Hybrid deep learning for predicting software defects in agile environments

The rapidly changing requirements, along with short iterations in agile environments, make predicting requirement changes crucial for accurately detecting software defects in time. Most conventional machine learning techniques (SVM, Random Forest, Logistic Regression) depend on manually engineered features and do not model temporal, contextual, and semantic dependencies. In contrast, many current deep learning models tend to consider either structural metrics or process characteristics in isolation. Also, class imbalance and low interpretability are further impeding practical adoption in agile workflows. To overcome these gaps, we propose AgileDefectAI, an explainable defect-prediction framework that leverages a new deep learning model, HybridBugNet. HybridBugNet merges convolutional layers for local pattern recognition, Bi-LSTM layers for modelling sequential feature dependencies in commit and process histories, and attention-based layers based on the Transformer for identifying long-range contextual dependencies. Static code metrics, process dynamics, and semantic embeddings are integrated via a weighted feature fusion mechanism, while class imbalance is jointly mitigated using SMOTE and focal loss. SHAP analysis and attention visualisations for explainability to derive actionable insights into high-risk modules. HybridBugNet achieves F1-scores above 0.80 and AUC-ROC around 0.89 on benchmark datasets (NASA MDP, PROMISE, and agile GitHub repositories), outperforming classical ML baselines by ∼8–12% in F1-score and single-backbone deep models by ∼4–6%. Ablation studies confirm the complementary benefits of hybrid modelling and feature fusion; the framework novelty and agile defect prediction aptness are discussed.