Sparsity-based evolutionary multi-objective feature selection for multi-label classification

Abstract

Feature selection (FS) is typically an essential pre-processing step for many machine learning tasks. However, most existing FS approaches focus on single-label classification, whereas multi-label classification (MLC) is an emerging topic where each instance can be assigned to more than one class label. Sparsity-based FS is an efficient and effective method that can be applied to MLC. However, existing sparsity-based approaches based on gradient descent can get trapped at local optima, and cannot optimise the selected number of features and classification performance simultaneously that are often in conflict, thus making FS a multi-objective problem. Evolutionary multi-objective optimisation (EMO) can be applied to FS due to its potential global search ability. EMO-based algorithms have not been utilised in sparsity-based approaches for FS in MLC. This paper proposes a novel sparsity-based MLC FS method based on Multi-objective Evolutionary Algorithm with Decomposition (MOEA/D). The experimental results indicate improvement in the MLC performance in comparison to a existing multi-objective FS algorithms.