Task-tailored Pre-processing: Fair Downstream Supervised Learning

TL;DR

This work tackles fairness in supervised learning by proposing task-tailored pre-processing that yields fair, utility-preserving transformed data for arbitrary downstream models. It replaces traditional data-fairness regularization with a supervision-aware objective based on the Hirschfeld–Gebelein–Rényi (HGR) correlation, optimized via a min-max bilevel framework where the outer problem learns a pre-processing map ${G^*}$ and the inner problem learns an upstream predictor ${\tilde{h}}^*}$. The authors derive theoretical guarantees linking upstream fairness and utility to downstream performance, including consistency guarantees across end users and invariant fairness under feature engineering, and extend the analysis to separation via conditional HGR. Empirically, the framework demonstrates improved fairness-utility trade-offs on tabular datasets and CelebA images, with robust performance across diverse downstream models and meaningful downstream fairness improvements over competing methods. The approach offers practical advantages for data distributors and legacy systems by enabling proactive fairness control in a model-agnostic downstream setting, while highlighting considerations like covariate-shift handling and parameter tuning for real-world deployments.

Abstract

Fairness-aware machine learning has recently attracted various communities to mitigate discrimination against certain societal groups in data-driven tasks. For fair supervised learning, particularly in pre-processing, there have been two main categories: data fairness and task-tailored fairness. The former directly finds an intermediate distribution among the groups, independent of the type of the downstream model, so a learned downstream classification/regression model returns similar predictive scores to individuals inputting the same covariates irrespective of their sensitive attributes. The latter explicitly takes the supervised learning task into account when constructing the pre-processing map. In this work, we study algorithmic fairness for supervised learning and argue that the data fairness approaches impose overly strong regularization from the perspective of the HGR correlation. This motivates us to devise a novel pre-processing approach tailored to supervised learning. We account for the trade-off between fairness and utility in obtaining the pre-processing map. Then we study the behavior of arbitrary downstream supervised models learned on the transformed data to find sufficient conditions to guarantee their fairness improvement and utility preservation. To our knowledge, no prior work in the branch of task-tailored methods has theoretically investigated downstream guarantees when using pre-processed data. We further evaluate our framework through comparison studies based on tabular and image data sets, showing the superiority of our framework which preserves consistent trade-offs among multiple downstream models compared to recent competing models. Particularly for computer vision data, we see our method alters only necessary semantic features related to the central machine learning task to achieve fairness.

Figures (20)

• Figure 1: A data manager has the pre-processing map $G^*$ that transforms original data ${\bf D}$ to pre-processed data ${\tilde{\bf D}}$ and the upstream supervised model ${\tilde{h}}^*$ that satisfies a certain level of fairness and accuracy on ${\tilde{\bf D}}$. Then ${\tilde{\bf D}}$ is distributed to end users who can fit their own downstream supervised models ${\tilde{h}}_{k}^*$, not sticking to ${\tilde{h}}^*$, but the manager wants their downstream models to satisfy similar levels of fairness and accuracy with the upstream model ${\tilde{h}}^*$. The subscript $k(i)$ implies the model of the $i$th end user.
• Figure 2: A Toy example with data generated from $Y = (2A-1)\sin(X)+ 2AX + \epsilon$ where $A \sim {\rm Ber}(0.5)$, $X \sim {\rm N}(A,1^2)$, and $\epsilon \sim {\rm N}(0,0.1^2)$. The first row illustrates the original and transformed data. For both Feldman et al. (2015) feld:etal:15 and Gordaliza et al (2019) gord;etal;19, $X$ is transformed to $\tilde{X}$ to maximally remedies discrimination, while $\tilde{Y}=Y$. Our method yields $({\tilde{X}},{\tilde{Y}})$ having a more efficient accuracy and fairness trade-off curve, as shown in the second row. For figures in the second row, the dot is the average, with the bars of 2$\times$standard error. The line styles distinguish the choices of fairness degree: weak (dashed line), medium (dash-dot line), and strong (dotted line). In our case, the transformed data under different $A$ overlap with each other, both showing a smooth intermediate functional relationship in contrast to the existing methods.
• Figure 3: Comparison for statistical parity in handling $A_{\text{bin}}$: The closer to the upper left the scores are, the higher performance is achieved. Downstream and Upstream indicate ${\tilde{h}}_k^*$ and ${\tilde{h}}^*$ respectively. The dot and each bar imply the average and $2\times$standard error in each axis. The line styles distinguish Budgets 1 (dashed line), 2 (dash-dot line), and 3 (dotted line).
• Figure 4: Comparison of HV scores: The hypervolume indicators are calculated based on the three score pairs (by the three budgets) for each competing method and downstream model. 2$\times$Standard errors are drawn from the 5 independent runs. A higher score implies that its corresponding method tends to have a more efficient and diverse trade-off curve. See also Figure \ref{['fig:hv-indicator-ks']} for the KS-type fairness score.
• Figure 5: Consistency of AUC and SP in $A_{\text{bin}}$: Boxplots compare consistency scores across competing methods and the inclusion of the boosting model in calculating the scores.

Theorems & Definitions (21)

• Definition 1: Independence
• Definition 2: Separation
• Definition 3: HGR correlation
• Proposition 1
• Theorem 1
• Corollary 1
• Remark 1
• Lemma 1
• Theorem 2
• Remark 2