Lookahead Counterfactual Fairness

TL;DR

This work identifies a key limitation of static counterfactual fairness: it often fails to guarantee fair downstream outcomes when individuals adapt to ML decisions. It introduces lookahead counterfactual fairness (LCF), which enforces fairness on the future status after an individual's response, and extends to path-dependent variants. The authors provide theoretical conditions under which LCF can be satisfied, construct a predictor g(\check{Y},U) that achieves perfect or relaxed LCF in a linear causal setting, and generalize to path-dependent fairness. Empirical validation on synthetic data and the Law School Success dataset shows that LCF-based predictors reduce disparities in future outcomes (AFCE) with competitive or improved predictive performance (MSE) compared to baselines. The results demonstrate a practical route to mitigating downstream inequities in systems where decision-makers influence future behavior.

Abstract

As machine learning (ML) algorithms are used in applications that involve humans, concerns have arisen that these algorithms may be biased against certain social groups. \textit{Counterfactual fairness} (CF) is a fairness notion proposed in Kusner et al. (2017) that measures the unfairness of ML predictions; it requires that the prediction perceived by an individual in the real world has the same marginal distribution as it would be in a counterfactual world, in which the individual belongs to a different group. Although CF ensures fair ML predictions, it fails to consider the downstream effects of ML predictions on individuals. Since humans are strategic and often adapt their behaviors in response to the ML system, predictions that satisfy CF may not lead to a fair future outcome for the individuals. In this paper, we introduce \textit{lookahead counterfactual fairness} (LCF), a fairness notion accounting for the downstream effects of ML models which requires the individual \textit{future status} to be counterfactually fair. We theoretically identify conditions under which LCF can be satisfied and propose an algorithm based on the theorems. We also extend the concept to path-dependent fairness. Experiments on both synthetic and real data validate the proposed method.

Figures (6)

• Figure 1: Causal graph in Example \ref{['example:1']}.
• Figure 2: A causal graph and individual responses with two features $X_1,X_2$. The black arrows represent the connections described in structural functions. The red arrows represent the response process. The green dash arrows are the potential connection to prediction $\hat{Y}$.
• Figure 3: Density plot for $Y'$ and $\check{Y}'$ in synthetic data. For a chosen data point, we sampled a batch of $U$ under the conditional distribution of it and plot the distribution of $Y'$ and $\check{Y}'$.
• Figure 4: Causal model for the Law School Dataset.
• Figure 5: Density plot for $F'$ and $\check{F}'$ in law school data. For a chosen data point, we sampled $K$ from the conditional distribution of $K$ and plot the distribution of $F'$ and $\check{F}'$.

Theorems & Definitions (21)

• Example 3.1: Law school success
• Example 3.2
• Definition 3.1: Individual response
• Definition 4.1
• Example 4.1
• Theorem 4.1: Violation of LCF under CF predictors
• Definition 5.1: Counterfactual random variable
• Theorem 5.1: Predictor with perfect LCF
• Definition 5.2: Relaxed LCF
• Corollary 5.1: Relaxed LCF with predictor in equation \ref{['decision']}