Feature selection using Fisher score and multilabel neighborhood rough sets for multilabel classification

Abstract

In recent years, feature selection for multilabel classification has attracted attention in machine learning and data mining. However, some feature selection methods ignore the correlations among labels, resulting in low performance, and most of them face challenges in determining an appropriate neighborhood radius for neighborhood systems and suffer from expensive time cost. To overcome the issues, we propose a novel feature selection method using Fisher score and multilabel neighborhood rough sets (MNRS) in multilabel neighborhood decision systems. First, to identify the correlations between labels under a binary distribution, two types of new mutual information between labels are considered, and their balance coefficients are defined. By enhancing strong correlations and weakening weak correlations between labels, a mutual information-based Fisher score model with a second-order correlation between labels is designed to fit multilabel data. Second, to address the problem of automatically choosing a neighborhood radius, a subset of heterogeneous and homogeneous samples is employed to develop a new classification margin as a neighborhood radius, and some concepts of neighborhood, neighborhood class, and upper and lower approximations are formulated for multilabel neighborhood decision systems. The weight and dependency degree are presented to effectively measure the uncertainty of samples in multilabel neighborhood decision systems. Thus, we further present a new classification margin-based MNRS model. Finally, a filter-wrapper preprocessing algorithm for feature selection using the improved Fisher score model is proposed to decrease the spatiotemporal complexity of multilabel data, and a heuristic feature selection algorithm is designed for improve classification performance on multilabel datasets. Experimental results on thirteen multilabel datasets show that the proposed algorithm is effective in selecting significant features, demonstrating its excellent classification ability in multilabel datasets.