Home > Research > Publications & Outputs > Multi-view Bayesian spatio-temporal graph neura...

Electronic data

  • accepted version

    Accepted author manuscript, 1.8 MB, PDF document

    Available under license: Other


Text available via DOI:

View graph of relations

Multi-view Bayesian spatio-temporal graph neural networks for reliable traffic flow prediction

Research output: Contribution to Journal/MagazineJournal articlepeer-review

E-pub ahead of print
  • J. Xia
  • S. Wang
  • X. Wang
  • M. Xia
  • K. Xie
  • J. Cao
<mark>Journal publication date</mark>20/10/2022
<mark>Journal</mark>International Journal of Machine Learning and Cybernetics
Publication StatusE-pub ahead of print
Early online date20/10/22
<mark>Original language</mark>English


Accurate traffic flow prediction is critically essential to transportation safety and Intelligent Transportation Systems (ITS). Existing approaches generally assume the traffic data are complete and reliable. However, in real scenarios, the traffic data are usually sparse and noisy due to the unreliability of the road sensors. Meanwhile, the global semantic traffic correlations among the road links over the road network are largely ignored by existing works. To address these issues, in this paper we study the novel problem of reliable traffic prediction with noisy and sparse traffic data and propose a Multi-View Bayesian Spatio-Temporal Graph Neural Network (MVB-STNet for short) to effectively address it. Specifically, we first construct the traffic flow graphs from two views, the structural traffic graph based on the topological closeness of the road sensors, and the semantic traffic graph which is constructed based on the traffic flow correlations among all the road sensors. Then the features of the two views are learned simultaneously to more broadly capture the spatial correlations. Inspired by the effectiveness of Bayesian neural networks in handling data uncertainty, we design the Bayesian Spatio-Temporal Long Short-Term Memory Net layer to more effectively learn the spatio-temporal features from the sparse and noisy traffic data. Extensive evaluations are conducted over two real traffic datasets. The results show that our proposal significantly improves current state-of-the-arts in terms of traffic flow prediction with sparse and noisy data.