acceptodds
Withdrawn from ICLR 2027

SI-Fair: Adaptive Stochastic Interpolants for Fair Graph Representation Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Fairness in graph neural networks is particularly challenging because sensitive information can be encoded not only in node attributes but also in graph topology. This creates a tension between individual fairness, which encourages similar predictions for structurally similar nodes, and counterfactual fairness, which requires predictions to remain invariant to changes in sensitive attributes. To address this issue, we propose SI-Fair, an adaptive stochastic transport framework that jointly models these two fairness objectives through continuous distribution transformation. SI-Fair first disentangles node representations into unbiased structural information, bias-associated components, sensitive attributes, and auxiliary features, thereby separating task-relevant structural semantics from sensitive information. It then employs stochastic interpolants to construct a continuous transport path from the original biased representation distribution toward a fairness-aware target distribution. A dynamic adjustment mechanism adaptively modulates the stochasticity of the transport process according to the relative optimization progress of the two fairness objectives, enabling a flexible balance between structural consistency and counterfactual invariance. We further establish a theoretical connection between the learned transport process and fairness preservation by deriving an upper bound on the Kullback–Leibler divergence to the target distribution and relating its components to individual and counterfactual fairness objectives. Experiments on multiple real-world graph datasets demonstrate that SI-Fair consistently improves complementary fairness measures while preserving competitive predictive performance. These results indicate that continuous stochastic transport provides an effective framework for reconciling multiple fairness requirements in graph representation learning.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Graph Neural Networks, Algorithmic Fairness, Individual Fairness, Counterfactual Fairness, Stochastic Interpolants

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_DpjMOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes SI-Fair, an adaptive stochastic interpolants framework designed to reconcile the inherent conflict between individual fairness and counterfactual fairness in graph representation learning. The authors design three modules: (1) a structure-attribute disentanglement module, which decomposes node representations into unbiased structural features SuS_u, bias-associated components SbS_b, sensitive attributes AA, and auxiliary features CC; (2) a stochastic interpolants transport module, which constructs a transport path from the biased distribution ρ0\rho_0 to the ideal fair distribution ρ1\rho_1 by means of a continuous-time stochastic process; and (3) a dynamic fairness adjustment module, which adaptively modulates the stochasticity of the transport process according to the optimization progress of the two fairness objectives. Theoretically, the authors derive an upper bound on the KL divergence, decompose it into independent terms associated with the velocity-field error (corresponding to individual fairness) and the score-function error (corresponding to counterfactual fairness), and claim that under optimality conditions the transport path is equivalent to the Schrödinger Bridge solution. Experiments are conducted on four real-world datasets and one synthetic dataset, demonstrating improvements in the fairness metrics.

Critical strengths
  1. The authors carry out a series of theoretical analyses, which to a certain extent offer a viable line of thought for fair graph representation learning.
  2. The experiments are relatively comprehensive. The authors report multiple fairness evaluation metrics on four real-world datasets as well as on a self-constructed dataset, and the accompanying hyperparameter analyses and ablation studies further corroborate the performance of the proposed model.
Critical weaknesses
  1. The disentanglement mechanism described in Section 3.1 does not appear to hold, thereby introducing a rather serious logical flaw into the theoretical contributions of the paper. The theoretical contributions of this work are built upon the premise stated in Section 3.1 that the "structure-attribute disentanglement holds perfectly." However, a fundamental problem arises here. Equation (2) imposes only geometric orthogonality in the vector space. Orthogonality is not equivalent to statistical independence, still less to causal independence: two orthogonal vectors may nevertheless exhibit highly nonlinear dependence. The paper lacks any conditional independence constraint or any actual optimization of a mutual information upper bound. Furthermore, Equation (1) asserts that I(Su;Y)≥η⋅I(X;Y)I(S_u;Y)\geq \eta\cdot I(X;Y) and I(Sb;A)≤δbI(S_b;A)\leq \delta_b, yet the paper does not specify which mutual information estimator is employed. This suggests that Equation (1) constitutes merely a theoretical assumption rather than an actual optimization objective. Orthogonal regularization alone, together with mutual information constraints that are never actually optimized, is insufficient to achieve semantic-level disentanglement in deep non-convex networks.
  2. The wording of the paper is overly strong. Statements such as "the proposed disentanglement rigorously satisfies both fairness criteria" appear throughout. Given that the disentanglement mechanism does not hold, the use of such phrasing is inappropriate.
  3. The performance annotations in the main experiments are problematic. The boldface annotations in Table 1 (denoting the best performance) contain several evident factual errors, which seriously undermine the credibility of the experimental section. For example, for the SAKL metric on Facebook100, GEAR reports 0.121 whereas SI-Fair reports 0.129. Since lower SAKL is preferable, SI-Fair is not the best result in this column; bolding SI-Fair here constitutes direct misrepresentation of the data.
  4. At the beginning of Appendix A, the text states: "This appendix contains additional details for the ICML 2026 submission," whereas the paper is under review at ICLR 2027. This inconsistency should be corrected.
Questions for authors
  1. Could the authors clarify whether the mutual information constraints in Equation (1) are actually optimized during training? If so, which mutual information estimator is employed?
  2. Could the authors provide quantitative validation of the disentanglement quality, in order to substantiate the effectiveness of the proposed disentanglement?
Reviewer_L59tOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper investigates a conflict between individual fairness and counterfactual fairness in graph neural networks (GNNs). To mitigate this conflict, the paper proposes an adaptive stochastic transport framework called SI-Fair. The core idea of SI-Fair is to jointly model these two fairness notions through continuous distribution transformation. Specifically, SI-Fair first disentangles node representations into different components and then leverages stochastic interpolants to transfer the biased node distribution into an ideal fair distribution. Furthermore, the paper offers a theoretical analysis to connect the learned transport process and fairness preservation. Experiments on multiple real-world datasets demonstrate that SI-Fair improves fairness while preserving the downstream performance.

Critical strengths
  1. The paper investigates the fairness of graph neural networks, which is an increasingly important topic and will have a real impact on society.

  2. The paper identifies the conflict between individual fairness and counterfactual fairness, which is overlooked by previous studies.

  3. The paper provides both empirical and theoretical evidence on the effectiveness of the proposed method. In the empirical evidence, the paper compares the proposed method with 7 baselines on individual and counterfactual fairness. In the theoretical evidence, the paper proves the upper bound on the KL divergence of the model with the ideal fair distribution, the relationship between individual and counterfactual fairness with the KL divergence bound, and the convergence of Lv\mathcal{L}_v and Ls\mathcal{L}_s.

Critical weaknesses
  1. There are some key points that need to be further clarified:

    a. The paper claims that individual fairness requires similar outcomes for structurally similar nodes. However, in the general setting of graph-based individual fairness, similar nodes are always determined according to the needs of the task and may vary across tasks. For example, task A calculates similar nodes according to structural information, while task B leverages node feature information. Thus, it is not rigorous to directly define individual fairness as similar outcomes for structurally similar nodes.

    b. Although some group fairness methods are effective in improving counterfactual fairness, the paper cannot treat the literature review of group fairness methods as that of counterfactual fairness. Specifically, the fourth paragraph of the Introduction presents a literature review of group fairness and neglects existing studies of counterfactual fairness.

  2. The paper provides limited empirical or theoretical analysis on why there exists a conflict between individual fairness and counterfactual fairness. According to the example on social recruitment shown in the third paragraph of the Introduction, it makes sense that there is a conflict between these two fairness notions. However, as an academic paper, it is necessary to provide some evidence for this conflict. The authors should provide the counterfactual fairness (individual fairness) performance changes after individual fairness (counterfactual fairness) training.

  3. The Pareto frontiers in Section 4.3 lack a comparison with the individual fairness method. The Pareto frontiers are conducted to verify the trade-off performance between individual fairness and counterfactual fairness. The current experimental results in Figure 1 miss individual fairness baselines, resulting in a failure to verify a better individual-counterfactual fairness balance of the proposed method.

  4. Some components of the proposed method need further clarification, which limits reproducibility. Specifically, the paper fails to clarify how to obtain the fair target samples x1x_{1} shown in line 3 of Algorithm 4, and how to train the disentanglement module and downstream predictor

Questions for authors

Refer to Critical Weaknesses for more details.

Reviewer_NNTrOpenReview ↗
Rating
3
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes SI-Fair, a transmission-based framework that uses stochastic interpolants to unify individual fairness and counterfactual fairness in graph neural networks. The authors reformulate fairness as an optimal transmission problem that continuously transmits a biased node representation distribution to an ideal fairness distribution. First, a two-branch network decouples node representations into unbiased structural features, bias components, sensitive attributes, and auxiliary features; then, stochastic interpolants are used to construct continuous random paths connecting the two distributions.

Critical strengths
  1. Applying random interpolation theory to graph fairness and unifying the two conflicting fairness criteria into a continuous transmission problem is more persuasive than direct adversarial training.

  2. The experimental setup was well-designed, and the experimental data were comprehensive.

3.The upper bound for KL is explicitly decomposed into independent loss terms corresponding to the two fairness objectives, which allows the theory to align with the algorithmic design.

Critical weaknesses
  1. Key assumptions have not been clearly explained, such as how the ideal fair distribution is constructed and how paired samples are generated.

2.Random interpolation (augmentation) has been attempted in previous work, but the discussion of related work has been somewhat insufficient.

  1. The author should provide an explanation and analysis of the runtime.
Questions for authors
  1. Where exactly does the ideal fair distribution come from? If a fair representation has already been sampled, what problem does the transmission process solve?

2.What assumptions does SND rely on?

3.Compared to directly applying random perturbations or random substitutions to sensitive attributes, what are the algorithmic advantages and differences of this method? How much does the computational overhead increase?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.