What is a graph neural network?
A graph neural network (GNN) is a neural network for data shaped as a graph: entities as nodes and relationships as edges. Customers and the accounts they share, parts and the suppliers that make them, or molecules and their bonds all have this shape. Ordinary neural networks expect a fixed grid or sequence and lose the connection information. A GNN keeps it.
The mechanism is message passing. In each layer, every node collects the current representations of its neighbors, combines them with its own, and produces an updated representation. After several layers, each node’s embedding reflects its neighborhood several hops out. These embeddings then feed a prediction for each node, each edge, or the whole graph.
Graph neural networks in practice
A payments company fights fraud rings that share devices, addresses, and bank accounts across many apparently unrelated customers. A GNN trained on the transaction graph learns that an account connected to several known fraudulent accounts through shared attributes is itself high risk, even when its own transactions look ordinary. A manufacturer maps its multi-tier supply chain as a graph and uses a GNN to estimate how a disruption at one second-tier supplier propagates to finished goods. GNNs also appear in recommendation systems and in reasoning over knowledge graphs. They need the graph to be assembled first, which is usually the largest part of the work.