[Paper Review] Spatio-Temporal Dual Graph Neural Networks for Travel Time Estimation

Guangyin Jin, Huan Yan+3 more|arXiv (Cornell University)|May 28, 2021

Traffic Prediction and Management Techniques45 references4 citations

TL;DR

This paper proposes Spatio-Temporal Dual Graph Neural Networks (STDGNN), a novel end-to-end deep learning framework that jointly models road segments and intersections via node-wise and edge-wise graphs to improve travel time estimation. By integrating multi-scale spatio-temporal dual graph learning and multi-task learning, STDGNN achieves state-of-the-art performance on three real-world datasets, outperforming existing baselines significantly.

ABSTRACT

Travel time estimation is one of the core tasks for the development of intelligent transportation systems. Most previous works model the road segments or intersections separately by learning their spatio-temporal characteristics to estimate travel time. However, due to the continuous alternations of the road segments and intersections in a path, the dynamic features are supposed to be coupled and interactive. Therefore, modeling one of them limits further improvement in accuracy of estimating travel time. To address the above problems, a novel graph-based deep learning framework for travel time estimation is proposed in this paper, namely Spatio-Temporal Dual Graph Neural Networks (STDGNN). Specifically, we first establish the node-wise and edge-wise graphs to respectively characterize the adjacency relations of intersections and that of road segments. In order to extract the joint spatio-temporal correlations of the intersections and road segments, we adopt the spatio-temporal dual graph learning approach that incorporates multiple spatial-temporal dual graph learning modules with multi-scale network architectures for capturing multi-level spatial-temporal information from the dual graph. Finally, we employ the multi-task learning approach to estimate the travel time of a given whole route, each road segment and intersection simultaneously. We conduct extensive experiments to evaluate our proposed model on three real-world trajectory datasets, and the experimental results show that STDGNN significantly outperforms several state-of-art baselines.

Motivation & Objective

• To address the limitation of prior work that models only road segments or intersections in isolation, leading to suboptimal travel time estimation.
• To capture the joint spatio-temporal correlations between intersections and road segments, which are dynamically interdependent in real traffic networks.
• To develop a unified, end-to-end deep learning framework that estimates travel time for entire paths, individual road segments, and intersections simultaneously.
• To improve estimation accuracy by modeling both static attributes and dynamic interactions between intersections and road segments through dual graph structures.

Proposed method

• Construct two distinct graphs: a node-wise graph for intersections and an edge-wise graph for road segments, each encoding structural and static features.
• Design a spatio-temporal dual graph learning module that applies spatial graph convolution and temporal convolution to extract independent spatio-temporal patterns from both graphs.
• Introduce a dual graph interaction mechanism to aggregate and exchange latent representations between the node-wise and edge-wise graphs, enabling cross-level feature interaction.
• Employ a multi-scale network architecture to capture hierarchical and multi-level representations across different receptive fields.
• Utilize multi-task learning to jointly predict travel time for full paths, individual road segments, and intersections, enhancing feature sharing and model generalization.
• Incorporate a learnable P-matrix to model the spatial proximity between intersections and road segments, improving structural alignment in the dual graph framework.

Experimental results

Research questions

• RQ1 Can jointly modeling intersections and road segments via dual graph structures improve travel time estimation accuracy compared to modeling them separately?
• RQ2 How do the independent spatio-temporal dynamics of intersections and road segments interact and influence each other in traffic flow propagation?
• RQ3 What is the impact of the dual graph interaction mechanism on capturing cross-structure dependencies in urban road networks?
• RQ4 How do model hyperparameters such as the number of layers and latent dimensionality affect the performance and generalization of the STDGNN framework?
• RQ5 To what extent do the individual components (e.g., w/o intersections, w/o links, w/o P-matrix) contribute to the overall performance of the model?

Key findings

• STDGNN significantly outperforms state-of-the-art baselines on three real-world trajectory datasets, achieving lower RMSE, MAE, and MAPE across all metrics.
• The ablation study shows that removing intersections, road segments, or the P-matrix leads to substantial performance drops—up to 6.07% in MAE on the Porto dataset—demonstrating the critical role of both components.
• The optimal model configuration uses 3 layers and a latent dimension of 30 for the Chengdu dataset and 20 for the Porto dataset, indicating that larger dimensions do not always improve performance due to overfitting.
• Performance peaks at 3 layers; increasing to 4 layers causes performance degradation, suggesting that deeper networks may overfit or fail to generalize on the given data.
• The multi-scale architecture enhances feature representation, and the multi-task learning setup improves estimation accuracy by enabling shared knowledge across path, segment, and intersection predictions.