GGAT: Gravitation-Based Graph Attention Networks

Graph-structured data is an important data form that is widely used in the real world. It can effectively and abstractly express entities in information and the relationships between entities. The appearance of Graph Neural Networks (GNNs) provides a potent tool for dealing with nonlinear data structures, which mainly learns node representation through information propagation and aggregation on the nodes in the graph. However, existing GNNs fail to adequately and efficiently integrate the topological structure of the network and node features during information propagation, resulting in an insufficient capture of the complex influence relationships between nodes. The limitation constrains the expression ability of the models and seriously impacts their performance in node classification tasks. To overcome this issue, we propose a Gravitation-based Graph Attention Network (GGAT) for node classification. Firstly, we define a novel similarity measurement method based on the formula of universal gravitation, which combines node information entropy and spatial distance. This method overcomes the limitation of existing similarity measurements that focus solely on the topological structure or node features, achieving a more comprehensive similarity assessment. Then, we apply it to the graph attention network as a novel attention mechanism. Compared with the traditional attention mechanisms based on learning, our proposed mechanism not only thoroughly considers the topological structure and node features to allocate the weights of neighbor nodes but also makes the calculation of attention weights more transparent with an intuitive physical significance, thereby improving the stability and interpretability of the model. Finally, the experiments are carried out on various real datasets, and the results show that GGAT is superior to the existing popular models in node classification performance.