Counterfactual Reasoning Algorithms For Bias Mitigation In Automated Recruitment Systems

Abstract

As automated recruiting algorithms employ machine learning models in the process of candidate ranking and filtering, their propensity to learn and perpetuate historical biases toward gender, ethnicity, and academic education is common. Using counterfactual thinking as a theoretical base, one can pinpoint and fix discrimination by exploring what the model would have predicted if the sensitive attributes were altered. In this paper, study present a new method that incorporates counterfactual data augmentation, optimization under fairness constraints, and individual measures of counterfactual fairness to form a seamless pipeline that can be implemented in production for hiring purposes. The experiments carried out on two benchmark hiring datasets show that the proposed algorithm decreases demographic disparity difference by up to 34.7%, as well as lowers equalized odds disparity by 28.9% against state-of-the-art methods of debiasing, without compromising the model's accuracy, which remains within 2.1% of the initial value. The comparative table and ROC-curve visualization provide additional evidence that counterfactual-based fine-tuning surpasses pre-processing and post-processing approaches with regard to performance on all protected groups.