[RecSys 2014] Deviation-Based and Similarity-Based Contextual SLIM Recommendation Algorithms
4 likes3,695 views
The document discusses the research of Yong Zheng on context-aware collaborative filtering and contextual slim (cslim) recommendation algorithms as part of his doctoral symposium at DePaul University. It covers the motivation, methodologies, and challenges associated with integrating contextual data into recommendation systems, highlighting both deviation-based and similarity-based approaches. The document also outlines ongoing work and future directions in improving effectiveness, scalability, and explainability of these algorithms.
[RecSys 2014] Deviation-Based and Similarity-Based Contextual SLIM Recommendation Algorithms
- 1. Deviation-Based and Similarity-Based Contextual SLIM Recommendation Algorithms Yong Zheng, DePaul University, Chicago, USA Oct 10, Doctoral Symposium @
- 2. Self-Introduction • Yong Zheng, Ph.D. Candidate, DePaul University • Research: Context-aware Collaborative Filtering • Supervisor: Dr. Bamshad Mobasher • Currently 5th Year at DePaul • Expected Graduation: Summer, 2015 2Center for Web Intelligence, DePaul University, Chicago, USA
- 3. Outline • Context-aware Recommender Systems (CARS) • Context-aware Collaborative Filtering • Contextual SLIM (CSLIM) Algorithms • Current Work and Ongoing Work • Challenges and Work in the Future 3Center for Web Intelligence, DePaul University, Chicago, USA
- 4. Outline • Context-aware Recommender Systems (CARS) • Context-aware Collaborative Filtering • Contextual SLIM (CSLIM) Algorithms • Current Work and Ongoing Work • Challenges and Work in the Future 4Center for Web Intelligence, DePaul University, Chicago, USA
- 5. Context-aware Recommender Systems (CARS) 5Center for Web Intelligence, DePaul University, Chicago, USA • What is Context? “Any information that can be used to characterize the situation of an entity”, Abowd and Dey, 1999 Sample of popular contextual variables in Recommender Systems (RS): Time (weekday/weekend/holiday), Location, Companion (kids/family), Mood (happy/sad/excited/upset), Weather, etc
- 6. Context-aware Recommender Systems (CARS) 6Center for Web Intelligence, DePaul University, Chicago, USA
- 7. Context-aware Recommender Systems (CARS) 7Center for Web Intelligence, DePaul University, Chicago, USA • Motivation Behind Users’ preferences change from contexts to contexts • Basic Assumptions RS should learn users’ preferences in contexts c from others’ preferences in the same c, e.g. context-aware collaborative filtering. • Challenges How to incorporate contexts into RS? E.g. CF, MF, etc How to develop effective CARS? How to Interpret contextual effects from the model? Sparsity problems in CARS.
- 8. Outline • Context-aware Recommender Systems (CARS) • Context-aware Collaborative Filtering • Contextual SLIM (CSLIM) Algorithms • Current Work and Ongoing Work • Challenges and Work in the Future 8Center for Web Intelligence, DePaul University, Chicago, USA
- 9. Context-aware Collaborative Filtering (CACF) 9Center for Web Intelligence, DePaul University, Chicago, USA • Collaborative Filtering (CF) CF is one of the most popular recommendation algorithms in the traditional RS. 1). Memory-based CF, e.g., ItemKNN CF 2). Model-based CF, e.g., matrix factorization 3). Hybrid CF • Context-aware Collaborative Filtering (CACF) CF CACF Contexts
- 10. Intro. Collaborative Filtering (CF) 10Center for Web Intelligence, DePaul University, Chicago, USA There are three series of most popular CF algorithms • Neighborhood-based CF ItemKNN-based CF, Sarwar, et al, 2001 UserKNN-based CF, Resnick, et al, 1994 • Matrix factorization (MF)-based CF Matrix Factorization, Sarwar, et al., 2000; Koren, et al., 2009 Tensor Factorization, Symeonidis, et al., 2008 • Sparse Linear Method (SLIM), Ning, et al., 2011
- 11. Intro. Collaborative Filtering (CF) 11Center for Web Intelligence, DePaul University, Chicago, USA • Neighborhood-based CF ItemKNN-based CF (ItemKNN) , Sarwar, et al, 2001 UserKNN-based CF (UserKNN) , Resnick, et al, 1994 M1 M2 M3 M4 U1 2 ? 5 4 U2 4 3 5 3 U3 4 3 1 4
- 12. Intro. Collaborative Filtering (CF) 12Center for Web Intelligence, DePaul University, Chicago, USA • ItemKNN, Sarwar, et al, 2001 𝑃𝑢,𝑖 = 𝑗∈𝑁 𝑖 𝑅 𝑢,𝑗 × 𝑠𝑖𝑚(𝑖, 𝑗 𝑗∈𝑁 𝑖 𝑠𝑖𝑚(𝑖, 𝑗 Rating Prediction in ItemKNN: Pros: 1). Straightforward 2). The explainability of the results Cons: 1). Sparsity problem 2). Predictions rely on similarity (from co-ratings)
- 13. Intro. Collaborative Filtering (CF) 13Center for Web Intelligence, DePaul University, Chicago, USA • Matrix Factorization, Sarwar, et al., 2000; Koren, et al., 2009 Both users and items are represented by a vector of weights on each Latent factors. User vector How users like some features Item vector How items obtain/capture those features
- 14. Intro. Collaborative Filtering (CF) 14Center for Web Intelligence, DePaul University, Chicago, USA • Matrix Factorization, Sarwar, et al., 2000; Koren, et al., 2009 Pros: 1). Work effectively and efficiently (e.g.,MapReduce) 2). Being easy/flexible to incorporate side information Cons: 1). Cold-start problem 2). Difficult to interpret the Latent factors
- 15. Intro. Collaborative Filtering (CF) 15Center for Web Intelligence, DePaul University, Chicago, USA • Sparse Linear Model (SLIM), Ning, et al., 2011 𝑆𝑖,𝑗 = 𝑅𝑖,: ⋅ 𝑊:,𝑗 = ℎ=1,ℎ≠𝑗 𝑁 𝑅𝑖,ℎ 𝑊ℎ,𝑗Ranking Prediction in SLIM-I: Matrix R = User-Item Rating matrix; W = Item-Item Coefficient matrix
- 16. Intro. Collaborative Filtering (CF) 16Center for Web Intelligence, DePaul University, Chicago, USA • Sparse Linear Model (SLIM), Ning, et al., 2011 𝑆𝑖,𝑗 = 𝑅𝑖,: ⋅ 𝑊:,𝑗 = ℎ=1,ℎ≠𝑗 𝑁 𝑅𝑖,ℎ 𝑊ℎ,𝑗 Ranking Prediction in SLIM-I: Results of Comparison between ItemKNN and SLIM-I: 1). Coefficients in W are similar to Item-item similarities in ItemKNN; 2). SLIM-I removed the normalization function; 3). The item-item coefficients DO NOT rely on co-ratings; 𝑃𝑢,𝑖 = 𝑗∈𝑁 𝑖 𝑅 𝑢,𝑗 × 𝑠𝑖𝑚(𝑖, 𝑗 𝑗∈𝑁 𝑖 𝑠𝑖𝑚(𝑖, 𝑗 Rating Prediction in ItemKNN:
- 17. Intro. Collaborative Filtering (CF) 17Center for Web Intelligence, DePaul University, Chicago, USA • Sparse Linear Model (SLIM), Ning, et al., 2011 SLIM-I: Matrix W is an Item-Item Coefficient Matrix ItemKNN SLIM-U: Matrix W is a User-User Coefficient Matrix UserKNN Pros: 1).Avoid unreliable similarity calculations in ItemKNN/UserKNN 3.Work effectively and efficiently; Obtain explainability. Cons: Cold-start problems
- 18. CF and Context-aware CF (CACF) 18Center for Web Intelligence, DePaul University, Chicago, USA CF Neighborhood-based Collaborative Filtering Matrix Factorization SLIM Pros Explainability Effectiveness Efficience Effectiveness Explainability Cons Sparsity Problem Weak Explainability Cold-start Problem CACF Differential Context Modeling (DCM),2012 Tensor F, 2010 Context-aware MF (CAMF),2011 CSLIM, 2014 Incorporate Contexts
- 19. CACF: Differential Context Modeling (DCM) 19Center for Web Intelligence, DePaul University, Chicago, USA DCM incorporates contexts into neighborhood-based CF (e.g., UserKNN, ItemKNN, Slope One) by applying contextual constraints to different functional components in the algorithms. For example, the similarity of items can be calculated within the same (or similar) contexts, instead of calculations without considering contexts. Drawbacks: Overfitting in Top-N Recommendations
- 20. CACF: Context-aware MF (CAMF) 20Center for Web Intelligence, DePaul University, Chicago, USA CAMF incorporates contexts by adding contextual rating deviations, which is actually a dependent way to model the contextual effects. bi is item’s rating bias, which is replaced by an aggregation of item’s rating biases in different contextual situations. Drawbacks: Difficult to interpret latent factors. 𝑅𝑖,𝑗 = 𝜇 + 𝑏 𝑢 + 𝑏𝑖 + 𝑝 𝑢 𝑇 𝑞𝑖 𝑅𝑖,𝑗,{𝑐1,𝑐2,...,𝑐 𝑁 = 𝜇 + 𝑏 𝑢 + 𝑗=1 𝑁 𝐵𝑖𝑗𝑐 𝑗 + 𝑝 𝑢 𝑇 𝑞𝑖 Rating Prediction in MF: Rating Prediction in CAMF:
- 21. CACF: Tensor Factorization (TF) 21Center for Web Intelligence, DePaul University, Chicago, USA TF is an independent way, which directly considers each contextual variable as an individual context in the multi-dimensional space. Drawbacks: 1). Computational costs increase exponentially with the number of contexts increases. 2). Contexts are usually depdendent rather than fully independent.
- 22. Outline • Context-aware Recommender Systems (CARS) • Context-aware Collaborative Filtering • Contextual SLIM (CSLIM) Algorithms • Current Work and Ongoing Work • Challenges and Work in the Future 22Center for Web Intelligence, DePaul University, Chicago, USA
- 23. CACF: Contextual SLIM (CSLIM) Algorithms 23Center for Web Intelligence, DePaul University, Chicago, USA CSLIM: Incorporate contexts into SLIM 𝑆𝑖,𝑗 = 𝑅𝑖,: ⋅ 𝑊:,𝑗 = ℎ=1,ℎ≠𝑗 𝑁 𝑅𝑖,ℎ 𝑊ℎ,𝑗Ranking Prediction in SLIM-I: CSLIM has a uniform ranking prediction: CSLIM aggregates contextual ratings with item-item coefficients. However, there are two main points: 1).The rating to be aggregated should be placed under same c; 2).Accordingly, W indicates coefficients under same contexts; Incorporate Contexts 𝑆𝑖,𝑗,𝑐 = ℎ=1,ℎ≠𝑗 𝑁 𝑅𝑖,ℎ,𝑐 𝑊ℎ,𝑗
- 24. CACF: Contextual SLIM (CSLIM) Algorithms 24Center for Web Intelligence, DePaul University, Chicago, USA Ranking prediction in CSLIM: The remaining problem is how to calculate , it is because context-aware data set is usually sparse – it is not guaranteed that user also rated other items under the same contexts c. We use an estimation to obtain : 1).Deviation-Based CSLIM 2).Similarity-Based CSLIM 𝑆𝑖,𝑗,𝑐 = ℎ=1,ℎ≠𝑗 𝑁 𝑅𝑖,ℎ,𝑐 𝑊ℎ,𝑗 𝑅𝑖,ℎ,𝑐 𝑅𝑖,ℎ,𝑐
- 25. Outline • Context-aware Recommender Systems (CARS) • Context-aware Collaborative Filtering • Contextual SLIM (CSLIM) Algorithms • Current Work and Ongoing Work • Challenges and Work in the Future 25Center for Web Intelligence, DePaul University, Chicago, USA
- 26. Deviation-Based CSLIM Algorithms 26Center for Web Intelligence, DePaul University, Chicago, USA Research Status: Completed. Publications: • Y. Zheng, B. Mobasher, R. Burke. "CSLIM: Contextual SLIM Recommendation Algorithms", ACM RecSys, Silicon Valley, USA, 2014 • Y. Zheng. "Deviation-Based and Similarity-Based Contextual SLIM Recommendation Algorithms", ACM RecSys, Silicon Valley, USA, 2014 • Y. Zheng, B. Mobasher, R. Burke. "Deviation-Based Contextual SLIM Recommenders", ACM CIKM, Shanghai, China, 2014
- 27. Deviation-Based CSLIM Algorithms 27Center for Web Intelligence, DePaul University, Chicago, USA Ranking prediction in CSLIM: 𝑆𝑖,𝑗,𝑐 = ℎ=1,ℎ≠𝑗 𝑁 𝑅𝑖,ℎ,𝑐 𝑊ℎ,𝑗 𝑅𝑖,𝑗,𝑐 = 𝑅𝑖,𝑗 + 𝑙=1 𝐿 𝐷𝑗,𝑙 𝑐𝑙Example: CSLIM-I-CI, R = non-contextual Rating Matrix D = Contextual Rating Deviation Matrix W = Item-item Coefficient Matrix C = a binary context vector, as below Weekend Weekday At Home At Park 1 0 0 1
- 28. Deviation-Based CSLIM Algorithms 28Center for Web Intelligence, DePaul University, Chicago, USA Several CSLIM algorithms can be built: 1). Build upon ItemKNN or UserKNN ItemKNN CSLIM-I models, W is item-item matrix UserKNN CSLIM-U models, W is user-user matrix 2). How to define Contextual Rating Deviations? Dependent with items D is a Context-Item (CI) matrix Dependent with users D is a CU matrix Individual devations D is a vector of deviations So, there are six models in total: CSLIM-I-CI, CSLIM-I-CU, CSLIM-I-C; CSLIM-U-CI, CSLIM-U-CU, CSLIM-U-C;
- 29. Deviation-Based CSLIM Algorithms (Optional) 29Center for Web Intelligence, DePaul University, Chicago, USA Stay tuned: Y. Zheng, B. Mobasher, R. Burke. "Deviation-Based Contextual SLIM Recommenders", ACM CIKM, Shanghai, China, 2014 New Updates: 1). Discover patterns how to select appropriate CSLIM algorithms in advance based on data characteristics; 2). Develop general CSLIM algorithms, so that deviations can be estimated from any contexts;
- 30. Deviation-Based CSLIM Algorithms 30Center for Web Intelligence, DePaul University, Chicago, USA Deviation-Based CSLIM algorithms have been demonstrated to outperform the state-of-the-art CACF algorithms in terms of Top-N evaluation metrics, e.g., precision, recall, MAP, NDCG, etc.(over 5 data sets)
- 31. Similarity-Based CSLIM Algorithms 31Center for Web Intelligence, DePaul University, Chicago, USA Research Status: Ongoing. Ranking prediction in CSLIM: , in this case, will be estimated by fusing with similarity of two contexts. Of course, there could be many more ways to model sim. 𝑆𝑖,𝑗,𝑐 = ℎ=1,ℎ≠𝑗 𝑁 𝑅𝑖,ℎ,𝑐 𝑊ℎ,𝑗 𝑅𝑖,ℎ,𝑐 𝑅 𝑢,𝑖,𝑐−1 = 𝑅 𝑢,𝑖,𝑐−2 × 𝑆𝑖𝑚(𝐾𝑖𝑑𝑠, 𝑃𝑎𝑟𝑡𝑛𝑒𝑟 × 𝑆𝑖𝑚(𝑊𝑒𝑒𝑘𝑑𝑎𝑦, 𝑊𝑒𝑒𝑘𝑒𝑛𝑑
- 32. Outline • Context-aware Recommender Systems (CARS) • Context-aware Collaborative Filtering • Contextual SLIM (CSLIM) Algorithms • Current Work and Ongoing Work • Challenges and Work in the Future 32Center for Web Intelligence, DePaul University, Chicago, USA
- 33. Challenges 33Center for Web Intelligence, DePaul University, Chicago, USA 1. Effectiveness Several reviewers argued about scalability of CSLIM. There are several solutions to reduce costs: If Large # users/items KNN, select Top-K neighbors; If Large # contexts select relevant ones only; Besides, it is possible to use coordinate descent for optimization instead of gradient descent. 2. Similarity of contexts Most related work on the calculations of similarity contexts rely on the existing contextual ratings.
- 34. Pros and Cons 34Center for Web Intelligence, DePaul University, Chicago, USA 1. Pros Effectiveness and also Explainability 2. Cons Not all cells can be learned! Sol: Factorize contexts/items to represent them by vectors
- 35. Future Work 35Center for Web Intelligence, DePaul University, Chicago, USA 1. Explainability Try to explore and discover the explainability of CSLIM algorithms, e.g., discovering emotional effects compared with other CARS algorithms (Zheng, et al, 2013, RecSys) 2. Scalability Try to evaluate CSLIM on large data sets. Notice: there are no real+large data in CARS domain! 3. Similarity-Based CSLIM models Try to model and learn similarity of contexts in different ways to develop similarity-based CSLIM models.
- 36. Acknowledgement • NSF Student Funding • ACM RecSys Doctoral Symposium • Mentors: Dr. Pablo Castells and Dr. Irwin King • My supervisor: Dr. Bamshad Mobasher • My parents 36Center for Web Intelligence, DePaul University, Chicago, USA
- 37. Any Questions or Suggestions? Yong Zheng, DePaul University, Chicago, USA Oct 10, Doctoral Symposium @