Message passing is a fundamental paradigm in current GNN and graph representation learning. However, it has several drawbacks. Theoretically, the power of message passing is bounded by Weisfeiler-Lehman (WL) isomorphism test, is limited to simple convolution filters, and is restricted to tree-structured computation graphs. Empirically, researchers find that message passing suffer from over-smoothing, which is a result of neighbor aggregation. The author believe that there is deeper connections between these problems, and a new paradigm is needed to overcome the theoretical bottle neck.

University of Notre Dame

There are many important sub-area in graph representation learning. However, the author believe that there’re two key areas that have the potential of pushing fundamentals of graph representation learning forward:

1)	Latent graph inference

2)	Breaking the bottleneck of message passing


Key Area For Future Graph Model Development

Hamilton, W.L. Graph Representation Learning. *Synthesis Lectures on Artificial Intelligence and Machine Learning*. 2020, 14(3), 1–159. https://doi.org/10.2200/S01045ED1V01Y202009AIM046

Graph Representation Learning by William Hamilton

Latent graph inference is the task of infering graph or relational relations from unstructured or semi-structured data. It’s potentially very useful because it allows us to use GNN type of methods on data even without graph structure input. There has been initial work in this area, and current findings suggest that it can improve model performance even on data with graph relation input.

Learn Before

Related