[2301.06650] Enhancing Deep Site visitors Forecasting Fashions with Dynamic Regression

View a PDF of the paper titled Enhancing Deep Site visitors Forecasting Fashions with Dynamic Regression, by Vincent Zhihao Zheng and a couple of different authors

View PDF
HTML (experimental)

Summary:Deep studying fashions for site visitors forecasting typically assume the residual is unbiased and isotropic throughout time and house. This assumption simplifies loss capabilities akin to imply absolute error, however real-world residual processes typically exhibit important autocorrelation and structured spatiotemporal correlation. This paper introduces a dynamic regression (DR) framework to reinforce current spatiotemporal site visitors forecasting fashions by incorporating structured studying for the residual course of. We assume the residual of the bottom mannequin (i.e., a well-developed site visitors forecasting mannequin) follows a matrix-variate seasonal autoregressive (AR) mannequin, which is seamlessly built-in into the coaching course of by way of the redesign of the loss perform. Importantly, the parameters of the DR framework are collectively optimized alongside the bottom mannequin. We consider the effectiveness of the proposed framework on state-of-the-art (SOTA) deep site visitors forecasting fashions utilizing each velocity and movement datasets, demonstrating improved efficiency and offering interpretable AR coefficients and spatiotemporal covariance matrices.