Abstract
Traffic congestion prediction is critical for urban planning and navigation. While most models focus on dense variables like average speed, congestion events are sparse and occur in continuous time. This paper introduces STGNPP, a framework that combines graph neural networks with neural point processes to predict the occurrence time and duration of congestion events. By modeling traffic as a continuous-time evolution across a road network, STGNPP outperforms state-of-the-art methods by approximately 10%.
Introduction
Traditional traffic prediction research focuses on "dense" variables—metrics like average speed or traffic flow measured at regular intervals (e.g., every 5 minutes). However, traffic congestion events are "sparse." They do not occur at every time step and exist on a continuous timeline rather than a discrete grid.
Existing models face two major hurdles:
- Spatio-Temporal Dependencies: Congestion propagates through a road network over time; a jam on one street often triggers a jam on an adjacent street.
- Hybrid Dynamics: Traffic states involve both gradual, continuous changes (normal flow) and sudden, instantaneous shifts (the formation of a bottleneck). Traditional discrete-time models struggle to capture this "hybrid" nature.
The STGNPP Framework
The authors propose STGNPP, which integrates several sophisticated components to bridge the gap between graph learning and event prediction.
Key Components
- Spatio-Temporal Graph Learning Module: This uses a Link-Wise Transformer to capture long-range temporal dependencies and a Graph Convolutional Network (GCN) with adaptive matrices to model how congestion spreads spatially across the road network.
- Spatio-Temporal Inquirer: A specialized module that extracts "contextual link representations" by mapping historical traffic states to specific past congestion events.
- Continuous GRU (GRU-flow): Instead of a standard discrete Gated Recurrent Unit, the authors use a "neural flow" architecture. This allows the model to represent the hidden state of a road as a continuous function of time, effectively modeling the evolution of traffic between events without the massive computational overhead of ODE-based solvers.
- Periodic Gated Intensity Function: The "intensity function" determines the likelihood of an event occurring. The authors improved this by adding a periodic gated mechanism that accounts for human-centric patterns, such as morning and evening rush hours.
Experimental Evaluation
The model was evaluated on two massive, real-world datasets collected from the Amap mobile application:
- Beijing: 573 road links, containing 249,464 congestion events.
- Chengdu: 435 road links, containing 204,768 congestion events. Both datasets cover a six-month period in 2021 and include link speeds, congestion labels, and periodic features (time of day, day of week).
Key Results
STGNPP consistently outperformed ten baselines, including simple statistical models (Historical Average), standard spatio-temporal models (DCRNN, Graph WaveNet), and existing neural point processes (NHTPP, RMTPP).
- Superior Accuracy: In Beijing, it achieved a Mean Absolute Error (MAE) for occurrence time of 16.95 minutes, compared to 18.86 minutes for the best neural point process baseline (FNN-TPP).
- Duration Prediction: It also excelled at predicting how long a jam would last, with a duration MAE of 13.15 minutes in Beijing, a significant improvement over traditional traffic models.
- Ablation Success: Removing the GCN or the Transformer components led to a noticeable drop in performance, proving that both spatial relationships and long-term temporal context are essential for accurate event forecasting.
Discussion and Future Work
While effective, the authors noted that increasing the number of GCN layers beyond three caused performance to degrade due to the "over-smoothing" problem common in graph neural networks. Additionally, the model's performance is sensitive to the number of flow layers in the continuous GRU; too many layers can lead to over-fitting.
Future directions include integrating external factors like weather conditions, holidays, or social events, which could further refine the "intensity function" of traffic congestion.
Conclusion
STGNPP shifts the paradigm from time-series forecasting to point process modeling for traffic congestion. By leveraging neural flows and spatial graph context, it provides a more accurate and continuous representation of road network dynamics, offering a robust solution for predicting sparse traffic events.
Summary updated