Explainable Deep Learning Framework for Automated Software Defect Prediction Using Repository Analytics

Abstract

Software defect prediction helps development teams identify high-risk modules before complete testing and release. This study presents an explainable deep learning framework that uses repository analytics for automated defect prediction. The framework combines static code metrics, change history, developer activity, ownership information, defect records, source-code embeddings, and commit-message embeddings. A dual-branch neural architecture processes numerical and semantic features, while Shapley-based explanations show how each feature influences the predicted defect probability. The model was evaluated against logistic regression, support vector machine, random forest, XGBoost, and a conventional deep neural network using accuracy, precision, recall, F1-score, and AUC. The proposed framework achieved 95.4% accuracy and an AUC of 0.97, outperforming all benchmark models. Code churn, previous defect count, cyclomatic complexity, commit frequency, and semantic features were the most influential predictors. The framework supports accurate, transparent, and practical software quality decisions by improving testing prioritization, maintenance planning, and resource allocation in development projects.