A hybrid of binary Particle Swarm Optimization and estimation distribution algorithm for feature selection

The risk of common diseases is likely determined by single nucleotide polymorphisms (SNPs). However, due to the tremendous number of candidate SNPs, there is a clear need to genotyping by selecting only a subset of all SNPs that are highly associated with a specific disease. In this paper, a new algorithm which is based on a hybrid of binary Particle Swarm Optimization (BPSO) and estimation distribution algorithms (EDA), named HBPSO, is proposed to search the optimal SNPs subset and Support Vector Machine (SVM) is adopted as the classifier. In addition, the concept of elite strategy is adopted in HBPSO. HBPSO not only eliminates the redundancy of feature, but also solves the problem of SVM´s parameters selection simultaneously. The proposed approach is tested on two datasets: Crohn´s disease and Lung cancer. The experimental results demonstrate that the performance of HBPSO is better than other methods.