Graph Convolution for Large-Scale Graph Node Classification Task Based on Spatial and Frequency Domain Fusion

In recent years, Graph Neural Networks (GNNs) have achieved significant success in graph-based tasks. However, they still face challenges in complex scenarios, particularly in integrating local and global information, enhancing robustness to noise, and overcoming the rigidity of graph structures. To address these issues, we propose a new GNN algorithm, LEGNN (Local and Global Enhanced Graph Neural Network), which introduces several key improvements over traditional GNN models such as GraphSAGE and GCN:Firstly, LEGNN combines a hybrid convolution method of spatial and spectral convolutions, enabling it to simultaneously capture local neighborhood relationships and global topological structures. This fusion mechanism effectively enhances the model’s ability to learn node representations, especially in complex graph structures, with accuracy improvements of 1%-2% compared to GCN and GraphSAGE. Secondly, LEGNN incorporates a noise prediction mechanism that injects controlled perturbations into the node representations, improving the model’s robustness, reducing overfitting, and enhancing generalization to unseen data. Finally, LEGNN introduces an adaptive graph structure adjustment mechanism, allowing the model to dynamically adjust the graph topology based on input data, enabling it to flexibly handle evolving data scenarios. In experiments on OGB datasets, LEGNN reduces training time by 55%-88% compared to GCN and GraphSAGE, and demonstrates higher stability when handling large-scale graph data. We validated the effectiveness of LEGNN through experiments on two OGB datasets, and the results show that LEGNN outperforms traditional GNN models in graph learning tasks, highlighting its potential in advancing the field.