Neural Message Passing

Initial node features are an important input for a graph neural network method, but sometimes the most important thing is the aggregation we perform during the training process. This aggregation is done during the message passing phase.

Message passing

This is the process of sending messages on graphs. Essentially, it's a way for GNNs to understand and make sense of information in a graph.

Let's simplify this. Imagine we have a bunch of friends and want to find out who is the most popular. We could go to each friend and ask how many friends they have and then use this information to identify the most popular one. This idea is similar to neural message passing. We're sending a message, i.e., asking about the number of friends, and receiving a message, i.e., the number of friends each person has.

With this process, we can understand the relationships between people and finally the answer to the question of who is most popular. This idea is illustrated in the image below in which the red node is found to be the most popular person using message passing: