An Adaptive Rank-Based Tensor Ring Completion Model for Intelligent Transportation Systems

Low rank tensor ring based data recovery algorithms have been widely used in vehicle road cooperation intelligent transportation system to recover missing data entries in the sensing data pre-processing stage. However, the existing tensor ring decomposition based methods often resolve the low rank optimization with predefined ranks, which often leads over-fitting when the selected rank is large. To overcome this challenge, we propose a Bayesian inference based tensor ring completion method which can automatically learn an optimal rank for the tensor ring completion. In this work, a likelihood Statistical model is firstly developed for low rank tensor ring approximation, and we impose a hierarchical sparse induced prior on the forward and horizontal slices of the kernel factor. Then, the Variational Bayesian algorithm is used to derive the parameters in the model, and the tensor ring rank can be achieved by gradually pruning the sparse horizontal and forward slice components in the factor. Finally, to elevate the proposal, numerous experiments have been conducted on two different intelligent transportation datasets, and the experimental results show that the proposed method can get the state-of-the-art performance in terms of recovery accuracy.