[Paper Review] Spatio-Temporal Point Processes with Attention for Traffic Congestion Event Modeling

Shixiang Zhu, Ruyi Ding+3 more|arXiv (Cornell University)|May 15, 2020

Traffic Prediction and Management Techniques43 references4 citations

TL;DR

This paper proposes a spatio-temporal point process model with attention mechanisms and a tail-up spatial correlation structure to model traffic congestion events using multi-modal data from traffic sensors and police reports. By leveraging adaptive attention for non-monotonic temporal dependence and directional road network structure, the model outperforms state-of-the-art methods in likelihood and prediction accuracy on real and synthetic data.

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.

Motivation & Objective

• To model traffic congestion events that result from both self-excitation (propagation of existing congestion) and exogenous triggering (traffic incidents reported by police).
• To address the limitations of traditional Hawkes processes that assume monotonic decay of influence over time and space, which does not reflect real traffic dynamics such as prolonged blockages.
• To incorporate the directional spatial correlation inherent in road networks, where upstream congestion affects downstream flow, using a tailored 'tail-up' spatial model.
• To develop a flexible, interpretable attention mechanism that learns non-homogeneous spatio-temporal dependencies between past and current congestion events.
• To improve prediction accuracy and likelihood of point process models by combining multi-modal data sources and advanced neural network-based attention.

Proposed method

• The model uses a neural network-based attention mechanism to learn dynamic, non-monotonic influence scores between past and current congestion events, replacing fixed parametric decay functions.
• An adaptive score function is introduced to measure the importance of past events based on non-homogeneous spatio-temporal distance in traffic networks, enhancing interpretability.
• The 'tail-up' model from spatial statistics is adapted to represent directional spatial dependence in road networks, capturing flow-dependent correlations between sensors.
• The framework integrates two data streams: traffic sensor counts (for congestion events) and police reports (for incident-triggered congestion), modeling both endogenous and exogenous effects.
• The attention mechanism is embedded within a point process framework, allowing the intensity of future events to depend on weighted historical events with learned attention weights.
• The model is trained end-to-end using likelihood maximization, enabling joint estimation of temporal dynamics and spatial structure.

Experimental results

Research questions

• RQ1 Can an attention-based mechanism effectively model non-monotonic temporal influence in traffic congestion events, particularly when congestion effects persist or re-emerge after delays?
• RQ2 How can the directional spatial structure of road networks be formally incorporated into a spatio-temporal point process to improve modeling of congestion propagation?
• RQ3 To what extent does combining traffic sensor data with police incident reports improve the predictive performance of congestion event models?
• RQ4 Can the attention scores learned by the model provide interpretable insights into the causal relationships between congestion events and incidents?
• RQ5 How does the proposed model compare to state-of-the-art methods in terms of likelihood and prediction accuracy on real-world traffic data?

Key findings

• The proposed attention-based spatio-temporal point process (APP) model achieves higher likelihood and prediction accuracy than state-of-the-art methods on both synthetic and real traffic data.
• The attention mechanism successfully captures non-monotonic influence patterns, such as delayed or sustained congestion effects after incidents, which standard models with exponential decay fail to represent.
• The learned attention scores reveal interpretable patterns, such as distinct 'network communities' corresponding to northbound and southbound highway directions, with low cross-directional influence.
• The tail-up spatial model effectively captures directional spatial correlations, showing stronger sensor correlations along the same traffic flow direction and dynamic changes during rush hours versus midnight.
• The covariance matrices learned by the tail-up model reveal that highway connectors (C1–C4) play a key role in structuring spatial dependence, especially during peak hours.
• The model demonstrates robustness and flexibility by outperforming baseline models across multiple evaluation metrics and data configurations.