Graph Neural Network Based on Spectral Parameter Estimation for Mobile Networks

For online learning, Graph Neural Networks(GNNs) often require real graph data for parameter fine-tuning, but obtaining the graph data is challenging. Currently, GNNs are trained based on existing graph data. When learning in dynamic networks, spatial-domain GNNs use the adjacency matrix of entire graphs as its input, which are extremely resource-intensive in terms of computation and storage while dealing with large-scale graph data. Moreover, spectral-domain GNNs suffer from high spectral decomposition complexity. To enhance the generalization of GNN models and achieve lightweight GNN models, this paper proposes a lightweight GNN model based on spectral domain, SDGAT(Spectral-Domain Graph Attention Network). By combining spectral domain graph convolution and attention mechanism, SDGAT can effectively capture the global structure and local features of a graph, which helps to suppress noise in the graph and improve the robustness of the model. Furthermore, a distributed estimation algorithm for dynamic network spectral parameters is proposed, which significantly reduces the computational overhead when processing large-scale graph data. Finally, authors validate the effectiveness and accuracy of SDGAT through classification and regression tasks, and explore the feasibility of extending this model to other GNNs.