In graph-based prediction settings, standard message passing in Graph Neural Networks often propagates correlations between neighborhoods and sensitive attributes, which results in biased node representations and unfair classification outcomes. In-processing mechanisms that modulate messages using protected-attribute relationships can enable fairness-aware representation learning by attenuating bias amplification during aggregation, as instantiated in this study through fairness-aware message corrections.

Graph Neural Networks (GNNs) are increasingly used when individuals, items, or entities are naturally connected. Typical examples include social graphs, interaction graphs, and other relational data used in user modeling. In these settings, the core strength of GNNs, i.e., learning from neighbors, also creates a structural channel through which sensitive attributes can influence predictions, even indirectly.

A difficulty is that fairness issues in graphs are not only a property of features, but also of topology. When nodes with similar sensitive attributes are more likely to connect, neighborhood aggregation can “echo” group membership and reinforce group-conditional patterns. As a consequence, interventions that merely remove sensitive attributes from node features may be insufficient, because group information can still be inferred from connectivity and local structure.

Hence, improving fairness should not come from severing relational signals altogether, but from controlling which relational signals are allowed to dominate representation learning.

In a study, in cooperation with Erasmo Purificato, Hannan Javed Mahadik, and Ernesto William De Luca, and published in the Proceedings of ACM UMAP ’25, we introduce FAME (Fairness-Aware MEssages), an in-processing bias mitigation approach that intervenes directly in the GNN message-passing mechanism.

The gap addressed is that, despite extensive work on fairness in machine learning, comparatively fewer approaches target how bias is propagated during message passing itself, and even fewer do so in a way that is explicitly tailored to common GNN layer families (convolution-based and attention-based).

High-level solution

Our central idea is to make the messages exchanged between connected nodes explicitly fairness-aware. Rather than treating all neighbor contributions as equally legitimate carriers of information, we modify message passing so that the influence of a neighbor is adjusted according to how the two nodes relate with respect to a sensitive attribute.

Concretely, we introduce (i) FAME, designed for graph convolutional layers, and (ii) A-FAME, designed for attention-based layers. Both variants instantiate the same conceptual lever: they reshape neighborhood aggregation so that the embedding update is less likely to amplify group-linked correlations that originate from homophily in the graph.

Our approach

Sensitive-attribute differences as an edge-level fairness signal

The primary problem is that bias can be transmitted through who is connected to whom, not only through explicit features. To address this, we introduce an abstraction that is simple but targeted: for every edge, we consider whether the endpoints differ with respect to the sensitive attribute.

This edge-level signal matters because it provides the model with a controlled way to treat certain relational paths as potentially bias-amplifying. It operationalizes a specific fairness hypothesis: when connectivity is strongly aligned with sensitive groups, uncorrected aggregation tends to over-reinforce group-conditional information.

Message modulation at the aggregation point where bias accumulates

A second mechanism is where we intervene: directly inside aggregation, i.e., at the point where neighbor information is mixed into a node’s representation across layers. Conceptually, we reweight (or reshape) neighbor contributions using the sensitive-attribute difference signal, so that representation learning becomes less dominated by same-group propagation patterns when those patterns are likely to encode discriminatory correlations.

Thanks to this, fairness interventions outside message passing can only indirectly affect what embeddings encode. By intervening at aggregation time, we target the dynamics through which biases are repeatedly reinforced across layers.

Architecture-aware instantiation for convolution and attention

A third mechanism is tailoring the same principle to different GNN families. In convolution-based GNNs, the intervention naturally takes the form of modifying how neighbor messages contribute during aggregation. In attention-based GNNs, the intervention acts on the attention mechanism itself, influencing which neighbors receive more weight.

This architecture-aware design matters for practical applicability: it acknowledges that “message passing” is implemented through different computational abstractions across GNNs, and fairness control needs to be applied at the relevant decision point (aggregation weights in convolutions, attention weights in attentional layers).

Findings and insights

Across three benchmark datasets spanning financial and social-network settings, we observe a consistent qualitative pattern: modifying message passing with a fairness-aware correction reduces group disparities under both prediction-parity and opportunity-style notions of fairness, while keeping predictive quality competitive.

Several insights emerge from these results. First, the improvements suggest that a meaningful portion of unfairness in GNN-based classification is attributable to the accumulation of biased neighborhood evidence across layers, rather than to isolated feature effects. Second, the comparison with “unaware” variants (where sensitive attributes are removed from inputs) supports the view that topology can remain a strong proxy for protected attributes, so fairness requires more than feature suppression. Third, the relative strength of the approach against other GNN-specific baselines indicates that where the intervention is placed (inside aggregation rather than around it) can be a decisive factor in mitigating bias amplification.

Overall, the evidence supports a conceptual takeaway: fairness constraints can be operationalized as controls on relational influence, and message passing provides a natural locus for implementing such controls.

Conclusion

This work contributes a message-level in-processing perspective on fairness in GNNs: rather than treating neighborhood aggregation as a fixed inductive bias, we make it a fairness-aware mechanism that can be shaped to reduce bias propagation in graph-structured data.

A natural next direction is to generalize the correction principle beyond the specific layer families considered here, aiming for a formulation that transfers across architectures without redesign. Additional research should also address richer sensitive-attribute settings (multi-group and intersectional attributes), partial or missing sensitive-attribute information, and broader downstream tasks where exposure and feedback loops may interact with representation bias. Finally, expanding evaluation to more graph domains and fairness notions can clarify when message-level interventions are sufficient, and when they should be combined with complementary controls at data, objective, or decision levels.