Non-Local Graph Attention Networks Optimized with Neighborhood Feature Smoothing

The attention mechanism effectively addresses the issues of anisotropy and long-range dependencies in graph neural networks (GNNs). However, when node feature correlations are weak, the edge weights learned by the model often fail to accurately reflect the true relationships between node class labels, particularly for long-range dependencies between nonneighboring nodes. Although some existing methods attempt to solve this problem, they typically suffer from quadratic computational complexity, limiting their applicability in largescale graph data. To address this, we propose a non-local graph attention network (NLGAT) framework that adaptively adjusts the feature smoothing of neighboring nodes. This framework integrates a local-to-global aggregation strategy, smoothing the feature representations of nodes within their local neighborhoods through adaptive smoothing loss. This approach enhances the embedding capabilities of the local aggregator. With these highquality local embeddings, the global aggregator can more effectively capture information from remote, similar nodes. Notably, the proposed global aggregation strategy has linear complexity. Experimental results demonstrate that NLGAT outperforms current popular graph Transformer methods on several benchmark datasets, especially for large-scale graph datasets.