Investigating the link between road network and congestion for highly congested cities

Nishant Kumar, Yatao Zhang, Nina Wiedemann, Jimi Oke, Martin RaubalResearch Square 2025 · DOI

Abstract

Policy transfer is an efficient strategy for urban policymaking. When transferring policies for mitigating urban traffic congestion, the unique city-specific characteristics of urban traffic necessitate a thorough investigation into the relationship between the local road network and urban traffic congestion. We investigate the role of road network features in explaining the spatial variation of congestion in seven highly congested cities worldwide. The cities are segmented into uniform square-shaped urban tiles at different spatial scales, and the road network features of these tiles are used to predict the spatial variation of congestion. We study the feature importance at each scale and determine the direction of influence of these features, thereby characterising their role as alleviating or exacerbating features. For a given city, the features which retain their alleviation or exacerbation role across scales are deemed actionable. While some features were found to be actionable across several cities, significant variations, including role reversal, were found across cities, thus underscoring the need to leverage the context-specificity of urban traffic to tailor our congestion mitigation efforts.

Introduction

Urban planners frequently attempt "policy transfers," adapting successful congestion-mitigation strategies from one metropolis to another. However, these efforts often fall short because every city possesses a unique "physical DNA"—a distinct fingerprint of geography, infrastructure, and driving behavior. This research addresses a fundamental question: How do specific road network features explain the variation in traffic congestion within a city, and how do these relationships shift across different spatial scales and congestion types?

Methodology

To bridge the gap between microscopic traffic events and macroscopic urban planning, we developed a multi-scale analytical framework. We analyzed seven diverse global cities: Auckland, Bogota, Cape Town, Istanbul, Mexico City, Mumbai, and New York City.

The multi-scale framework for identifying actionable road network features.

The methodology involved five key steps:

  1. Spatial Tiling: Cities were segmented into uniform square tiles ranging from 0.25 km² to 6.25 km².
  2. Feature Extraction: Using OpenStreetMap (OSM), we extracted 14 road network features across five categories: Classic Graph Features, Intersection Counts, Intersection Types, Network Structure, and Demand Attraction.
  3. Machine Learning: Random Forest (RF) models were trained to predict Recurrent Congestion (RC) and Non-Recurrent Congestion (NRC).
  4. Explainable AI (SHAP): We utilized SHAP values and a new metric, the Sensitivity Ratio (SR), to classify features as either "alleviating" (reducing congestion) or "exacerbating" (increasing it).
  5. Actionability Filter: A feature is deemed "actionable" only if it maintains a consistent directional influence across all spatial scales and both congestion types.

Results and Discussion

Our analysis revealed that road network features affect traffic in highly city-specific ways. Surprisingly, there were minimal similarities in feature importance across the studied cities. For instance, "average circuity" (road curviness) acts as an alleviating feature in most cities but exacerbates traffic in Mumbai, highlighting why generic policy transfers often fail.

Feature importance and model performance across seven cities and multiple spatial scales.

Key findings include:

  • Scale Complexity: Model performance (R²) systematically decreases at finer spatial scales (below 1 km²). At the neighborhood level, human factors and random incidents likely dominate, making road-network-based predictions more challenging.
  • Universal Exacerbators: Across nearly all cities and scales, the number of traffic lights and total crossings were identified as primary exacerbators of recurrent congestion, confirming their role as systemic bottlenecks.
  • RC vs. NRC Drivers: Recurrent congestion is heavily driven by supply-side constraints like intersections, while non-recurrent congestion is more closely linked to demand-side factors like local centrality.
  • Predictive Performance: The models performed best for Auckland and New York City (mean R² ~0.4–0.5) but struggled with Mumbai (R² < 0.1), likely due to Mumbai's extreme density and unique geographic constraints.

Conclusion

The study’s most significant contribution is the "actionable feature" concept. By filtering for scale-invariance, we identify the "safe bets" for urban intervention. While supply-side interventions must be hyper-local and tailored to a city's specific road geometry, demand-side interventions share more similarities across cities, making them better candidates for broader policy frameworks. Future work will incorporate temporal dynamics and the impact of new mobility solutions like autonomous vehicles to further refine these urban traffic fingerprints.

Summary updated

Abstract

<title>Abstract</title> Policy transfer is an efficient strategy for urban policymaking. When transferring policies for mitigating urban traffic congestion, the unique city-specific characteristics of urban traffic necessitate a thorough investigation into the relationship between the local road network and urban traffic congestion. We investigate the role of road network features in explaining the spatial variation of congestion in seven highly congested cities worldwide. The cities are segmented into uniform square-shaped urban tiles at different spatial scales, and the road network features of these tiles are used to predict the spatial variation of congestion. We study the feature importance at each scale and determine the direction of influence of these features, thereby characterising their role as alleviating or exacerbating features. For a given city, the features which retain their alleviation or exacerbation role across scales are deemed actionable. While some features were found to be actionable across several cities, significant variations, including role reversal, were found across cities, thus underscoring the need to leverage the context-specificity of urban traffic to tailor our congestion mitigation efforts.
BibTeX
@inproceedings{kumar2025investigating,
  author = {Nishant Kumar and Yatao Zhang and Nina Wiedemann and Jimi Oke and Martin Raubal},
  title = {{Investigating the link between road network and congestion for highly congested cities}},
  booktitle = {Research Square},
  year = {2025},
  doi = {10.21203/rs.3.rs-4952650/v1},
}
Figure