Research map / Graph learning / GNN architectures

FairSIN: Achieving Fairness in Graph Neural Networks through Sensitive Information Neutralization

AAAI 2024 · Graph learning › GNN architectures · approach: Robustness, privacy, and fairness of GNNs · 25 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes FairSIN, a neutralization-based fair GNN method that adds features of heterogeneous neighbors (estimated by an MLP) before message passing to cancel sensitive bias, in data-centric and model-centric variants.

Problem

GNN predictions can be biased by sensitive attributes; existing fair GNN methods filter out sensitive information via edge dropping or feature masking, which also removes useful non-sensitive information and hurts the accuracy-fairness trade-off.

Approach

Instead of filtering, inject Fairness-facilitating Features (F3)—the average features of each node's heterogeneous neighbors, estimated by an MLP from the node's own features—before message passing, theoretically shown to neutralize sensitive information leakage; implemented as edge reweighting (FairSIN-G), feature neutralization (FairSIN-F), and a model-centric per-layer version with adversarial

Key results

• On GCN, FairSIN reduces DP by 63.29% and EO by 33.82% relative to the best baseline while keeping accuracy high (e.g., Pokec-n DP 0.57 vs 1.16 for next best).
• FairSIN-F, a data-centric variant, is only slightly worse than the full model and outperforms prior SOTA methods on fairness.
• FairSIN has the lowest training time among compared methods on Bail and Credit; EDITS runs out of memory on Pokec datasets.

Measured effects reported in the paper

Surprising findings

• Adding a discriminator alone (without neutralization) often reduces predictive accuracy, so the neutralization features are the key component.
• Message passing amplifies sensitive bias even in neutralized features, but neutralized features leak much less.

Limitations

• Heterogeneous-neighbor estimation is infeasible or uncertain for nodes with few or no heterogeneous neighbors, motivating an MLP estimator with its own approximation error.
• Theory assumes binary sensitive attributes and a linear intensity-function predictor with Gaussian assumptions.
• Pokec datasets: EDITS runs out of memory, so comparisons are incomplete there.

Experimental setup

Datasets: Bail Pokec-n Pokec-z German Credit

Benchmarks: Bail Pokec-n Pokec-z

Metrics: ACC F1 DP EO

Baselines: FairGNN EDITS NIFTY FairVGNN vanilla GNN

Builds on: FairGNN FairVGNN NIFTY EDITS

Closest papers by meaning

• Chasing Fairness in Graphs: A GNN Architecture Perspective (AAAI 2024 · similarity 0.93)
• Fair Graph U-Net: A Fair Graph Learning Framework Integrating Group and Individual Awareness (AAAI 2025 · similarity 0.90)
• FairGP: A Scalable and Fair Graph Transformer Using Graph Partitioning (AAAI 2025 · similarity 0.88)
• Adversarial Attacks on Fairness of Graph Neural Networks (ICLR 2024 · similarity 0.88)
• FairNet: Dynamic Fairness Correction without Performance Loss via Contrastive Conditional LoRA (NeurIPS 2025 · similarity 0.88)
• Disentangling, Amplifying, and Debiasing: Learning Disentangled Representations for Fair Graph Neural Networks (AAAI 2025 · similarity 0.88)