A graph neural network (GNN) is a neural network that operates directly on data shaped as a graph: a set of nodes connected by edges. Where a convolutional network assumes data laid out on a regular grid (the pixels of an image) and a sequence model assumes data laid out in a line (the words of a sentence), a GNN makes no such assumption. It works on the irregular, relational structure of a graph, which makes it a natural fit for data such as molecules (atoms joined by bonds), road and social networks, knowledge graphs, and the grids used in physical simulation.
The core mechanism is message passing. Each node holds a vector of features; in each layer, a node gathers information from its neighbors, combines it with its own features, and updates its representation. Stacking layers lets information flow further across the graph, so a node’s final representation reflects not just itself but the structure around it. A widely cited formulation is the graph convolutional network of Thomas Kipf and Max Welling, “Semi-Supervised Classification with Graph Convolutional Networks” (arXiv 1609.02907, presented at ICLR 2017), which introduced an efficient layer that “scales linearly in the number of graph edges and learns hidden layer representations that encode both local graph structure and features of nodes.”
GNNs matter because a great deal of important data is relational and does not sit neatly on a grid or in a sequence. A prominent example in this library is GraphCast, DeepMind’s machine-learning weather model, which represents the globe as a graph and uses a GNN to forecast weather faster and more accurately than the leading conventional system (see 2023-graphcast). The same shape of model is used in drug and materials discovery, where molecules are graphs, and in recommendation and fraud systems, where users and items form large interaction graphs.
Why business readers should care: when a problem is naturally about things and the relationships between them - supply chains, networks, molecules, transactions - a graph neural network is often the model that fits the data without forcing it into a grid or a list, and it underlies several of the AI-for-science results that have drawn attention.