Long-term Prediction on Graph Data with Causal Network Construction

Duxin Chen

Abstract:

Graph data encodes unobservable complex spatiotemporal information of complex systems, which brings a great challenge for accurate and stable long-term prediction. Therefore, long-term prediction of such networked flow data has always been one of the bottlenecks of modern complex systems. In previous studies, most researchers only paid attention to short-term prediction of graph data. The accuracy of these models deteriorates rapidly when they are applied in long-term prediction. In this study, a causation based spatio-temporal feature
extraction method with a novel deep learning framework named pyramid spatio-temporal network (PSTN) is proposed for longterm
networked flow prediction tasks, which can achieve stable long-term spatio-temporal feature extraction with the constructed causal network, and the multiple temporal extraction mechanism. PSTN has achieved prior performance in the tested PeMSD7(M) and PeMS-Bay data sets, where the long-term prediction accuracy of PSTN is highly better than other widely-used baseline prediction models, which may provide some insight into spatiotemporal prediction researches.