To address the limitations of reduced accuracy in graph diffusion methods when handling complex edge relation-ships,this paper proposed a curvature-based graph diffusion neural network.The method introduced Ollivier-Ricci curvature to quantify edge curvature,providing a geometric measure of graph structure.The algorithm adjusted the weights of the random transition matrix using curvature,modifying them based on geometric relationships.It then combined the processed curvature matrix with the graph diffusion matrix to update the weight coefficients for model training.Experimental results show that the improved method maintains the advantages of smoothing graph signals effectively and reducing high-frequency noise.It in-creased accuracy by 0.3 to 2.0 percentage points on datasets with varying numbers of edges and nodes.The method optimized message aggregation in graph diffusion,utilizing node information and edge weights within the graph structure more effectively.This enhancement improves model performance in node classification tasks and provides a reliable method and experimental basis for future graph-based research.
维普
万方数据