Citation

Lyu C, Lu QL, Wu X, Antoniou C. Transp. Res. C Emerg. Technol. 2024; 160: e104502.

Copyright

(Copyright © 2024, Elsevier Publishing)

DOI

10.1016/j.trc.2024.104502

PMID

unavailable

Abstract

Missing values are prevalent in spatio-temporal traffic data, undermining the quality of data-driven analysis. While prior works have demonstrated the promise of tensor completion methods for imputation, their performance remains limited for complicated composite missing patterns. This paper proposes a novel imputation framework combining tensor factorization and rank minimization, which is effective in capturing key traffic dynamics and eliminates the need for exhaustive rank tuning. The framework is further supplemented with time series decomposition to account for trends, spatio-temporal correlations, and outliers, with the intention of improving the robustness of imputation results. A Bregman ADMM algorithm is designed to solve the resulting multi-block nonconvex optimization efficiently. Experiments on four real-world traffic state datasets suggest that the proposed framework outperforms state-of-the-art imputation methods, including the context of complex missing patterns with high missing rates, while maintaining reasonable computation efficiency. Furthermore, the robustness of our model in extreme missing data scenarios, as well as under perturbation in hyperparameters, has been validated. These results also underscore the potential benefits of incorporating temporal modeling for more reliable imputation.

Language: en

Keywords

Bregman ADMM; Changepoint detection; Data imputation; Tensor completion; Tucker factorization