JIT Software Bug Prediction using DBST Cleanup Mechanism with Explainable Backfit Transformer and Deep Kendall Analysis

Just-in-Time (JIT) bug prediction is crucial for improving software quality by identifying defects at the commit level. However, existing methods face significant challenges, including imbalanced datasets, inadequate representation of code modification structures, and limited explainability in fault models. To address these limitations, this study proposes a novel framework for JIT bug prediction, comprising three key components: a Density-aware Borderline Synthetic Tomek (DBST) cleanup mechanism, an Explainable Backfit structural transformer, and Deep Kendall Analysis. The DBST cleanup mechanism processes imbalanced JIT datasets by integrating Borderline SMOTE with Tomek links. It prioritizes boundary instances and enhances class separation using relative density and the taxicab metric. For feature extraction, the Explainable Backfit structural transformer captures syntactic and semantic features from code changes through dependency parsing, constituency parsing, and an equivariant semantic attention mechanism. Additionally, Deep Kendall Analysis employs a Deep Loopy Bi-LSTM to model long-term dependencies and Kendall rank correlation to discern relationships within the data. Finally, a global-local interpretable model-agnostic technique generates synthetic neighbours based on global and local relationships, providing transparent explanations for bug predictions. The proposed framework addresses critical gaps in JIT bug prediction by improving dataset processing, enhancing feature representation, and ensuring model explainability. This study contributes to the field by introducing a robust and interpretable approach to JIT bug prediction, with potential implications for advancing software quality assurance practices.