GCN

1. Node Classification

最佳入门： https://mlabonne.github.io/blog/intrognn/

• We can enrich our node representation by aggregating its features with those of its neighbors. This operation is called convolution, or neighborhood aggregation.
• If we added a second GCN layer, our model would not only aggregate feature vectors from the neighbors of each node, but also from the neighbors of these neighbors.
• but there's a catch: if we add too many layers, the aggregation becomes so intense that all the embeddings end up looking the same. This phenomenon is called over-smoothing and can be a real problem when you have too many layers.

2. Link Prediction

GraphSAGE ()

• Note that the weights used in the aggregation step are not specific to a node but only specific to iteration k. Therefore, they are shared by all the neighborhoods making generalization to unseen nodes possible.
• To sum up, you can consider GraphSAGE as a GCN with subsampled neighbors.

Heterogeneous Graphs