Towards Balanced Defect Prediction with Better Information Propagation

Defect prediction, the task of predicting presence defects in source code artifacts, has broad application software development. prediction faces two major challenges, label scarcity, where only a small percentage artifacts are labeled, and data imbalance, majority labeled non-defective. Moreover, current defect methods ignore impact information propagation among this negligence leads to performance degradation. In paper, we propose DPCAG, novel model address above three issues. We treat as nodes graph, learn propagate influence neighboring iteratively an EM framework. DPCAG dynamically adjusts contributions each node selects high-confidence for augmentation. Experimental results on real-world benchmark datasets show that improves compare state-of-the-art models. particular, achieves substantial superiority when measured by Matthews Correlation Coefficient (MCC), metric is widely acknowledged be most suitable imbalanced data.