A Feature Selection Technique Using Self-Organizing Maps for Software Defect Prediction

Abstract

An important goal of Software Defect Prediction is to help increase the efficiency of software development. Increase in defect prediction accuracy would result in less resource consumption, and hence, effective feature selection techniques are needed to provide better inputs to the classifier. A carefully selected subset of features can not only increase prediction accuracy, but also result in less expensive computation. Wrapper methods for feature selection are based on evaluating feature subsets with a predetermined criterion. In this paper, the authors suggest an approach for feature selection by carefully evaluating feature subsets using clustering-based method. Cost based feature selection method based on Self-Organizing maps (CFSSOM) can be divided into three steps. First, computing the subsets of the feature set which are to be considered for evaluation. Second, clustering each of those feature subsets into two clusters. We use an ANN-based learning algorithm called self-organizing maps for clustering feature subsets. Third, applying labels on those two clusters based on data representation and measuring how strongly they are related to original labels. We have successfully implemented and proved that this feature selection technique improves prediction results based on experiments on PROMISE repository datasets.