FairGuide: Graph Structural Fairness

• FairGuide is a graph-structural fairness framework that improves GNN fairness by strategically adding a few new links without removing any original edges.
• It leverages a pseudo downstream task using differentiable community detection and meta-gradient based link scoring with cross-group boosting to optimize fairness.
• Empirical results on datasets like Github and Pokec demonstrate significant reductions in fairness disparities while preserving predictive utility for downstream tasks.

Searching arXiv for the specified paper and nearby fairness-on-graphs work. FairGuide is a graph-structural fairness framework for graph neural networks (GNNs) that guides a biased user graph toward a more unbiased one by adding a limited number of new links rather than rewiring or removing existing edges. It is introduced in “Let’s Grow an Unbiased Community: Guiding the Fairness of Graphs via New Links” (Lu et al., 21 Aug 2025). The central premise is that real user graphs often contain structural barriers—tightly knit subgroups that mostly connect internally—and that message passing in GNNs amplifies these barriers into disparate downstream outcomes. FairGuide addresses this by optimizing the placement of new links so that the resulting graph fosters more unbiased communities, with the stated goal of improving fairness not only for a single task but across diverse downstream applications trained on the fairness-guided graph (Lu et al., 21 Aug 2025).

1. Problem setting and structural view of fairness

FairGuide is motivated by the observation that fairness problems in GNNs are not only a property of node features or model training objectives, but are also embedded in the graph structure itself. Real user graphs, including social-network-like graphs, may exhibit tightly connected internal subgroups and weak inter-group connectivity. The paper characterizes these as structural barriers that translate into inequitable access to resources and influence, and it argues that GNN message passing can amplify such bias into unfair downstream predictions, including recommendation bias against female authors or developers (Lu et al., 21 Aug 2025).

The framework therefore adopts a specifically additive intervention: rather than modifying the graph arbitrarily, it asks whether one can guide an existing graph toward an unbiased structure by introducing a bounded number of carefully selected new links. In the paper’s formulation, these added links are intended to “grow” the graph into more integrated communities, thereby reducing the structural conditions under which sensitive information propagates into unfair predictions.

A defining feature of FairGuide is its focus on structural fairness. Instead of equating fairness purely with a downstream classifier metric, it treats the bias of community assignments inferred from the graph as a proxy for structural bias. This choice is tied to the observation that many downstream labels correlate with latent communities; the paper therefore frames community debiasing as a mechanism for fairness generalization across tasks whose labels are not known in advance. This suggests a shift from task-specific fairness correction to graph-level fairness guidance.

2. Formal setup and fairness criteria

FairGuide operates on a user graph

$G=(V,E,X),$

where $V=\{v_1,\ldots,v_N\}$ is the node set, $E\subseteq V\times V$ the edge set, and $X\in\mathbb{R}^{N\times M}$ the node-attribute matrix, with $x_i\in\mathbb{R}^M$ the feature of node $v_i$. The adjacency matrix is

$$A\in\mathbb{R}^{N\times N},\qquad A_{ij}=1 \text{ if } (i,j)\in E, \text{ else } 0.$$

The sensitive attribute is binary, $s\in\{0,1\}$, and $S$ denotes the vector of sensitive attributes over nodes. The downstream tasks considered are node classification, with ground-truth labels $Y$ and predictions $V=\{v_1,\ldots,v_N\}$0, and unsupervised community detection (Lu et al., 21 Aug 2025).

The fairness metrics used for downstream node classification are Statistical Parity and Equal Opportunity. For binary classification,

$V=\{v_1,\ldots,v_N\}$1

and for multi-class predictions $V=\{v_1,\ldots,v_N\}$2,

$V=\{v_1,\ldots,v_N\}$3

Equal Opportunity is defined as

$V=\{v_1,\ldots,v_N\}$4

FairGuide’s key additional notion is structural fairness, measured by applying statistical parity to predicted community assignments $V=\{v_1,\ldots,v_N\}$5 and sensitive attributes $V=\{v_1,\ldots,v_N\}$6:

$V=\{v_1,\ldots,v_N\}$7

Lower values indicate more structurally fair communities. This embeds fairness in the graph topology as reflected by its community structure, rather than only in the outputs of a specific downstream classifier.

The overarching graph-editing problem is first written as a bi-level fairness-guidance objective:

$V=\{v_1,\ldots,v_N\}$8

where $V=\{v_1,\ldots,v_N\}$9 is a fairness metric such as $E\subseteq V\times V$0 or $E\subseteq V\times V$1, and $E\subseteq V\times V$2 is the link budget. The additive-only constraint $E\subseteq V\times V$3 is central: FairGuide preserves all original edges and only adds new ones.

3. Pseudo downstream task: differentiable community detection

FairGuide does not assume access to the ultimate downstream task during graph editing. Instead, it introduces a differentiable community detection task as a pseudo downstream task, with the argument that downstream labels often correlate with latent communities. The optimization objective is correspondingly reformulated as

$E\subseteq V\times V$4

This pseudo-task replaces explicit downstream training with community detection on the modified graph and evaluates fairness through the bias of those community assignments (Lu et al., 21 Aug 2025).

The pseudo-task pipeline is deliberately decoupled into feature-based initialization and structure-based propagation. First, community initialization is obtained from node attributes alone:

$E\subseteq V\times V$5

The number of communities $E\subseteq V\times V$6 is a hyperparameter. The paper describes the MLP stage as an MLP-based self-supervised autoencoder followed by K-means clustering.

Second, structure-aware label propagation with restart is applied over the normalized adjacency matrix $E\subseteq V\times V$7:

$E\subseteq V\times V$8

where $E\subseteq V\times V$9 is the number of propagation steps and $X\in\mathbb{R}^{N\times M}$0 is the restart probability. The paper emphasizes that this decoupling avoids differentiating through MLP training, because the initialization is feature-only. That design is reported to stabilize and accelerate meta-gradient computation.

Conceptually, the pseudo-task serves two roles. It supplies a differentiable surrogate objective through which graph edits can be optimized, and it anchors the fairness objective in a structural quantity—community bias—that is argued to transfer across downstream tasks. A plausible implication is that FairGuide is best suited to settings where community organization is a strong latent factor in the predictive structure of the data.

4. Meta-gradient link addition and algorithmic procedure

FairGuide selects new links through a meta-gradient computed with respect to the adjacency matrix. The pseudo loss is

$X\in\mathbb{R}^{N\times M}$1

and the corresponding meta-gradient is

$X\in\mathbb{R}^{N\times M}$2

Because $X\in\mathbb{R}^{N\times M}$3, the gradient does not propagate through the MLP-based initializer.

Candidate non-edges $X\in\mathbb{R}^{N\times M}$4 with $X\in\mathbb{R}^{N\times M}$5 are then scored by an adjusted gradient:

$X\in\mathbb{R}^{N\times M}$6

where $X\in\mathbb{R}^{N\times M}$7 encourages cross-group edges. The negative sign converts small meta-gradients that reduce $X\in\mathbb{R}^{N\times M}$8 more strongly into high scores. This cross-group boost operationalizes the paper’s intuition that inter-group links help break structurally biased bubbles.

Discrete edge selection under the link budget is performed with Gumbel-max sampling. Noisy scores are defined by

$X\in\mathbb{R}^{N\times M}$9

with temperature $x_i\in\mathbb{R}^M$0 and numerical-stability constant $x_i\in\mathbb{R}^M$1. The top-$x_i\in\mathbb{R}^M$2 edges under $x_i\in\mathbb{R}^M$3 are added:

$x_i\in\mathbb{R}^M$4

This cycle repeats until the link budget is exhausted.

The algorithmic procedure is therefore iterative. It initializes $x_i\in\mathbb{R}^M$5, computes $x_i\in\mathbb{R}^M$6, repeatedly propagates communities, evaluates $x_i\in\mathbb{R}^M$7, computes the meta-gradient, applies cross-group boosting, samples top-$x_i\in\mathbb{R}^M$8 candidate edges, and updates the adjacency matrix. Original edges are always preserved, candidate edges are all non-edges, and no link removals are allowed (Lu et al., 21 Aug 2025).

For suggesting one link for a specific node $x_i\in\mathbb{R}^M$9, the paper gives the per-link time complexity

$v_i$0

where $v_i$1 is average degree and $v_i$2 the number of communities. Reported runtime per added link is $v_i$3 on Github, $v_i$4 on Pokec-n, and $v_i$5 on Pokec-z.

5. Correlation-based fairness generalization

A distinctive element of FairGuide is its theoretical claim that optimizing fairness on the pseudo-task can generalize to downstream tasks. The analysis uses the Pearson correlation coefficient

$v_i$6

Let $v_i$7 be community labels, $v_i$8 the sensitive attribute, and $v_i$9 a downstream prediction. The main theorem assumes that community labels and downstream predictions are highly correlated:

$$A\in\mathbb{R}^{N\times N},\qquad A_{ij}=1 \text{ if } (i,j)\in E, \text{ else } 0.$$0

If pseudo-task training drives $$A\in\mathbb{R}^{N\times N},\qquad A_{ij}=1 \text{ if } (i,j)\in E, \text{ else } 0.$$1 close to $$A\in\mathbb{R}^{N\times N},\qquad A_{ij}=1 \text{ if } (i,j)\in E, \text{ else } 0.$$2, namely

$$A\in\mathbb{R}^{N\times N},\qquad A_{ij}=1 \text{ if } (i,j)\in E, \text{ else } 0.$$3

for small $$A\in\mathbb{R}^{N\times N},\qquad A_{ij}=1 \text{ if } (i,j)\in E, \text{ else } 0.$$4, then

$$A\in\mathbb{R}^{N\times N},\qquad A_{ij}=1 \text{ if } (i,j)\in E, \text{ else } 0.$$5

The interpretation given in the paper is that when correlation between the sensitive attribute and pseudo labels is reduced to near zero, the correlation between the sensitive attribute and downstream predictions is correspondingly bounded, thereby ensuring fairness generalization (Lu et al., 21 Aug 2025).

The proof sketch relies on three ingredients: a lemma equating Pearson correlation with cosine similarity between z-scores, the spherical law of cosines for angle sums and differences, and a theorem stating that if $$A\in\mathbb{R}^{N\times N},\qquad A_{ij}=1 \text{ if } (i,j)\in E, \text{ else } 0.$$6 and $$A\in\mathbb{R}^{N\times N},\qquad A_{ij}=1 \text{ if } (i,j)\in E, \text{ else } 0.$$7, then

$$A\in\mathbb{R}^{N\times N},\qquad A_{ij}=1 \text{ if } (i,j)\in E, \text{ else } 0.$$8

Applying this with $$A\in\mathbb{R}^{N\times N},\qquad A_{ij}=1 \text{ if } (i,j)\in E, \text{ else } 0.$$9 yields the stated bounds.

This analysis is conditional, not universal. The paper explicitly assumes that community labels correlate with downstream labels, and it later identifies weak community–label correlation as a limitation. The theoretical argument therefore supports FairGuide most directly in applications where community structure is a salient intermediate variable for prediction.

6. Empirical evaluation

FairGuide is evaluated on three graph datasets—Github, Pokec-n, and Pokec-z—using node classification and community detection fairness, with downstream GNN backbones including 2-layer GCN, 2-layer GraphSage, and 10-layer APPNP. Hidden dimension is 128, the learning rate is $s\in\{0,1\}$0, training runs for 1000 epochs, and results are reported over seeds $s\in\{0,1\}$1 as mean $s\in\{0,1\}$2 std (Lu et al., 21 Aug 2025).

Dataset	Scale	Sensitive attribute / label
Github	32,132 developers, 270,088 follower edges	Country grouped into developed vs developing / popularity $s\in\{0,1\}$3 followers$s\in\{0,1\}$4
Pokec-z	67,797 nodes, 882,765 edges	Region / job field
Pokec-n	66,569 nodes, 729,129 edges	Region / job field

The experimental baselines are Rand. Add, Link Pred., EDITS, Graphair, and Fairgen. Utility is measured by F1 and AUC for node classification, and fairness by $s\in\{0,1\}$5 and $s\in\{0,1\}$6 for node classification together with $s\in\{0,1\}$7 for community detection. Link budgets are matched across methods: 3% of original edges on Pokec-n, 1.5% on Pokec-z, and 4% on Github, with 100 edges added per iteration.

On node classification with GCN, FairGuide consistently reduces fairness gaps relative to the vanilla graph and also outperforms the best fairness baseline on the reported metrics. On Github, the vanilla graph yields F1 $s\in\{0,1\}$8, AUC $s\in\{0,1\}$9, $S$0 $S$1, and $S$2 $S$3, while FairGuide yields F1 $S$4, AUC $S$5, $S$6 $S$7, and $S$8 $S$9. On Pokec-n, FairGuide reduces $Y$0 from $Y$1 to $Y$2 and $Y$3 from $Y$4 to $Y$5, while slightly improving F1 from $Y$6 to $Y$7. On Pokec-z, it reduces $Y$8 from $Y$9 to $V=\{v_1,\ldots,v_N\}$00 and $V=\{v_1,\ldots,v_N\}$01 from $V=\{v_1,\ldots,v_N\}$02 to $V=\{v_1,\ldots,v_N\}$03, with small changes in F1 and AUC.

The community-detection results are consistent with the structural-fairness objective. Using Louvain communities, Github $V=\{v_1,\ldots,v_N\}$04 decreases from $V=\{v_1,\ldots,v_N\}$05 to $V=\{v_1,\ldots,v_N\}$06, Pokec-n from $V=\{v_1,\ldots,v_N\}$07 to $V=\{v_1,\ldots,v_N\}$08, and Pokec-z from $V=\{v_1,\ldots,v_N\}$09 to $V=\{v_1,\ldots,v_N\}$10. Preliminary analysis on Github additionally shows that random or predicted link addition already reduces community bias relative to the original graph, but FairGuide is designed to optimize this effect systematically through fairness-guided gradients.

The paper also reports generalization across architectures: when GraphSage and APPNP are trained on FairGuide-updated graphs, they exhibit consistent fairness gains, with reduced $V=\{v_1,\ldots,v_N\}$11 and only small changes in F1. Utility–fairness trade-off curves under varying link addition rates from 0.5% to 3% indicate that, for the same fairness level, FairGuide often attains higher F1, and for the same F1, lower $V=\{v_1,\ldots,v_N\}$12 than baselines.

Ablation studies further support the framework design. Replacing community detection with random labels (“FairGuide\C”) substantially worsens fairness, indicating that pseudo-task alignment matters. Removing dynamic sampling (“FairGuide\S”) also degrades fairness, supporting the iterative Gumbel-max selection mechanism. Hyperparameter sensitivity on Pokec-n with GCN shows that $V=\{v_1,\ldots,v_N\}$13 improves markedly when the number of communities $V=\{v_1,\ldots,v_N\}$14 and the cross-group boost $V=\{v_1,\ldots,v_N\}$15, while F1 remains stable within $V=\{v_1,\ldots,v_N\}$16–$V=\{v_1,\ldots,v_N\}$17 percentage points across tested values.

7. Relation to prior approaches, assumptions, and limitations

The paper positions FairGuide against both in-processing and pre-processing debiasing methods. In-processing GNN debiasing methods, including adversarial training and regularization, operate on fixed graphs; FairGuide instead changes the graph structure itself through link addition. Pre-processing debiasing methods often rewire or remove edges without strict constraints, whereas FairGuide preserves all original edges and imposes a limited additive budget, which the paper describes as more practical in real communities (Lu et al., 21 Aug 2025).

The method also differs from conventional link prediction. Traditional link prediction tends to favor homophilous edges, which may reinforce bubbles. FairGuide instead scores non-edges through structural fairness gradients and cross-group boosting, biasing the procedure toward links that more directly reduce community-level disparity.

The framework comes with explicit assumptions and limitations. It assumes a binary sensitive attribute and leaves extension to multi-attribute or continuous sensitive attributes as future work. It assumes that community labels correlate with downstream labels; where that correlation is weak, fairness gains may be smaller. It also presumes that adding edges is feasible in the application domain, whereas privacy or policy constraints may make such interventions difficult. Finally, the method debiases the current graph structure; significant future distribution shifts may alter the fairness–utility landscape.

These constraints delimit the domain in which the method’s theoretical and empirical claims should be interpreted. Within those boundaries, FairGuide presents a structurally grounded fairness paradigm: fairness is not treated merely as a post hoc property of a trained GNN, but as a property of the graph communities from which many downstream predictions inherit their bias.