Reducing Labeling Effort in Architecture Technical Debt Detection through Active Learning and Explainable AI

Active learning can significantly reduce the manual work required to identify architectural problems in software: The study found that using "Breaking Ties" active learning achieved a 49% reduction in annotation effort while maintaining strong performance (F1-score of 0.72), meaning organizations can identify architectural technical debt with less human labeling work.

Explainability tools help developers understand why AI systems flag architectural issues: The researchers used SHAP and LIME to explain the model's decisions, with expert evaluation showing both methods provided useful insights—LIME was particularly preferred for clarity—making AI-assisted detection more trustworthy and actionable for development teams.

Keyword-based filtering combined with AI can scale detection across large codebases: The approach identified over 103,000 candidate architectural debt issues across 10 open-source projects, demonstrating that hybrid human-AI methods can help organizations monitor technical debt at scale without overwhelming manual review processes.

Architectural technical debt is harder to detect than other types of technical debt: The paper emphasizes that architectural compromises are "abstract and context-dependent," which is why combining multiple techniques (filtering, active learning, and explainability) was necessary—suggesting that responsible AI for code quality requires thoughtful, multi-layered approaches rather than single solutions.