Low-Rank Autoregressive Tucker Decomposition for Traffic Data Imputation

Real traffic data is often missing due to diverse interference. However, uncompleted inputs will weaken the abilities of intelligent transportation systems. Therefore, it is of great interest that suitable imputation methods be designed. This paper proposes a low-rank autoregressive Tucker decomposi...

Full description

Saved in:
Bibliographic Details
Published in2024 29th International Conference on Automation and Computing (ICAC) pp. 1 - 6
Main Authors Lu, Jiaxin, Gong, Wenwu, Yang, Lili
Format Conference Proceeding
LanguageEnglish
Published IEEE 28.08.2024
Subjects
Accuracy
autoregressive model
Computational modeling
Correlation
Data models
Imputation
low rankness
Matrix decomposition
Numerical models
spatiotemporal correlations
Spatiotemporal phenomena
Tensors
Traffic data imputation
Tucker Decomposition
Online AccessGet full text

Cover

More Information
Summary:Real traffic data is often missing due to diverse interference. However, uncompleted inputs will weaken the abilities of intelligent transportation systems. Therefore, it is of great interest that suitable imputation methods be designed. This paper proposes a low-rank autoregressive Tucker decomposition (LATD) method by exploring the spatiotemporal correlations embedded in high-dimensional traffic data. The low-rank factor matrices and core tensor introduced by the Tucker decomposition allow us to better characterize the long-term trends of the traffic data. We incorporate an autoregressive model to extract the short-term patterns involved. Besides implementing differences between neighboring elements to promote smoothness, this regularization is also well interpretable for characterizing the spatiotemporal correlations. To solve the LATD model, we design a proximal alternating linear minimization algorithm to update each variable iteratively. Numerical experiments on two real traffic datasets indicate that our proposed model outperforms other imputation methods in achieving higher accuracy.
DOI:10.1109/ICAC61394.2024.10718844