Wrapper Feature Selection based Heterogeneous Classifiers for Software Defect Prediction

Abstract

The performance of Software Defect Prediction (SDP) models depends on the quality of dataset used for training the models. The high dimensionality of software metric features has been noted as a data quality problem which affects the performance of SDP models. This makes it crucial to apply feature selection (FS) to SDP since FS can remove irrelevant and redundant software metric features. In this study, the effect of wrapper-based FS methods on classification techniques in SDP was investigated. The wrapper FS methods were based on different search methods; Best First Search (BFS), Genetic Search (GS), Greedy Stepwise Search (GSS) and Multi-Objective Evolutionary Search (MOES) so as to investigate their respective effect on classifiers in SDP. Five (5) publicly available software defect datasets were used. These datasets were classified by the individual classifiers which were carefully selected based on their characteristics hence the heterogeneity. Naïve Bayes (NB) was selected from Bayes category Classifier, K-Nearest Neighbor (KNN) was selected from Instance-Based Learner category and (J48) Decision Tree from Trees Function classifier. The experimental results clearly showed that the application of wrapper FS method to datasets before classification in SDP is better and should be encouraged as NB with GS based Wrapper Method had the best accuracy performance. It can be concluded that FS methods are capable of improving the performance of predictive models in SDP.