TR-VBI: The Bayesian Tensor Ring Decomposition Model with Automatic Rank Determination for Spatiotemporal Traffic Data Imputation

Spatiotemporal traffic data imputation is important to leverage the role of traffic data in Intelligent Transportation Systems (ITSs), as missing data is inevitable during collection. Existing tensor factorization models for traffic data imputation commonly use decomposition methods with limited data information extraction capabilities, and they are susceptible to overfitting when confronted with large pre-defined ranks and high data missing rates. In this paper, a novel tensor factorization model is proposed. Firstly, to enhance data mining capability, tensor ring (TR) decomposition is used to approximate data, which enables the extraction of more latent information. Subsequently, to address the overfitting issue, an automatic determination mechanism of tensor ring ranks (TR-ranks) is proposed by placing sparsity-inducing hierarchical prior over the core factors. Finally, extensive experiments are conducted on four types of traffic data sets with different missing scenarios. Compared with some state-of-the-art baseline models, the proposed model shows the effectiveness in diverse missing scenarios.