Spatio-Temporal Traffic Forecasting with Neural Graph Cellular Automata

Petrônio C. L. Silva , Omid Orang , Lucas Astore, Frederico G. Guimarães

Transformer-based neural network cells with distributed federated learning for flexible cellular automata topologies, aiming for large-scale forecasting of complex spatiotemporal processes.

Key contributions:

• Each cell is a forecaster, which means that their states represent a one-step-ahead forecasting value for a specific spatial location
• Using attention/transformer architectures to allow flexible neighborhood
• Federated Learning is employed to allow the distributed, collaborative, and private training of cell model
• The inference can also be made in parallel and/or distributed methods

In case you have any questions, do not hesitate in contact us using the following e-mail: petronio.candido@ifnmg.edu.br