A novel supervised gene clustering approachby mining interdependent gene patterns

Abstract

This paper proposes a general methodology of gene clustering based on gene selection to improve the classification accuracy as well as to address the curse of high dimensional datasets. The proposed method first tries to empirically establish the gene clustering technique based on gene selection approach, where the dependency of the genes with respect to clusters have been measured and their categories are defined such as; dependent, independent, lighter dependent and partial dependent within a range of [1-0]. Those genes which fall under the category of lighter dependent [0-0.5] are again checked and get reassigned to the cluster by measuring the interdependency with respect to that cluster and finally all the genes are ranked within clusters. The top most ranked genes of each clusters are taken together to form a pool of genes giving rise to reduced form of dataset. The classification performance of the original datasets and reduced form of those datasets have been measured with mostly used Naïve Bayesian, Decision Tree, Neural Network, Nearest Neighbor and SVM classifiers. Additionally, the classification accuracy of the proposed model has also been verified with few existing feature/attribute selection as well as clustering methods such as; ACA, t-Test, k-means, SOM, MRMR etc. An evident finding is that, the proposed algorithm has shown best classification accuracy with excellent predictive capability.