What is a graph neural network (GNN)?
Length:
3 min
Published:
June 9, 2026
What is a graph neural network?
A graph neural network (GNN) is a type of neural network that works with data shaped as a graph: a set of nodes (people, products, molecules) joined by edges (friendships, purchases, chemical bonds). Most neural networks expect data laid out as a grid or a sequence. A GNN instead reads the connections, so the prediction about one node depends on its neighbours.
The core idea is message passing. Each node gathers information from the nodes it is linked to, updates its own picture, and repeats. After a few rounds, every node carries a summary of its local neighbourhood. The network then uses those summaries to classify a node, predict a missing link, or score the whole graph.
In plain words
Think of how you judge a stranger by the company they keep. You learn a little from the person directly, and a lot from who they hang out with. A GNN does the same: it figures out what each item is by passing notes back and forth between connected items until the picture settles.
Where you see it
- Fraud detection. Spotting suspicious accounts by how they transact with others, not just their own behaviour.
- Recommendations. Suggesting products or content based on the web of who-bought-what.
- Drug discovery. Predicting how a molecule behaves from the graph of its atoms and bonds.
- Logistics and networks. Routing, traffic forecasting, and modelling power or telecom grids.
Common pitfalls
- Good data means a good graph. If the edges are wrong or missing, the model learns the wrong relationships.
- They get expensive at scale. Huge graphs with billions of edges need sampling tricks and serious infrastructure.
- Over-smoothing. Pass messages too many times and every node starts to look the same, washing out the signal.
- Not every problem is a graph. If relationships do not matter, a simpler model is faster and easier to maintain.
Related articles:
- What is a neural network? - The broader family that a GNN belongs to.
- What is a convolutional neural network (CNN)? - A sibling architecture built for images instead of graphs.
- Machine learning vs deep learning - Where these networks sit and what makes one "deep".
Want to stay one step ahead?
Don't miss our best insights. No spam, just practical analyses, invitations to exclusive events, and podcast summaries delivered straight to your inbox.