Online learning for software defect prediction

Abstract

Just-in-Time Software Defect Prediction (JIT-SDP) aims to detect defect-inducing code changes at the moment they are committed, allowing developers to take proactive measures to ensure software quality. However, traditional JIT-SDP models struggle with concept drift and the need for large amounts of labeled data, making them less effective in dynamic software development environments. This master thesis introduces the Semi-supervised Stochastic Weight Averaging (S3WA) model, an adaptive learning approach that leverages both labeled and unlabeled data while dynamically adjusting to evolving data streams. The model is evaluated against state-of- the-art online learning techniques using both artificial and real-world datasets, with a particular emphasis on JIT-SDP scenarios. The results demonstrate that S3WA maintains higher predictive accuracy over time compared to existing models, effectively handling concept drift while reducing the reliance on labeled data. These findings highlight the potential of adaptive semi-supervised approaches to improve defect prediction in real-time software development workflows.