Transfer learning in collaborative filtering

Abstract

Transfer learning and collaborative filtering have been studied in each community separately since early 1990s and were married in late 2000s. Transfer learning is proposed to extract and transfer knowledge from auxiliary data to improve the target learning task and has achieved great success in text mining, mobile computing, bioinformatics, etc. Collaborative filtering is a major intelligent component in various recommender systems, like movie recommendation in Netflix, news recommendation in Google News, people recommendation in Tencent Weibo (microblog), advertisement recommendation in Facebook, etc. However, in many collaborative filtering problems, we may not have enough data of users’ preferences on items, which is known as the data sparsity problem. Transfer learning in collaborative filtering (TLCF) is studied to address the data sparsity problem in the user-item preference data in recommender systems. In this thesis, we develop this new multidisciplinary area mainly from two aspects. First, we propose a general learning framework, study four new and specific problem settings for movie recommendation and people recommendation, and design four novel TLCF solutions correspondingly. We transfer knowledge from different types of auxiliary data based on a general regularization framework, and design batch algorithms, stochastic algorithms and distributed algorithms to solve the optimization problems. Second, we survey and categorize traditional transfer learning works into model-based transfer, instance-based transfer and feature-based transfer, and build a relationship between traditional transfer learning algorithms and TLCF solutions from a unified view of model-based transfer, instance-based transfer, and feature-based transfer.