough? b) is there something wrong with the dataset? c) current user-similarity is not strong enough to capture the best recommender relationships? [NO TRUST]
24. trust enables interaction in environments that are characterised by uncertainty read more: N. Lathia “Learning to Trust on the Move” International Workshop on Trust in Mobile Environments (TIME). IFIPTM, June 2008, Trondheim, Norway.
25. filtering is a trust problem weight users based on how like-minded you have been to them in the past (similarity) to: weight users based on the quality of the opinions you receive from them (trust)
26. 1. find neighbours 2. make predictions 4. get feedback 3. recommend trust engine risk engine source: N. Lathia, S. Hailes, L. Capra “Trust-Based Collaborative Filtering” Joint iTrust and PST Conference (IFIPTM), June 2008, Trondheim, Norway.
29. (trust-based) select valuable neighbours a-symmetric value-added reward information
30.
31.
32.
33.
34. no profile? trust bootstrapping [see: Massa et al] uncertain? award potential recommenders extensions? Note (a): a [computed] subset of the community of users seems to be good recommenders for everybody. analysis?
35. a user kNN algorithm generates a graph source: N. Lathia, S. Hailes, L. Capra “kNN CF: A Temporal Social Network” To appear, ACM RecSys ’08. October 2008, Lausanne, Switzerland
51. changing collaborative filtering now: find best data source ? user-dependent problem can provably generate an RMSE < 0.856! Goal: adaptive collaborative filtering