Transfer learning in heterogeneous collaborative filtering domains
Download as PPTX, PDF2 likes3,382 views
This document presents a method for transferring knowledge between heterogeneous collaborative filtering domains to address data sparsity problems. The authors propose a transfer collective factorization (TCF) framework that learns shared latent features from auxiliary "like" data and applies them to predict ratings in a target domain. TCF extends probabilistic matrix factorization and collective matrix factorization models. Experimental results on Movielens and Netflix datasets show TCF improves rating predictions over baseline methods, especially at high sparsity levels. The authors conclude TCF is an effective transfer learning approach for collaborative filtering but note future work could focus on addressing the cold start problem for users without any ratings.
![Rating-matrix generative model (ICML’09)
• RMGM is derived and extended from FMM generative model,
which can be formulated as
2013/3/27
• The difference:
• It learns (U, V) and (U3, V3) alternatively.
• A soft indicator matrix is used. E.g., U [0, 1]n d.
8](https://image.slidesharecdn.com/transferlearninginheterogeneouscollaborativefilteringdomains-130327193849-phpapp02/85/Transfer-learning-in-heterogeneous-collaborative-filtering-domains-8-320.jpg)
Transfer learning in heterogeneous collaborative filtering domains
- 1. 2013/3/27 Transfer learning in heterogeneous collaborative filtering domains Authors/ Weike Pan and Qiang Yang Affiliation/ Dept. of CSE, Hong Kong University of Science and Technology Source/ Journal of Artificial Intelligence (2013) Presenter/ Allen Wu 1
- 2. Outline • Introduction • Heterogeneous collaborative filtering problems 2013/3/27 • Transfer by collective factorization • Experimental results • Conclusion 2
- 3. Introduction • Data sparsity is a major challenge in collaborative filtering (CF). • Overfitting can easily happen for prediction. 2013/3/27 • Some auxiliary data of the form “like” or “dislike” may be more easily obtained. • It’s more convenient for users to express preference. • How do we take advantage of auxiliary knowledge to alleviate the sparsity problem? • Most existing transfer learning methods in CF consider auxiliary data from several perspectives. • User-side transfer, item-side transfer, knowledge-transfer. 3
- 4. Probabilistic Matrix Factorization (NIPS’08) • 2013/3/27 4
- 5. Social Recommendation (CIKM’08) • 2013/3/27 5
- 6. Collective Matrix Factorization (KDD’08) • 2013/3/27 6
- 7. CodeBook Transfer (IJCAI’09) • 2013/3/27 7
- 8. Rating-matrix generative model (ICML’09) • RMGM is derived and extended from FMM generative model, which can be formulated as 2013/3/27 • The difference: • It learns (U, V) and (U3, V3) alternatively. • A soft indicator matrix is used. E.g., U [0, 1]n d. 8
- 9. Heterogeneous collaborative filtering problems • • 2013/3/27 9
- 10. Challenges • 2013/3/27 10
- 11. Overview of solution • 2013/3/27 11
- 12. Model formulation • Assume a user u’s rating on an item i in the target data, rui, is generated from 2013/3/27 • user-specific latent feature vector Uu 1 d, where u=1,…,n. • item-specific latent feature vector Vi 1 d, where i=1,…,m. • some data-dependent effect denoted as B d d. 12
- 13. Model formulation (Cont.) • Likelihood: • Prior: 2013/3/27 • Posterior Likelihood Prior (Bayesian inference) • Log(Posterior)= Log(Likelihood Prior) 13
- 14. Model formulation • 2013/3/27 14
- 15. Learning the TCF 2013/3/27 15
- 16. Learning U and V in CMTF • Theorem 1. Given B and V, we can obtain the user-specific latent matrix U in a closed form. 2013/3/27 16
- 17. Learning U and V in CSVD • 2013/3/27 17
- 18. Learning U and V in CSVD (Cont.) 2013/3/27 18
- 19. • 2013/3/27 19
- 20. Algorithm of TCF 2013/3/27 20
- 21. Data sets • 2013/3/27 21
- 22. Evaluation metrics • Summary of Data sets 2013/3/27 • Evaluation metrics 22
- 23. Baselines and parameter settings • 2013/3/27 23
- 24. Performance of Moviepilot data 2013/3/27 24
- 25. Performance of Netfliex data 2013/3/27 25
- 26. Performance on Netflix at different sparsity levels • SCVD performs better than CMTF in 2013/3/27 all cases. 26
- 27. Conclusion • This paper investigate how to address the sparsity problem in CF via a transfer learning solution. 2013/3/27 • The TCP framework is proposed to transfer knowledge from auxiliary data to target data to alleviates the data sparsity. • Experimental results show that TCP performs significantly better than several state-of-the-art baseline algorithms. • In the future, the “pure” cold-start problem for users without any rating is needed to be addressed via transfer learning. 27