Source-linked AI summary
AST-GCN: Attribute-Augmented Spatiotemporal Graph Convolutional Network for Traffic Forecasting
Jiawei Zhu, Chao Tao, Hanhan Deng, Ling Zhao, Pu Wang, Tao Lin, Haifeng Li
TL;DR
Traffic forecasting requires more than historical traffic data because weather and surrounding POIs also affect traffic states, while existing methods rarely integrate such external factors. AST-GCN models these factors as dynamic and static road attributes and integrates them into a spatiotemporal graph convolution model. On real data, its prediction results outperform baselines, with experiments also examining information type, prediction length, interpretability, and robustness.
Problem
Accurate traffic forecasting remains difficult because traffic states depend on historical information and external factors, while few existing methods integrate those factors comprehensively.
Method
AST-GCN models weather and surrounding POIs as dynamic and static road-section attributes and integrates them with traffic features in a spatiotemporal graph convolution model.
Results
AST-GCN prediction results outperform baselines on real data, and adding both dynamic and static information lowers prediction error by 0.21% and 0.31%.
Takeaways & Limitations
Modeling external information can facilitate traffic forecasting and supports examining how weather and POI attributes contribute to prediction.
Takeaways & Limitations
The study is limited by its data sources and uses only POI distributions and weather conditions to represent external factors; high and low speed peaks show considerable deviations.
Abstract
from arXiv · showhide
Traffic forecasting is a fundamental and challenging task in the field of intelligent transportation. Accurate forecasting not only depends on the historical traffic flow information but also needs to consider the influence of a variety of external factors, such as weather conditions and surrounding POI distribution. Recently, spatiotemporal models integrating graph convolutional networks and recurrent neural networks have become traffic forecasting research hotspots and have made significant progress. However, few works integrate external factors. Therefore, based on the assumption that introducing external factors can enhance the spatiotemporal accuracy in predicting traffic and improving interpretability, we propose an attribute-augmented spatiotemporal graph convolutional network (AST-GCN). We model the external factors as dynamic attributes and static attributes and design an attribute-augmented unit to encode and integrate those factors into the spatiotemporal graph convolution model. Experiments on real datasets show the effectiveness of considering external information on traffic forecasting tasks when compared to traditional traffic prediction methods. Moreover, under different attribute-augmented schemes and prediction horizon settings, the forecasting accuracy of the AST-GCN is higher than that of the baselines.
1 INTRODUCTION
Traffic forecasting supports intelligent transportation management but remains difficult because traffic states depend on historical conditions and diverse external factors. AST-GCN incorporates dynamic and static external information into a spatiotemporal graph convolution model, and experiments report better performance than baselines.
- Traffic forecasting provides a basis for urban transportation management, planning, and traffic-flow guidance.Predicted traffic states can support advance deployment and guidance of traffic flows.
- Historical traffic information alone is insufficient because static and dynamic external factors also affect future traffic states.Examples include surrounding restaurants and changing weather conditions.
- AST-GCN models external factors as road-section attributes and jointly represents them with traffic features.The resulting augmented feature vectors are designed to enhance the model’s perception of external information.
- The model integrates both dynamic factors such as weather and static factors such as surrounding POIs within a spatiotemporal graph convolution model.
- Experiments on real data report that AST-GCN prediction results outperform the baselines, supporting the effectiveness of modeling external information.
2 RELATED WORKS
Traffic forecasting research progressed from statistical and machine-learning methods to deep spatiotemporal models, but integrating external influences remains an open problem. AST-GCN addresses this gap by treating external factors as road-segment attributes and combining them with traffic characteristics.
- Early traffic forecasting methods included historical averages, time-series models, and machine-learning algorithms such as KNN, Bayesian inference, and SVM.Time-series approaches rely on stability assumptions and may not capture traffic-flow mutations.
- Deep learning methods were adopted for their ability to capture nonlinear and complex traffic-flow patterns.
- CNN-LSTM combinations improved prediction accuracy by modeling spatial and temporal dependencies, but CNNs are limited by transportation networks’ non-Euclidean topology.
- Traffic forecasting must account for external factors such as weather conditions and surrounding POI distributions in addition to historical and spatial information.
- AST-GCN treats external factors as road-segment attributes and integrates them with traffic characteristics to enhance external-information perception.
3 METHOD
AST-GCN augments spatiotemporal graph convolution with static and dynamic external attributes, including POIs and weather, to model spatial-temporal traffic dependencies and predict future traffic states.
- Traffic forecasting formulation: Traffic forecasting learns future traffic states from road-network topology G, traffic features X, and auxiliary attribute matrix K.The task uses historical traffic speeds and external factors affecting traffic.
- External factors: Static POI distributions influence traffic states, while dynamic weather conditions can directly affect traffic speed over time.The paper illustrates both effects using differences between road sections and weather conditions across days.
- Attribute augmentation: The attribute augmentation unit models external factors as static attributes S and dynamic attributes D, then derives augmented matrices from traffic features and attributes.Static attributes remain constant across timestamps; dynamic attributes use a window of m + 1 time points to represent cumulative effects.
- Spatiotemporal graph convolution: GCN layers generate road-section representations capturing spatial dependencies, and GRU layers process these time-varying representations to capture temporal dependencies.The GCN and GRU combination forms the spatiotemporal modeling component.
- AST-GCN model: AST-GCN integrates the attribute augmentation unit with a spatiotemporal graph convolution model, using the enhanced matrix E as input for final prediction.The framework includes data preprocessing, attribute augmentation, spatiotemporal dependency modeling, and prediction.
4.1 Datasets
The experiments use Shenzhen taxi speeds, road-section POI distributions, and time-varying weather observations from January 2015.
- The SZ taxi dataset contains trajectories from January 1–31, 2015, covering 156 major Luohu District road sections.Road-section connectivity is represented by a 156*156 adjacency matrix, with traffic speeds organized by section and timestamp.
- The SZ POI dataset represents each road section using the POI category with the largest local proportion.It distinguishes nine POI categories, including catering, shopping, transportation, education, medical services, and accommodations.
- The SZ Weather dataset records five weather categories every 15 minutes in January 2015.These observations form a dynamic attribute matrix with size 156*2,976.
4.2 Evaluation metrics
Model performance is evaluated with error and predictive-ability metrics, where lower error values and higher predictive-ability values indicate better results.
- RMSE measures prediction error, with smaller values indicating smaller error and better model performance.
- MAE measures the average absolute difference between predicted results and ground truth.
- R2 measures predictive ability, with larger values indicating better prediction results.
- Explained variation measures the proportion of variation in real traffic states accounted for by the proposed model.It is used to measure predictive ability.
4.3 Parameter Settings
The study fixes several training settings and searches training epochs and hidden units experimentally, selecting values where evaluation metrics become stable.
- The learning rate, batch size, and training-set proportion are manually fixed at 0.001, 64, and 0.8, respectively.
- Training epochs are tested over [500, 1,000, 1,500, 2,000, 3,000, 3,500], with stability turning at 3,000 epochs.
- With training epochs fixed at 3,000, hidden units are selected from [8, 16, 32, 64, 100, 128], and stability occurs at 100 units.
4.4 Baselines
AST-GCN is compared with statistical, regression, graph-convolutional, recurrent, and spatiotemporal traffic forecasting baselines.
- The baselines are ARIMA, SVR, GCN, GRU, TGCN, and DCRNN.Their hyperparameters follow the original papers or released code.
4.5 Experimental Results
AST-GCN integrates static POI and dynamic weather information into spatiotemporal traffic forecasting. Experiments compare baselines, attribute configurations, prediction horizons, robustness, and interpretability.
- Experimental design: AST-GCN combines spatiotemporal modeling with external information, evaluating static POI distributions and dynamic weather conditions.The experiments include baseline comparisons and ablations with neither, one, or both attribute types.
- Forecasting comparison with baselines: 44.19% and 40.49% RMSE reductions are reported against SVR and ARIMA, respectively, while AST-GCN also outperforms GCN and GRU.Against GCN and GRU, RMSE is reduced by approximately 28.58% and 20.44%, respectively.
- Ablation experiments: 0.98% and 10.46% RMSE reductions are reported for AST-GCN with weather+POI versus TGCN and DCRNN, respectively.Weather-only and POI-only variants also reduce RMSE relative to these spatiotemporal baselines.
- Ablation experiments: Dynamic weather information performs better than static POI information, while combining both reduces prediction error by 0.21% and 0.31% relative to single-attribute models.The results are interpreted as indicating that weather and POI information provide complementary external factors.
- Prediction horizons: 0.64% and 11.64% lower RMSE than TGCN and DCRNN, respectively, is reported for the 60-minute horizon; performance remains good across tested horizons.The tested horizons are 15, 30, 45, and 60 minutes.
- Robustness: AST-GCN’s evaluation metrics change negligibly across Gaussian and Poisson noise settings, supporting the reported robustness analysis.The perturbation analysis varies Gaussian σ across 0.2, 0.4, 0.6, 0.8, 1, and 2, and Poisson λ across 1, 2, 4, 8, and 16.
- Interpretation: External information brings predictions closer to ground truth at peaks and turning points, with combined dynamic and static attributes producing smaller deviations than either alone.The comparisons include models enhanced with POI information, weather information, or their combination against forecasting without external information.
5 CONCLUSION
The paper concludes that AST-GCN addresses the limited treatment of traffic-affecting external factors by integrating static and dynamic information. Baseline comparisons support the importance of external information, and perturbation analysis tests robustness.
- Conclusion: AST-GCN integrates both static external information and dynamic external data into an attribute-enhanced spatiotemporal graph convolution model.The conclusion frames this integration as the paper’s response to incomplete consideration of traffic-affecting external factors.
- Conclusion: Baseline comparisons verify the importance of considering external information in traffic forecasting tasks.The conclusion also identifies perturbation analysis as the robustness test for the model.