Statistical traffic state analysis in large‐scale transportation networks using locality‐preserving non‐negative matrix factorisation

Yufei Han, Fabien Moutarde

Abstract

Traditional traffic management often focuses on link-level analysis, failing to capture the global dynamics of large-scale urban networks. This paper presents a methodology for network-level traffic state analysis using Locality-Preserving Non-negative Matrix Factorization (LPNMF). By reducing the dimensionality of traffic data from over 13,000 links to just 7 latent dimensions, the approach identifies typical spatio-temporal congestion patterns. Experimental results on simulated data for Paris demonstrate that LPNMF outperforms standard PCA by preserving the topological structure of traffic states, enabling more compact and meaningful clustering of global traffic dynamics.

Introduction: The Challenge of Scale

Modern urban traffic management systems frequently suffer from "tunnel vision," excelling at local link analysis but struggling to interpret the "big picture" of metropolitan congestion. In a complex network like Paris, which comprises over 13,000 interconnected links, the resulting data volume creates a "curse of dimensionality." Modeling every link simultaneously is computationally prohibitive and prone to noise. To address this, there is a critical need for dimensionality reduction techniques that can extract the "latent signatures" of city-wide traffic while preserving the global configuration and physical relationships between adjacent road segments.

Methodology: Locality-Preserving Non-negative Matrix Factorization

The proposed framework leverages Locality-Preserving Non-negative Matrix Factorization (LPNMF) to compress high-dimensional traffic states into a manageable subspace.

Methodological framework for network-level traffic state analysis using LPNMF and K-means clustering.

LPNMF and Graph Regularization

Unlike standard Principal Component Analysis (PCA), LPNMF enforces non-negativity constraints, which align with the physical reality of traffic indices. The core innovation lies in the integration of a Graph Laplacian penalty into the NMF objective function. This ensures that if two high-dimensional traffic states are similar, their low-dimensional projections remain close, thereby preserving the geometric manifold of the data.

Similarity and Clustering

A weighted fusion scheme is employed to calculate similarity between states, accounting for the correlation between a link and its immediate neighbors. Following dimensionality reduction, the K-means algorithm is applied to the low-dimensional projections to categorize typical spatial configurations of congestion and global traffic dynamics.

Experimental Setup

The methodology was validated using the IAURIF database, generated via the Metropolis traffic simulation software.

  • Network Scale: 13,627 links in Paris and its suburbs.
  • Data Volume: 146 simulated sequences of 8-hour morning peak periods.
  • Scenarios: Three demand settings were analyzed: Isotropic (uniform congestion), Anisotropic (localized bottlenecks), and Extreme (systemic gridlock due to accidents).

Key Insights and Results

Dimensionality and Reconstruction

Analysis of reconstruction errors revealed that the 13,627-dimensional state space could be effectively represented in just 7 dimensions. This massive compression significantly reduces the computational burden for subsequent analysis and forecasting.

LPNMF vs. PCA

While PCA focuses on maximizing variance, it often fails to distinguish the specific topological locations of congestion. LPNMF, by preserving local structures, successfully differentiates between spatially distinct but statistically similar congestion levels (e.g., congestion in the North vs. the Center). Furthermore, LPNMF produced more compact clusters with lower sample variance compared to PCA.

Identified Traffic States

The model successfully identified five distinct global traffic states:

  1. Free Flowing Cluster (FFC)
  2. Light Isotropic Congestion Cluster (LICC)
  3. Heavy Congestion Cluster (HCC)
  4. Peak Congestion Cluster (PCC)
  5. After-peak Congestion Cluster (APCC)

By tracking the temporal trajectories through these states, the authors could categorize the underlying demand settings. For instance, "Extreme" scenarios were characterized by trajectories that failed to return to the Free Flowing state within the observation window.

Conclusion and Future Directions

This research demonstrates that LPNMF is a powerful tool for distilling complex urban traffic data into actionable global patterns. Future work will focus on transitioning from offline batch training to on-line NMF schemes to handle real-time data streams. Additionally, the authors aim to utilize these low-dimensional projections as inputs for long-term forecasting models to predict systemic gridlock before it occurs.

Summary updated

Abstract

Statistical traffic data analysis is a hot topic in traffic management and control. In this field, current research progresses focus on analysing traffic flows of individual links or local regions in a transportation network. Less attention are paid to the global view of traffic states over the entire network, which is important for modelling large‐scale traffic scenes. Our aim is precisely to propose a new methodology for extracting spatiotemporal traffic patterns, ultimately for modelling large‐scale traffic dynamics, and long‐term traffic forecasting. The authors attack this issue by utilising locality‐preserving non‐negative matrix factorisation (LPNMF) to derive low‐dimensional representation of network‐level traffic states. Clustering is performed on the compact LPNMF projections to unveil typical spatial patterns and temporal dynamics of network‐level traffic states. The authors have tested the proposed method on simulated traffic data generated for a large‐scale road network, and reported experimental results validate the ability of our approach for extracting meaningful large‐scale space‐time traffic patterns. Furthermore, the derived clustering results provide an intuitive understanding of spatial‐temporal characteristics of traffic flows in the large‐scale network and a basis for potential long‐term forecasting.
BibTeX
@inproceedings{han2013statistical,
  author = {Yufei Han and Fabien Moutarde},
  title = {{Statistical traffic state analysis in large‐scale transportation networks using locality‐preserving non‐negative matrix factorisation}},
  booktitle = {IET},
  year = {2013},
  doi = {10.1049/iet-its.2011.0157},
}
Figure