A Hybrid Encoding Based Particle Swarm Optimizer for Feature Selection and Classification

Feature selection (FS) is an important issue for classification, which aims to search the optimal feature subset to assist the classification task. Bio-inspired algorithms, such as particle swarm optimization (PSO), have shown superior performances in dealing with feature selection. However, current methods still suffer from local optimal and lack of efficient encoding manner for particles, which results in limited classification accuracy. In this paper, we proposed a hybrid encoding based PSO, HE-PSO for wrapper-based FS, where a novel encoding consisting of both integer and categorical value is applied. The new encoding way considerably takes the interactions between different features into account. In addition, a new updating strategy for particles' positions is developed, which is able to explore and search more promising and better solutions. Experimental results on benchmark data sets validate the effectiveness of our proposed approach in classification accuracy.