An adaptive approach with active learning in software fault prediction

Background: Software quality prediction plays an important role in improving the quality of software systems. By mining software metrics, predictive models can be induced that provide software managers with insights into quality problems they need to tackle as effectively as possible. Objective: Traditional, supervised learning approaches dominate software quality prediction. Resulting models tend to be project specific. On the other hand, in situations where there are no previous releases, supervised learning approaches are not very useful because large training data sets are needed to develop accurate predictive models. Method: This paper eases the limitations of supervised learning approaches and offers good prediction performance. We propose an adaptive approach in which supervised learning and active learning are coupled together. NaiveBayes classifier is used as the base learner. Results: We track the performance at each iteration of the adaptive learning algorithm and compare it with the performance of supervised learning. Our results show that proposed scheme provides good fault prediction performance over time, i.e., it eventually outperforms the corresponding supervised learning approach. On the other hand, adaptive learning classification approach reduces the variance in prediction performance in comparison with the corresponding supervised learning algorithm. Conclusion: The adaptive approach outperforms the corresponding supervised learning approach when both use Naive-Bayes as base learner. Additional research is needed to investigate whether this observation remains valid with other base classifiers.