Effective memetic algorithm for multilabel feature selection using hybridization-based communication

Abstract

Multilabel feature selection, which is used to select features relevant to multiple labels, has demonstrated its effectiveness in prediction time and learning accuracy. A memetic algorithm that uses a feature subset refining process has been verified to outperform existing genetic algorithms in identifying an optimal feature subset. However, as the refinement process is consistently applied to all solutions of feature subsets, similar feature subsets can be over-produced, thereby limiting the synergy of hybridization. Here, we propose an evolutionary multilabel feature selection algorithm that searches the final feature subset using multiple populations to prevent limiting the synergy of hybridization. A new hybridization-based communication process refines solutions originated from each best solution of multiple populations, then, randomly distributes the produced solutions. With this approach, the proposed method circumvents the degradation of search capability and keeps the synergy of hybridization. Our experimental results indicate that the proposed method could identify better feature subsets than conventional methods.