Counterfactual Fairness in ML

• Counterfactual fairness is a criterion defined through structural causal models that ensures an individual’s prediction would remain unchanged if their sensitive attribute were altered in a counterfactual scenario.
• It leverages interventions on sensitive attributes within a causal framework, requiring analysis of latent variables, mediation, and causal pathways to isolate unfair effects.
• Algorithmic approaches include SCM-based inference, adversarial techniques, and plug-in methods, each offering trade-offs between rigorous fairness constraints and predictive utility.

Counterfactual fairness is a causal-individual fairness criterion demanding that a decision or prediction for an individual would remain invariant had the individual's protected attribute (e.g., race, gender) been different in a hypothetical “counterfactual world.” Formulated rigorously within the structural causal model (SCM) framework of Pearl, counterfactual fairness has shaped both theoretical and algorithmic developments in fairness-aware machine learning. The notion is distinctively individual-level and rooted in interventions over the sensitive attribute, requiring complex reasoning about latent (exogenous) variables, mediation, and causal pathways.

1. Formal Definition and Causal Foundations

Counterfactual fairness is defined in the context of SCMs, typically comprising a set of exogenous (unobserved) variables $U$, endogenous (observed) variables $V$ (including features $X$, outcomes $Y$, and sensitive attributes $A$), and a collection of structural equations $F$. Given this, a predictor $\hat Y$ satisfies counterfactual fairness if, for any individual with observed $(A=a, X=x)$ and for all alternative values $a'$ of the sensitive attribute, the distribution over predictions under interventions is invariant: $$P\bigl(\hat Y_{A\leftarrow a}(U)=y \mid A=a, X=x\bigr) = P\bigl(\hat Y_{A\leftarrow a'}(U)=y \mid A=a, X=x\bigr)\;\;\forall y, a, a'$$ $\hat Y_{A\leftarrow a'}(U)$ denotes the counterfactual prediction under an intervention $A \leftarrow a'$, holding exogenous variables $U$ fixed—often interpreted via the three-step process: abduction (inferring $U$ from observed data), action (modifying $A$ in $F$), and prediction (computing counterfactual outcomes) (Kusner et al., 2017).

A sufficient (but not necessary) condition for counterfactual fairness is to train $\hat Y$ only on non-descendants of $A$ in the causal DAG, since descendants may encode information causally downstream of the sensitive attribute (Zuo et al., 2023). Path-specific counterfactual fairness and individual-level regularizers have also been proposed to further refine the effect of allowed versus forbidden causal pathways.

2. Relationship to Group Fairness and Observational Parity

Counterfactual fairness is fundamentally an individual-level property, contrasting with group fairness definitions such as demographic parity (DP), equalized odds (EO), and calibration. Recent work has rigorously established the conditions under which counterfactual fairness coincides with group fairness metrics; e.g., when the causal DAG blocks all paths from $A$ to $Y$ except those allowed for DP, EO, or calibration, then enforcing the corresponding group metric ensures counterfactual fairness (Anthis et al., 2023, Rosenblatt et al., 2022). Notably, for a large class of SCMs with $U \perp A$, every counterfactually fair predictor is demographically fair and vice versa (Rosenblatt et al., 2022). However, in general, group fairness may not guarantee individual-level counterfactual invariance, especially when there is direct or indirect effect of $A$ that is not blocked by group-level constraints.

3. Algorithmic Approaches for Learning Counterfactually Fair Models

SCM-based and Latent Variable Methods

The canonical learning procedures for counterfactual fairness require either explicit functional forms for structural equations (as in the original FairLearning pipeline) or suitable latent variable models (e.g., VAEs approximating latent $U$). Practical algorithms proceed by (i) inferring the posterior over exogenous variables $U$ given observed $(A, X)$, (ii) generating counterfactual instances via intervention on $A$, and (iii) ensuring $\hat Y$ is invariant (in distribution) across counterfactual worlds (Kusner et al., 2017, Zuo et al., 2023). In high-dimensional or partially unknown causal settings, posterior inference is commonly amortized via auto-encoding frameworks.

Adversarial and Data Augmentation Techniques

Representation learning frameworks penalize statistical dependence between learned representations and $A$ using MMD, adversarial discriminators, or invariance penalties (e.g., CLAIRE, INVFAIR). Counterfactual data augmentation is widely used: models are trained to minimize the discrepancy between predictions on factual and counterfactual instances generated by learned VAEs or GANs (Ma et al., 2023, Grari et al., 2020, Ma et al., 2023). These frameworks dominate in regimes where the true causal model is noisy, uncertain, or partially misspecified.

Plug-in, Preprocessing, and Plug-and-Play Methods

Algorithmic simplifications have led to plug-in methods such as the Fair Learning through dAta Preprocessing (FLAP) algorithm, which preprocesses covariates to remove $A$-dependence, allowing any supervised learner to be used downstream (Chen et al., 2022). Similarly, plugin counterfactual fairness (PCF) and double machine learning (DML Fairness) operate without full SCM specification by combining predictions on factual and counterfactual representations, with error bounds characterized by the quality of counterfactual estimation (Zhou et al., 2024, Rehill, 2023).

4. Extensions and Practical Considerations

Graph and Structured Data

Extending counterfactual fairness to relational settings, frameworks such as GEAR define graph counterfactual fairness, wherein predictions for each node must remain invariant under arbitrary counterfactual assignments to both the node's and its neighbors' sensitive attributes (Ma et al., 2022). VAE-based data augmentation and distance penalties on node embeddings are used to enforce invariance across a combinatorially large space of counterfactual graphs.

Path-Specific and Lookahead Counterfactual Fairness

Recent research addresses the need for path-dependent counterfactual fairness (PCF), where only certain causal pathways from $A$ to $Y$ are deemed unfair and blocked in counterfactual generation (Zuo et al., 2023, Zuo et al., 2024). Lookahead counterfactual fairness (LCF) further advances the paradigm by incorporating downstream adaptation: predictions are required to enforce future counterfactual invariance over status variables $Y'$ that evolve in response to the model's outputs and individual strategic behavior (Zuo et al., 2024).

5. Fairness-Utility Tradeoffs and Theoretical Guarantees

The imposition of counterfactual fairness invariably incurs a trade-off in predictive utility, formalized in terms of excess risk. For regression, the excess risk is proportional to $$\text{Var}(A) \cdot \mathbb{E}_{U}[(\mathbb{E}[Y|U,A=0]-\mathbb{E}[Y|U,A=1])^2]$$, while for classification it is given by the conditional mutual information $I(A;Y|U)$ (Zhou et al., 2024). Algorithmic frameworks such as PCF are constructed to provide Bayes-optimal (risk-minimizing) counterfactually fair predictions, typically by averaging over factual and counterfactual outputs weighted by population priors.

6. Tensions, Limitations, and Current Debates

A series of studies have critically examined both the conceptual underpinnings and practical limitations of counterfactual fairness:

• Requirement of a Well-specified Causal Model: Most algorithms assume correctness of the causal DAG and often specific exogeneity or independence assumptions (e.g., $S$ is exogenous). Misspecification can break fairness guarantees or degrade performance, motivating frameworks like CLAIRE and INVFAIR that require only relaxed invariance assumptions (Ma et al., 2023, Duong et al., 2023).
• Unobservability of Individual Counterfactuals: By definition, counterfactual worlds are unobservable; validation relies on the plausibility and adequacy of the SCM.
• Fairness-Group Parity Coincidence: Under plausible independence, group-level fairness (e.g., demographic parity) and counterfactual fairness coincide (Anthis et al., 2023, Rosenblatt et al., 2022), questioning the necessity of complex causal machinery in certain regimes.
• Interpretability and Stakeholder Transparency: Removal of direct effects via counterfactual fairness can perturb within-group orderings, raising concerns regarding the semantic meaning and interpretability of "fair" decisions (Rosenblatt et al., 2022). Some advocate for order-preserving algorithms to complement counterfactual invariance with transparency in individual rankings.
• Ethical and Social Validity: Applying interventions to ill-defined social attributes (e.g., "what if this individual were of a different race?") poses ethical, conceptual, and identification challenges. Additionally, the strict removal of sensitive effects may disproportionately penalize high-performing individuals in marginalized groups (Rehill, 2023).

7. Empirical Evaluations and Open Directions

Empirical work consistently demonstrates that enforcing counterfactual fairness reduces measures such as Wasserstein distance and MMD between factual and counterfactual prediction distributions, often at modest cost to utility in regression or classification tasks on real and synthetic datasets (Kim et al., 15 Apr 2025, Zuo et al., 2023). Recent advances—such as EXOC (auxiliary variable causal reasoning), FairPFN (transformers trained on synthetic SCM data), and GAN/VAE-based augmentation—offer robustness against misspecified causal graphs or limited background knowledge (Tian et al., 2024, Robertson et al., 2024, Ma et al., 2023).

Open challenges include tractable enforcement in high-dimensional or structured domains; designing interventions that capture dynamic, sequential, or strategic feedback; reconciling group and individual fairness under data limitations; and arriving at fairness constraints that align with both legal doctrines and social expectations. Practical implementation benefits from frameworks that flexibly interpolate between strict SCM identification, regularized invariance, and black-box plug-in strategies depending on domain, risk profile, and epistemic assumptions.