Kwon, Soo Min: Learning Predictors from Multidimensional Data with Tensor Factorizations
Title: Learning Predictors from Multidimensional Data with Tensor Factorizations
Name: Soo Min Kwon
Major: Electrical and Computer Engineering
School affiliation: School of Engineering
Programs: James J. Slade Scholars Program
Other contributors: Anand D. Sarwate
Abstract: Tensor factorizations and decompositions are important tools in many applications. These techniques aim to represent multidimensional data (tensors) using latent factors modeled as low rank matrices. Recent work has shown that imposing factorization structures on the coefficients within a statistical regression paradigm works efficiently on tensor data, in the sense that the number of parameters, and hence the amount of data needed to estimate them, is greatly reduced. Imposing such structures in a machine learning setting makes more efficient use of the inherent spatial and temporal structure of the tensor. In this paper, we explore classification problems in which we estimate predictors using a CANDECOMP/PARAFAC factorization structure. Our approach works well when the true tensor parameter exhibits a low rank structure. We use an alternating minimization algorithm to estimate the factor matrices and compare the prediction performance to approaches which ignore the tensor structure by vectorizing the data. We perform experiments in which the parameter is exactly low rank as well as approximately low rank. Our experimental results show in both cases that imposing structures on standard classification algorithms improves accuracy for prediction.