Integrating Defect Data, Code Review Data, and Version Control Data for Defect Analysis and Prediction

In this thesis, we present a new approach to integrating software system defect data: Defect reports, code reviews and code commits. We propose to infer defect types by keywords. We index defect reports into groups by the keywords found in the descriptions of those reports, and study the properties of each group by leveraging code reviews and code commits. Our approach is more scalable than previous studies that consider defects classified by manual inspections, because indexing is automatic and can be applied uniformly to large defect dataset. Also our approach can analyze defects from programming errors, performance issues, high-level design to user interface, a more comprehensive variety than previous studies using static program analysis. By applying our approach to Honeywell Automation and Control Solutions (ACS) projects, with roughly 700 defects, we found that some defect types could be five times more than other defect types, which gave clues to the dominant root causes of the defects. We found certain defect types clustered in certain source files. We found that 20%-50% of the files usually contained more than 80% of the defects. Finally, we applied a known defect prediction algorithm to predict the hot files of the defects for the defect types of interest. We achieved defect hit rate 50%-90%.