Fast and Accurate Tensor Completion with Tensor Trains: A System Identification Approach

04/17/2018

∙

by   Ching-Yun Ko, et al.

∙

0

∙

share

We propose a novel tensor completion approach by equating it to a system identification task. The key is to regard the coordinates and values of the known entries as inputs and outputs, respectively. By assuming a tensor train format initialized with low-rank tensor cores, the latter are iteratively identified via a simple alternating linear scheme to reduce residuals. Experiments verify the superiority of the proposed scheme in terms of both speed and accuracy, where a speedup of up to 23× is observed compared to state-of-the-art tensor completion methods at a similar accuracy.