2012 kdd-com soc：adaptive transfer of user behaviors over composite social network - slides
- 1. ComSoc: Adaptive Transfer of User Behaviors over Composite Social Network Erheng Zhong1, Wei Fan2, Junwei Wang3, Lei Xiao3 and Yong Li3 1. Department of Computer Science and Engineering, HKUST 2. IBM T.J. Watson Research Center 3. Tencent Inc.
- 2. User Online Behaviors Instant Messaging Watching Movies Listening to Music Browsing News Playing Games Posting Messages Shopping Online
- 3. User Behavior Prediction User behavior prediction is important! Behavior 1 ? 1 ? 1 User 1 ? ? 1 ? 1 Data is extremely sparse! (e.g., >99.9% empty) Overfitting!
- 4. Composite Social Network (CSN) Users engage in multiple social • Multiple single networks networks nowadays. • Overlapping users • Different levels of influence • Different aspects of users’ behaviors Considering overlapping users as bridges...
- 5. Problem Behavior 1 ? ? 1 ? 1 User 1 ? ? 1 ? 1
- 6. SN for User Behavior Prediction Connected users share similar interests. Users share information across networks.
- 7. User Behavior Prediction Based on LDA Social Regularization (Blei et al., 2003) (Chang et al., 2009) (Erosheva et al., 2004)
- 8. ComSoc: Adaptive Transfer How to transfer knowledge? Select networks for different users adaptively!
- 9. ComSoc: A Synthetic Example (a) (b) (c) (d) (e) (f)
- 10. ComSoc: Generative Process User-Topic User-Topic Distribution Distribution Select Network Generate Topics Generate Generate Behaviors Relations
- 11. ComSoc: Inference Gibbs Sampling based on Posterior Distributions Topic Assignments for User-Item Interactions Topic Assignments for User-User Relations the number of times that the topic t is assigned to user i in the network s without current assignment the number of times that the network s is assigned to user i without current assignment the number of times that the topic t is assigned to item k without current assignment the number of times that items are assigned to the topic t without current assignment
- 12. ComSoc: Large-scale Implementation Distributed inference user-topic count of each user in the network s item-topic count of each item topic assignments for each user-item interaction
- 13. Experiments: Setting Evaluation Metrics Baselines LDA (Blei et al., 2003), LinkLDA (Erosheva et al., 2004), RTM (Chang et al., 2009) one-single-network vs. naïve combined network Held-out set according to temporal information
- 15. Experiments: Performance (Douban) No Network Composit Single Network Naïve Combined Network
- 16. Experiments: Performance (Tencent) No Network Composite Single Network Naïve Combined Network
- 18. Experiments: Parameter Analysis Correspondence between social networks and the interaction network
- 19. Conclusion Motivations: User behavior prediction is important but data is sparse. Users engage in composite social network. Users have different levels of trust on different networks. Models: ComSoc: A relational topic model with network selection A large-scale implementation based on Map/Reduce Experiments: Large-scale data collections from Tencent and Douban Improvement is as high as 0.03 on MAP The code is available upon request.
- 20. ComSoc: Adaptive Transfer of User Behaviors over Composite Social Network We thank the support of Hong Kong RGC GRF Projects 621010 and 621211.