Graph neural networks
We introduced in Chapter 3, Neural Networks and Graphs, the concept of GNNs, a deep learning method that works on graph-structured data. It is now time to go into the details of how GNNs can be used for unsupervised learning. Similar to other techniques seen in this chapter (e.g., matrix factorization), GNNs will also allow us to obtain embeddings of our graph elements (nodes, edges, and graphs). However, it is worth noting that different from shallow embeddings, some of the GNN techniques are inductive, and therefore allow the embeddings to also be applied to unseen data.
As shown in the previous chapter, Convolutional Neural Networks (CNNs) are widely used in images to extract multi-scale localized spatial features that are exploited by deeper layers to construct more complex and highly expressive representations.
In recent years, it has been observed that concepts such as multi-layer and locality are also useful for processing graph-structured data...
