Principled Regularization for Matrix Factorization

• Activity Number: 656
• Type: Invited
• Date/Time: Thursday, August 8, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Robert M. Bell*+ and Suhrid M. Balakrishnan
• Companies: AT&T Labs-Research and AT&T Labs-Research
• Keywords: high dimensional data ; recommender systems ; penalized least squares ; matrix completion

Abstract

Regularized matrix factorization has proven to be a very effective tool for dealing with high dimensional data, particularly in building recommender systems. L2 regularization is implemented by adding to the optimization criterion one or more terms that are proportional to the L2 norm of various sets of parameters. Implementation requires specifying the structure of these regularization terms and values for regularization parameters used as multipliers. We focus on the question of how many distinct regularization parameters should be used for a typical matrix factorization model. Theory and a simulation experiment suggest using more regularization parameters than are often used in practice.