An adaptive synthetic sampling and batch generation-oriented hybrid approach for addressing class imbalance problem in software defect prediction

Abstract Learning classifiers with uneven class distribution datasets poses a significant challenge in software defect prediction. This problem arises when the number of samples representing one class is significantly smaller than the others, leading to weak classification performance, particularly for minority class instances. Traditional classification models assuming equal class instances can result in low prediction accuracy and decision-making precision for minority class instances, raising concerns about identifying such instances accurately. To overcome this issue, this research proposes a hybrid technique that combines the Adaptive Synthetic Sampling (ADASYN) approach with a batch generator named the HADAB technique. ADASYN generates synthetic samples for the minority class, balancing the dataset and improving prediction accuracy. Conversely, the batch generator feeds data to the model in batches, enhancing training efficiency. The Multi-Layer Perceptron (MLP) serves as the base classifier in this study. The proposed HADAB technique significantly improves prediction accuracy and training efficiency without requiring additional parameter tuning, algorithm modification, or increasing complexity. We validate the performance of HADAB using publicly available NASA datasets encompassing diverse types. The results demonstrate the superiority of HADAB over traditional prediction accuracy methods. In conclusion, the proposed HADAB technique offers a practical and effective solution for handling class imbalance in software defect prediction, leading to improved prediction accuracy.