An Inductive Multi-Scale Graph Memory Network for Spatial–Temporal Traffic Data Kriging
IEEE Access · 2026 · 14 · 78297-78318
DOI: 10.1109/access.2026.3694513 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Adnan Zeb und 4 weiteren Autor:innen. Die Arbeit nennt 39 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Complete traffic data collection is essential for the effective operation of intelligent transportation systems. However, deploying sensors at every location for complete coverage is not economically and geographically feasible, resulting in incomplete traffic data. Such sparsity of data adversely affects downstream operations like traffic monitoring and control. To overcome this issue, spatial-temporal kriging methods have recently become popular in inferring missing sensor data from observed locations. However, existing methods are limited in modeling the complex multi-scale spatial dependencies within traffic networks and often fail to effectively retain and leverage historical traffic patterns, which ultimately limits their generalization capabilities to unobserved locations. In this paper, we propose a novel spatial-temporal traffic data kriging model, the Inductive Multi-Scale Graph Memory Network (IMSGMN), which advances beyond state-of-the-arts in both inference capabilities and computational efficiency. Specifically, IMSGMN integrates two innovative components: a Multi-scale Spatial Diffusion (MSD) module that generates and processes multiple dynamic graph structures to capture complex, time-varying, sample-level spatial dependencies at local, intermediate, and global scales, and a Kriging Memory Bank (KMB) that stores and retrieves relevant historical traffic patterns to effectively model wide-range temporal dependencies. A Spatial-Temporal Fusion (STF) module then combines the outputs of both modules to generate the final output. Extensive evaluations on standard traffic datasets demonstrate that IMSGMN achieves significant improvements in interpolation performance while substantially reducing computational time by 70% and parameter count by 54.2% compared to recent state-of-the-art deep learning-based approaches.
Zitieren
APA
Zeb, A., Shah, P. M., Chen, J., Peng, P., & Zheng, J. (2026). An Inductive Multi-Scale Graph Memory Network for Spatial–Temporal Traffic Data Kriging. IEEE Access, 14, 78297-78318. https://doi.org/10.1109/access.2026.3694513
Vancouver
Zeb A, Shah PM, Chen J, Peng P, Zheng J. An Inductive Multi-Scale Graph Memory Network for Spatial–Temporal Traffic Data Kriging. IEEE Access. 2026;14:78297-78318. doi:10.1109/access.2026.3694513
BibTeX
@article{Zeb2026,
author = {Zeb, Adnan and Shah, Pir Masoom and Chen, Junde and Peng, Peipei and Zheng, Jianying},
title = {An Inductive Multi-Scale Graph Memory Network for Spatial–Temporal Traffic Data Kriging},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {78297--78318},
doi = {10.1109/access.2026.3694513}
}
EUVIMED