Class Balancing Approaches in Dataset for Software Defect Prediction: A Systematic Literature Review

Defects cause a great impact on software reliability, and therefore also reduce software quality. To address this, the area of software defect prediction has emerged. This is an area that allows software engineers to predict when a module is susceptible to defects that may cause an error in the future, and thus focus efforts on defect-prone modules. One of the problems with this process is that the datasets with which a classification algorithm is trained are not in balance, which translates into results that are biased towards the majority class (in this context, the error-free modules). This paper reports a Systematic Literature Review based on the Kitchenham and Charter method to find out which class balancing algorithms have been used in the context of software defect prediction, the metrics with which their performance is evaluated and the characteristics that define the training datasets on which these algorithms are applied. This research revealed that while oversampling algorithms such as SMOTE are commonly used for class balancing, combining them with undersampling techniques can yield better performance. Additionally, no single evaluation metric is superior; each has its situational use. The study also found that static code metrics, object-oriented design metrics, and process metrics are important software characteristics. This suggests that class balance in training datasets for software defect prediction algorithms impacts multiple areas of the system.