TrafficStream: A Streaming Traffic Flow Forecasting Framework Based on Graph Neural Networks and Continual Learning

Xu Chen, Junshan Wang, Kunqing XieIJCAI 2021 · DOI

TL;DR

This paper introduces TrafficStream, a novel framework for traffic flow forecasting that addresses the challenge of evolving traffic networks and flow patterns over long periods. By combining Graph Neural Networks (GNNs) with Continual Learning, TrafficStream effectively integrates new traffic patterns while retaining historical knowledge, leading to accurate and efficient predictions in dynamic environments.

What Problem Does This Paper Address?

Accurate traffic flow forecasting is crucial for efficient transportation systems. However, existing methods often struggle with the dynamic nature of real-world traffic networks. These networks are not static; they expand with new roads and sensors, and traffic flow patterns themselves evolve over time due to changing human behavior and urban development. Traditional approaches that assume static network structures or require complete retraining for each new period are computationally expensive and fail to adapt to these long-term changes. This paper tackles the problem of efficiently learning and adapting forecasting models in such "streaming" traffic networks where both the network topology and traffic patterns are in constant flux.

Methods

TrafficStream is built upon two core ideas: new traffic pattern fusion and historical knowledge consolidation, both implemented within a GNN framework enhanced by Continual Learning.

  • New Traffic Pattern Fusion:
    • Expanding Network Training: To handle network expansion, TrafficStream constructs sub-graphs that include newly added nodes and their neighbors within a 2-hop radius. This approach reduces computational complexity compared to retraining on the entire expanded graph.
  • Evolution Pattern Detection: To identify evolving traffic patterns in existing nodes, the paper proposes a method using Jensen-Shannon Divergence (JSD) to measure the similarity between historical and current traffic flow distributions. Nodes with high JSD scores are flagged as having significantly changed patterns, and their data is used to update the model.
  • Historical Knowledge Consolidation: To prevent "catastrophic forgetting" (where learning new information erases old knowledge), TrafficStream employs two strategies:
  • Information Replay: Stable historical traffic data from nodes with low JSD scores (indicating consistent patterns) is replayed during training to reinforce learned knowledge.
  • Parameter Smoothing: Elastic Weight Consolidation (EWC) is used to impose smoothing constraints on model parameters. This technique selectively restricts changes to parameters deemed important for preserving historical knowledge, while allowing more flexibility for parameters that need to adapt to new patterns.

Datasets

The paper introduces and uses the PEMS3-Stream dataset. This dataset is derived from the California Transportation Agencies' Performance Measurement System (PEMS) and contains real-time traffic flow data collected every 30 seconds, aggregated into 5-minute intervals. The dataset spans from 2011 to 2017, capturing the expansion of the traffic network over these years. The authors specifically select data from July 10th to August 9th each year to ensure a consistent temporal window for analysis.

Key Insights and Ideas

  • The "Streaming" Nature of Traffic Networks is Key: The paper highlights that real-world traffic forecasting is not a static problem. Networks grow, and patterns change, necessitating models that can adapt continuously rather than relying on periodic retraining.
  • Continual Learning is a Natural Fit: The principles of Continual Learning are well-suited for this problem, enabling models to learn sequentially from new data without forgetting past knowledge.
  • JSD for Pattern Evolution: Using Jensen-Shannon Divergence to quantify changes in traffic patterns is an elegant way to identify nodes that require special attention for model updates.
  • Balancing New Learning and Knowledge Retention: The core challenge is to learn new patterns effectively without losing previously acquired knowledge. TrafficStream's dual approach of pattern fusion and knowledge consolidation directly addresses this trade-off.
  • Efficiency Matters for Real-World Deployment: The paper emphasizes that while accuracy is paramount, computational efficiency is equally critical for practical applications, especially with large-scale, continuously updating datasets.

Key Results

  • Effectiveness: TrafficStream consistently outperforms traditional online training methods (Static-STModel, Expansible-STModel) in terms of prediction accuracy (MAE, RMSE). It achieves performance comparable to the upper bound of retraining models annually (Retrained-STModel, STGCN, STSGCN).
  • Efficiency: TrafficStream demonstrates significantly lower total training times and per-epoch training times compared to models that require full retraining each year. This efficiency makes it more suitable for real-world deployment in dynamic environments.
  • Ablation Studies:
    • Pattern Fusion: Both integrating new node patterns and detecting evolving patterns in existing nodes are crucial. TrafficStream's comprehensive fusion approach yields the best results.
  • Knowledge Consolidation: Both information replay and parameter smoothing contribute to preventing knowledge forgetting. Their combined use further enhances predictive accuracy, with data replay showing a slightly better individual impact.

Limitations and Future Work

  • Limitations:
    • The paper focuses on traffic flow forecasting. While the framework is general, its direct application to other traffic-related predictions (e.g., speed, occupancy) might require adjustments.
  • The choice of the top 5% of nodes with the highest JSD for pattern evolution detection is a hyperparameter that might need tuning for different datasets or scenarios.
  • Future Work:
    • The authors suggest exploring more sophisticated methods for pattern fusion and knowledge consolidation.
  • Investigating the framework's applicability to other types of dynamic graph learning tasks beyond traffic forecasting is a potential avenue.
  • Further research into optimizing the computational efficiency and scalability of the framework for even larger and more complex networks could be beneficial.

Summary updated

BibTeX
@inproceedings{chen2021trafficstream,
  author = {Xu Chen and Junshan Wang and Kunqing Xie},
  title = {{TrafficStream: A Streaming Traffic Flow Forecasting Framework Based on Graph Neural Networks and Continual Learning}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.24963/ijcai.2021/498},
}
Figure