Two Improvements to the Matrix Factorization Approach for Recommender Systems

• Activity Number: 454 - Recommender Systems and Large-Margin Machines: From Statistics Perspectives
• Type: Topic Contributed
• Date/Time: Wednesday, July 31, 2019 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Mu Zhu*
• Companies: University of Waterloo
• Keywords: alternating optimization; coordinate descent; saddle points

Abstract

I describe two improvements to the matrix factorization (MF) approach for making personalized recommendations, popularized by the 2006-2009 Netflix Contest. First, content information can be incorporated directly into the MF approach. This not only improves recommendation accuracy, but also provides useful insights about the contents themselves, as well as makes the resulting recommendations more easily interpretable. Second, the underlying optimization problem driving the MF approach is nonconvex, with many traps in practice --- such as inferior local solutions and saddle points, but the search space of the optimization algorithm can be expanded in order to avoid some of them. (Based on joint works with Peter Forbes, Jennifer Nguyen, and W. James Murdoch.)