Chaotic Harmony Search based Multi-objective Feature Selection for Classification of Gene Expression Profiles

How to effectively select a subset of discriminant features from the high-dimensional low- sample-size microarray gene expression profiles remains crucial and meaningful for the bioinformatics analysis tasks such as locating disease genes and building classifiers for cancer diagnosis. Though meta-heuristic harmony search algorithm has been used for feature selection, it suffers from entrapment in local optima and low convergence speed. To this end, we propose a hybrid chaotic harmony search based multi-objective feature selection method, which uses the chaotic map to replace the parameter of harmony search during the optimization process. Specifically, the minimum redundancy maximum relevancy feature selector is first used to pre-select a subset of relevant features. Then, the chaotic harmony search is employed on the reduced feature set to find an optimal feature subset, where the fitness of a candidate solution is evaluated by a multi-objective formulation. Finally, extensive comparative experiments against its competitors, including six filter and four wrapper feature selection methods, are conducted on six public microarray datasets. Results show that the proposed method obtains higher classification accuracy. Besides, the convergence analysis indicates its efficiency.