Deep spatial-temporal bi-directional residual optimisation based on tensor decomposition for traffic data imputation on urban road network

The capacity of fully exploiting underlying spatial-temporal dependencies holds the key for missing traffic data imputation, however, previous studies have neglected the residual information from recovery models. To refine this task, we propose a spatial-temporal bi-directional residual optimisation (ST-BiRT) model on the basis of tensor decomposition to effectively improve the imputation performance. The novelty of our approach concentrates on a well-designed bi-directional residual structure, which reduces model errors dramatically. We can greatly exploit the potential of the optimisation structure by dynamically stacking massive residual units, thereby significantly enhancing the recovery capability. When faced with various combinations of missing scenario and missing rate problems, ST-BiRT model can perform with better accuracy and robustness. Here, the experiments on the Guangzhou traffic speed dataset demonstrate that the proposed ST-BiRT model outperforms the state-of-the-art baseline models. In addition, the mechanism of the bi-directional residual optimisation can address extreme cases and their evaluation metrics can reach acceptable values even when the missing rate exceeds 90%. Finally, the superiority of the ST-BiRT model in repairing loss or low-quality traffic data is confirmed by visualising the experimental results.

Automated summary

The ST-BiRT model proposed in this study effectively enhances the imputation performance of traffic data through a bi-directional residual structure, demonstrating a significant advantage on the Guangzhou traffic speed dataset.