Abstract
Software Fault Prediction (SFP) is an important task of software engineering, which attempts to discover the faulty modules in source code proactively to improve quality and reducing maintenance costs. The machine learning (ML) models has significantly enhanced fault prediction capability, yet these methods are black box, and limits in interpretability and trust. Explainable AI (XAI) approaches, such SHapley Additive exPlanations (SHAP), are useful to get insights, and quantify the impact of specific characteristics to the model decisions. The study proposed a robust and interpretable framework for SFP which integrates class imbalance handling (SMOTE), optimisation (using hyperparameter tuning (SMOHY)), SHAP-based explainability and Feature Sensitivity (FS) analysis to produce accurate and transparent models using 41 ML classifiers against 41 opensource java projects. SMOHY improved the efficiency and predictability of the SFP model by 22.38% compared to ORGD and 6.5% compared to SMOTE, attaining highest median AUC value of 0.82. ExtraTrees and RandomForest algorithms achieve the highest median accuracies of 87% and 86%. The SHAP values revealed RFC as the most significant feature, offering a visual depiction of its influence on model predictions. FS analysis demonstrated that RFC metrics are the most important metric affecting SFP model performance, validating SHAP’s applicability in SFP.
Paper Type
Short Paper
DOI
10.62036/ISD.2026.62
Elaborating Explainable AI for Software Fault Prediction: Interpretability Techniques and Performance Insights
Software Fault Prediction (SFP) is an important task of software engineering, which attempts to discover the faulty modules in source code proactively to improve quality and reducing maintenance costs. The machine learning (ML) models has significantly enhanced fault prediction capability, yet these methods are black box, and limits in interpretability and trust. Explainable AI (XAI) approaches, such SHapley Additive exPlanations (SHAP), are useful to get insights, and quantify the impact of specific characteristics to the model decisions. The study proposed a robust and interpretable framework for SFP which integrates class imbalance handling (SMOTE), optimisation (using hyperparameter tuning (SMOHY)), SHAP-based explainability and Feature Sensitivity (FS) analysis to produce accurate and transparent models using 41 ML classifiers against 41 opensource java projects. SMOHY improved the efficiency and predictability of the SFP model by 22.38% compared to ORGD and 6.5% compared to SMOTE, attaining highest median AUC value of 0.82. ExtraTrees and RandomForest algorithms achieve the highest median accuracies of 87% and 86%. The SHAP values revealed RFC as the most significant feature, offering a visual depiction of its influence on model predictions. FS analysis demonstrated that RFC metrics are the most important metric affecting SFP model performance, validating SHAP’s applicability in SFP.
Recommended Citation
Rathi, S., Neti, L.B.M., Misra, S. & Kumar, L.(2026). Elaborating Explainable AI for Software Fault Prediction: Interpretability Techniques and Performance Insights. In M. Valenta, B. Mannová, R. Pergl, A. Przybylek, M. Lang, H. Linger, C. Schneider, N. Iivari, & E. Insfran (Eds.), Making ISD Sustainable: Reloaded with AI and Automation (ISD2026 Proceedings). Prague, Czech Republic: Czech Technical University in Prague. ISBN: 978-80-01-07585-2. https://doi.org/10.62036/ISD.2026.62