Multi-objective particle swarm optimization based on particle contribution and mutual information for feature selection method

As a method to address feature selection, multi-objective particle swarm optimization (MOPSO) algorithm can effectively balance the two objectives of feature subset size and classification error rate. However, with the increase of the dimension of the feature space, the expansion of the search space scale makes MOPSO easy to converge to the local optimum prematurely due to the insufficient search ability. In this paper, we propose a global optimum selection strategy based on the contribution of particles, which divides the population into regions with different contribution types through the dominance relationship, and selects the appropriate global optimum for each region to improve the search ability of the algorithm. The schematic diagram of the strategy is shown in Fig. 1. In addition, mutual information, as a kind of prior information, can obtain a set of feature subsets with strong performance in the feature selection process. Different from most methods which only use mutual information in the initialization, this paper combines mutual information with the iterative process, and uses a new adaptive threshold strategy to balance the influence of mutual information on the population evolution in the iterative process to avoid its strong guiding effect which makes the particles converge rapidly into the local optimum. In summary, this paper proposes a multi-objective particle swarm optimization based on particle contribution and mutual information for feature selection method (PCMOPSO-MI-FS). Experimental results on eight UCI benchmark classification datasets and two gene expression datasets show that PCMOPSO-MI-FS achieves satisfactory results.