Graph Counterfactual Fairness Overview

• Graph counterfactual fairness is a causal framework that guarantees prediction invariance by applying hypothetical interventions on sensitive attributes in graph-structured data.
• It integrates methods such as neural causal modeling, counterfactual data augmentation, and adversarial debiasing to audit and enforce fairness in graph neural networks.
• Recent advances demonstrate significant bias reductions (up to 60–80%) with minimal impact on predictive accuracy, underscoring practical benefits in real-world applications.

Graph counterfactual fairness is a rigorous causal notion aiming to ensure that predictions on graph-structured data are invariant under hypothetical interventions on sensitive attributes, while accounting for the propagation of bias through complex dependencies in the structural graph. Operationalizing this notion requires explicit modeling of the underlying causal mechanisms—with uncertainty in graph structure, message-passing dynamics, and constraints arising from real-world feasibility or absence of sensitive attribute annotations. Recent advances integrate neural causal modeling, counterfactual data augmentation, adversarial debiasing, edge-based interventions, causal discovery, and information-theoretic regularization to both audit and enforce counterfactual fairness across diverse graph learning scenarios.

1. Formal Definition and Causal Characterizations

Let $G=(V, E, X, S, Y)$ denote a graph with node set $V$, edge set $E$, features $X$, sensitive attribute(s) $S$ (at node or graph level), and targets $Y$. The encoding of counterfactual fairness on graphs relies on adapting Pearl's three-step abduction–action–prediction protocol to structural causal models (SCMs) predicated upon the observed $G$ or a family of plausible SCMs:

• Individual Counterfactual Fairness: For node $i$, the prediction $\hat{Y}_i$ (from a GNN or downstream model) is counterfactually fair if, for all relevant values $s, s'$ of the sensitive attribute, the distribution of predicated outcomes is invariant under $V$0 and $V$1, conditional on all observed non-descendants:

$V$2

This definition extends to graph-wide settings where interventions on $V$3 may have cascading effects through $V$4, $V$5, and propagation effects in message-passing (Kher et al., 18 Feb 2025, Guo et al., 2023, Ma et al., 2022).

• Graph Counterfactuals: For node-level representations or predictions, this requires generating, for each $V$6, counterfactual graphs $V$7 via intervention(s) on their sensitive attributes and/or those of their neighbors, then enforcing invariance of model outputs across such counterfactual and factual graphs (Ma et al., 2022).
• Group Counterfactuals: In settings such as recommender systems or recourse, counterfactual fairness is assessed by constructing group-level counterfactual scenarios (e.g., via edge addition/deletion or feature perturbation), and auditing the disparity of outcomes or recourse cost across protected groups (Fragkathoulas et al., 2024, Medda et al., 2023, Boratto et al., 2023).

2. Causal and Algorithmic Frameworks for Graph Counterfactual Generation

Enforcing graph counterfactual fairness requires explicit modeling of how sensitive attributes affect graph features, structure, and predictive outcomes.

• Neural Causal Models (NCMs): For node-level causal models with known graph structure $V$8, structural equations of the form $V$9 are parameterized with neural nets. Dual mapping architectures $E$0 (forward generative) and $E$1 (posterior abductor) are jointly trained to model both factual and counterfactual generation pathways. Explicit kernel least-squares (conditional MMD) losses are introduced to enforce level-3 (marginal law) consistency for counterfactual sampling, ensuring $E$2 in RKHS (Kher et al., 18 Feb 2025).
• Counterfactual Data Augmentation: Approaches such as GEAR generate counterfactual subgraphs by perturbing central and/or neighbor node sensitive attributes in synthetic local contexts, then force the learned node representations to be invariant under these interventions using a Siamese-style or adversarial loss (Ma et al., 2022).
• Counterfactual Selection from Embedding Space: The CAF and Fairwos frameworks avoid unrealistic synthetic counterfactuals by searching for real nodes or subgraphs that serve as close proxies with differing sensitive embeddings. CAF splits node embeddings into content and environment, and selects nearest counterfactuals with matching non-sensitive components, enforcing consistency and orthogonality. Fairwos further supports training without access to true sensitive attributes using pseudo-sensitive embeddings (Guo et al., 2023, Wang et al., 2024).
• Recourse and Feasibility Constraints: FGCE formalizes feasible group counterfactual explanations, constructing subgroups of data instances reachable by feasible operations under linear and monotonicity constraints, and optimizing for cost-coverage trade-offs in counterfactual recourse (Fragkathoulas et al., 2024).

3. Counterfactual Fairness in Graph Neural Networks and Generative Graph Models

Graph neural networks (GNNs) instantiate specific challenges due to the entanglement of node features, neighborhood structure, and bias propagation.

• Fair-ICD: Constructs counterfactual neighborhoods by rewiring same-sensitive attribute edges to most-similar nodes of opposite attribute, performs unbiased aggregation via MLP-predicted neighborhood means, and applies adversarial training to scrub sensitive information from learned embeddings. The fairness-accuracy tradeoff is optimized via a minimax objective (Wo et al., 20 Aug 2025).
• Graph Diffusion Models (FairGDiff): Defines a simple SCM, $E$3, where $E$4 represents homophilic (sensitive-induced) link formation, and introduces counterfactual interventions at the “treatment” level. Forward and backward diffusion processes are conditioned on both factual and matched counterfactual treatments, with losses that jointly encourage independence of the generated graph from the sensitive attribute, maintaining graph utility (Wang et al., 2 Mar 2026).
• Metrics and Auditing: Standard and counterfactual fairness metrics used include statistical parity (Δ_DP), equal opportunity (Δ_EO), topology-bias ratios, group-utility disparities (e.g., ΔNDCG in recommender systems), as well as confidence bounds and switch-rate metrics when causal graph structure is uncertain (Valério et al., 6 Jan 2026, Medda et al., 2023, Boratto et al., 2023).

4. Uncertainty and Partial Knowledge of Causal Graph Structure

Real-world applications rarely admit full knowledge of the underlying causal graph; hence, robustness to graph uncertainty is critical.

• CF-GU (Counterfactual Fairness with Graph Uncertainty): Rather than relying on a single DAG, CF-GU bootstraps causal discovery (with domain knowledge constraints), enumerates all plausible DAGs, quantifies per-edge and subgraph-level entropy (normalized Shannon entropy $E$5), and computes confidence bounds on counterfactual fairness metrics by evaluating given classifiers under this “bag” of SCMs (Valério et al., 6 Jan 2026).
• Partial Graph Knowledge (PDAG/MPDAG): For partially directed acyclic graphs, efficient algorithms exist to determine features that are definitely non-descendants of sensitive nodes in all Markov-equivalent DAGs by critical set analysis of b-possibly causal paths, thus ensuring predictors built from these features are counterfactually fair even under graph uncertainty (Zuo et al., 2022).

Approach/Framework	Causal Model Usage	Counterfactual Sample Construction	Fairness Enforcement/Regularization
NCM (Neural Causal Model) (Kher et al., 18 Feb 2025)	Known DAG, neural parameterization	Learned forward + abduction networks, MMD L3 loss	MMD loss on output distributions
CAF (Guo et al., 2023)	SCM disentangling S/C/E; causal node embeddings	Latent nearest neighbor search for realistic CFs	Invariance and orthogonality penalties
Fair-ICD (Wo et al., 20 Aug 2025)	Structural equation for S→(X,A)→Z→Y	Counterfactual neighbor replacement and data augmentation	Adversarial training plus MLP for unbiased aggregation
FGCE (Fragkathoulas et al., 2024)	Edge-based feasibility constraints	Feasible graph edits, group-level optimization	Trade-off metrics: cost, coverage, kAUC/dAUC/cAUC
CF-GU (Valério et al., 6 Jan 2026)	Causal discovery under constraint, bootstrapped SCMs	Multiple SCMs from equivalence class	Confidence bounds on PSR/NSR, entropy analysis
GEAR (Ma et al., 2022)	Local SCMs, message-passing aug.	GraphVAE-based counterfactual subgraph gen.	Siamese/invariance loss on node embeddings

5. Consumer and Group-Level Counterfactual Fairness in Recommendation

In graph-based recommendation, counterfactual fairness is instantiated by identifying manipulations (edge deletions/additions) that close utility gaps between demographic groups.

• GNNUERS: Explains user-side unfairness by optimizing edge-deletion masks in the bipartite user-item graph so as to minimize group disparity (e.g., in NDCG) while preserving general utility, yielding explanations as sets of links whose removal would balance utility across groups. Explains observed disparities by linking them to graph-structural statistics such as degree, density, and intra-group distance (Medda et al., 2023).
• Counterfactual Graph Augmentation: Minimally augments the user-item graph with new edges exclusively for disadvantaged groups, optimizing a differentiable objective that combines reduction in group-utility disparity and a soft constraint on the number of edge additions. The resulting set of edges constitutes a counterfactual graph where the system would deliver fairer recommendations, and reveals which users/items were structurally disadvantaged (Boratto et al., 2023).

6. Practical Implementation, Empirical Results, and Impact

Empirical studies across a variety of data domains (credit, bail, NBA and Pokec social networks, recommender datasets) consistently show:

• Joint optimization over factual and counterfactual consistency (with proper regularization) yields fair–accurate Pareto frontiers superior to statistical or adversarial-only proxies (Kher et al., 18 Feb 2025, Ma et al., 2022).
• Realistic counterfactuals selected from the observed graph (as in CAF, Fairwos) produce more reliable fairness–utility trade-offs than synthetic perturbations or feature masking (Guo et al., 2023, Wang et al., 2024).
• Systems such as Fair-ICD and FairGDiff achieve substantial reductions (often 60–80%) in bias metrics with negligible or positive impact on predictive accuracy on large graphs (Wo et al., 20 Aug 2025, Wang et al., 2 Mar 2026).
• Auditing with graph uncertainty (CF-GU) exposes both the fragility and robustness of conclusions under alternative causal assumptions, and enables practitioners to quantify the reliability of fairness assessments via entropy-based metrics and confidence bounds (Valério et al., 6 Jan 2026, Zuo et al., 2022).
• The burden of recourse and coverage attributes can be directly quantified (FGCE), revealing disparities not just in average prediction but in the real-world feasibility and cost of achieving fair outcomes (Fragkathoulas et al., 2024).

7. Limitations and Ongoing Directions

• Many frameworks inherently assume binary or categorical sensitive attributes; generalization to continuous or multi-valued cases remains a nontrivial extension (Wo et al., 20 Aug 2025).
• The quality of counterfactual generation (realism, feasibility) is highly contingent on the accuracy and identifiability of the underlying SCM; uncertainties in graph discovery and structural equations propagate to fairness claims (Valério et al., 6 Jan 2026, Zuo et al., 2022).
• Scalability remains a bottleneck in full counterfactual enumeration on massive graphs; recent approaches use efficient nearest-neighbor search, pseudo-sensitive proxies, or local augmentation (Wang et al., 2024, Guo et al., 2023).
• Evaluating fairness in composite or dynamic graph settings (e.g., streaming, evolving graphs, multi-relational) raises new challenges for both modeling and inference, under active study.

Graph counterfactual fairness unifies causal modeling, counterfactual computation, and algorithmic fairness enforcement on graphs. The integration of faithful counterfactual sample generation, explicit causal regularization, realistic audited metrics, and principled handling of structural uncertainty establishes this area as central to trustworthy and equitable graph-based machine learning.