Improving graph neural network via complex-network-based anchor structure

The technique of graph/network embedding in artificial intelligence, which embeds graph data into a low-dimensional vector space in the form of machine learning, can help computer to efficiently process and analyze the complex graph data by using the vector operations. Graph Neural Network (GNN) and neural network, an end-to-end graph embedding technique based on Graph Signal Processing (GSP), which aggregates the topological information of the neighborhoods of each node in a graph, has attracted wide attention. However, most of the existing GNN models are limited to local structure information, and the location differences of nodes in the global topology are not sufficiently considered. This leads to that many nodes with the similar local topology are very difficult to distinguish. To address this problem, we propose an anchor-structure-aware GNN (AS-GNN) model to implement more accurate node distinguishment by capturing the global topology information based on the characteristics of complex networks. Anchor structure is defined as a key sub-graph composed of key nodes and edges in a graph. Taking it as a location reference, we can get the location information of each node in the global topology of graph and carry it into the embedding of nodes. By this way, the node vectors including richer topology information of graph are produced by GNN, and thus the nodes with similar local topologies can be distinguished well. To evaluate the performance of AS-GNN, we compare AS-GNN with some existing baseline models by the experiments of the classic GNN application tasks of link prediction and pairwise node classification on five real-world datasets. The experimental results have confirmed the above claims. (c) 2021 Elsevier B.V. All rights reserved.