Human-in-the-Loop Fault Localisation Using Efficient Test Prioritisation of Generated Tests

Many existing fault localisation techniques become less effective or even inapplicable when not adequately supported by a rich test suite. To overcome this challenge, we present a human-in-the-loop fault localisation technique, QFiD, that works with only a small number of initial failing test cases. We augment the failing test cases with automatically generated test data and elicit oracles from a human developer to label the test cases. A new result-aware test prioritisation metric allows us to significantly reduce the labelling effort by prioritising the test cases to achieve maximum localisation accuracy. An evaluation with EvoSuite and our test prioritisation metric shows that QFiD can significantly increase the localisation accuracy. After only ten human labellings, QFiD can localise 27% and 66% of real-world faults in Defects4J at the top and within the top ten, respectively. This is a 13 and 2 times higher performance than when using the initial test cases. QFiD is also resilient to human errors, retaining 80% of its acc@1 performance on average when we introduce a 30% error rate to the simulated human oracle.