Spatio-Temporal Point Processes with Attention for Traffic Congestion Event Modeling

Shixiang Zhu, Ruyi Ding, Minghe Zhang, Pascal Van Hentenryck, Yao Xie

TL;DR

This paper introduces a novel spatio-temporal point process model that uses an attention mechanism to predict traffic congestion events. By integrating traffic sensor data with police incident reports, the model captures both the self-exciting nature of congestion and the external triggers from accidents, outperforming existing methods in prediction accuracy.

What Problem Does This Paper Address?

Predicting traffic congestion is crucial for modern transportation systems, but it's a complex challenge. Existing models often struggle to capture the intricate spatio-temporal dynamics of congestion. Specifically, they fail to adequately model two key aspects:

  1. Non-monotonic influence: The impact of a past event (like an accident) on future congestion might not simply decay over time; it can persist or even increase for a period.
  2. Directional spatial correlation: Traffic congestion propagates along road networks in a directional manner. Congestion upstream can affect downstream traffic, and sensors on the same road segment in the same direction are more correlated than those in different directions or on unconnected roads.

Furthermore, traditional models often rely on parametric assumptions for event influence, which may not be flexible enough to capture real-world traffic dynamics. The paper aims to address these limitations by developing a more expressive and flexible model.

Methods

The core of the proposed solution is a Spatio-Temporal Point Process (STPP) enhanced with a novel attention-based mechanism.

  • Attention Mechanism: Borrowing from natural language processing, an attention mechanism is used to model the non-linear and long-term dependencies between past traffic congestion events and the current event. This allows the model to "focus" on past events that are most relevant to predicting future congestion, moving beyond simple temporal decay assumptions.
  • Multi-Head Attention: To capture more complex triggering effects, the model employs a multi-head attention mechanism.
  • Adaptive Score Function: A new score function is introduced to measure the importance between past and current events. This function is learned by a neural network and adapts to non-homogeneous spatio-temporal distances in traffic networks, going beyond simple dot-product scores.
  • Tail-Up Model for Spatial Correlation: To account for the directional spatial dependence induced by road networks, the paper adapts the "tail-up" model from spatial statistics. This model uses "stream distance" along the road network rather than Euclidean distance and imposes directional correlations.
  • Exogenous Promotion: Police reports of traffic incidents (like accidents) are incorporated as an exogenous factor that can trigger congestion. The model uses a parameter $\gamma$ to control the decay rate of this influence.
  • Online Algorithm: To handle streaming data efficiently, an online algorithm is proposed that retains only the most informative past events for attention computation, ensuring constant memory usage.

The conditional intensity function of the point process is modeled as a sum of three components: a background intensity ($\mu_0$), exogenous promotion from police incidents ($\mu_1$), and endogenous self-excitation from past congestion events ($\lambda'$).

Datasets

The model is evaluated on the Atlanta traffic dataset, which comprises:

  1. Traffic Congestion Events: Collected from traffic sensors (14 sensors on I-75 and I-85 highways) recording traffic counts. Congestion events are detected when traffic counts exceed a threshold.
  2. 911 Calls-for-Service: Used to extract traffic incidents reported by police. This data includes call time, location, and processing time, serving as exogenous triggers.
  3. Atlanta Traffic Network: Road network information extracted from OpenStreetMap to model spatial dependencies.

The paper also uses synthetic datasets generated from four different point process models (Hawkes, self-correcting, and two non-homogeneous Poisson processes) to validate the model's ability to capture various temporal patterns.

Key Insights and Ideas

  • Dual Triggering Effects: Traffic congestion is not just self-exciting; it's also significantly influenced by external events like traffic accidents. Modeling both simultaneously is crucial for accurate prediction.
  • Attention for Non-Monotonicity: The attention mechanism is a powerful tool for capturing complex temporal dependencies that don't follow simple decay patterns, which is common in traffic phenomena.
  • Directional Spatial Modeling: The "tail-up" model, adapted for stream distance, effectively captures how congestion propagates along specific road directions, a key aspect often missed by models using simple Euclidean distance.
  • Interpretability of Spatial Correlation: The learned spatial correlation matrices from the tail-up model reveal "network communities" and directional influences, providing insights into how different parts of the road network interact during congestion.
  • Efficiency for Streaming Data: The online attention mechanism is vital for real-world applications where data arrives continuously, allowing for efficient processing without retaining all historical data.

Key Results

  • Superior Performance: The proposed Attention-based Point Process (APP) model consistently outperforms baseline methods (LSTM, NHP, RMTPP, SAHP, Hawkes Process) on both synthetic and real Atlanta traffic data.
  • Higher Log-Likelihood: The APP model achieves significantly higher average maximum log-likelihood values, indicating a better fit to the data. For the Atlanta traffic dataset, the APP+Tailup model achieved a max log-likelihood of 458.5 (time & space) compared to others like RMTPP at 490.1 (time only) and APP+Euclidean at 570.7 (time & space).
  • Improved Prediction Accuracy: The model demonstrates higher location prediction accuracy (37.6% for APP+Tailup) and lower time prediction error (3.7 MSE for APP+Tailup) compared to baselines.
  • Validation of Components: Experiments show that incorporating both the tail-up spatial correlation and the exogenous promotion from police incidents significantly improves performance. Removing these components (e.g., APP-μ₁+Tailup) leads to a drop in performance.
  • Qualitative Insights: Visualizations of conditional intensities and spatial correlations provide qualitative evidence that the model captures realistic traffic dynamics, including rush hour patterns and the impact of network structure.

Limitations and Future Work

  • Computational Cost: While the online attention mechanism improves efficiency, the full APP model can still be computationally intensive, especially for very large datasets or real-time applications requiring extremely low latency.
  • Data Granularity: The current model focuses on discrete congestion events. Future work could explore modeling continuous traffic flow metrics or finer-grained event types.
  • Generalizability: While tested on Atlanta data, further validation on diverse urban traffic networks would be beneficial to assess the model's generalizability.
  • Integration of More Data Sources: The authors suggest that incorporating additional data sources, such as weather information or public transit schedules, could further enhance prediction accuracy.
  • Real-time Deployment: The paper mentions the need for efficient online algorithms, hinting at future work focused on robust real-time deployment and operationalization of the model.

Summary updated

Abstract

We present a novel framework for modeling traffic congestion events over road networks. Using multi-modal data by combining count data from traffic sensors with police reports that report traffic incidents, we aim to capture two types of triggering effect for congestion events. Current traffic congestion at one location may cause future congestion over the road network, and traffic incidents may cause spread traffic congestion. To model the non-homogeneous temporal dependence of the event on the past, we use a novel attention-based mechanism based on neural networks embedding for point processes. To incorporate the directional spatial dependence induced by the road network, we adapt the “tail-up” model from the context of spatial statistics to the traffic network setting. We demonstrate our approach’s superior performance compared to the state-of-the-art methods for both synthetic and real data.
BibTeX
@article{zhu2020spatio,
  author = {Shixiang Zhu and Ruyi Ding and Minghe Zhang and Pascal Van Hentenryck and Yao Xie},
  title = {{Spatio-Temporal Point Processes with Attention for Traffic Congestion Event Modeling}},
  journal = {IEEE Trans. Intell. Transp. Syst.},
  year = {2020},
  doi = {10.1109/tits.2021.3068139},
}
Figure