DBOS_US: a density-based graph under-sampling method to handle class imbalance and class overlap issues in software fault prediction

Improving software quality by predicting faults during the early stages of software development is a primary goal of software fault prediction (SFP). Various machine learning models help to predict software faults. However, the imbalanced class distribution in the datasets may challenge some traditional learning approaches as they are more biased toward the majority class. The existence of class overlapping makes the prediction difficult owing to learning the minority class inaccurately. In addition to that, high data dimensionality makes the classification process complex and time-consuming. To enhance the performance of the classifier, handling these data quality issues is a big concern. This paper proposes a hybrid density-based method DBOS_US to address the class imbalance, noise, and class overlap in SFP. Initially, the density-based overlap removal (DBO) clustering algorithm is proposed to filter noisy and overlapped instances. Then, a graph-based algorithm, ShapeGraph, is adapted to handle imbalanced classes. The objective of the proposed method DBOS_US is to improve the performance of the traditional SFP classifiers. The experiments are conducted on 11 benchmark datasets from the PROMISE repository using six machine learning models (SVM, DT, KNN, NB, RF, and boosting). The experimental findings and statistical analysis revealed that the proposed method outperforms seven state-of-the-art techniques in terms of Area Under the Curve (AUC), G-mean, Recall (PD), and Probability of False alarms (PF). The proposed method improves the average values of G-mean, Recall, and AUC by at least 2.5%, 8.8%, and 1.2%, respectively.