Feature selection inspired by human intelligence for improving classification accuracy of cancer types

Feature selection is an essential task to predict clinical risk and biomarkers from the gene expression data. For practical matters, to choose the significant genes, researchers have been addressed several classical feature selection problems over the past decades for subsequent classification of genomics datasets with large ambient dimensionality but a small number of observations. To overcome high dimensionality and overfitting issues, in this paper, we developed a new gene selection technique by combination of minimum redundancy maximum relevance (mRMR) and teaching learning‐based optimization for accurate cancer prediction. Firstly, in the proposed approach, mRMR is applied to find the most discriminative genes from the original feature sets, and then a precise teaching learning‐based optimization with opposition‐based learning approach further refines the reduced feature set that can contribute to identifying the type of cancers. In addition, a new activation function is also investigated for effective gene selection, which is applied to convert continuous to binary search space. Support vector machine (SVM) is used as a fitness function in the proposed method to select relevant features that can help to estimate the predictive accuracy and classify cancer accurately. Attempts have made to increase the performance of SVM classifier by tuning penalty factor, kernel parameter, and tube size parameter with the help of proposed method. In order to testify computational efficiency of proposed algorithm, we have collected six gene expression datasets. Experimental results demonstrated that proposed method by utilizing SVM with Radial Basis Function kernel function is able to significantly reduce the irrelevant genes and outperform the conventional wrapper methods in terms of accuracy and model interpretation.