Hierarchical Feature Selection Based on Instance Correlation and Label Semantic Structure

Abstract

Hierarchical classification learning aims to exploit the hierarchical relationship between data categories. Making full use of the hierarchical structure between class labels can effectively reduce the number of categories for each classification task and improve the accuracy of classification. For hierarchical feature selection, usually the more similar two labels are, the more features they share. However, existing hierarchical feature selection algorithms often ignore this. In addition, current hierarchical feature selection algorithms do not deeply consider the semantic structure between labels when exploiting label correlations. In this article, we propose a hierarchical feature selection based on instance correlation and label semantic structure. This algorithm expresses the correlation between instances with the help of Laplacian matrix. Then, the instance correlation is combined with the semantic relationship between labels in the hierarchical structure to construct a hierarchical feature selection model. To prove the effectiveness of the proposed algorithm, a large number of experiments are conducted on hierarchical datasets in different fields, and multiple hierarchical feature selection are compared. The experimental results demonstrate that the proposed algorithm has significant performance superiority.