Graph node classification with soft-flow convolution and linear-complexity attention mechanism

Abstract

Traditional Graph Neural Networks (GNNs) typically use a message-passing mechanism to aggregate information from neighboring nodes. This message-passing mechanism is analogous to diffusing messages, often resulting in the homogenization of node features. GNNs also tend to be ineffective at capturing features from distant nodes and learning the global structure of the graph, which can reduce performance in node classification tasks. To address these issues, this paper proposes a novel model—Enhanced Soft-Flow Graph Convolutional Network (ESAGCN) based on a global attention mechanism. This model defines a learnable, parameterized phase angle that allows the edge directions between nodes to change continuously, enabling features to flow between nodes. Additionally, it incorporates the self-attention mechanism from Transformers to capture global information within the graph network, enhancing the global representation of nodes. We also employ a simple kernel trick to reduce the complexity of the model’s global attention mechanism to linear complexity. Experimental results demonstrate that the integration of global and local information in graphs is crucial for the learning process of GNNs, especially in directed graphs, significantly improving the accuracy of node classification.