Mining Action Rules for Defect Reduction Planning

Explainable AI for software has a faithfulness problem: The paper identifies that post-hoc explanations from black-box machine learning models don't always accurately reflect what those models actually compute, raising concerns about relying on such explanations for critical software quality decisions.

Simpler, more transparent approaches can outperform complex AI systems: CounterACT generates defect reduction plans without using black-box models, yet achieves higher overlap with actual developer changes (95% at release level, 85.97% at commit level) compared to six established competing approaches, suggesting that explainability and effectiveness aren't sacrificed by avoiding opaque AI.

Counterfactual explanations provide actionable guidance: Rather than just explaining why code is buggy, CounterACT generates specific "courses of action" showing what code changes would reduce defects, making the AI output directly useful for developers rather than merely interpretable.

Transparency enables better quality assurance: By providing plans with measurable precision, recall, and F1-scores (median 88.12%), the approach allows software teams to evaluate and trust the recommendations before implementing them, supporting more responsible deployment of AI in critical software development processes.

Responsible AI in software engineering requires moving beyond black boxes: The research demonstrates that generating defect reduction plans through interpretable methods rather than explaining opaque models may be a more trustworthy path for using AI to improve software quality.