A0230
Title: A strategy for improving the speed in tensor decomposition analysis
Authors: Michele Gallo - University of Naples L'Orientale (Italy) [presenting]
Abstract: Tensors are powerful algebraic objects for handling huge amounts of data. To extract information from data or just for compression of data, several methods can be used to process the tensors. Here, we consider the higher-order singular value decomposition also known as the Tucker tensor decomposition. Due to the huge number of elements, large collections of deterministic and randomized algorithms have been proposed, that are supposed to be faster and/or cheaper. The goal is to propose an efficient algorithm that provides the same results as HOSVD ones. The approach proposed is based on rearranging data in a higher-order tensor, and applying eigenvalues analysis on the Gram unfolding tensor. To show the accuracy and efficiency of the algorithm proposed, we perform numerical experiments on synthetic and real-world data to directly show the strengths and weaknesses of our proposal.
