Contextual and time-based developer metrics in software defect prediction

In today's world, software products play an important role in making our lives easier in our business and social life. This leads to an increase in interest in software products and to a growing variety of software products, increasing their size and complexity. The concept of quality in software has been emphasized since there is a competition between software products and ever increasing in software costs. In order to increase the quality of the software, error prediction models are used which produce data- driven results. These kinds of data can be obtained from the software process, source codes, product outputs, and organizational structures. Additionally, there are many studies that examine the people factors and their relationship to software defects. Developer related metrics can be picked up by different perspectives, such as different codding skills and techniques, experience levels, ownership of the related modules, organization structure of companies. These kinds of metrics have been captured the in one or a limited number of snapshots of source codes, although the experience and knowledge of a developer are constantly changing depending on time and content of the codebase. In this study, we have offered to measure time-based developer experience with specified context are files and directories. We have captured time-based experience metrics obtaining to the previous development activities on specified context and examine the relationship of these metrics with the software defects. In our study, we have recalculated the developer metrics in each commit for each file and directory on two granularity levels: file level and commit level. We have used static code metrics and time-based and contextual developer experience metrics together with our models. To evaluate our defect prediction models, we have used Jackrabbit and Lucene projects. In order to define a classification method, we have compared the five the most preferred classification algorithms. We have identified two algorithms that perform best for our data sets. The performances of our models are evaluated according to the results of precision, recall, false positive rate and f-measure. We have used correlation-based feature selection method for specifying the best explanatory metrics of software defects. According to model outputs, time-based developer metrics which are obtained from a file are good merits for software defect prediction. In addition, static code metrics and time-based developer metrics must be used together since the developer's experience information cannot be calculated in the early versions of the source code.