Weighted low rank approximations with provable guarantees

Abstract

The classical low rank approximation problem is: given a matrix A, find a rank-k matrix B such that the Frobenius norm of A - B is minimized. It can be solved efficiently using, for instance, the Singular Value Decomposition (SVD). If one allows randomization and approximation, it can be solved in time proportional to the number of non-zero entries of A with high probability. Inspired by practical applications, we consider a weighted version of low rank approximation: for a non-negative weight matrix W we seek to minimize Σi,j(Wi,j· (Aij-Bi,j))2. The classical problem is a special case of this problem when all weights are 1. Weighted low rank approximation is known to be NP-hard, so we are interested in a meaningful parametrization that would allow efficient algorithms. In this paper we present several efficient algorithms for the case of small k and under the assumption that the weight matrix W is of low rank, or has a small number of distinct columns. An important feature of our algorithms is that they do not assume anything about the matrix A. We also obtain lower bounds that show that our algorithms are nearly optimal in these parameters. We give several applications in which these parameters are small. To the best of our knowledge, the present paper is the first to provide algorithms for the weighted low rank approximation problem with provable guarantees. Perhaps even more importantly, our algorithms proceed via a new technique, which we call "guess the sketch". The technique turns out to be general enough to give solutions to several other fundamental problems: adversarial matrix completion, weighted non-negative matrix factorization and tensor completion.