Abstract
Graph Neural Networks (GNNs) are increasingly adapted via the "pre-train and prompt" paradigm. However, existing prompting methods often neglect demographic biases inherent in graph data. We introduce ADPrompt, a framework that achieves fairness through hierarchical debiasing. By implementing Adaptive Feature Rectification at the input and Adaptive Message Calibration during propagation, ADPrompt reduces bias without compromising model utility.
Introduction
Graph Neural Networks (GNNs) have revolutionized learning on networked data. The recent shift toward "pre-train and prompt" strategies allows for efficient adaptation of large-scale models to downstream tasks. Despite their success, GNNs are susceptible to two primary forms of bias: attribute bias, where node features correlate with sensitive attributes, and structural bias, where "echo chambers" in the graph topology amplify prejudices during message passing.
As shown in Figure 3, real-world datasets like the German Credit and Pokec networks exhibit significant disparities in label distributions and neighbor characteristics based on sensitive attributes (e.g., gender or region). Existing graph prompting methods prioritize performance, often at the cost of fairness. ADPrompt addresses this by integrating fairness directly into the prompting mechanism.
Methodology: The ADPrompt Framework
ADPrompt employs a "Dual Prompting" strategy to intervene at critical stages of the GNN’s information flow. Unlike previous "hard" interventions that delete edges and risk destroying useful data, ADPrompt uses "soft" interventions to re-weight information flow.
Adaptive Feature Rectification (AFR)
AFR targets attribute bias at the input layer. It utilizes a self-gating mechanism to identify and suppress feature dimensions that carry sensitive information. By projecting node features through an attribute projector, the model generates a node-specific prompt that purifies the data before it enters the GNN.
Adaptive Message Calibration (AMC)
To combat structural bias, AMC introduces learnable structure prompts at every GNN layer. These prompts act as dynamic weights, adjusting the influence of neighboring nodes during message passing. This prevents the amplification of biased signals while preserving the graph's fundamental topology.
Adversarial Optimization and Theoretical Guarantees
The framework is trained using an adversarial approach where a "fairness police" layer attempts to predict sensitive attributes from node representations. The prompts are optimized to minimize this predictor's accuracy. Mathematically, the authors prove that this dual-prompting mechanism tightens the upper bound of Generalized Statistical Parity, ensuring a reduced treatment gap between demographic groups.
Experimental Evaluation
The researchers evaluated ADPrompt on four datasets: Credit Defaulter, German Credit, Pokec-z, and Pokec-n. The framework was compared against seven baselines, including state-of-the-art fairness-aware methods like FPrompt, across various pre-training strategies such as GraphCL and GAE.
Key Results
- Superior Fairness: ADPrompt consistently achieved the lowest scores in fairness metrics like Statistical Parity ($Δ SP$) and Equal Opportunity ($Δ EO$). On the German Credit dataset, it reduced these bias metrics by approximately 2% compared to standard methods.
- No "Fairness Tax": ADPrompt bucked the trend of accuracy-fairness trade-offs, actually improving accuracy by 3% on the German Credit dataset while simultaneously increasing fairness.
- Robustness: The method proved effective regardless of the underlying GNN architecture, performing well with GCN, GraphSAGE, and GAT backbones.
Discussion and Conclusion
ADPrompt demonstrates that hierarchical intervention—addressing both attributes and structure—is essential for effective debiasing in GNNs. By using parameter-efficient "soft" interventions, the framework preserves graph integrity while muting discriminatory signals.
While ADPrompt offers significant advancements, future research should address scalability for billion-node graphs and extend the framework to heterogeneous graphs with complex relationship types. Overall, ADPrompt provides a robust, mathematically grounded approach to building fairer Graph AI.