Title: Advancing Equity in Graph Prompting: A Dual-Prompt Strategy to Counteract Structural and Attribute Bias

Abstract: Leveraging self-supervised pre-training on unlabeled graph data has emerged as a standard approach for Graph Neural Networks (GNNs). Nevertheless, a disconnect frequently persists between these pre-training goals and the requirements of downstream applications. Graph prompting techniques aim to close this divide by adjusting frozen, pre-trained GNNs to specific tasks via learnable prompts. While effective, current methods predominantly prioritize performance metrics, often neglecting fairness issues. Since downstream graph datasets naturally embed biases within both node attributes and network structures, pre-trained GNNs may generate representations that vary significantly across different demographic groups. To resolve this issue, we introduce Adaptive Dual Prompting (ADPrompt), a framework designed for fairness-aware adaptation of pre-trained GNNs. ADPrompt integrates two synergistic elements: Adaptive Feature Rectification, which employs personalized attribute prompts to filter out sensitive information at the input stage, and Adaptive Message Calibration, which utilizes layer-specific structure prompts to dynamically control information flow from adjacent nodes. Through the joint optimization of these components, ADPrompt not only adapts the pre-trained GNN but also alleviates biases occurring at both the attribute and structural levels. Evaluations across four benchmark datasets using various pre-training strategies reveal that ADPrompt consistently surpasses seven strong baseline models in node classification tasks.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC