RAES: Redundancy-Aware Evolutionary Search for Individual Fairness Testing

Authors:

Abstract
Context: Individual Fairness Testing (IFT) identifies discriminatory instances (IDIs) where individuals who differ only in protected attributes receive different predictions from a software system. Population-based IFT methods can repeatedly evaluate identical inputs, consuming a fixed testing budget without expanding the tested input set. Objectives: This work introduces RAES (Redundancy-Aware Evolutionary Search), an evolutionary framework that operationalises redundancy reduction as an online design requirement for population-based IFT. Methods: RAES redesigns mutation, crossover, and adaptive mutation-rate adjustment to reduce test redundancy and increase unique IDI detection during search. Unit-step mutation reduces zero-effect perturbations in integer-encoded features, half-uniform crossover recombines only loci on which parents differ, and diversity-driven adaptation raises the mutation rate when population diversity declines. RAES is evaluated on five tabular datasets and four classifier families. Results: RAES detects 29.85%–320.08% more unique IDIs at 1.02×–28.04× higher throughput than the evaluated baselines, with 22.32×–43.32× lower redundancy. In multi-attribute testing, RAES detects 1.49×–7.53× more IDIs with 6.42×–231.50× lower redundancy. Retraining with RAES generated IDIs reduces discrimination from 0.045 to 0.016 while maintaining model accuracy. Conclusion: Controlling redundancy during test generation improves IDI detection and throughput for the tabular classifiers evaluated in this study.

This publication has 24 references indexed in Scilit: