Empirical Study of Meta-learning-Based Approach for Predictive Mutation Testing

One of the thrust areas in software testing is design of good quality test suites. Mutation testing is a widely acknowledged technique for validating test suites. Mutation testing works by altering the original program so as to introduce faults intentionally. The altered program or mutant is tested against the test suite. If the test suite is able to detect or kill all the intentionally introduced faults, it is considered adequate. The adequacy of test suite is measured in terms of mutation score. However, the cost of executing mutant programs is a big deterrent to adoption of mutation testing in practice. Therefore, research community has proposed a Machine Learning (ML) based strategy called Predictive Mutation Testing (PMT). PMT allows mutation testing outputs to be predicted without actually executing mutant programs. The research objective undertaken in this empirical study is to propose a meta-learning-based approach for PMT. The proposed approach is compared with classical machine learning and other state of art ensemble-based approaches. Analysis of the results based on statistical tests shows that the proposed meta-learning-based approach for PMT outperforms classical machine learning as well as ensemble approaches like XGBoost, CatBoost, and random forests.