Evaluating SHAP-based explanations for defect predictions In software engineering : a multi-stakeholder user study and sustainability consideration

Abstract

This research provides the first comprehensive empirical evaluation of SHAP-based explanations in software defect prediction systems, systematically investigating stakeholder perception patterns, practical effectiveness in decision-making outcomes, and sustainability implications through measurement of energy consumption.

The prediction models of software defect have gained sophistication yet their “blackbox” nature often leaves different stakeholders involved in software development struggling to understand and trust their recommendations. Despite the explainable AI techniques like SHAP (SHapley Additive ExPlanations) promise to alleviate this shortcoming and although we are still not in possession of a finished product information regarding how the different stakeholders actually think and employ such explanations in real life software development scenarios. In addition, there is the cost of producing environmental cost of generating that kind of explanation is still an unexplored territory, raising important questions about the sustainability of explainable AI in practice.

The research uses a sequential mixed-methods research methodology, which incorporates qualitative stakeholder analysis, quantitative effectiveness testing and empirical sustainability research. A defect prediction model based on Random Forest, augmented with SHAP explainers was created and tested in the following three ways: (1) semi-structured interviews and thematic analysis of 21 software engineering professionals (12 developers and 9 managers) to gain understanding of stakeholder perception and use patterns; (2) within-participants A/B testing of 11 developers regarding decision-making performance with and without SHAP explainers across realistic defect resolution situations; and (3) systematic energy consumption measurement using CodeCarbon instrumentation to quantify the computational overhead of explanation generation.

The study shows a large discrepancy between the stakeholders in how they perceive and use explanations, with developers showing an appreciation of technical information by far the most (78%) relative to managers (33%), a 45%point advantage that raises questions over how to effectively design the explainability of AI solutions to suit all stakeholders. Thematic analysis identifies three fundamental patterns: Technical Transparency Paradox (mathematical transparency can reduce practical understanding for non-technical stakeholders), Contextual Translation Needs (raw SHAP outputs require domain-specific interpretation), and Trust Through Understanding (stakeholder trust correlates with explanation comprehension at appropriate expertise levels). A controlled experimental analysis shows that SHAP explanations have a substantial positive effect on decision-making effectiveness producing a mean accuracy gain of 10.6 percentage points (p = 0.016, Cohen d = 0.87) and boosting confidence levels with no efficiency trade-offs. Nevertheless, the sustainability analysis shows that there is significant computational overhead, where the generation of SHAP explanations consumes 40.7% extra energy (0.220 kWh) and 95.5 gCO2eq more emissions compared to those observed in the baseline defect prediction operations.

The research has three major contributions to explainable AI theory and practice. First, it provides an empirical foundation that there is indeed a role-specific difference in the perception and utilization of XAI, demonstrating that explanation effectiveness depends fundamentally on stakeholder expertise and decision context rather than technical sophistication alone. Second, it is the first quantitative empirical validation of SHAP explanation effectiveness in software engineering contexts, moving beyond computational metrics to demonstrate measurable improvements in real-world decision-making scenarios. Third, it adds sustainability concerns to XAI evaluation schemes by measuring the environmental of explanation generation, which demonstrates a high-energy cost that must be weighed against the interpretability value.

These findings are used in the design of stakeholder-adaptive explanation systems, which find the right balance between interpretability, effectiveness, and environmental responsibility and define methodological grounds that holistically evaluate the overall effectiveness of an XAI by taking into consideration both technical performance and human factors and sustainability constraints. considerations.