ANALYZING THE PERFORMANCE OF HYBRID MODEL RECONSTRUCTION APPROACH FOR BALANCED AND UNBALANCED DATASET

Abstract

Software defect prediction helps in identifying fault prone classes of a software in the early phases of software development life cycle. This helps in efficient resource allocation, since more resources should be allocated to such fault prone classes. Defect prediction(DP) models are developed from different machine learning methodologies in which the models are trained using data from previous releases of a software. However, while developing DP models, researchers have to deal with certain issues. Two such critical issues are addressed in this study a) unavailability of historical data of a software project, and b) imbalanced nature of training data set, where the distribution of classes is highly skewed. In order to deal with scarcity of historical data of a project, literature studies use cross project defect prediction(CPDP). Though, a number of literature studies have suggested methods for improving the accuracy of DP models with imbalanced datasets, but the imbalanced issue has not been properly addressed in the scenario of CPDP.Thus, this study analyzes the performance of a CPDP model HYDRA(Hybridized moDel Reconstruction Approach), developed by Xia et al. using imbalanced data. The results of the study are empirically validated using twenty open source software projects, where ten software projects are of imbalanced nature. Furthermore, we also suggest variation in the fitness function of HYDRA for improving its performance. The results of the study are statistically assessed using Wilcoxon test.