A network for data shaped like a network: molecules, road systems, social connections, supply chains. Rather than assuming a grid like an image or a sequence like a sentence, it works with whatever connections actually exist and lets information flow along them.
Technical view
Operates by message passing, where each node aggregates transformed representations of its neighbours over several rounds, giving each node a view of its k-hop neighbourhood. Variants differ in aggregation, including graph convolution, attention-based and sampling-based approaches. Known limitations are over-smoothing with depth, where node representations converge, and limited expressive power bounded by the Weisfeiler-Lehman test.
Related
Part of AI Terminology Index, section: How the Machinery Works.