Source-linked AI summary
Subgraph Filtering for Fair Graph Neural Networks
Haohui Lu, jiyuan Tian, Fangyu Zhou, Shahadat Uddin
TL;DR
GNNs can encode sensitive-group information through topology and message passing even without sensitive node features, while many existing methods do not directly control local propagation pathways. SF-GNN filters bias-prone edges using sensitive homophily and structural amplifiers, and experiments across five benchmark datasets show improved fairness with competitive predictive performance. The paper therefore supports targeted structural control as a practical direction for fairness-aware graph learning, while identifying broader fairness settings for future work.
Problem
Existing fairness-aware GNN methods mainly constrain representations or prediction distributions globally rather than directly controlling local structural pathways through which biased information propagates.
Method
SF-GNN combines sensitive homophily with hub participation and triadic closure to identify bias-prone edges, then stochastically drops or downweights them during message passing.
Results
Across five benchmark datasets, SF-GNN consistently reduces statistical parity disparity while maintaining competitive predictive performance against standard GNNs, graph transformers, and fairness-aware baselines.
Takeaways & Limitations
Targeted regulation of bias propagation paths provides an effective and practical direction for fairness-aware graph learning.
Takeaways & Limitations
Future work must extend SF-GNN to equalized odds, multi-class and multi-attribute settings, and adaptive filtering for heterophily and temporal graphs.
Abstract
from arXiv · showhide
Graph neural networks (GNNs) can exhibit unfair behavior even when sensitive attributes are excluded from node features, because graph topology and message passing propagate group-correlated signals under sensitive homophily. Existing fairness-aware GNN methods mainly constrain representations or prediction distributions at a global level, without explicitly controlling the local structural pathways through which biased information propagates during aggregation. We propose Subgraph Filtering for Fair Graph Neural Networks (SF-GNN), a lightweight and architecture-agnostic framework that mitigates structural bias at its source. SF-GNN identifies bias-prone edges by combining sensitive homophily with structural propagation amplifiers, including hub participation and triadic closure. It then incorporates stochastic edge filtering into each message-passing step to selectively downweight or remove these edges while preserving the remaining graph structure. Training further incorporates a statistical-parity regularizer with a warm-up schedule to stabilize optimization. Experiments on five benchmark datasets show that SF-GNN achieves consistent fairness improvements while maintaining competitive predictive performance, leading to a better fairness--accuracy trade-off than recent fairness-aware GNN baselines.
1 Introduction
GNN topology and message passing can propagate sensitive-group information into predictions, while existing fairness methods often leave these local pathways insufficiently controlled. SF-GNN filters bias-prone edges during aggregation and improves fairness–accuracy trade-offs across five benchmark datasets.
- Sensitive homophily lets message passing amplify group-correlated signals even when sensitive attributes are excluded from node features.
- Existing fairness-aware GNN methods mainly constrain representations or prediction distributions globally, leaving local structural propagation pathways insufficiently regulated.
- SF-GNN identifies bias-prone edges by combining sensitive homophily with structural amplifiers such as hub participation and triadic closure.
- SF-GNN applies stochastic filtering during message passing to selectively drop or downweight bias-prone edges while preserving remaining graph structure.
- Across five benchmark datasets, SF-GNN consistently reduces statistical parity disparities while maintaining competitive predictive performance.
- The framework is evaluated against standard GNNs, graph transformers, and fairness-aware baselines, with ablations and sensitivity analyses supporting targeted structural filtering.
2 Related Work
Fairness-aware GNN research addresses sensitive information that can propagate through graph structure, but existing approaches vary in whether they modify representations, propagation, graph structure, or architectures. Many methods may also add computational overhead and complexity.
- Sensitive attributes can indirectly influence node predictions through graph structure and message passing, even when sensitive features are absent from node inputs.
- Standard GNN backbones and graph transformers provide strong representation or long-range dependency modeling but may amplify structural bias under sensitive homophily.
- FairGNN, NIFTY, GraphAir, and BIND address unfairness through attribute estimation, counterfactual stability, graph augmentation, or distillation-based regularization.
- Structural and aggregation-aware methods include sensitive-information neutralization, fairness-aware propagation, and fairness-aware structural encodings in graph transformers.
- Many existing methods rely on adversarial optimization, graph editing, adaptive reweighting, or architectural modifications that can introduce computational overhead and complexity.
3 Preliminary
The paper formulates fairness-aware node classification on attributed graphs, where message passing can reinforce topology-linked sensitive information. It evaluates accuracy and group fairness using statistical parity and equality of opportunity metrics.
- 3.1 Problem Statement: The task is node classification on an attributed graph with node features, binary labels, and binary sensitive attributes.
- 3.1 Problem Statement: The learning goal is accurate node-label prediction while reducing disparities between sensitive groups.
- 3.2 Fairness Concepts in GNNs: Message passing aggregates neighboring information, repeatedly reinforcing correlations between sensitive attributes and graph connectivity across layers.
- 3.2 Fairness Concepts in GNNs: A general message-passing layer combines neighborhood messages through an aggregation operator and an update function.
- 3.2 Fairness Concepts in GNNs: Sensitive homophily produces neighborhoods dominated by the same group, allowing learned representations to encode sensitive-group information without sensitive input features.
- 3.2 Fairness Concepts in GNNs: Structural bias propagation describes disparities arising from biased topology interacting with iterative aggregation rather than from feature-level bias alone.
- 3.3 Evaluation Metrics: Evaluation uses accuracy, statistical parity difference, and equality of opportunity, seeking lower ∆SP and ∆EO without substantially sacrificing ACC.
- 3.3 Evaluation Metrics: Lower ∆SP indicates fairer positive prediction rates across sensitive groups, with ∆SP = 0 representing perfect statistical parity.
4 Method
SF-GNN identifies sensitive-homophilic edges that structural features make more likely to amplify group-correlated information, then filters those edges during message passing while retaining the rest of the graph. Training combines predictive loss with statistical parity, and the method is designed for efficient integration into GNN backbones.
- Bias-Prone Subgraph Identification: SF-GNN targets homophilic edges that are structurally positioned to amplify sensitive-group information, rather than treating all homophilic edges equally.The mask combines sensitive homophily with hub participation or triadic closure, and can be computed once before training.
- Bias-Prone Subgraph Identification: A GNN’s multilayer propagation makes node influence depend on the number and strength of paths between nodes.The normalized adjacency includes self-loops, while X denotes node features and W denotes learnable transformations.
- Subgraph Filtering Message Passing: SF-GNN introduces edge-dependent filtering weights that selectively drop or downweight bias-prone edges during aggregation.For bias-prone edges, the weight is stochastically set to zero or to an attenuation factor γ with 0 < γ < 1.
- Subgraph Filtering Message Passing: The filtering mechanism attenuates propagation pathways while preserving the remaining graph structure.Residual connections and normalization layers support expressive capacity and training stability.
- Fairness-Aware Training Objective: Training balances predictive accuracy and group fairness using cross-entropy and a statistical-parity regularizer with a warm-up schedule.The fairness coefficient increases linearly from zero to its target value, and sensitive attributes are used during training but not required during inference.
- Computational Complexity: SF-GNN performs weighted edge aggregation in O(md) per layer, with overall training complexity O(Lmd + Tprep) and memory complexity O(nd + m).Bias-prone edge detection is preprocessed once before training.
5 Experiments
SF-GNN is evaluated on five benchmark node-classification datasets against standard GNNs, graph transformers, and fairness-aware baselines. It consistently improves fairness while preserving competitive predictive performance, with ablations and sensitivity analyses supporting the design.
- Experimental setup: Experiments use five fairness-aware graph-learning benchmarks with node classification, designated sensitive attributes, and 50%/25%/25% train-validation-test splits.The datasets are NBA, German, Credit, Income, and Pokec_n; experiments use a two-layer GCN backbone and MLP classifier.
- Baselines: SF-GNN is compared with standard GNNs, GraphTrans, and fairness-aware methods using common splits, implementations, hyperparameter settings, and evaluation metrics.The evaluation covers GCN, GAT, GraphTrans, FairGNN, NIFTY, BIND, and GraphAir.
- Overall results: SF-GNN reduces fairness disparities while maintaining competitive or improved classification accuracy across the five datasets.The reported fairness metrics include statistical parity disparity and equality-of-opportunity disparity.
- Dataset results: On NBA, SF-GNN raises ACC from 71.65 for FAIRGT to 72.15 while reducing ∆SP from 11.16 to 2.29; on Pokec_n, it achieves ACC 91.31 and ∆SP 1.11.These are the strongest reported cross-dataset examples, with SF-GNN attaining the highest ACC on both datasets.
- Dataset results: On Income, SF-GNN reaches ACC 74.58 with ∆SP 3.38, between GraphTrans at ACC 77.45 and ∆SP 22.49 and FairGNN at ACC 45.56 and ∆SP 2.65.The comparison illustrates the reported fairness–accuracy trade-off across methods.
- Ablation and sensitivity analysis: Ablations indicate that structural filtering and fairness warm-up contribute to the fairness–utility trade-off, while moderate hyperparameter settings are generally stable.Overly aggressive filtering can remove useful structural information, and stronger fairness regularization can trade small accuracy reductions for fairness gains.
6 Conclusion
The paper frames fairness in GNNs as a structural problem and proposes SF-GNN to filter bias-prone propagation routes. Across five benchmarks, the framework reports favorable fairness–accuracy trade-offs, while future work extends its scope beyond the evaluated settings.
- Conclusion: SF-GNN identifies bias-prone edges by combining sensitive homophily with structural propagation amplifiers and attenuates them through stochastic edge filtering during aggregation.The framework uses edge dropping and downweighting while preserving the remaining graph structure.
- Conclusion: The framework is lightweight and architecture-agnostic, uses one-time preprocessing, and applies a warm-up statistical-parity regularizer for stable training.These design choices are presented as part of SF-GNN’s practical implementation.
- Conclusion: Across five benchmark datasets, SF-GNN consistently reduces statistical parity disparity while maintaining competitive predictive performance against multiple baseline categories.The conclusion attributes the gains to targeted structural filtering and the joint design of edge identification and training dynamics.
- Future work: Future work includes equalized odds, multi-class and multi-attribute settings, and adaptive filtering for heterophilous and temporal graphs.These extensions define the stated scope boundaries of the current work.