Improving Fairness in Graph Neural Networks via Mitigating Sensitive Attribute Leakage

• GNNs have achieved strong results in node representation learning, yet their predictions have also been shown to perpetuate discrimination.Their use of feature propagation helps capture higher-order neighborhood information, multi-hop dependencies, and local topology.Source: §1 · p. 1
• Historical discriminatory decisions and homophilous network connections can both contribute to sensitive-group separation in learned representations.The paper describes feature propagation as smoothing neighboring representations while separating distant ones.Source: §1 · p. 1
• Feature propagation can change feature-channel correlations, allowing initially innocuous channels to encode more sensitive information, a leakage route overlooked by prior approaches.The paper notes that many earlier methods borrow non-graph fair-representation techniques or debias features and topology without accounting for this correlation variation.Source: §1 · p. 1
• FairVGNN learns fair feature views by masking sensitive-correlated channels and adaptively clamping encoder weights; experiments report a better utility–fairness trade-off.The introduction presents these as its algorithmic contribution and summarizes its evaluation on real-world datasets.Source: §1 · p. 2
• The paper’s approach separates feature-view generation from adaptive weight clamping to address leakage from both the input and model perspectives.The view generator learns fair feature views; the clamping module targets sensitive-correlated encoder channels.Source: §1 · p. 2

2.1 Notations

• An attributed graph is represented as G = (V, E, X, A), with V denoting nodes, E edges, and X the node-feature matrix.The feature vector of node v_i is X_i, and X_:j denotes the jth feature channel.Source: §2.1 · p. 2

2.2 Fairness in Machine Learning

• The paper focuses on group fairness with a binary sensitive feature, while noting that extension to multiple sensitive groups is future work.It distinguishes group fairness—avoiding discriminatory outcomes for demographic groups—from individual fairness, which requires similar individuals to be treated similarly.Source: §2.2 · p. 2
• Fairness is evaluated using statistical parity and equal opportunity differences between sensitive groups, with lower Δsp and Δeo indicating fairer models.Δeo compares true-positive rates across groups; Δsp compares prediction independence from the sensitive feature.Source: §2.2 · p. 2

• Masking non-sensitive channels can reduce discrimination, and masking channels with higher sensitive correlation shows a decreasing trend in Δsp and Δeo.The investigation also reports that masking the sensitive channel itself does not eliminate discrimination.Source: §3 · p. 2
• Propagation can alter sensitive correlations across channels; on German, the 7th channel shifts from originally irrelevant to highly correlated with the sensitive feature.The section presents correlation changes on Credit as smaller than those on German.Source: §3 · p. 2, Figure 1 · p. 3
• On German, propagated-correlation masking (S2) achieves lower Δsp and Δeo than original-correlation masking (S1) for MLP and GCN; on Credit, S1 and S2 perform the same.The section attributes the Credit result to smaller correlation variations there.Source: Table 1 · p. 3, §3 · p. 3
• The S0, S1, and S2 strategies compare no masking with masking top-ranked channels by original or propagated sensitive correlation; masking improves fairness but can reduce utility.The section reports that S2’s fairness advantage over S1 on German reflects changes in correlation rank after propagation.Source: §3 · p. 3
• The correlation variation that determines which channels to mask depends heavily on the GNN propagation mechanism.Source: §3 · p. 3
• The stated objective is to generate a feature view that preserves task-related information while discarding sensitive information for downstream node classification.Source: §3 · p. 3

4.1 Generative Adversarial Debiasing

• The generator samples feature masks to create multiple views, which the GNN encoder processes into node representations for a classifier and sensitive discriminator.Adversarial learning updates the generator toward views that preserve task-relevant information while discarding bias-relevant information.Source: §4.1 · pp. 3–4

4 FRAMEWORK

• FairVGNN has two modules: adversarial debiasing learns fair feature views, and adaptive weight clamping targets sensitive-related encoder channels.The framework overview describes both modules.Source: §4 · p. 3, Figure 2 · p. 4

4.1.1 Fairness-aware View Generator.

• Because feature propagation can change correlations, the generator learns a data-driven feature mask instead of selecting features by their original sensitive correlations.The feature views are generated conditioned on the original features; masking probabilities are learnable and trained with gradient-based techniques.Source: §4.1.1 · p. 4

4.1.2 GNN-based Encoder.

• The encoder uses an L-layer GNN to represent node features and topology, while MLP-based classifier and discriminator predict labels and sensitive attributes.GCN, GraphSAGE, and GIN are listed as encoder backbones.Source: §4.1.2 · pp. 4–5

4.1.4 Adversarial Training.

• Adversarial training seeks fair views whose predictions do not reveal sensitive groups, while classification loss preserves model utility.The stated equivalence between the adversarial objective and the statistical-parity objective assumes α = 0.Source: §4.1.4 · p. 5

4.2 Adaptive Weight Clamping

• Adaptive clamping uses learned feature-keeping probabilities to weaken encoder weights for frequently masked channels; the analysis links clamping to an upper bound on group fairness.The stated bound depends on network homophily, initial feature difference, and masking probability.Source: §4.2 · pp. 5–6

4.3 Training Algorithm

• FairVGNN's additional mask-generation cost is described as preserving O(|V|+|E|) time complexity, with model complexity matching that of other L-layer GNN backbones.The method precomputes K masks per epoch; K is constant relative to the training steps described.Source: §4.3 · p. 6

5.1.1 Datasets.

• The evaluation uses three benchmark datasets, with their basic statistics reported in Table 2.Source: §5.1.1 · p. 6, Table 2 · p. 6

5.1.2 Baselines.

• The baselines include augmentation-based NIFTY and EDITS and adversarial-based FairGNN, each evaluated with GCN, GIN, and GraphSAGE backbones.Source: §5.1.2 · p. 7

5.1.3 Setup.

• Experiments tune hyperparameters individually, use the prescribed data split, and average results over five runs with different seeds.The setup fixes the number of generated fair feature views per epoch at K=10.Source: §5.1.3 · p. 7

5.2.1 Performance comparison.

• FairVGNN achieves the best average rank across all datasets and evaluation metrics, with a better utility–fairness trade-off than other bias-alleviating methods.The comparison is reported in Table 3.Source: §5.2.1 · p. 7, Table 3 · p. 7

5.2.2 Ablation study.

• Removing either fair-view generation or adaptive weight clamping worsens performance, and removing both produces still worse results.The two modules address discrimination from different aspects, with effects that can accumulate; without fair-view generation, replacing the mask-weighted threshold with a higher threshold permits more sensitive information leakage.Source: §5.2.2 · pp. 7–8, Table 4 · p. 8

5.3.1 Does adversarial training work?

• Removing the discriminator or generator raises model bias; retaining the discriminator without the generator yields lower bias than removing both in most cases.The authors attribute this to competition between the discriminator and generator, which supports sensitive-feature recognition and fair-view generation.Source: §5.3.1 · p. 8, Figure 3 · p. 8

5.3.2 Does adaptive weight clamping work?

• Except for GIN, adaptive weight clamping achieves lower bias than non-adaptive clamping and spectral normalization without hurting utility as much.Table 5 compares these weight-regularization approaches.Source: §5.3.2 · p. 8, Table 5 · p. 8

• FairGNN, NIFTY, and EDITS use distinct adversarial, triplet-based, and Wasserstein-based approaches, while REF-EREE provides structural explanations of GNN bias.Source: §6 · p. 8
• FairVGNN targets sensitive leakage into innocuous features as propagation varies correlations; bias from node degree, graph condensation, and class distribution is left for future work.Source: §6 · p. 9

7 CONCLUSION

• FairVGNN masks sensitive-correlated feature channels while accounting for correlation variation after propagation, then adaptively clamps encoder weights to absorb less sensitive information.The paper reports a better utility–fairness trade-off than other baselines.Source: §7 · p. 9
• Future work includes analyzing links among propagation, homophily, and correlation variation, using self-supervised learning to constrain representation bias, and considering multiple sensitive groups.The conclusion also notes that correlation variation depends on the dataset and group fairness is related to network homophily.Source: §7 · p. 9

A SUMMARY OF NOTATIONS

• A notation table summarizes commonly used symbols and their descriptions.Source: A SUMMARY OF NOTATIONS · p. 10, Table 7 · p. 10

B.1 Detailed Model Architecture

• The template combines neighborhood propagation and feature transformation; stacking L layers aggregates neighborhood information up to L hops, and FairVGNN’s encoder follows this template.This work considers one graph convolution for the described backbones.Source: §B.1 · p. 10

B.2 Hyperparameter for Each Baseline

• Baseline hyperparameters differ by method, with NIFTY, EDITS, and FairGNN each assigned their own configurations.Source: §B.2 · p. 10

C DATASET DETAILS

• German nodes are bank clients linked by similar credit accounts; the task predicts high or low credit risk, with gender as the sensitive feature.The node attributes include gender, loan amount, and other account-related details.Source: C DATASET DETAILS · p. 10
• Bail nodes are defendants released during 1990–2009, linked by similar criminal records and demographics; the task predicts violent or nonviolent crime, with race as sensitive.Source: C DATASET DETAILS · p. 10
• Credit nodes are credit-card users linked by similar purchase or payment patterns; the task predicts default, with age as the sensitive feature.Figure 5 shows correlation variation for German and Credit, with network homophily shown in parentheses.Source: C DATASET DETAILS · p. 10, Figure 5 · p. 11

D.1 Effect of Weight Clamping

• Adaptive weight clamping generally achieves a better utility–fairness trade-off than weight clamping and spectral normalization.It clamps weights more on sensitive-related features, minimally removing information beneficial for classification.Source: §D.1 · p. 11
• On Credit, GCN with adaptive weight clamping has higher bias than with direct weight clamping.Blind clamping can remove sensitive-overlapping information and reduce bias, but can also remove class-related information and lower utility.Source: §D.1 · p. 11
• On the Credit dataset, GCN accuracy with adaptive weight clamping was 77.79, below 78.04 with direct weight clamping, an exception to its generally better utility-fairness trade-off.Source: §D.1 · p. 11

D.2 Detailed Correlation Variation of German and Credit Datasets

• Credit’s feature correlations change more slowly than German’s after propagation; Credit has homophily 0.9595 versus German’s 0.8048.Source: §D.2 · p. 11
• On German, some feature channels quickly become more correlated with the sensitive channel while others become less correlated.Source: §D.2 · p. 11
• On Credit, original and propagated sensitive-correlation rankings yield the same S1 and S2 performance in Table 1.In the extreme case where node features strictly follow network homophily, propagation would not change feature distributions or correlations.Source: §D.2 · p. 11