We can consider the failure cases of the WL algorithm to motivate provably more powerful graph neural networks. However, message passing approaches generally fail to identify closed triangles in a graph, which can be considered to be a critical limitation. This paper suggests that we use the relational pooling approach, which considers augmenting message passing graph neural networks with unique node ID features.

Relational pooling approach benefits

Even though the relational pooling approach can produce graph neural networks models that are more powerful than the WL algorithm, there are limitations to the relational pooling approach. The two key limitations are that the full algorithm is computationally intractable and there is no way to characterize how much powerful this approach can make graph neural networks in comparison to the WL algorithm.

Relational pooling approach limitations

It is important to research on how we can make graph neural networks more powerful than WL algorithm because there are some research that indicates that models such as message-passing graph neural networks are no more powerful than the WL algorithm.

University of Michigan - Ann Arbor

Graph neural networks capture the internal graph structures of natural language, like syntactic and semantic parse trees.

Graph Neural Networks

For most current applications, graph neural networks are used in one of three tasks: node classification, graph classification, or relation prediction.

Learn Before

Related

Learn After