Spatio-Temporal Meta-Graph Learning for Traffic Forecasting

Traffic forecasting as a canonical task of multivariate time series has been significant research topic in AI community. To address the spatio-temporal heterogeneity and non-stationarity implied traffic stream, this study, we propose Spatio-Temporal Meta-Graph Learning novel Graph Structure mechanism on data. Specifically, implement idea into Convolutional Recurrent Network (MegaCRN) by plugging Learner powered Meta-Node Bank GCRN encoder-decoder. We conduct comprehensive evaluation two benchmark datasets (i.e., METR-LA PEMS-BAY) new large-scale speed dataset called EXPY-TKY that covers 1843 expressway road links Tokyo. Our model outperformed state-of-the-arts all three datasets. Besides, through qualitative evaluations, demonstrate our can explicitly disentangle slots with different patterns be robustly adaptive to any anomalous situations. Codes are available at https://github.com/deepkashiwa20/MegaCRN.