Fast hierarchical tucker decomposition with single-mode preservation and tensor subspace analysis for feature extraction from augmented multimodal data

Abstract

Tensor decomposition is a valuable and robust method for multilinear feature extraction and the dimensionality reduction of multiway data with a wide range of applications. Various tensor network (TN) models have been developed to extract features, and to relax the dimensionality and storage complexity of highly dimensional data. In this study, we extend the family of TNs and propose the hierarchical Tucker decomposition model with single-mode preservation (HTDMP). Various tensor augmentation strategies are suggested to enrich existing data information. These strategies are applied in combination with the HTDMP and multimodal tensor subspace analysis for image classification. The numerical experiments conducted confirm that the proposed method can outperform well-known tensor decomposition algorithms.