Predicting Traffic Congestion Evolution: A Deep Meta Learning Approach

Yidan Sun, Guiyuan Jiang, Siew Kei Lam, Peilan He

Abstract

Traffic congestion is a dynamic phenomenon that propagates and decays across urban road networks. Traditional traffic prediction models often fail to capture the complex spatiotemporal dependencies and the asymmetric nature of congestion flow. This paper introduces the Deep Meta Learning Model (DMLM), a novel framework designed to predict the evolution of traffic congestion. By leveraging Dynamic Attributed Networks (DANs) and meta-learning techniques, DMLM adaptively incorporates both static road attributes and dynamic environmental factors. Experimental results on real-world data from Singapore demonstrate that DMLM significantly outperforms state-of-the-art baselines in predicting both congestion propagation and decay.

Introduction

Predicting urban traffic is not merely about identifying current bottlenecks; it requires understanding how congestion evolves over time and space. When one road segment becomes congested, will the traffic spill over to adjacent roads (propagation), or will the jam dissipate (decay)? Addressing this "evolution" is critical for effective traffic management and navigation.

Existing approaches typically rely on pattern mining from historical data. However, these methods face significant challenges:

  1. Spatiotemporal Complexity: Traffic patterns vary wildly across different times and locations.
  2. Asymmetric Transitivity: Traffic flow is directional; congestion moving from Road A to Road B does not imply the same pattern in reverse.
  3. Contextual Dynamics: Static features like schools or stadiums have time-varying impacts on traffic (e.g., school hours vs. midnight). Traditional models often treat these as static or ignore them entirely.

The Deep Meta Learning Model (DMLM)

The proposed DMLM framework addresses these challenges by modeling the road network as a Dynamic Attributed Network (DAN). In this network, nodes represent road segments with both static attributes (e.g., length, lanes, nearby POIs) and dynamic attributes (e.g., speed, weather, time).

The DMLM framework overview, showing the construction of the attributed graph and the core model components.

The architecture consists of several innovative components:

  • Asymmetric Initial Representations: Using Non-negative Matrix Factorization (NMF), the model generates distinct "source" and "target" representations for each road segment to capture directional traffic flow.
  • Meta-LSTM Module: Unlike standard LSTMs with fixed weights, the Meta-LSTM uses small "meta-learner" networks to generate weights based on the specific attributes of a road segment. This allows the model to adapt its predictive logic to the unique "personality" of different locations.
  • Meta-Attention Mechanism: This module dynamically weighs different temporal scales (e.g., recent vs. weekly patterns) based on the current context, ensuring the most relevant historical data informs the prediction.
Detailed architecture of the Meta-LSTM and Meta-Attention components, illustrating how meta-learners generate weights from road attributes.

Experimental Evaluation

The DMLM was evaluated using a comprehensive dataset from Singapore, including GPS-based bus trajectories, road network data from OpenStreetMap, and POI data from government sources. The dataset covered 1,858 road segments over four months.

The model demonstrated superior performance across two primary tasks:

  • Propagation Prediction: DMLM achieved 86-88% accuracy, a substantial improvement over traditional mining methods like STC (~66%).
  • Decay Prediction: The model reached even higher accuracy levels of 89-94%.

Ablation studies confirmed that the "meta" components are essential; removing them led to significant performance degradation, highlighting the importance of context-aware weight generation.

Conclusion and Future Outlook

The DMLM represents a significant step forward in traffic forecasting by treating static attributes as dynamic influences through meta-learning. It successfully captures the asymmetric and context-dependent nature of traffic evolution.

Future work will focus on addressing potential biases in bus-derived traffic data, reducing the computational overhead of meta-learning architectures, and scaling the system to handle the complexity of larger mega-cities with tens of thousands of road segments.

Summary updated

Abstract

Many efforts are devoted to predicting congestion evolution using propagation patterns that are mined from historical traffic data. However, the prediction quality is limited to the intrinsic properties that are present in the mined patterns. In addition, these mined patterns frequently fail to sufficiently capture many realistic characteristics of true congestion evolution (e.g., asymmetric transitivity, local proximity). In this paper, we propose a representation learning framework to characterize and predict congestion evolution between any pair of road segments (connected via single or multiple paths). Specifically, we build dynamic attributed networks (DAN) to incorporate both dynamic and static impact factors while preserving dynamic topological structures. We propose a Deep Meta Learning Model (DMLM) for learning representations of road segments which support accurate prediction of congestion evolution. DMLM relies on matrix factorization techniques and meta-LSTM modules to exploit temporal correlations at multiple scales, and employ meta-Attention modules to merge heterogeneous features while learning the time-varying impacts of both dynamic and static features. Compared to all state-of-the-art methods, our framework achieves significantly better prediction performance on two congestion evolution behaviors (propagation and decay) when evaluated using real-world dataset.
BibTeX
@inproceedings{sun2021predicting,
  author = {Yidan Sun and Guiyuan Jiang and Siew Kei Lam and Peilan He},
  title = {{Predicting Traffic Congestion Evolution: A Deep Meta Learning Approach}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.24963/ijcai.2021/417},
}
Figure