CoATR: A Convolutional Autoregressive Tensor-Ring decomposition method for traffic data

Spatio-temporal traffic data collected by various sensing systems have chronic issues of missing and corruption, thus accurate data imputation and prediction have been extensively researched. Previous most methods can be divided into two categories: model-driven methods with physical explanations and data-driven deep learning methods. These methods all achieve promising results due to their respective advantages. However they have some limitations: model-driven methods are often linear and may not accurately model the spatiotemporal complexity, while deep learning methods often lack the physical reality making them probe to over-fitting. Inspired by both, we propose anew Convolutional-based generalized Autoregressive Tensor- Ring decomposition method (CoATR) for the completion of spatio-temporal data. CoATR not only retains the advantages of the tensor-ring (TR) decomposition model for global modeling of spatio-temporal data but also exploits the ability of deep networks to model nonlinear features. To be specific, we introduce TR decomposition to capture the global low-rankness and employ multilayer convolutional neural networks to model the global complex interactions among the TR factors for exploring the nonlinear features of the spatio-temporal data. Moreover, we design a new autoregressive network to further explore the local temporal variation in the data. Extensive experiments on a variety of common traffic datasets have validated the effectiveness and superiority of the CoATR over classical model-driven methods and other state-of-the-art data-driven deep learning methods.