Decision-centric fairness: Evaluation and optimization for resource allocation problems

TL;DR

This work addresses fairness in score-based resource allocation by introducing decision-centric fairness, which enforces demographic parity only within the decision-making region defined by deployment thresholds. It formalizes the concept, proposes a learning objective that penalizes unfairness only on the top-score subset using the $1$-Wasserstein distance, and develops decision-centric performance metrics like $\text{AUC-PR}_{\tau}$ alongside fairness measures $\text{ABPC}_{\tau}$ and $\text{ABCC}_{\tau}$. Through semi-synthetic experiments on multiple datasets, the authors show that decision-centric fairness can achieve better trade-offs between fairness and predictive performance than global fairness, especially under dynamic resource constraints. The findings have practical implications for credit risk, churn management, and other resource-allocation contexts where decision thresholds vary over time, and point to future work in uplift modeling, ranking-based approaches, and multi-attribute fairness with legal considerations.

Abstract

Data-driven decision support tools play an increasingly central role in decision-making across various domains. In this work, we focus on binary classification models for predicting positive-outcome scores and deciding on resource allocation, e.g., credit scores for granting loans or churn propensity scores for targeting customers with a retention campaign. Such models may exhibit discriminatory behavior toward specific demographic groups through their predicted scores, potentially leading to unfair resource allocation. We focus on demographic parity as a fairness metric to compare the proportions of instances that are selected based on their positive outcome scores across groups. In this work, we propose a decision-centric fairness methodology that induces fairness only within the decision-making region -- the range of relevant decision thresholds on the score that may be used to decide on resource allocation -- as an alternative to a global fairness approach that seeks to enforce parity across the entire score distribution. By restricting the induction of fairness to the decision-making region, the proposed decision-centric approach avoids imposing overly restrictive constraints on the model, which may unnecessarily degrade the quality of the predicted scores. We empirically compare our approach to a global fairness approach on multiple (semi-synthetic) datasets to identify scenarios in which focusing on fairness where it truly matters, i.e., decision-centric fairness, proves beneficial.

Figures (14)

• Figure 1: Densities of predicted scores $\tilde{y}$ for two demographic groups (with protected attributes $s = 0$ and $s = 1$, in blue and red, respectively), along with the corresponding demographic parity ($DP$) across all possible thresholds. By inducing decision-centric fairness, we aim to achieve demographic parity in the decision-making region, i.e., where $\tilde{y} > \tau$, to ensure a proportionally equal number of positive outcomes across the two groups at all thresholds within this region.
• Figure 2: The decision-centric fairness approach for different values of $\lambda$. Score distributions (PDFs) on a test set after training each model for 30 epochs are shown, split by the protected attribute ($s = 0$ and $s = 1$). Solid lines represent the distributions of all predicted scores, while dotted lines represent the top-$k\%$ score distributions. The decision threshold $\tau$ is shown as a vertical line. An animated version of these plots is available in the https://github.com/SimonDeVos/DCF (github.com/SimonDeVos/DCF).
• Figure 3: An example partial precision-recall curve and its corresponding $\text{AUC-PR}_\tau$, which summarizes precision and recall for all decision thresholds within the decision-making region $[\tau,1]$. This region corresponds to the top-left segment of the PR curve starting at $\tau$.
• Figure 4: Pareto fronts illustrating the trade-off between predictive performance and fairness ($\text{AUC-PR}_{\tau}$ vs. $\text{ABPC}_{\tau}$) for the decision-centric (orange) and global (blue) fairness induction approaches for three datasets.
• Figure 5: Pareto fronts illustrating the trade-off between predictive performance and fairness ($\text{AUC-PR}_{\tau}$ vs. $\text{ABCC}_{\tau}$) for the decision-centric (orange) and global (blue) fairness induction approaches for three datasets.