Physical-Virtual Collaboration Modeling for Intra- and Inter-Station Metro Ridership Prediction

Lingbo Liu, Jingwen Chen, Hefeng Wu, Jiajie Zhen, Guanbin Li, Liang LinIEEE Trans. ITS 2022 · DOI

TL;DR

This paper introduces PVCGN, a novel graph neural network model that combines physical metro network topology with virtual relationships derived from passenger flow data to predict station-level ridership. By leveraging both real-world connections and learned passenger behavior patterns, PVCGN significantly improves prediction accuracy, outperforming existing methods on large-scale datasets.

What Problem Does This Paper Address?

Accurately predicting metro ridership is crucial for efficient urban transportation management, but it's a challenging task. Traditional methods often overlook the complex, non-Euclidean structure of metro systems or rely solely on physical connections, failing to capture the nuanced evolution of passenger flow. This paper addresses the need for a more sophisticated model that can effectively learn from both the inherent structure of the metro network and the dynamic patterns of passenger movement.

Methods

The core of this work is the Physical-Virtual Collaboration Graph Network (PVCGN). It constructs three types of graphs to represent the metro system:

  • Physical Graph: Directly models the real-world connections between metro stations.
  • Similarity Graph: Captures passenger flow similarity between stations using Dynamic Time Warping (DTW) on historical flow series.
  • Correlation Graph: Models passenger flow correlation based on origin-destination (OD) distribution data.

These graphs are fed into a Graph Convolution Gated Recurrent Unit (GC-GRU) for learning spatial-temporal representations. Additionally, a Fully-Connected Gated Recurrent Unit (FC-GRU) is used to capture global evolution tendencies. The entire system is framed as a Seq2Seq model for sequential prediction.

Datasets

The authors introduce two large-scale datasets for their experiments:

  • SHMetro: Derived from Shanghai Metro transaction records, covering 288 stations and 958 physical edges over a specific period.
  • HZMetro: Based on Hangzhou Metro transaction records, featuring 80 operational stations and 248 physical edges.

Both datasets provide inflow and outflow data at 15-minute intervals.

Key Insights and Ideas

  • Beyond Physical Topology: The paper strongly advocates that relying solely on the physical metro network is insufficient for accurate ridership prediction. Virtual graphs derived from passenger behavior are essential for capturing complex dynamics.
  • Complementary Nature of Graphs: The combination of physical, similarity, and correlation graphs provides a more comprehensive understanding of the metro system than any single graph alone. The physical graph grounds the model in reality, while the virtual graphs capture emergent patterns.
  • Human Domain Knowledge is Key: The construction of similarity and correlation graphs leverages human domain knowledge (passenger flow patterns and OD distributions), demonstrating its value in enhancing machine learning models for real-world problems.
  • Universality of the Approach: The PVCGN framework is shown to be effective not only for station-level prediction but also for the more challenging task of online OD ridership prediction, highlighting its broad applicability.

Key Results

The proposed PVCGN consistently outperforms existing state-of-the-art methods across various metrics (RMSE, MAE, MAPE) on both the SHMetro and HZMetro datasets.

  • Overall Performance: PVCGN achieves significant improvements, reducing MAPE by at least 1% on SHMetro and outperforming methods like DCRNN and GCRNN by a large margin on HZMetro.
  • Rush Hour Prediction: During peak hours (7:30-9:30 and 17:30-19:30), PVCGN demonstrates superior accuracy, with MAPE improvements over DCRNN and GCRNN.
  • High-Ridership Stations: For stations with high passenger traffic, PVCGN consistently ranks first, showing its robustness and effectiveness in critical scenarios. For example, it achieves a MAPE of 10.62% for 15-minute predictions on SHMetro, outperforming others.
  • Component Analysis: Experiments show that incorporating both physical and virtual graphs (P+S+C-Net) yields the best results, confirming the effectiveness of the collaborative approach. Even without the physical graph, the virtual graphs (S+C-Net) provide competitive performance.

Limitations and Future Work

  • Incomplete OD Data: The paper acknowledges that in online metro systems, complete OD data is not immediately available. While PVCGN handles this by using incomplete OD vectors, future work could explore better ways to estimate or infer the missing OD information.
  • External Factors: The authors suggest incorporating external factors like weather and holiday events to further improve ridership prediction accuracy, as these can significantly influence passenger behavior.
  • Periodic Nature: The periodic nature of metro ridership (e.g., daily or weekly patterns) is mentioned as a factor that could be better leveraged for representation learning.
  • Data Requirements: The construction of similarity and correlation graphs relies on historical passenger flow and OD data, which may not always be readily available or as granular in all metro systems.

Summary updated

BibTeX
@article{liu2022physical,
  author = {Lingbo Liu and Jingwen Chen and Hefeng Wu and Jiajie Zhen and Guanbin Li and Liang Lin},
  title = {{Physical-Virtual Collaboration Modeling for Intra- and Inter-Station Metro Ridership Prediction}},
  journal = {IEEE Trans. Intell. Transp. Syst.},
  year = {2022},
  doi = {10.1109/tits.2020.3036057},
}
Figure