Gene selection and classification using correlation feature selection based binary bat algorithm with greedy crossover

Abstract

DNA microarray analysis plays a prominent role in classifying genes related to cancer. The dimension of the data is high and difficult to handle during classification. Hence, the dimension has to be reduced and highly predictive gene features must be obtained without affecting the accuracy. Previous studies concentrated either on improving the classification accuracy or reduction of gene features. Here, the multi‐objective problem of obtaining reduced gene features with high classification accuracy is addressed using the proposed correlation feature selection filter and binary bat algorithm (BBA) with greedy crossover. The gene feature subsets are obtained using the correlation based feature selection filter and optimized using the BBA. Suboptimal solutions obtained due to pre‐convergence of BBA are reset using the proposed greedy crossover. Highly predictive genes features are obtained and evaluated with support vector machine 10‐fold cross‐validation. An average classification accuracy of 95.85% with predictive gene features <1% of the total dataset was obtained when applied on cancer microarray datasets. The solution for the multi‐objective problem of obtaining high classification accuracy with minimal number of genes is achieved with better performance over the existing algorithms. Also, the problem of pre‐convergence with suboptimal solutions in optimization algorithms is overcome.

Gene Selection Based Cancer Classification With Adaptive Optimization Using Deep Learning Architecture

“…The surveyed method by Saberi-Movahed et al [16] does not describe a suitable solution for feature consideration. In Akila and Allin [17], the drawbacks of overfitting issues and increased false positives are observed. The transparency and interpretability are less in Azadifar et al [18] because of degraded feature learning capability.…”

“…Preceding works focussed either on enhancing the categorization accuracy or lessening gene attributes. In Akila and Allin [17], the major objective was high accuracy attainment with fewer features. It considered filter and wrapper-based techniques for feature selection.…”