ACONA: active online model adaptation for predicting continuous integration build failures

Continuous Integration (CI) reduces risk in software development, but a CI build usually brings huge time and resource consumption. Machine learning methods have been employed to cut the expenses of CI and provide instant feedback by predicting CI results. Nevertheless, effective learning requires massive training data which is not available for a new project. Moreover, due to the diversified characteristics of different projects, reusing models built on other projects leads to poor performance. To address this problem, we propose a novel active online model adaptation approach ACONA, which dynamically adapts a pool of classifiers trained on various projects to a new project using only a small fraction of new data it actively selects. With empirical study on Travis CI, we show that ACONA achieves an improvement of F-Measure by 40.0% while reducing Accumulated Error by 63.2% and the adapted model outperforms existing approaches.