Enhancing Algorithmic Fairness: Integrative Approaches and Multi-Objective Optimization Application in Recidivism Models

The fairness of Artificial Intelligence (AI) has gained tremendous attention within the criminal justice system in recent years, mainly when predicting the risk of recidivism. The primary reason is attributed to evidence of bias towards demographic groups when deploying these AI systems. Many proposed fairness-improving techniques applied at each of the three phases of the fairness pipelines, pre-processing, in-processing and post-processing phases, are often ineffective in mitigating the bias and attaining high predictive accuracy. This paper proposes a novel approach by integrating existing fairness-improving techniques: Reweighing, Adversarial Learning, Disparate Impact Remover, Exponential Gradient Reduction, Reject Option-based Classification, and Equalized Odds optimization across the three fairness pipelines simultaneously. We evaluate the effect of combining these fairness-improving techniques on enhancing fairness and attaining accuracy. In addition, this study uses multi- and bi-objective optimization techniques to provide and to make well-informed decisions when predicting the risk of recidivism. Our analysis found that one of the most effective combinations (i.e., disparate impact remover, adversarial learning, and equalized odds optimization) demonstrates a substantial enhancement and balances achievement in fairness through various metrics without a notable compromise in accuracy.