[Paper Review] DDP-GCN: Multi-Graph Convolutional Network for Spatiotemporal Traffic Forecasting

Kyung‐Eun Lee, Wonjong Rhee|arXiv (Cornell University)|May 29, 2019

Traffic Prediction and Management Techniques4 citations

TL;DR

DDP-GCN proposes a multi-graph convolutional network that integrates distance, direction, and positional relationship as spatial dependencies for spatiotemporal traffic speed forecasting. By modeling these non-Euclidean relationships through graph elements and partition filters, the method achieves significant performance gains in long-term forecasting on complex urban networks, especially during peak hours.

ABSTRACT

Traffic speed forecasting is one of the core problems in transportation systems. For a more accurate prediction, recent studies started using not only the temporal speed patterns but also the spatial information on the road network through the graph convolutional networks. Even though the road network is highly complex due to its non-Euclidean and directional characteristics, previous approaches mainly focused on modeling the spatial dependencies using the distance only. In this paper, we identify two essential spatial dependencies in traffic forecasting in addition to distance, direction and positional relationship, for designing basic graph elements as the fundamental building blocks. Using the building blocks, we suggest DDP-GCN (Distance, Direction, and Positional relationship Graph Convolutional Network) to incorporate the three spatial relationships into deep neural networks. We evaluate the proposed model with two large-scale real-world datasets, and find positive improvements for long-term forecasting in highly complex urban networks. The improvement can be larger for commute hours, but it can be also limited for short-term forecasting.

Motivation & Objective

• Address the limitation of existing GCN-based traffic forecasting models that rely solely on Euclidean distance for spatial dependencies.
• Identify and incorporate two non-Euclidean spatial dependencies—direction and positional relationship—into graph-based deep learning for traffic forecasting.
• Develop a multi-graph convolution framework that effectively encodes complex road network topologies, including directional flow and spatial proximity.
• Improve long-term traffic speed forecasting accuracy in highly complex urban environments where traditional models struggle.
• Investigate the contribution of each spatial dependency and the impact of architectural components like partition filters through ablation studies.

Proposed method

• Define three distinct graph elements: distance (W_D), direction (W_θ), and positional relationship (W_P), each representing a different spatial dependency.
• Construct a multi-graph convolutional network where each graph element is processed independently before being fused through a stacked architecture.
• Introduce partition filters T_M and T'_M to sub-divide directional and distance-based graphs into components with similar characteristics, enhancing representation learning.
• Use K-polynomial ChebNet for message passing over graphs, with the stacked multi-graph design replacing the need for large K-hop convolutions.
• Apply residual connections and gated mechanisms to capture long-range temporal dependencies, integrating temporal modeling with spatial graph learning.
• Train the end-to-end model using historical speed sequences, optimizing for forecasting accuracy on future time steps.

Experimental results

Research questions

• RQ1 How do non-Euclidean spatial dependencies—direction and positional relationship—impact traffic speed forecasting accuracy compared to distance-only modeling?
• RQ2 Can multi-graph convolution with distinct spatial graph elements (distance, direction, positional relationship) outperform single-graph or K-hop GCN baselines in long-term traffic forecasting?
• RQ3 What is the relative contribution of each spatial dependency (distance, direction, positional relationship) to model performance, as measured by ablation?
• RQ4 How do partition filters improve the modeling of spatial relationships in directed road networks?
• RQ5 Does the stacked multi-graph design surpass the performance of standard K-hop GCN convolutions, especially for long-term predictions?

Key findings

• DDP-GCN significantly improves long-term traffic speed forecasting performance on two large-scale real-world urban datasets, particularly during peak commute hours.
• Ablation studies show that removing the positional relationship graph element causes the largest performance degradation, indicating its critical role in modeling spatial dependencies.
• Replacing directional or positional relationship graphs with distance-only graphs leads to performance drops, confirming that distance alone is insufficient for complex urban networks.
• The use of partition filters consistently improves model performance, demonstrating their effectiveness in refining spatial graph representations.
• DDP-GCN(Stacked) outperforms K-hop GCN models even when those models use larger K values, indicating that stacking distinct spatial graph elements is more effective than increasing receptive field size via K-hops.
• The model shows limited improvement in short-term forecasting, suggesting its primary advantage lies in modeling long-range spatial and temporal dependencies in complex urban road networks.