Most machine learning practitioners spend their days with data that fits neatly into rows, sequences, or pixel grids. Text has an order. Images have a layout. Graphs? They refuse to behave. A graph is just a collection of nodes connected by edges — think users in a social network, atoms in a molecule, or papers linked by citations. But unlike a sentence, shuffling a graph’s nodes doesn’t create a new graph. It’s the same structure wearing a different hat. That property, called permutation invariance, quietly upends a lot of standard ML intuition.
The tasks you can run on graphs split cleanly by scope. At the graph level, you might generate new molecules for drug discovery or predict how a network evolves over time. At the node level, AlphaFold’s protein structure predictions hinge on estimating 3D coordinates for individual atoms within a molecular graph. Edge-level work covers everything from predicting drug side effects to recommending connections in social platforms. Google Maps even uses subgraph property prediction to estimate arrival times. The authors of this Hugging Face course introduction emphasize a practical split: transductive learning operates on one fixed graph, while inductive learning trains across multiple separate graphs. The transductive route is trickier than it looks — carving train, validation, and test splits from a single graph is genuinely subtle.
Representation matters too. You can store a graph as an edge list or an adjacency matrix, but both are potentially misleading. Most real-world graphs are sparse, so adjacency matrices waste enormous amounts of memory. More importantly, treating a graph’s representation as if it were ordered — like a sentence or image — is a category error. The course walks through pre-neural methods before arriving at Graph Neural Networks and eventually graph Transformers, the natural endpoint for anyone already comfortable with attention mechanisms.
This primer is the first in what looks to be a broader curriculum on graph machine learning. If you’ve only worked with CNNs on images or transformers on text, graphs will force you to rethink some assumptions you didn’t know you had. The payoff is access to a whole class of problems — molecules, social structures, knowledge graphs — that simply don’t fit into rectangular tensors without serious contortions.
