Multi-label Feature Selection Using Particle Swarm Optimization: Novel Initialization Mechanisms

In standard single-label classification, feature selection is an important but challenging task due to its large and complex search space. However, feature selection for multi-label classification is even more challenging since it needs to consider not only the feature interactions but also the label interactions. Particle Swarm Optimization (PSO) has been widely applied to select features for single-label classification, but its potential has not been investigated in multi-label classification. Therefore, this work proposes PSO-based multi-label feature selection algorithms to investigate the importance of population initialization in multi-label feature selection. Particularly, the discriminative information is utilized to let the swarm start with more promising feature combinations. Results on eight real-world datasets show that the new strategies can reduce the number of features and improve classification performance over using all features and standard PSO-based multi-label feature selection.