SAGA: A Hybrid Technique to handle Imbalance Data in Software Defect Prediction

Software defect prediction has been a concurrent topic in software quality-based research. Predictive models that identify defect prone parts of Software can be evolved from defect data and software metrics. Various studies conducted in the past have explored Machine Learning-based approaches for this purpose but the problem of handling imbalanced defect data without compromising on the model's performance remains at large. In this work, we have proposed, compared, and analyzed a hybrid technique, SAGA(SMOTE + AdaSS + Genetic Algorithm), for solving the imbalance problem faced in software defect prediction. SAGA employs ensemble classification based on feature space partitioning in conjunction with the Synthetic Minority Oversampling technique. Various parameters related to feature space partitioning are optimized using the Genetic Algorithm. The values of ROC-AUC, G-mean, Balance, and Accuracy obtained on open-source datasets confirm the effectiveness of the proposed technique.