Using machine learning to classify test outcomes

When testing software it has been shown that there are substantial benefits to be gained from approaches which exercise unusual or unexplored interactions with a system - techniques such as random testing, fuzzing, and exploratory testing. However, such approaches have a drawback in that the outputs of the tests need to be manually checked for correctness, representing a significant burden for the software engineer. This paper presents a strategy to support the process of identifying which tests have passed or failed by combining clustering and semi-supervised learning. We have shown that by using machine learning it is possible to cluster test cases in such a way that those corresponding to failures concentrate into smaller clusters. Examining the test outcomes in cluster-size order has the effect of prioritising the results: those that are checked early on have a much higher probability of being a failing test. As the software engineer examines the results (and confirms or refutes the initial classification), this information is employed to bootstrap a secondary learner to further improve the accuracy of the classification of the (as yet) unchecked tests. Results from experimenting with a range of systems demonstrate the substantial benefits that can be gained from this strategy, and how remarkably accurate test output classifications can be derived from examining a relatively small proportion of results