Delab_link_prediction_for faloutsos.pptx

DATA

1. Link Prediction based on Multi-modal Social Networks Panagiotis Symeonidis, PhD Christos Perentis, MSc Delab, Department of Informatics, Aristotle University of Thessaloniki

2. Problem Definition Online Social Networks such as Facebook enable users to form explicit networks expanding their social circle, by adding new friends. Online Social Rating Networks such as Epinions or Flixter allow users to form implicit networks, by rating, tagging or liking items. Exploiting all interactions in multi-modal networks allow us to: Model network evolution Link Prediction Identify communities, trends, anomalies Recommend friends, items, products, groups, places

3. P opular web systems such as Epinions , Flixter , etc. allow users to form: 1. Explicit social networks 2. Implicit social network What are Multi-modal Social Networks? U 1 U 3 I 1 I 2 3 5 4 2 1 U 2

4. Related Work Link Prediction Graph Proximity Models Node Based Methods (FOAF, Adamic /Adar, Jaccard …) Path Based Methods (Katz, tKatz , RWR..) Collaborative Filtering User-based CF Item-based CF Latent Factors Models (SVD…)

5. Motivation-Contribution Exploit auxiliary sources - implicit networks to provide: Enhanced recommendation in the explicit network, using multiple sources Recommendation in the absence of information from the implicit network Real and Synthetic Data sets Epinions , Flixter xSocial Generator (Network buddy, Network co-comment)

6. Proposed Method Capturing similarities using tKatz of varying lengths l=2,3… sKatz (A): using single friendship network source cKatz (A+B): using multiple sources from the user-friendship and co-comment networks User Item User A B Item B^T C

7. Some Results.. Type Nodes Edges Buddy Network 100.000 388.458 Co-comment Network 233.820 467.640

8. Future Work 1 Friend Recommendation with more than two networks User {U1, U2,U3 ...U4} Item {I1, I2,I3 …I4} Tag {T1, T2,T3 …T4} User User x User User x Item User x Tag Item Item x User Item x Item Item x Tag Tag Tag x User Tag x Item Tag x Tag Online Social Network Online Rating Network Semantic Orientation

9. Future Work 2 Item recommendation based on multi-modal networks U 1 U 3 U 2 User-User Network Example Adjacency matrix similarity measure User-User similarity matrix I 1 I 2 U 1 4 5 U 2 2 1 U 3 3 3.66 (0.5*5+0.25*1)/(0.5+0.25)=3.66 Rating Prediction Equation U 1 U 2 U 3 U 1 1 1 U 2 1 U 3 1 U 1 U 2 U 3 U 1 1 0.5 0.5 U 2 0.5 1 0.25 U 3 0.5 0.25 1

10. Future Work 3 Cross-domain Recommendation (1/2) Challenges Imbalance of data size among domains Sparseness of data not a common user set among domains

11. Future Work 3 Cross-domain Recommendation (2/2) Proposed solution.. Domain 1 … Domain s Domain 1 D 1,1 … D s,1 … … ….. … Domain s D 1,1 D s,s