Critical Structure-aware Graph Neural Networks

Graph neural networks (GNNs) have achieved significant success in many real-world applications by performing message-passing between nodes to embed graph data into low-dimensional and dense vector space. The capacity (depth and width) of the captured structure limits the expressive power of graph neural networks. However, existing GNN models mainly assume that the graph structure perfectly reflects the relationship between nodes and aggregate local neighbor information based on the original graph structure, ignoring the complex semantic information of critical structures. To address these challenges, we propose a series of innovations on the critical structures in graph data from three typical scales "connection, local structure, and higher-order structure", and propose a series of critical structure-aware GNNs for better representation quality and robustness.