Exploiting mutant’s relationship with code, faults, and patches for higher efficacy of mutation analysis

Abstract

Mutation testing has long been used as a powerful testing technique to evaluate the test suite quality as well as to analyse the program under test. However, it still has issues of scalability and applicability due to the high cost of running tests against mutants. To improve its efficacy even on large programs, this dissertation explores mutant's relationship with code, faults, and patches that have not been explored much but are likely to be useful. First, we investigate the mutant's relationship with code to assist predicting the kill of a mutant without running the tests. By exploring the natural language channel of the code around the mutant, we build a deep neural network to learn and predict the mutant's killability. Second, we explore the mutant's relationship with the faults to localise the faults. We propose several statistical inference techniques that can learn the mutant-fault relationships in advance, in terms of their similarity of test executions. Lastly, we investigate the mutant's relationship with the patches. We hypothesise that the mutants (faults) and patches do not syntactically differ from each other and conduct various empirical studies to show how much they are similar. Our empirical results suggest that exploiting those relationships makes mutation analysis more viable and effective.