Advancing Software Defect Prediction through Ensemble XAI Methods: Insights and Performance Evaluation

Abstract

This doctoral thesis presents a comprehensive investigation into enhancing the interpretability and transparency of Machine Learning (ML) models in the domain of Software Defect Prediction (SDP) through Model-Agnostic eXplainable Artificial Intelligence (XAI) methods. The primary objective is to elucidate the decision-making processes of ML models, both at individual (local) and global levels, thus bridging the gap between predictive power and comprehensibility demanded by stakeholders in the SDP domain.

The methodological approach adopted involves an iterative and exploratory process, utilizing XAI techniques such as ELI5, SHAP, and LIME, among others. These techniques are systematically applied across multiple case studies, each focusing on specific aspects of model interpretability and transparency in SDP. Through iterative refinement and exploration, the research uncovers insights into the importance of features, contributions to individual predictions, and overall model decisions. Furthermore, ensemble modeling techniques are integrated to amalgamate feature importance scores obtained from diverse XAI methods, thereby optimizing predictive accuracy while simultaneously preserving interpretability.

This research significantly contributes to the field of SDP by furnishing a thorough understanding of ML model decision-making processes. It enhances model interpretability and transparency, effectively addressing critical gaps in traditional feature selection and outlier detection methodologies. Moreover, it offers valuable insights into ensemble modeling approaches, elucidating their role in optimizing predictive accuracy while maintaining interpretability. Validation of the developed methodologies is conducted through rigorous empirical studies and comparative analyses, thus ensuring their effectiveness and usability in real-world SDP scenarios.