Procedural Fairness and Its Relationship with Distributive Fairness in Machine Learning

TL;DR

This work addresses the gap between procedural fairness and distributive fairness in ML by introducing an in-training objective based on $GPFFAE$ to enforce fair decision logic. It replaces costly SHAP explanations with a gradient-based FAE and adds a regularization term to the loss, yielding procedurally fair models that also improve distributive fairness across seven datasets with only ~ $0.9\%$ average accuracy loss. The paper also investigates how inherent dataset bias and procedural fairness influence distributive fairness, showing that data bias and model-era fairness can either compound or cancel biases in outcomes, and that procedural fairness optimization generally reduces unfairness at the source. A key takeaway is that attaining distributive fairness via data debiasing plus procedural fairness training offers a root-cause solution, while optimizing distributive fairness alone can produce fair outcomes but leave underlying processes or data biased. The findings have practical implications for designing fair ML pipelines and motivate future multi-objective optimization to balance accuracy, procedural fairness, and distributive fairness.

Abstract

Fairness in machine learning (ML) has garnered significant attention in recent years. While existing research has predominantly focused on the distributive fairness of ML models, there has been limited exploration of procedural fairness. This paper proposes a novel method to achieve procedural fairness during the model training phase. The effectiveness of the proposed method is validated through experiments conducted on one synthetic and six real-world datasets. Additionally, this work studies the relationship between procedural fairness and distributive fairness in ML models. On one hand, the impact of dataset bias and the procedural fairness of ML model on its distributive fairness is examined. The results highlight a significant influence of both dataset bias and procedural fairness on distributive fairness. On the other hand, the distinctions between optimizing procedural and distributive fairness metrics are analyzed. Experimental results demonstrate that optimizing procedural fairness metrics mitigates biases introduced or amplified by the decision-making process, thereby ensuring fairness in the decision-making process itself, as well as improving distributive fairness. In contrast, optimizing distributive fairness metrics encourages the ML model's decision-making process to favor disadvantaged groups, counterbalancing the inherent preferences for advantaged groups present in the dataset and ultimately achieving distributive fairness.

Figures (11)

• Figure 1: Relationship graph for the synthetic dataset.
• Figure 2: Comparing the SHAP value of the sensitive attribute of the ML model obtained by the two methods on each dataset. Each red and blue dot represents a data point for an advantaged and disadvantaged group, respectively. The horizontal coordinate values are the results of the SHAP method's explanation of the sensitive attribute of the corresponding data point. Larger absolute SHAP values represent a greater impact on the decision, while a positive/negative sign indicates a positive/negative benefit to the decision. By employing our approach, the influence of sensitive attributes on the decision-making process is diminished to nearly zero, ensuring procedural fairness.
• Figure 3: The trends of the dataset bias (Dataset DP), distributive fairness (DP), and procedural fairness (GPFFAE) metrics with the increase of parameter $p$ which controls the degree of bias in the dataset.
• Figure 4: Relationship graph for the LFR-TSA and LFR-FSA datasets.
• Figure 5: The trends of GPFFAE (procedural fairness) and DP (distributive fairness) metrics with changes in $w_s$. A larger absolute value of $w_s$ means that the decision-making process of the ML model is more unfair, and a positive/negative sign indicates that the ML model is biased in favor of the advantaged/disadvantaged group. Figures show mean values over 10 random runs with a $1-\sigma$ error-bar. Here, $\sigma$ is the standard deviation.