Multi-Label Feature Selection Using Particle Swarm Optimization: Novel Local Search Mechanisms

Multi-label feature selection has become an indispensable preprocessing step of a multi-label classification problem which can reduce the number of features while maintaining or even improving the classification performance. Particle swarm optimization (PSO) has been widely applied to feature selection, but mainly for single-label classification. In comparison with single-label feature selection, multi-label feature selection is a more challenging task due to the interaction between the class labels. On such a large and complex search space, PSO usually loses its population diversity and converges to a local optimal quickly. PSO usually maintains the best position discovered during its evolutionary process, called gbest which plays an essential role in the search of PSO. We propose a novel local search strategy which improves the gbest with an expectation of preventing the premature convergence problem. The proposed local search employs a flipping o perator to search f or better feature subsets surrounding the current gbest. Experimental results on eight real-world datasets show that the proposed local search can assist PSO to evolve better feature subsets.