Just-in-time software defect prediction method for non-stationary and imbalanced data streams

Compared to traditional software defect prediction, Just-In-Time Software Defect Prediction (JIT-SDP) is a more fine-grained software defect prediction method used for defect prediction at the software change level. However, JIT software defect datasets in online data stream scenarios suffer from issues like validation delay, concept drift, and class imbalance evolution, which severely impact the predictive performance of JIT-SDP. This paper introduces a just-in-time software defect prediction method for non-stationary and imbalanced data streams, JNAI (JIT-SDP method for Non-stationary And Imbalanced data streams). This method solves validation delays, concept drifts, and class imbalance issues in existing JIT software defect processing technology. It proposes a validation delay framework to correct data labels, and a concept drift adaptation mechanism that combines intra-project and cross-project data filtering to mitigate concept drift while avoiding prediction bias caused by cross-project data. Next, a dynamic classifier selection method integrating a tiered AdaBoost is designed, using classifiers trained on preceding data to predict subsequent data labels iteratively, thereby addressing the issue of class distribution imbalance in data streams. Finally, the Hoeffding Tree is selected as the base classifier, and the processed dataset is used to train it, forming the final model of the just-in-time software defect prediction method. Experiments were conducted on six public JIT-SDP datasets and ten open-source GitHub projects, and the results show that JNAI effectively improves the predictive performance of just-in-time software defect prediction.