Revisiting Virtual Nodes in Graph Neural Networks for Link Prediction

It is well known that the graph classification performance of graph neural networks often improves by adding an artificial virtual node to the graphs, which is connected to all nodes in the graph. Intuitively, the virtual node provides a shortcut for message passing between nodes along the graph edges. Surprisingly, the impact of virtual nodes with other problems is still an open research question. In this paper, we adapt the concept of virtual nodes to the link prediction scenario, where we usually have much larger, often dense, and more heterogeneous graphs. In particular, we use multiple virtual nodes per graph and graph-based clustering to determine the connections to the graph nodes. We also investigate alternative clustering approaches (e.g., random or more advanced) and compare to the original model with a single virtual node. We conducted extensive experiments over different datasets of the Open Graph Benchmark (OGB) and analyze the results in detail. We show that our virtual node extensions yield rather stable performance increases and allow standard graph neural networks to compete with complex state-of-the-art models, as well as with the models leading the OGB leaderboards.