Stochastic Gradient Descent on Tensors with Missing Data
Keywords: tensor recovery, stochastic gradient descent, missing data
TL;DR: We present a SGD-based method for tensor recovery with missing data.
Abstract: Large tensor linear systems of equations pose a challenge due to the sheer amount of data stored. They quickly become difficult to solve, due to either time constraints or memory constraints, and they become even more challenging when some of the data is missing. We present a stochastic iterative method to address both of these issues by adapting Stochastic Gradient Descent to the tensor case and adding a correction term to debias the gradient for missing data. We prove convergence results for our method and experimentally verify these results on synthetic data.
Submission Number: 17
Loading