Source-linked AI summary
A3T-GCN: Attention Temporal Graph Convolutional Network for Traffic Forecasting
Jiawei Zhu, Yujiao Song, Ling Zhao, Haifeng Li
TL;DR
Accurate real-time traffic forecasting is difficult because traffic flows have complex spatial and temporal dependencies. A3T-GCN combines GCNs, GRUs, and attention to model these dependencies and global temporal variation. On two real datasets, it outperformed the compared models across prediction horizons.
Problem
Real-time traffic forecasting requires modeling complex temporal dependencies and spatial correlations induced by urban road-network topology.
Method
A3T-GCN uses GCNs for road-network spatial dependencies, GRUs for sequential traffic-speed variation, and attention to capture and assemble global temporal trends.
Results
A3T-GCN was superior to HA, ARIMA, SVR, GCN, GRU, and T-GCN in prediction precision across different prediction horizons on the SZ-taxi and Los-loop datasets.
Takeaways & Limitations
A3T-GCN is applicable to short-term and long-term traffic forecasting tasks in urban road networks.
Takeaways & Limitations
The study defines traffic state only as traffic speed, rather than traffic flow, speed, and density generally.
Abstract
from arXiv · showhide
Accurate real-time traffic forecasting is a core technological problem against the implementation of the intelligent transportation system. However, it remains challenging considering the complex spatial and temporal dependencies among traffic flows. In the spatial dimension, due to the connectivity of the road network, the traffic flows between linked roads are closely related. In terms of the temporal factor, although there exists a tendency among adjacent time points in general, the importance of distant past points is not necessarily smaller than that of recent past points since traffic flows are also affected by external factors. In this study, an attention temporal graph convolutional network (A3T-GCN) traffic forecasting method was proposed to simultaneously capture global temporal dynamics and spatial correlations. The A3T-GCN model learns the short-time trend in time series by using the gated recurrent units and learns the spatial dependence based on the topology of the road network through the graph convolutional network. Moreover, the attention mechanism was introduced to adjust the importance of different time points and assemble global temporal information to improve prediction accuracy. Experimental results in real-world datasets demonstrate the effectiveness and robustness of proposed A3T-GCN. The source code can be visited at https://github.com/lehaifeng/T-GCN/A3T.
1 INTRODUCTION
Traffic forecasting must model complex temporal dependencies and road-network spatial dependencies. A3T-GCN combines GCNs, GRUs, and attention to capture these dependencies and assemble global temporal information.
- 1 INTRODUCTION: Traffic forecasting is challenging because traffic states exhibit complex temporal dependencies and depend on road-network topology.Temporal dependence includes periodicity and tendency, while spatial dependence reflects upstream-to-downstream transmission and downstream-to-upstream effects.
- 1 INTRODUCTION: Traditional parametric models rely on stationarity and cannot represent nonlinear, uncertain traffic states or random-event interference.
- 1 INTRODUCTION: RNNs, LSTMs, and GRUs model temporal dependence but neglect spatial dependence, motivating models that address both dimensions.
- 1 INTRODUCTION: CNNs can characterize spatial dependence, but traffic networks have non-Euclidean structures that limit direct CNN applicability.
- 1 INTRODUCTION: Because recent time points are not always more informative, global-correlation mechanisms are needed to distinguish the importance of different times.
- 1 INTRODUCTION: A3T-GCN combines GCNs and GRUs with attention to capture spatiotemporal dependencies and assemble global variation information for urban-road traffic forecasting.
2 A3T-GCN
A3T-GCN represents urban traffic forecasting as learning future speeds from a road-network graph and historical speed features. It combines GCNs for spatial dependence, GRUs for local temporal variation, and attention for global temporal trends.
- Problem formulation: Traffic forecasting predicts future traffic speed from historical speeds on urban roads, using the road network and traffic features as model inputs.The road network is represented as G = (V, E), while historical traffic speeds form the node feature matrix.
- Spatial modeling: GCNs capture spatial dependence by encoding road-network topology together with the attributes of connected road sections.The model uses graph structure and adjacent-node relationships to characterize spatial features.
- Temporal modeling: A GRU captures local temporal tendencies by using the previous hidden state and current traffic speed through gated state updates.Reset and update gates control which previous information is neglected or carried into the current state.
- Attention modeling: The attention model assigns weights to hidden states at different times and combines them into a context vector describing global traffic variation.A multilayer perceptron is used as the scoring function, and the context vector is computed from weighted hidden states.
- A3T-GCN model: A3T-GCN combines GCNs, GRUs, and attention to capture spatial dependence, local temporal tendencies, and global traffic variation for forecasting.The attention mechanism re-weights historical traffic states, extending the earlier T-GCN model.
- Training objective: Training minimizes prediction error between real and predicted road speeds while adding a regularization term to help avoid overfitting.λ controls the regularization contribution in the objective function.
3 EXPERIMENTS
Experiments on SZ-taxi and Los-loop evaluate A3T-GCN across prediction precision, spatiotemporal modeling, long-term horizons, attention, and noise robustness. The model is reported to perform well across both real-world traffic-speed datasets and multiple forecasting settings.
- Experimental setup: Experiments use SZ-taxi and Los-loop traffic-speed datasets, with 80% of data for training and forecasts at 15, 30, 45, and 60 minutes.SZ-taxi covers 156 major roads in Shenzhen’s Luohu District; the study compares A3T-GCN with HA, ARIMA, SVR, GCN, and GRU.
- Prediction precision: A3T-GCN achieves the optimal prediction performance across all metrics on both real-world datasets.The reported metrics include RMSE, MAE, accuracy, R2, and explained variance.
- Spatiotemporal prediction: Compared with GCN, A3T-GCN reduces RMSE by approximately 31.11%, 31.08%, 30.94%, and 30.78% at 15, 30, 45, and 60 minutes, respectively.These results are reported for SZ-taxi and support the model’s spatial-characteristics capture.
- Spatiotemporal prediction: Compared with GRU, A3T-GCN reduces RMSE by approximately 2.51%, 4.19%, 4.99%, and 2.55% at 15, 30, 45, and 60 minutes, respectively.The comparison is reported for SZ-taxi and supports the model’s ability to capture temporal dependence.
- Long-term prediction: A3T-GCN has the lowest RMSE across 15-, 30-, 45-, and 60-minute horizons on SZ-taxi, while results on Los-loop show consistent long-term prediction behavior.Its forecasting results change slightly as the time-series length changes, and the paper reports applicability to short-term and long-term traffic forecasting.
- Attention mechanism: Against T-GCN, A3T-GCN lowers RMSE by approximately 0.86%, 1.31%, 1.14%, and 0.99% across 15-, 30-, 45-, and 60-minute series.Accuracy is reported as approximately 0.32%, 0.48%, 0.43%, and 0.37% higher at the same horizons.
- Robustness analysis: Under Gaussian and Poisson perturbations, evaluation metrics remain basically unchanged as noise levels vary, indicating stable forecasting results under the tested perturbations.The perturbation analysis uses Gaussian σ values from 0.2 to 2 and Poisson λ values from 1 to 16; consistent results are reported on both datasets.
4 CONCLUSIONS
A3T-GCN jointly models spatial correlations and global temporal dynamics for traffic forecasting. It represents roads as graph nodes, uses GCN and GRU components, and applies attention to capture global temporal variation.
- A3T-GCN captures global temporal dynamics and spatial correlations simultaneously for traffic forecasting.The model is evaluated on urban road-network-based forecasting tasks.
- The urban road network is represented as a graph, with traffic speed on roads encoded as node attributes.
- GCN captures spatial dependencies from road-network topology, while GRU captures dynamic variation in historical traffic speeds.
- An attention mechanism captures and assembles global temporal variation in the traffic state.
- A3T-GCN was tested on two real datasets, SZ-taxi and Los-loop, for urban road-network-based traffic forecasting.