Improved multi-layer binary firefly algorithm for optimizing feature selection and classification of microarray data

Abstract

Gene microarray technology can detect many gene expressions simultaneously, which is essential for disease diagnosis. However, microarray data are usually characterized by small samples and high dimensionality, which requires obtaining the most matching genes to the disease before building a classification model. In this paper, we propose an improved multilayer binary firefly-based method that is divided into two phases. The first stage reduces the dimensionality space by improved Max-Relevance and Min-Redundancy (mRMR). The second stage refines the feature space from coarse-grained level to fine-grained level by improved multilayer binary firefly algorithm (MBFA) at each recursive step. We built a stacking classifier model based on attention mechanism in publicly available gene microarray data and evaluated them compared to advanced hybrid feature selection methods. The experimental results show that compared with the classical chi square test, variance, logistic regression and decision tree methods, the proposed method can achieve higher classification accuracy with fewer features. In addition, the proposed method also shows excellent results in comparison with different published hybrid feature selection methods.