[Paper Review] 3D Graph Convolutional Networks with Temporal Graphs: A Spatial Information Free Framework For Traffic Forecasting

Bing Yu, Mengzhang Li+2 more|arXiv (Cornell University)|Mar 3, 2019

Traffic Prediction and Management Techniques19 references42 citations

TL;DR

The paper introduces 3D-TGCN, a spatio-temporal graph neural network that learns temporal similarities among roads to construct a temporal adjacency graph and performs 3D graph convolutions to forecast traffic speeds without relying on predefined spatial graphs.

ABSTRACT

Spatio-temporal prediction plays an important role in many application areas especially in traffic domain. However, due to complicated spatio-temporal dependency and high non-linear dynamics in road networks, traffic prediction task is still challenging. Existing works either exhibit heavy training cost or fail to accurately capture the spatio-temporal patterns, also ignore the correlation between distant roads that share the similar patterns. In this paper, we propose a novel deep learning framework to overcome these issues: 3D Temporal Graph Convolutional Networks (3D-TGCN). Two novel components of our model are introduced. (1) Instead of constructing the road graph based on spatial information, we learn it by comparing the similarity between time series for each road, thus providing a spatial information free framework. (2) We propose an original 3D graph convolution model to model the spatio-temporal data more accurately. Empirical results show that 3D-TGCN could outperform state-of-the-art baselines.

Motivation & Objective

• Motivate accurate traffic forecasting amidst complex spatio-temporal dynamics and limited reliable spatial topology information.
• Propose a data-driven graph construction based on temporal similarity to capture cross-road patterns.
• Develop a 3D graph convolutional architecture that jointly learns spatial-temporal dependencies.
• Show that temporal-similarity graphs can outperform traditional spatial adjacency in real-world datasets.

Proposed method

• Construct a temporal similarity graph by applying Dynamic Time Warping (DTW) to road time series and connect each road to its top similar roads.
• Develop a 3D graph convolution operation that jointly aggregates over spatial neighbors and temporal windows in a single layer.
• Stack multiple 3D graph convolution blocks with layer normalization and GLU activations to form an end-to-end network.
• Train with a combined L2 and L1 loss, optimized via Adam, for short-term traffic forecasting horizons.
• Evaluate against baselines on real-world datasets PeMSD7 and PEMS-BAY, comparing against STGCN, DCRNN, and other models.

Experimental results

Research questions

• RQ1 Can traffic speed prediction be improved by learning a temporal rather than spatial adjacency among roads?
• RQ2 Does a 3D graph convolution that fuses spatial and temporal information outperform models that treat these dimensions separately?
• RQ3 Is a purely data-driven, spatial-information-free graph construction viable at scale for real-world traffic networks?
• RQ4 How does 3D-TGCN perform in short- and long-term forecasting tasks compared with state-of-the-art baselines?

Key findings

• 3D-TGCN achieves state-of-the-art performance on two large real-world datasets without requiring a spatial topology graph.
• Temporal similarity-based graphs (DTW) outperform traditional spatial adjacency in modeling road relationships for forecasting.
• The 3D graph convolution layer jointly captures spatial and temporal dependencies and offers faster training and convergence than some RNN-based approaches.
• On PeMSD7 and PEMS-BAY, 3D-TGCN shows strong predictive accuracy for both short-term and longer horizons, outperforming baselines like STGCN and DCRNN in several settings.
• The model is scalable and adaptable to other spatio-temporal tasks beyond traffic forecasting, thanks to its graph-learning and joint 3D convolution design.