Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Trust and Recommender Systems

1,850 views

Published on

A brief introduction of the knowledge of Trust and Recommender systems.

Published in: Education, Technology
  • Be the first to comment

Trust and Recommender Systems

  1. 1. ZhaYefei 2013.6.24 1 Trust and Recommender System
  2. 2. Outline  Recommender System  Trust Models  Trust in Recommender System  Conclusion
  3. 3. Recommender System  Information overload  Classified catalogue  Search  Ask for friends  Two-win  Info Producer  Info Consumer  Benefit Long tail Why ?
  4. 4. Recommender System Application Amazon More than 35% sale are from Recommender System! Rating Explainatio n
  5. 5. douban FM hulu Like ? 60% users benefit! Recommender System
  6. 6. Recommender System Collaborative Filtering Content-based Filtering Algorithm Item-basedUser-based 1st 2nd 3rd
  7. 7. Recommender System Content- based Filtering Movie A Movie B Movie C Like Like Like Movie A Type : Love; Romantic Movie B Type : Horror;Thriller Movie C Type : Love; Romantic similar User A User B User C
  8. 8. User-based Filtering Recommender System Item A Item B Item C Item D Like Recommend User A User B User C
  9. 9. Item-based Filtering Item A Item B Item C Like Recommend similar Recommender System User A User B User C
  10. 10. Local Trust PageRank Models Mole Trust Tidal Trust 1st 2nd 3rd Trust Global Trust
  11. 11. Paolo Massa  Italy  SAC 2005 (Symposium on Applied computing. ACM, 2005) A Trust-enhanced Recommender System application: Moleskiing MoleTrust
  12. 12. MoleTrust G H I A B C D E F 0 1 2 3 dist 0 A 1 B C D 2 E F 3 G H I
  13. 13. MoleTrust A B C D E F  Setp1 --(BFS)  dist=0,1,2  user[dist] user[dist-1] dist=0, user[0]= A dist=1, user[1]=B,C,D dist=2, user[2]=E,F  Setp2  trust(A)=1  For each dist =1,2,…
  14. 14. ( ) ( ) ( ( )* ( , )) ( ) ( ) i pre u i pre u trust i edge i u trust u trust i = = = ∑ ∑  Setp2  For each u in user[dist]  trust(i=pre(u)) >=0.6 eg. A B C D E F 0.8 0.7 0.5 0.8 0.7 0.7 0.8 dist=1 : Trust(B)=0.8; Trust(C)=0.7; Trust(D)=0.5; dist=2: Trust(E)=(0.8*0.6+0.7*0.7)/(0.8+0.7)=0.65 Trust(F)= (0.7*0.7)/0.7=0.7 MoleTrust
  15. 15. Jennifer Ann Golbeck University of Maryland  Ph.D thesis 2005 Computing and Applying Trust in Web-base Social Networks TidalTrust
  16. 16. TidalTrust G H I A B C D E F ( ) | ( ) | js j adj i is t t adj i ∈ = ∑ 1st : the trust rating from node i to node jijt eg. 2 AB AC AE t t t + = 2 AE AF AG t t t + = ( )is jst f t=
  17. 17. TidalTrust G H I A B C D E F : the trust rating from node i to node j ijt 2nd ( ) ( ) ij js j adj i is ij j adj i t t t t ∈ ∈ = ∑ ∑ 3rd ( ) max ( ) max ij ij ij js j adj i t is ij j adj i t t t t t ∈ ∩ ≥ ∈ ∩ ≥ = ∑ ∑
  18. 18. TidalTrust S c 9 8 1 0 9 9 S k 8 6 8 8 9 9 9 10 10 9 Choose The Max as Threshold 2nd Maxim 9 8 1 0 8 9 10 9 1s t Min=8 Min=8 Min=9 9  Setp1 --(BFS)
  19. 19. TidalTrust S c 9 8 1 0 9 9 S k 8 6 8 8 9 9 9 10 10 9 Choose The Max as Threshold  The shortest path Num=3  Setp2 Max( Strength Paths to Sink ) Max(9,9)=9
  20. 20. MoleTrust VS. TidalTrust G H I A B C D E F MoleTrust: Trust(AG) => Trust(AE)Trust(EG) A B E G TidalTrust: Trust(AG) => Trust(AB)Trust(BG) A B E G
  21. 21. PageRank A C B D E 1 1 ( )( ) ( ) (1 ) ( ... ) ( ) ( ) n n PR tPR t PR A d d C t C t = − + + + eg. ( ) ( ) ( ) (1 0.85) 0.85*( ) 1 3 PR B PR C PR A = − + + 1.. 1.. ( )* ( ) ( ) (1 ) ( ) ( ) i i i n i i n C t RP t RP A d d C t = = = − + ∑ ∑ ?
  22. 22. Trust-aware Recommender Systems  Trust in Recommender Systems  Paolo Massa  Italy RecSys2007 John O’Donovan  University College Dublin(Ireland)  IUI2005 (International Conference on Intelligent User Interfaces) Trust in Recommender System
  23. 23. Trust Trust in Recommender System Collaborative Filtering Data sparsity Be easily attacked
  24. 24. Trust in Recommender System ( ) ( ) ( , )( ( ) ) ( ) | ( , ) | p P i p P i sim c p p i p c i c sim c p ∈ ∈ − = + ∑ ∑ Pure Collaborative Filtering: 1st . User Similarity 2nd. Rating Predictor P(i): User similarity of c c(i): Rating predicted for item i by c p(i): Rating for item i by a producer p sim(c, p):Similarity between c and p
  25. 25. Trust [N*N] Rating [N*M] Input N: Users M: Items Trust Metric Estimated Trust[N*N ] Similarity Metric User [N*N] Similarity Rating Predictor Predicted Rating [N*M] Output First step Second step Pure Collaborative Filtering Trust in Recommender System
  26. 26. From the Epinions.com Web site 49,290 users who rated a total of 139,738 different items at least once, writing 664,824 reviews. 487,181 issued trust statements. Consists of 2 files Ratings data Trust data Experimental Analysis Dataset
  27. 27. Experimental Analysis
  28. 28. Experimental Analysis
  29. 29.  Introduce Recommender System 、 MoleTrust 、 TidalTrust 、 PageRa nk  Trust is very effective in alleviating RSs weaknesses:  Data sparsity;  Be easily attacked;  Cold-start.  Trust propogation is a tradeoff in terms of Accuracy and Coverage; Conclusion
  30. 30. Thanks for your attention !

×