Title: Advancements in Bug Prediction: A Focus on Ticket-Level Analysis
The ongoing challenge of minimizing software defects has led to the development of various bug prediction methodologies aimed at optimizing testing efforts. Traditionally, these approaches have concentrated on identifying software fragments—such as classes, methods, commits, or lines of code—that are most prone to contain bugs. However, existing bug prediction models primarily assist developers in fixing already identified issues rather than preventing the introduction of new ones. A novel strategy introduced to address this issue is Ticket-Level Prediction (TLP), a framework that seeks to identify tickets—requests to implement changes in software—that are likely to introduce new bugs once executed.
TLP operates through a comprehensive analysis based on three crucial temporal stages in a ticket’s lifecycle: Open, In Progress, and Closed. This approach is grounded in two primary assertions: first, that the accuracy of TLP increases as tickets progress toward the closed stage due to enhancements in feature reliability over time; and second, that the predictive power of the various features used changes at each temporal point.
To facilitate this, the TLP framework assesses 72 features derived from six distinct families: code metrics, developer characteristics, external and internal temperature metrics, intrinsic characteristics, a ticket-to-tickets relationship, and just-in-time (JIT) analysis. The evaluation of TLP employs a sliding-window methodology, carefully balancing feature selection with three machine-learning classifiers, drawn from a dataset comprising approximately 10,000 tickets from two Apache open-source projects.
Preliminary results from the TLP evaluation reveal a notable trend: accuracy in prediction increases significantly as tickets proceed closer to the closed stage. This finding underscores a critical trade-off between the early detection of potential bugs and the accuracy of such predictions. Furthermore, an analysis of feature families indicated that no single family predominated across all stages of the ticket lifecycle. Early in the process, developer-centric features provided the most insightful information, while code metrics and JIT indicators held greater significance as tickets approached closure. Interestingly, temperature-related features maintained a consistent, complementary role throughout the ticket’s lifecycle.
The implications of these findings are substantial, suggesting that defect prediction methods can, and should, be implemented earlier in the software development lifecycle. By shifting the focus upstream, TLP offers significant potential for risk-aware ticket triaging and informs developer assignments before any actual code is written. This advancement not only enhances the bug prediction landscape but also promises to improve overall software reliability and performance.
As the industry continues to embrace more proactive approaches in software development, TLP stands poised to make a significant impact, fostering an environment of quality assurance that prioritizes prevention alongside detection.
