Graph Negative Feedback Bias Correction Framework for Adaptive Heterophily Modeling

Researchers propose Graph Negative Feedback Bias Correction (GNFBC), a framework that addresses limitations in Graph Neural Networks when processing heterophilic graphs where connected nodes have different characteristics. The method uses negative feedback mechanisms to correct bias from homophily assumptions and can be integrated into existing GNN architectures with minimal computational overhead.

• →Conventional Graph Neural Networks suffer from homophily bias, performing poorly on heterophilic graphs where connected nodes differ significantly.
• →GNFBC introduces a negative feedback loss mechanism that penalizes prediction sensitivity to label autocorrelation.
• →The framework incorporates graph-agnostic model outputs as feedback terms to counteract correlation-induced bias.
• →GNFBC can be seamlessly integrated into existing GNN architectures without significant computational overhead.
• →The approach leverages Dirichlet energy to guide the use of independent node feature information.