Efficiency of oversampling methods for enhancing software defect prediction by using imbalanced data

Software defect prediction (SDP) is essential to analyze and identify defects present in a software model in early stages of software development. The identification of these defects and their early removal provides cost-efficient software. Machine learning (ML) techniques have been successfully used for developing defect prediction models. However, these techniques deliver off-target results when implemented on imbalanced datasets. For example, a dataset with unequal class distribution is technically imbalanced. Thus, ML techniques on such imbalanced data lead to a biased prediction of minority class instances, which are more important than majority class instances. Therefore, the imbalanced data problem must be resolved to successfully develop an efficient SDP model. In this study, we evaluated the prediction capability of ML classifiers for software defect prediction on nine imbalanced NASA datasets by applying oversampling methods. In addition, we considered five oversampling methods to synthesize minority class instances and make the datasets balanced. Dataset imbalance was eliminated using the five oversampling techniques. The oversampling techniques replicated or synthesized the instances of minority classes to balance the datasets. When the datasets were balanced, the ML classifiers were used to develop a defect prediction model. The experimental results acquired by applying ML classifiers on the imbalanced and balanced data showed an enhancement in the learning capability of ML techniques with the implementation of sampling techniques. Oversampling methods considerably improved the prediction performance of the ML classifiers.