Performance analysis of relief and mRMR algorithm combination for selecting features in lupus Genome-Wide Association Study

This paper describes research on the use of feature selection techniques to find correlation between single-nucleotide-polymorphism (SNP) in genes with the lupus disease in Genome-Wide Association (GWA) study. Feature selection is the process of selecting features that are correlated and discarding features that have no correlation in data mining. In this research, feature selection techniques will be applied on 262, 264 SNPs. SNP is a variation of nitrogen base pairs in human DNA. SNP number is very large so we needed feature selection techniques when performing data mining. Various feature selection techniques have been proposed with different accuracy for different types of data. This research uses a combination of Relief and minimal-redundancy-maximal relevance (mRMR) algorithms as a feature selection method. Classification methods, including decision tree, SVM, and Naive Bayes, are applied to the selected SNPs. We compare the results with Chi-Squared Test which is used commonly in GWA. We also compare the results with composer of feature selection algorithms: Max-Relevance (Mutual Information), Relief, and mRMR. We found that the combination algorithm does not yield in good performance for selecting SNPs in genome-wide association study with lupus disease. We also found that mRMR algorithm gives best result for selecting feature which gives very good classification accuracy.