A many-objective feature selection for multi-label classification

Abstract

Feature selection is an important task in machine learning. As multi-label classification tasks appear in various fields, researchers have investigated multi-label feature selection algorithms to reduce data dimensions. Most of the existing wrapper multi-label feature selection algorithms use multi-objective method to obtain the selected features. However, there are multiple criteria to measure the quality of multi-label classification results. In view of this, this study presents a many-objective optimization based multi-label feature selection algorithm (MMFS). To improve the diversity and convergence of NSGA III, we propose an improved NSGA III algorithm with two archives. In this algorithm, new crossover and mutation operators for feature selection are designed to improve the exploration capability, and the influence of the selection threshold θ on feature scale and multi-label classification performance in real number coding is studied. Finally, we conduct experiments on 11 multi-label datasets. The experiments demonstrate that MMFS can balance multiple objectives, remove irrelevant and redundant features, and obtain satisfactory classification results.