Scalable Spatiotemporal Graph Neural Networks

Neural forecasting of spatiotemporal time series drives both research and industrial innovation in several relevant application domains. Graph neural networks (GNNs) are often the core component architecture. However, most GNNs, computational complexity scales up to a quadratic factor with length sequence times number links graph, hence hindering these models large graphs long temporal sequences. While methods improve scalability have been proposed context static graphs, few efforts devoted case. To fill this gap, we propose scalable architecture that exploits an efficient encoding spatial dynamics. In particular, use randomized recurrent network embed history input into high-dimensional state representations encompassing multi-scale Such then propagated along dimension using different powers graph adjacency matrix generate node embeddings characterized by rich pool features. The resulting can be efficiently pre-computed unsupervised manner, before being fed feed-forward decoder learns map predictions. training procedure parallelized node-wise sampling without breaking any dependency, thus enabling networks. Empirical results on datasets show our approach achieves competitive art, while dramatically reducing burden.