A filter-based bare-bone particle swarm optimization algorithm for unsupervised feature selection

Due to good exploration capability, particle swarm optimization (PSO) has shown advantages on solving supervised feature selection problems. Compared with supervised and semi-supervised cases, unsupervised feature selection becomes very difficult as a result of no label information. This paper studies a novel PSO-based unsupervised feature selection method, called filter-based bare-bone particle swarm optimization algorithm (FBPSO). Two filter-based strategies are proposed to speed up the convergence of the algorithm. One is a space reduction strategy based on average mutual information, which is used to remove irrelevant and weakly relevant features fast; another is a local filter search strategy based on feature redundancy, which is used to improve the exploitation capability of the swarm. And, a feature similarity-based evaluation function and a parameter-free update strategy of particle are introduced to enhance the performance of FBPSO. Experimental results on some typical datasets confirm superiority and effectiveness of the proposed FBPSO.