A Survey on Handling Imbalance Data on Software Defect Prediction

Various machine learning techniques for defect software prediction have been used. The learning model is based on the training dataset. The labelled dataset consists of independent variables like difficulty, design complexity, the line of codes, maintenance severity, and dependent variables, either true or false. However, the defect dataset has a problem with class-imbalanced data. Software defects are often rare events compared to non-defective instances. The improvement of the accuracy of defect prediction depends on data quality. Handling imbalanced data is crucial because an imbalanced dataset can lead to biased predictions, where the model fails to accurately predict the minority class, which is the defective software. This study found that most methods used to handle imbalanced data in SDP rely on oversampling, with 13 studies employing this approach. Five studies utilized undersampling, while the remainder combined both techniques or employed alternative methods.