A Particle Swarm Optimisation Based Multi-objective Filter Approach to Feature Selection for Classification

Feature selection (FS) has two main objectives of minimising the number of features and maximising the classification performance. Based on binary particle swarm optimisation (BPSO), we develop a multi-objective FS framework for classification, which is NSBPSO based on multi-objective BPSO using the idea of non-dominated sorting. Two multi-objective FS algorithms are then developed by applying mutual information and entropy as two different filter evaluation criteria in the proposed framework. The two proposed multi-objective algorithms are examined and compared with two single objective FS methods on six benchmark datasets. A decision tree is employed to evaluate the classification accuracy. Experimental results show that the proposed multi-objective algorithms can automatically evolve a set of non-dominated solutions to reduce the number of features and improve the classification performance. Regardless of the evaluation criteria, NSBPSO achieves higher classification performance than the single objective algorithms. NSBPSO with entropy achieves better results than all other methods. This work represents the first study on multi-objective BPSO for filter FS in classification problems.