Feature selection via maximizing inter-class independence and minimizing intra-class redundancy for hierarchical classification

Abstract

Hierarchical feature selection has proven to be significant in reducing classification difficulty. Many existing hierarchical feature selection methods use the hierarchy in the class space as structural information to improve performance. However, these methods focus on the inter-class structural relations and ignore the intra-class feature correlations, which is challenging as the numbers of classes and features in current datasets are both growing rapidly. In this paper, we propose a hierarchical feature selection method that maximizes inter-class independence and minimizes intra-class redundancy using structure and feature relations. First, we investigate the class hierarchy dependency in the tree structure as the structural relation regularization, which maximizes the independence between unrelated classes in the class structure. Second, we transform the feature correlations into a mathematical representation as feature relation regularization, which minimizes the redundancy in each class on the premise of sparsity. Finally, we unify the two regularization terms into a hierarchical feature selection method to trade off the structural and feature relations. Our method exhibits excellent efficiency and effectiveness compared with other feature selection methods.