Active learning empirical research on cross-version software defect prediction datasets

Abstract

Software quality plays an important part in software engineering. Active learning is introduced to conduct supervised learning classifier because labeling cost is very high. However, in the real software quality assurance process, there are fewer labeled instances in the initial stage of software development, and there may be a historical data set developed by the same team. Therefore, learning from the historical data set can be used for an active learning query strategy. In our empirical study, we design and conduct experiments on promise datasets, which are gathered from real open-source projects. We find that the meta active learning query strategy can perform better than the commonly used query strategy when a little data is labeled.