[icml2022] using virtual nodes to promote graph structure learning

With the development of graph kernel and graph representation learning , Many better methods have been proposed to deal with the scalability and over smoothing problems in graph structure learning . However , Most of these strategies are designed based on practical experience rather than theoretical analysis . In this paper , We use a specific virtual node to connect to all existing vertices , Without affecting the attributes of the original vertices and edges . We further prove that this kind of virtual node can help to establish an effective single state edge vertex transformation and state inversion to restore the original graph . It also shows that adding virtual nodes can maintain the local and global structure , So as to obtain better graph representation learning . We use virtual nodes to extend graph kernel and graph neural network , The graph classification and subgraph isomorphic matching tasks are tested . Experimental results show that , The graph with virtual nodes as input significantly improves the graph structure learning , Similar results can be obtained using their edge to vertex graphs . We also discuss the expression ability obtained from false points in neural networks .

Thesis link ：

https://arxiv.org/abs/2206.08561