Spatio-Temporal Graph Neural Network for Urban Spaces: Interpolating Citywide Traffic Volume

Silke K. Kaiser, Filipe Rodrigues, Carlos Lima Azevedo, Lynn H. Kaack · 2025

Read Spatio-Temporal Graph Neural Network for Urban Spaces: Interpolating Citywide Traffic Volume by Silke K. Kaiser, Filipe Rodrigues, Carlos Lima Azevedo, Lynn H. Kaack (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-06292.txt.

Introduction

Silke Kaiser, Filipe Rodrigues, Carlos Lima Azevedo, and Lynn Kaack present a model in 2025 for filling gaps in city traffic counts. Sensors cover city streets sparsely, yet planners need volumes everywhere. Urban networks are messy, counts are spiky with many zeros, and past models leaned on highways.

Their graph network learns by hiding known counts and guessing them back. It reads each node's role and uses a loss tuned for zero-heavy data. Two open benchmarks, Berlin bike data and New York taxi data, test the method. Even with ninety-nine percent of sensors gone, error grew only modestly, beating recent rivals.

Catalog record

CREATORSilke K. Kaiser, Filipe Rodrigues, Carlos Lima Azevedo, Lynn H. Kaack
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE73 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-06292

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Spatio-Temporal Graph Neural Network for Urban Spaces: Interpolating Citywide Traffic Volume Silke K. Kaiser, Filipe Rodrigues, Carlos Lima Azevedo, Lynn H. Kaack Abstract Reliable street-level traffic volume data, covering multiple modes of transportation, helps urban planning by informing decisions on infrastructure improvements, traffic management, and public transportation. Yet, traffic sensors measuring traffic volume are typically scarcely located, due to their high deployment and maintenance costs. To address this, interpolation methods can estimate traffic volumes at unobserved locations using available data. Graph Neural Networks have shown strong performance in traffic volume forecasting, particularly on highways and major arterial networks. Applying them to urban settings, however, presents unique challenges: urban networks exhibit greater structural diversity, traffic volumes are highly overdispersed with many zeros, the best way to account for spatial dependencies remains unclear, and sensor coverage is often very sparse. We introduce the Graph Neural Network for Urban Interpolation (GNNUI), a novel urban traffic volume estimation approach. GNNUI employs a masking algorithm to learn interpolation, integrates node features to capture functional roles, and uses a loss function tailored to zero-inflated traffic distributions. In addition to the model, we introduce two new open, large-scale urban traffic volume benchmarks, covering different transportation modes: Strava cycling data from Berlin and New York City taxi data. GNNUI outperforms recent, some graph-based, interpolation methods across metrics (MAE, RMSE, true-zero rate, Kullback-Leibler divergence) and remains robust from 90% to 1% sensor coverage. On Strava, for instance, MAE rises only from 7.1 to 10.5, on Taxi from 23.0 to 40.4, demonstrating strong performance under extreme data scarcity, common in real-world urban settings. We also examine how graph connectivity choices influence model accuracy. 1 Introduction Reliable data on citywide traffic volumes is critical for transportation and urban planning decisions. Urban traffic data across modes can provide insights for targeted infrastructure improvements, traffic management, or improving the efficiency of public transportation ( Leduc 2008 ; Zheng et al. 2014 ) . Traffic volume data is collected by sparsely distributed traffic sensors that count the number of vehicles passing by. Extending this monitoring to all street segments (a street section between two intersections) is largely infeasible due to the high costs associated with deploying and maintaining such devices. We explore how traffic volumes in locations without sensors can be interpolated employing Graph Neural Network s ( GNN s). Urban traffic volume is spatio-temporal, following a network-based structure. GNN s have gained significant attention for traffic volume prediction due to their ability to model non-Euclidean spatial dependencies ( Jiang and Luo 2022 ) . When combined with temporal learning mechanisms, these models hold great promise for various urban applications ( Jin et al. 2024 ) . (a) Street network in Manhattan, NYC, comprising 8,156 street segments. The 21 traffic sensors are marked with dots. (b) Distribution of taxi trip volume per street segment in Manhattan, NYC. (c) Cycling network in Berlin, comprising 4,958 street segments. The 40 bicycle sensors are marked with dots. (d) Distribution of Strava cycling volume per street segment in Berlin. Figure 1: Urban street networks (left) and corresponding distributions of street-level traffic volume (right). It is important to note that the traffic and cyclist sensors count all passing motorized vehicles and cyclists, respectively—not just taxis or Strava users. Taxi and Strava data represent only subsets of total traffic. Nonetheless, we include the sensor locations to highlight the extent of sensor sparsity in urban settings. Previous work has shown that interpolation with GNN s performs well for highways and similar networks ( Yao and Huang 2023 ; Zheng et al. 2023 ; Wu et al. 2021a ; Wu et al. 2021b ; Appleby et al. 2020 ) . However, these networks differ profoundly from urban street networks, and correspondingly, the existing models overlook several critical aspects of urban contexts. Urban street networks exhibit greater infrastructural diversity than highways, making it essential to incorporate additional features that capture these differences. For example, the number of shops or the speed limit can influence traffic volumes. Moreover, the urban traffic data can be highly overdispersed, as in the case of specific transportation services and modes (e.g., cycling), with a significant proportion of zero values (Figure 1 ). Most existing models rely on the assumption of normally distributed data and adopt corresponding loss functions, which may not work for urban data. Furthermore, the higher density and structural complexity of urban networks complicate spatial dependency modeling. In poly-centric urban forms, for example, similar traffic patterns can emerge even between distant areas. Since GNN s are highly sensitive to graph construction, a systematic comparison of graph configurations is needed. Lastly, urban networks tend to exhibit far lower sensor density relative to the number of street segments when compared to highways. For instance, Manhattan’s street network comprises 8,156 segments, yet only 21 sensors recorded motorized traffic in 2023 ( NYC Open Data 2024 ) . Berlin’s cycling network spans 4,958 segments, but cyclist volume was collected by just 22 permanent and 18 temporary sensors ( Senate Department for the Environment, Mobility, Consumer and Climate Protection Berlin 2024b ) (see Figure 1 ). This is in stark contrast to highway networks, such as California’s highway network, commonly used through the PeMS benchmark dataset, with over 40,000 sensors ( State of California ) . To address these challenges, we propose Graph Neural Network for Urban Interpolation ( GNNUI ), a model that interpolates traffic volumes while accounting for urban-specific characteristics. It captures spatial dependencies through graph structure. To model temporal dependencies and enable effective interpolation, we adopt a masking algorithm and time window approach from the IGNNK model by Wu et al. 2021a . To better capture the heterogeneity of urban environments, GNNUI incorporates diverse street-level features, such as infrastructure, points of interest, and weather. To more accurately model urban traffic distributions, we introduce a zero-inflated negative binomial distribution ( Zhuang et al. 2022 ) . Since interpolation targets are known in advance, the full graph, including nodes with missing values, is included during training, allowing the model to leverage their features and context from the outset. We share the full implementation of our approach in the accompanying repository for transparency and reproducibility https://github.com/silkekaiser/GNNUI.git , and to support it, we assemble and publish two large-scale, street-level urban traffic datasets: Strava cycling data from Berlin and New York City taxi data, each enriched with auxiliary urban features and available for download ( Kaiser 2025 ) . In addition, we empirically address two further questions: What is the most effective graph representation for interpolation? and How well can traffic volumes be predicted under varying data scarcity? To answer the first, we compare different adjacency matrix configurations. For the second, we evaluate model performance across a range of sensor coverage levels, from sparse to near-complete. Most existing research focuses on settings with high sensor availability; however, real-world urban networks are characterized by extreme data sparsity, making performance under low coverage particularly important. In summary, our main contributions are the following: • We introduce GNNUI , a spatio-temporal GNN which interpolates citywide traffic volume while handling the unique challenges of urban contexts: infrastructural diversity through the inclusion of node features and handling of the skewed zero-inflated ground truth distribution through a zero-inflated negative

Read the complete text