(A) hybrid instance selection using nearest-neighbor framework for cross-project defect prediction with consideration of class imbalance

Abstract

Software defect prediction can help to allocate testing resources on fault-prone modules. Typically, local data within a company are used to build classifiers. In contrast to such Within-Project Defect Prediction (WPDP), there may exist some cases, e.g., pilot projects, that lack past data. Cross-project defect prediction (CPDP) using data from other projects can be useful in such cases. The major challenge of CPDP is different distributions in the training and test data. To tackle this, instances of the source data similar to the target data are selected to build classifiers. Software defect datasets have a class imbalance problem, i.e., the size ratio of the defective class to the clean class is very low. It usually lowers the performance of classifiers. In the presence of irrelevant or redundant information, prediction performance may be degraded as well. To address all the above issues, we propose a Hybrid Instance Selection using Nearest-Neighbor (HISNN) framework. It performs a hybrid classification that selectively learns local knowledge (via k-Nearest Neighbor) and global knowledge (via naive Bayes). Instances that have strong local knowledge are identified via nearest-neighbors with the same class label. To identify the optimal feature selection technique, we compare 9 feature selection techniques in cross-project settings. After features are chosen, classifiers are built, tested, and later evaluated based on the statistical significance test and the effect size test. The results show that the predictive performances of HISNN are comparable to those of WPDP. Using HISNN, companies without local data can predict defects with high performance until sufficient data are collected. Consequently, software quality can be managed effectively.