Part 1

Graph Neural Networks (GNNs) have emerged as powerful tools for modeling and analyzing graph-structured data, finding applications across diverse fields like social network analysis, bioinformatics, and recommender systems. Despite their success, scaling GNNs to handle larger and more complex graphs remains a significant challenge. Graph pooling plays a crucial role in enhancing their performance and making them more applicable to a larger class of datasets and applications.

In this post I will give a gentle introduction to pooling in GNNs and an overview of the main families of existing graph pooling approaches. Then, I will present some possible ways to evaluate how good a pooling method is. As the focus in on a tutorial exposition, I will cover only a representative subset of all methods that exist in the literature. I also have tried to keep the notation and the nomenclature as uniform as possible, even if there is quite some variance in the literature.

In this part, I will give a brief introduction to GNNs, graph pooling, and to a general framework for expressing a generic graph pooling operator.

Let’s start!

❶ An introduction to pooling in GNNs

Before talking about pooling in GNNs let’s look at how pooling works in more traditional deep learning architectures, such as a Convolutional Neural Network (CNN).