GIN: How to Design the Most Powerful Graph Neural Network
Graph Neural Networks are not limited to classifying nodes.
One of the most popular applications is graph classification. This is a common task when dealing with molecules: they are represented as graphs and features about each atom (node) can be used to predict the behavior of the entire molecule.
However, GNNs only learn node embeddings. How to combine them in order to produce an entire graph embedding? In this article, we will:
-
See a new type of layer, called “global pooling”, to combine node embeddings;
-
Introduce a new architecture called Graph Isomorphism Network (GIN), designed by Xu et al. in 2018.
We’ll detail the advantages of GIN in terms of discriminative power compared