Process metrics, valued for their language independence and ease of collection, have been shown to outperform product metrics in defect prediction. Among these, change entropy (Hassan, 2009) is widely used at the file level and has proven highly effective. Additionally, past research suggests that co-change patterns provide valuable insights into software quality. Building on these findings, we introduce Co-Change Graph Entropy, a novel metric that models co-changes as a graph to quantify co-change scattering.

Co-changes refer to the phenomenon wherein two or more software entities are modified together within the same commit or when they are changed to accomplish a specific task or functionality. The key to accurate co-change prediction lies in effectively predicting unseen co-changes-those occurring between entities that have not been co-changed before. However, despite considerable research on co-change patterns and prediction, there remains a significant gap in understanding unseen co-changes, including their prevalence, complexity, and predictability. We model co-changes as a graph, treating unseen co-change prediction as a link prediction task. Our method leverages file proximity measures derived from both homogeneous and heterogeneous networks, alongside other similarity measures, to predict these unseen co-changes. Analysis of 14 Apache Software Foundation projects revealed a significantly higher prevalence of unseen co-changes (up to 23x more frequent in specific projects and 7x on average) compared to recurrent co-changes. Interestingly, comparisons of co-change complexity based on file distance in the directory structure revealed no decisive differences between the two types. Our graph-based approach achieved good accuracy in predicting unseen co-changes (average AUROC of 0.84, with some projects reaching up to 0.98). Our method achieved significantly better performance than the baseline approach, demonstrating an average recall of 90% and a precision of 38%. While the precision value might seem modest, our approach achieves very high precision@k values (near 100% for 13 out of 14 projects up to$k=100$), underlining its effectiveness in real-world applications.