Tensor Factorization Using Auxiliary Information

Abstract. Most of the existing analysis methods for tensors (or multi-way arrays) only assume that tensors to be completed are of low rank. However, for example, when they are applied to tensor completion prob-lems, their prediction accuracy tends to be significantly worse when only limited entries are observed. In this paper, we propose to use relation-ships among data as auxiliary information in addition to the low-rank assumption to improve the quality of tensor decomposition. We introduce two regularization approaches using graph Laplacians induced from the relationships, and design iterative algorithms for approximate solutions. Numerical experiments on tensor completion using synthetic and bench-mark datasets show that the use of auxiliary information improves com-pletion accuracy over the existing methods based only on the low-rank assumption, especially when observations are sparse.