A Feedback-Enhanced Effort-Aware Just in Time Defect Prediction Pipeline

Résumé fourni par la source

Software defect prediction aims to identify faults in code before they cause failures. It helps developers locate problematic modules early in the development cycle. Early detection reduces the cost of fixing defects and prevents them from affecting users. Just-in-Time defect prediction is the concept of detecting buggy commits early on in the development stage rather than in the release. While recent research focuses on improving the accuracy of pre-diction faulty commits, fewer studies explore how to do that task with minimal effort. In this study, we propose an effort-aware JIT defect prediction pipeline that utilizes code pre-trained models finetuned with Low Rank Adaptation (LoRA) and a feedback loop mechanism. The aim is to improve fault detection while reducing effort. Our experiments evaluate four different transformer-based models: CodeBERT, JavaBERT, UniXcoder, and RoBERTa on ApacheJIT. Our results illustrate CodeBERT and UniXcoder are benefiting the most from the feedback loop with the highest gain in F1 score, where CodeBERT highest gain is 0.14 and UniXcoder is 0.02. RoBERTa performed the best across experiments with an F1 score of 0.73. However, Popt metric remains static with limited gain that reached maximum 0.02, suggesting that further experiments are needed to improve performance.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.