Alternating minimization for recommender systems with sparse data

Abstract

Alternating minimization is a popular approach to solving low-rank matrix completion problems. This algorithm forms the backbone of many recommendation systems, and is particularly successful on large datasets whose computation can be parallelized. However the conditions necessary for the theoretical guarantees of alternating minimization result in poor performance on very sparse datasets. This thesis studies alternating minimization in the context of such datasets, and works to provide a better algorithm for use in recommendation systems with sparse data. We discuss the properties necessary to ensure the theoretical guarantees of the algorithm, and explore an alteration of a standard implementation of alternating minimization that performs better on sparse datasets.