Addressing diversity in hiring procedures: a generative adversarial network approach

Abstract

The combination of machine learning and organizational psychology has led to innovative methods to address the diversity-validity dilemma in personnel selection, which is the tradeoff between selecting valid predictors of job performance while minimizing adverse impact. Recent technological advancements provide new strategies to mitigate gender biases while preserving the ability to predict job performance accurately. Our research introduces a novel framework consisting of three blocks: a gating block to filter user data, a bias measurement block using an adversarial network for detecting gender bias, and a feature importance block, identifying and removing biased, non-contributory performance features. We applied this model architecture to both simulated datasets and real-world hiring scenarios, with a particular emphasis on personality-based algorithms, aiming to refine the hiring predictive models to be gender fair and to meet the EEOC standards. In simulated environments, 70% of the predictive models get their impact ratio improved, approaching the ideal ratio by 22.73% while only incurring a slight 4.16% decrease in performance predictability. Real-world data testing yielded similar improvements, with 71% of the models showing an increased impact ratio, 18.8% closer to the ideal, and a 2.18% increase in predictive accuracy for job performance. The findings suggest that the application of neural networks can be an effective strategy for enhancing fairness in hiring practices with only minimal loss in predictive accuracy. Future research directions should explore the refinement of these models and the implications of their deployment in high-stakes hiring environments.