A Joint Learning Framework for Bridging Defect Prediction and Interpretation

Understanding why defect predictors classify software components as defective or clean is essential for software engineers that helps identify the root causes of defects and develop actionable bug-fixing plans. Existing solutions employ various explainable artificial intelligence methods to clarify the decision-making processes of defect predictors. However, these post hoc explanation techniques have two main limitations: first, the interpretation results do not accurately reflect the model’s decision logic, and second, they do not contribute to improving the performance of defect predictors. To address these limitations, we treat defect prediction and its corresponding interpretation as two distinct but closely related tasks, proposing a joint learning framework that trains the predictor and its interpreter simultaneously. The novelty of our approach lies in two main aspects: first, we design a feedback loop that transfers decision logic from the predictor to the interpreter, ensuring a high degree of conciseness for both components. Second, we incorporate interpretation results as a penalty term in the loss function of the joint learning framework. This not only enhances the accuracy of the predictor but also proposes a stronger constraint on the reliability of the interpreter. We validate our method against several existing explainable software defect predictors across multiple datasets. The results demonstrate its effectiveness.