2 Graph embeddings
This chapter covers
- Exploring graph embeddings and their importance
- Creating node embeddings using non-GNN and GNN methods
- Comparing node embeddings on a semi-supervised problem
- Taking a deeper dive into embedding methods
Graph embeddings are essential tools in graph-based machine learning. They transform the intricate structure of graphs—be it the entire graph, individual nodes, or edges—into a more manageable, lower-dimensional space. We do this to compress a complex dataset into a form that’s easier to work with, without losing its inherent patterns and relationships, the information to which we’ll apply a graph neural network (GNN) or other machine learning method.
Graphs, as we’ve learned, encapsulate relationships and interactions within networks, whether they’re social networks, biological networks, or any system where entities are interconnected. Embeddings capture these real-life relationships in a compact form, facilitating tasks such as visualization, clustering, or predictive modeling.
There are numerous strategies to derive these embeddings, each with its unique approach and application: from classical graph algorithms that use the network’s topology, to linear algebra techniques that decompose matrices representing the graph, and more advanced methods such as GNNs [1]. GNNs stand out because they can integrate the embedding process directly into the learning algorithm itself.