- 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 , bias-associated components , sensitive attributes , and auxiliary features ; (2) a stochastic interpolants transport module, which constructs a transport path from the biased distribution to the ideal fair distribution 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
- 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.
- 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
- 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 and , 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.
- 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.
- 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.
- 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
- 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?
- Could the authors provide quantitative validation of the disentanglement quality, in order to substantiate the effectiveness of the proposed disentanglement?