Randomized Nonnegative Tensor Factorization for Feature Extraction from High-dimensional Signals

Tensor decomposition methods are well-known tools for multilinear feature extraction from multi-way arrays with many important applications in signal processing and machine learning. Nonnegative Tensor Factorization (NTF) is a particular case of such methods, mostly addressed for processing nonnegat...

Full description

Saved in:
Bibliographic Details
Published in:2018 25th International Conference on Systems, Signals and Image Processing (IWSSIP) pp. 1 - 5
Main Authors: Zdunek, Rafal, Fonal, Krzysztof
Format: Conference Proceeding
Language:English
Published: IEEE 01.06.2018
Subjects:
Computational complexity
Correlation
Dimensionality Reduction
Hierarchical Alternating Least Squares Algorithm
Machine learning algorithms
Matrix decomposition
Multilinear Feature Extraction
Nonnegative Tensor Factorization
Randomized Approximation Algorithm
Signal processing algorithms
Symmetric matrices
Tensile stress
ISSN:2157-8702
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Tensor decomposition methods are well-known tools for multilinear feature extraction from multi-way arrays with many important applications in signal processing and machine learning. Nonnegative Tensor Factorization (NTF) is a particular case of such methods, mostly addressed for processing nonnegative multi-way arrays, such as hyperspectral observations or a set of images. One of the most efficient algorithms for NTF is the Hierarchical Alternating Least Squares (HALS) algorithm that belongs to a family of coordinate gradient descent updates. Despite its very good numerical properties, its computational complexity is quite large for large-scale datasets. In this study, we propose the randomized extension of the HALS, which considerably decreases its computational complexity with respect to the standard HALS. The numerical experiments, performed for various large-scale observations, confirm that the proposed algorithm is much faster than the standard one at the cost of slightly decreased performance.
ISSN:2157-8702
DOI:10.1109/IWSSIP.2018.8439450