Towards Reliable Explainable AI: Mitigating Disagreements in Software Defect Prediction with Hybrid Solutions

Machine learning models have revolutionized software defect prediction by improving reliability and reducing maintenance costs. However, their "black-box" nature often leaves developers struggling to trust and interpret predictions, limiting their practical use. While Explainable AI (XAI) techniques like LIME, SHAP, and BreakDown aim to make predictions more interpretable, they often produce inconsistent and conflicting explanations, making it difficult for developers to rely on their outputs. This research introduces a hybrid solution to tackle these challenges by combining optimized feature engineering with an innovative aggregation mechanism. Feature engineering refines the input data to reduce discrepancies in feature importance rankings across different XAI methods. Simultaneously, the aggregation mechanism consolidates explanations from multiple techniques, focusing on the most consistent and actionable features. Tests on real-world datasets show that this approach effectively minimizes disagreements while enhancing the clarity and reliability of predictions. By bridging the gap between model accuracy and interpretability, this work offers a practical framework that empowers developers with trustworthy, actionable insights for software defect prediction, paving the way for more transparent and reliable AI solutions in software engineering.