Introducing GraphSAGE

Hamilton et al. introduced GraphSAGE in 2017 (see item [1] of the Further reading section) as a framework for inductive representation learning on large graphs (with over 100,000 nodes). Its goal is to generate node embeddings for downstream tasks, such as node classification. In addition, it solves two issues with GCNs and GATs – scaling to large graphs and efficiently generalizing to unseen data. In this section, we will explain how to implement it by describing the two main components of GraphSAGE:

• Neighbor sampling
• Aggregation

Let’s take a look at them.

Neighbor sampling

So far, we haven’t discussed an essential concept in traditional neural networks – mini-batching. It consists of dividing our dataset into smaller fragments, called batches. They are used in gradient descent, the optimization algorithm that finds the best weights and biases during training. There are three types of gradient descent:

...