Spatiotemporal GNNs
GNN: Graph Neural Network
Overview
Spatiotemporal graph neural networks represent entities such as traffic sensors, body joints, or mesh elements as nodes, and use their connections to model how information changes over time. Common approaches combine graph convolutions or attention with recurrent networks, temporal convolutions, Transformers, or continuous-time models; some also learn changing graph structure or incorporate physical constraints. They are used for traffic forecasting, human-motion prediction, environmental modeling, and simulation, with evaluation ranging from traffic datasets such as METR-LA and PEMS to motion benchmarks such as Human3.6M. A central challenge is balancing accuracy and long-horizon stability with computational cost, while ensuring models generalize across regions, changing conditions, and irregular or sparse observations.
Latest papers
148 papers15 papers in the last four weeks, level with the four weeks before. 0.1% of all new papers.