Integrating neural mutation into mutation-based fault localization: A hybrid approach

Abstract

Fault localization is an important part of software testing and debugging, helping improve the process of fixing faults. Mutation-Based Fault Localization (MBFL) is widely used, but the reliance of Traditional-MBFL on syntactical mutants often limits its accuracy. To address this, we propose Neural-MBFL, which introduces neural mutation to generate semantically richer mutants using deep learning to better mimic real faults. Additionally, we present NeuraIntegra-MBFL, which combines neural and traditional mutation strategies through mutant combination and suspiciousness aggregation. Experiments on 835 faulty programs from the Defects4J benchmark show that Neural-MBFL improves fault localization compared to Traditional-MBFL, with a 35.50% relative improvement in MAP and 127 more faults localized at TOP-5, while maintaining acceptable computational cost. Compared to Neural-MBFL, NeuraIntegra-MBFL further enhances performance, particularly with suspiciousness aggregation, achieving an additional 11.96% MAP improvement and localizes 45 more faults at TOP-5, demonstrating the effectiveness of integrating suspiciousness scores. Using overlap and correlation analyses, we confirmed the complementarity between Neural-MBFL and Traditional-MBFL. Neural-MBFL is more effective at localizing faults that require understanding deep code semantics, while Traditional-MBFL performs better at handling rule-based modifications. NeuraIntegra-MBFL successfully integrates the strengths of both methods, offering better performance than either approach alone.