Dynamic Graph Neural Networks with Neighborhood Constraints

Graph data are well suited to represent complex interactions and dependencies owing to its inherent relational structure. Graph neural networks (GNNs) can effectively explore the topological structure and node feature information of graph data to perform perceptual inference tasks. In GNNs, the state of each node is iteratively updated by aggregating information from its neighbors. Note that some noise may be introduced due to the information aggregation of irrelevant neighbor nodes, leading to poor performance with higher computational overhead. This paper proposes a dynamic graph generation method based on neighborhood constraints, which learns the degrees of the nodes and imposes the neighborhood constraints to limit the neighbor-hood space of the nodes, resulting in a sparse graph with dynamic neighborhoods. Based on the sparse graph, the dynamic graph neural network selectively aggregates the information of the most important neighbors to improve the performance of the model. Furthermore, to ensure neighborhood constraints satisfaction during the learning process, a loss function is designed to penalize the nodes whose degrees exceed a specific threshold. The evaluation results demonstrate that our proposed method achieves a higher inference accuracy with the same or higher sparsity compared to state-of-the-art methods.