Greedy fuzzy vaguely quantified rough approach for cancer relevant gene selection from gene expression data

Abstract

Gene selection is an important technique to remove irrelevant genes and handle the problem of curse-of-dimensionality issue. In other words the objective of the gene selection problem is to find (a small number of) cancer responsible genes (called biomarkers) from large number of genes, which have highest class discernable property. Traditional gene selection techniques are often not scalable on large number of genes and they are not able to handle the problem of vagueness, indiscerniblity, ambiguity, overlappiness complex cancer subtypes classes as usually present in the microarray gene expression data. In this context, a novel greedy fuzzy vaguely quantified rough approach for feature (gene) selection (GFVQRFS) is proposed that handles curse-of-dimensionality issue, vagueness, indiscerniblity, ambiguity, overlapping and complex cancer subtypes classes. The proposed method is evaluated on eight publicly available microarray gene expression datasets and the results are compared with four other state-of-the-art methods namely, CFS-GA, CON-GA, CON-GS and FRFS-GA using three classifiers (viz., KNN, SVM and NB). Six different validity measures (viz., accuracy, precision, recall, macro average $F_1$-measures, micro average $F_1$-measures and kappa) are used to access the performance of the proposed GFVQRFS method with respect to the compared methods. The proposed method selects very less number of genes compared to those selected by the other counterpart methods. The experimental results reveal the edge of the proposed method over other counterpart methods for most of the datasets.