Supervised Structural Sparse Subspace Learning Based on Hierarchical Locality Preservation for Multimodal and Mixmodal Data

Qi Zhang,Tianguang Chu,Chaoyi Shi

doi:10.1109/access.2020.2980757

Abstract

We study the multimodal and mixmodal data-driven supervised structural sparse subspace learning problem in this paper, and present the $\alpha $ -structural regularization based hierarchical locality analysis ( $\alpha $ -SRHLA) model. Unlike most existing sparse subspace learning models that merely constrain the cardinalities of the subspace basis vectors, the present $\alpha $ -SRHLA model takes into account the structural correlations of the original data variables and generates “variable groups” for sparse subspace learning. As a result, the sparsity is induced in the scale of the variable group instead of the single variable, i.e., “structural sparsity”. In addition, the $\alpha $ -SRHLA considers the “hierarchical locality” of multimodal and mixmodal data, and derives the weighted local affinity correlations in both data-level and class-level. This helps to reveal the intrinsic distribution characteristics of the considered multimodal and mixmodal manifold structures. A series of experiments on normal and multimodal data classification, multimodal and mixmodal digit as well as face recognition verify the effectiveness of the present $\alpha $ -SRHLA model in dealing with both multimodal and mixmodal data.

Highlights

Supervised learning, as an efficient way that incorporates the information of label supervision with data distribution, can decrease the temporal and computational burden of the learning processes
In the past few years, numbers of supervised learning models have been presented such as [1]–[7], but most existing models might not always obtain satisfying results when dealing with multimodal and mixmodal data, as the hierarchical distribution properties exhibited in
Here we propose an α-structural regularization based hierarchical locality analysis (α-SRHLA) model that incorporates the ‘‘hierarchical locality’’ with ‘‘structural sparsity’’ and is capable of dealing with the multimodal and mixmodal cases

Summary

Introduction

Supervised learning, as an efficient way that incorporates the information of label supervision with data distribution, can decrease the temporal and computational burden of the learning processes. In the past few years, numbers of supervised learning models have been presented such as [1]–[7], but most existing models might not always obtain satisfying results when dealing with multimodal (i.e., samples within the same class have separated clusters) and mixmodal (i.e., some samples from different classes have relatively closer distances than those from the same class) data, as the hierarchical distribution properties (i.e., different distribution properties are shown in within-class and between-class scales) exhibited in. Separating odd and even numbers from 0-9 upon the handwritten digit images follows between-class multimodal properties as images of different odd/even numbers have the same labels. For the between-class mixmodality, when the face images have noises (e.g., different illuminations or decorations), data samples from different classes might be located more adjacently in the Euclidean space, whereas those of the same class might be separated with each other [8].

Objectives

Methods

Conclusion