GNN Expressivity
GNN: Graph Neural Network
Overview
The field asks whether a graph neural network can distinguish non-isomorphic graphs or represent targets such as cycles, motifs, long-range relations, and node roles; this matters because benchmark accuracy alone does not reveal what information a model can capture. For standard message-passing networks, the one-dimensional Weisfeiler–Leman (1-WL) test is a key reference point, with higher-order WL tests, logic, graph homomorphisms, and circuit complexity providing tools to characterize expressive power. To exceed standard message passing, researchers study higher-order and subgraph methods, positional and spectral encodings, graph transformers, recurrent updates, and structural or topological features, often trading greater discrimination for computational cost and scalability. A central tension is that stronger theoretical expressivity does not guarantee that useful information can be learned reliably, while efficient models may retain WL-style blind spots; current analyses work to separate architectural limits from optimization and data effects.
Latest papers
55 papers10 papers in the last four weeks, up 67% on the four weeks before. 0.1% of all new papers.