Causality Induced Distributed Spatio-temporal Feature Extraction

Various real world data contains complex coupling spatio-temporal information, which brings a huge challenge for prediction, especially long-term prediction. Therefore, in this study, we propose a causality induced spatiotemporal feature extraction method and a novel deep learning framework for long-term strongly coupling data prediction tasks, which can effectively extract long-term spatio-temporal dependence of the time series data through causal network, geographic network and multiple time extraction mechanism. The proposed algorithm has achieved outstanding prediction performance in the widely- used test data set of traffic flow, where the long-term prediction accuracy of is nearly 30% better than other state-of-the-art currently-used spatio-temporal prediction models.