Abstract
Uncertainty is an essential consideration for time series forecasting tasks.In this work, we specifically focus on quantifying the uncertainty of trafficforecasting. To achieve this, we develop Deep Spatio-Temporal UncertaintyQuantification (DeepSTUQ), which can estimate both aleatoric and epistemicuncertainty. We first leverage a spatio-temporal model to model the complexspatio-temporal correlations of traffic data. Subsequently, two independentsub-neural networks maximizing the heterogeneous log-likelihood are developedto estimate aleatoric uncertainty. For estimating epistemic uncertainty, wecombine the merits of variational inference and deep ensembling by integratingthe Monte Carlo dropout and the Adaptive Weight Averaging re-training methods,respectively. Finally, we propose a post-processing calibration approach basedon Temperature Scaling, which improves the model's generalization ability toestimate uncertainty. Extensive experiments are conducted on four publicdatasets, and the empirical results suggest that the proposed methodoutperforms state-of-the-art methods in terms of both point prediction anduncertainty quantification.