Dynamic Tucker Decomposition for Traffic Data Completion

The missing data issue is common in practical traffic data collection, making accurate completion critical for various applications such as transport planning and route navigation. A lot of approaches have been proposed to address this issue, but many challenges remain, especially in designing efficient online completion algorithms that adapt to the streaming nature of traffic data. In this paper, to jointly aggregate multidimensional correlations, we utilize streaming tensors to represent traffic data and propose a dynamic tensor completion model based on Tucker decomposition to estimate missing values. Furthermore, we develop an efficient online algorithm to track the optimal core tensor and factor matrices, thanks to an alternating optimization framework. Experiments are conducted on two public traffic datasets to evaluate our method against some state-of-the-art baseline approaches. The results show that our method achieves completion performance closer to that of non-dynamic methods in a running time close to that of dynamic methods.