Introduction to Graph Neural Networks (GNN)

Graph Neural Networks on Undirected Numerical Graphs

Undirected Graph

In this section we will discuss following methods

• Global Pooling
• Dummy Node
• Hierarchical Pooling

Method 1 : Global Pooling

Step 1 : Generating Mature Node Embeddings using Graph Neural Networks (GNN)

Hence this idea gave rise to Graph Neural Network (GNN) which is described below !!

a. Vector Implementation of Element — Wise Summation

b. Vector Implementation of Element — Wise Average

Step 2 : Aggregation — Combining Mature Node Embeddings to get Graph Embeddings

There are several ways that you can use to aggregate to get the entire graph numerical representation, few prominent onces have been shown below

Method 2 : Dummy / Virtual / Super Node

• Now we apply Graph Neural Network on the new graph which contains super node.
• After GNN computations we can say that matured embedding of the super node is the entire graph embedding since super node contains information from all other nodes !!

Method 3 : Hierarchical Pooling

This method is preferred over global pooling method because it is compute efficient due to the drop layer added between GNN layers