Sparse Low Rank Pursuit for Collaborative Filtering

• Activity Number: 105
• Type: Invited
• Date/Time: Monday, July 30, 2012 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Xiaotong Shen*+ and Yunzhang Zhu and Changqing Ye
• Companies: University of Minnesota-Twin Cities and University of Minnesota and University of Minnesota
• Keywords: Nonconvex ; Rank minimization ; Matrix completion

Abstract

Collaborative filtering concerns personalized information filtering, which predicts the likeability of information based on the opinions of users who think alike. In this paper, we propose methods to predict individual outcomes for missing values, where values are missing due to incomplete information of individual users. The proposed methods use constraints to seek the best low-rank approximation that is the sparsest, and a method block-wise coordinate descent for a scalable problem. Our experiments show potential applications of our methods for a significant improvement of several scale methods. Finally, a connection with rank minimization will be discussed.