University of Saskatchewan
Reliable and Human-Centered Explainable AI for Software Defect Prediction
Abstract
dc:description.abstractMachine learning (ML)-based software defect prediction has become an important technique for improving software quality by identifying defect-prone modules before testing and maintenance. However, despite achieving high predictive performance, these models are rarely adopted in practice because developers often cannot understand, trust, or effectively use their predictions. Existing post-hoc explainability techniques, including LIME, SHAP, and breakDown, frequently generate inconsistent, unstable, and difficult-to-interpret explanations, limiting both their reliability and practical usefulness. Consequently, developing reliable and human-centered explainable AI (XAI) has become a critical challenge for software engineering. This dissertation addresses this challenge through a series of interconnected studies that improve the reliability, usability, and practical integration of explainable AI for software defect prediction. The research first investigates developers' perceptions, expectations, and real-world challenges in adopting explainable AI through qualitative, scenario-based, and empirical studies, demonstrating that explanation consistency, trust, interpretability, and workflow integration are fundamental requirements for practical adoption. Motivated by these findings, the dissertation systematically analyzes disagreement among widely used explainability techniques and shows that different explainers often produce conflicting interpretations for identical predictions. To improve explanation reliability, this dissertation proposes multiple aggregation strategies that reconcile complementary evidence from different explainers to generate more stable, consistent, and trustworthy explanations. Building upon these methods, it introduces XMENTOR, a developer-centered Visual Studio Code extension that integrates aggregated explanations directly into software development workflows, enabling developers to access reliable explanations within their natural programming environment. The dissertation further improves explanation stability through a rank-aware aggregation framework and introduces CoScoreX, a contextual and contrastive explainability framework that incorporates software context and comparative reasoning to generate more faithful and developer-oriented explanations. Recognizing the growing adoption of deep learning in software engineering, this dissertation further extends explainability to transformer-based models, including CodeBERT, GraphCodeBERT, and CodeT5, through a comprehensive benchmarking framework that evaluates explanation techniques using fidelity, stability, and human-centered evaluation metrics. Finally, the dissertation advances explainable AI toward intelligent developer assistance by proposing an agent-based framework in which autonomous agents dynamically integrate multiple sources of evidence, adapt explanations to developer needs, incorporate feedback, and generate context-aware, actionable explanations throughout the software development lifecycle. Overall, this dissertation demonstrates that reliable explainability requires explanation aggregation, contextual reasoning, and adaptive intelligence, while human-centered explainability requires explanations to be understandable, trustworthy, actionable, and seamlessly integrated into developers' workflows. Together, these contributions advance Reliable and Human-Centered Explainable AI for Defect Prediction, promoting the practical adoption of trustworthy AI-assisted software engineering.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy (Ph.D.)
- Level thesis:degree_level
- Doctoral
- Discipline thesis:degree_discipline
- Computer Science
- Department thesis:degree_department
- Computer Science
- Grantor
- University of Saskatchewan
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Roy, Saumendu
- Advisors dc:contributor.advisor
-
- Roy, Banani
- Roy, Chanchal K.
- Committee members dc:contributor.committeemember
-
- Wang, Song
- Wahid, Khan Arif
- Rochan, Mrigank
- Gutwin, Carl
Subjects
dc:subject × 1Rights
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/10388/19005