Running summarizing tools on a new article

Abstract: Spatiotemporal traffic data, such as speed and flow measurements, often suffer from random and structured missingness caused by detector failures, communication interruptions, and outlier removal, which degrade the reliability of traffic-state estimation, forecasting, and control. Existing traffic data imputation methods typically model cross-sensor correlation, day-to-day similarity, and temporal dependence within a unified matrix or third-order tensor framework, which limits their ability to represent heterogeneous dependence patterns with different spectral compressibility. To address this issue, this study proposes a traffic data imputation model, termed BTR-TS, based on bilateral tensor ring decomposition. Specifically, the original third-order traffic tensor is reshaped into a fourth-order tensor of sensor $\times $ day $\times $ sub-slot $1\times $ sub-slot 2. Two coupled tensor-ring branches are then used to model the cross-sensor/day-to-day structure and the intra-day periodic structure, respectively. In addition, a one-dimensional temporal-difference smoothing term is imposed along the merged intra-day time axis. The resulting model is solved by a proximal alternating minimization (PAM) algorithm. The experimental results on the Guangzhou, PeMSD7(M), and PeMSD8 datasets show that BTR-TS achieves superior performance under both random and structured missingness, with particularly clear advantages in challenging missing scenarios. The code is available at https://github.com/demonrth/BTR-TS