Mutation testing offers a rigorous test efficacy criterion but is inherently expensive due to the large number of test goals (mutants), many of which have low utility—they are unproductive and not actionable. Recent work has begun to address this shortcoming and demonstrated that a mutant’s program context, extracted from a program’s abstract syntax tree (AST), is predictive of the utility of that mutant.

This paper makes three key contributions: (1) It proposes a new definition of mutant utility based on properties of the tests elicited by a mutant, rather than properties of the mutant itself; (2) It ex-plains how a-priori selection of mutants corresponds poorly to mutation testing in practice, and offers an alternative approach; (3)It presents and evaluates a machine learning-based mutant prioritization approach, showing that it can assist engineers to test more efficiently—frequently with half as much work for the same test completeness.