Predicting Defective Lines Using a Model-Agnostic Technique

Defects are concentrated in tiny code sections: The research found that only 1-3% of lines within a file typically contain defects, meaning developers waste significant effort reviewing entire files when they should focus on specific problem areas.

Explainable AI improves software quality assurance: The study used LIME, an explainable AI technique, to identify which specific code tokens (small units of code) contribute to defect predictions, making the model's reasoning transparent rather than treating it as a "black box."

The LINE-DP framework outperforms existing methods: The proposed approach achieved better results than six baseline approaches, with a recall rate of 0.61 (meaning it correctly identified 61% of actual defective lines) while maintaining practical efficiency at around 10 seconds of computation time.

False alarms remain a practical challenge: The framework flagged non-defective lines as problematic about 47% of the time, suggesting developers still need to exercise judgment rather than blindly trusting automated predictions.

Model-agnostic techniques enable practical software testing: By using explainable AI methods that work across different types of models, this approach can be adapted to various software systems and development contexts, making responsible AI tools more accessible to real-world development teams.