Abstract
One of the Greater London Authority's (GLA) response to the COVID-19 pandemic brings together multiple large-scale and heterogeneous datasets capturing mobility, transportation and traffic activity over the city of London to better understand 'busyness' and enable targeted interventions and effective policy-making. As part of Project Odysseus we describe an early-warning system and introduce an expectation-based scan statistic for networks to help the GLA and Transport for London, understand the extent to which populations are following government COVID-19 guidelines. We explicitly treat the case of geographically fixed time-series data located on a (road) network and primarily focus on monitoring the dynamics across large regions of the capital. Additionally, we also focus on the detection and reporting of significant spatio-temporal regions. Our approach is extending the Network Based Scan Statistic (NBSS) by making it expectation-based (EBP) and by using stochastic processes for time-series forecasting, which enables us to quantify metric uncertainty in both the EBP and NBSS frameworks. We introduce a variant of the metric used in the EBP model which focuses on identifying space-time regions in which activity is quieter than expected.
Introduction
During the COVID-19 pandemic, policymakers required precise insights into population adherence to lockdown guidelines and the dynamics of activity recovery. Traditional "busyness" monitoring relies on spatial scan statistics to identify anomalous clusters. However, standard methods typically employ a "planar" approach, treating urban environments as flat 2D planes. This is often a poor fit for urban mobility, which is constrained by road networks rather than moving "as the crow flies." A grid-based scan might identify a busy area but fails to pinpoint specific arterial roads.
To address this, researchers from The Alan Turing Institute and the University of Warwick developed an early-warning system that moves analysis from a 2D grid onto the actual graph of city streets. This system enables the identification of regions where traffic is significantly higher or lower than expected with high spatial resolution.
Methodology
The core of the system is the Expectation-Based Network Scan Statistic (NET). The workflow comprises three primary stages:
- Forecasting: To establish a baseline for "normal" activity, the authors utilize Gaussian Processes (GP). GPs model daily and weekly periodicity while quantifying uncertainty, providing a more robust baseline than simpler methods like Holt-Winters.
- Network Search Windows: Instead of rectangular windows, the algorithm searches over overlapping paths on the road network $\mathcal{G} = (V, E)$. This respects the actual geometry of the city, allowing for more localized detection.
- Likelihood Metrics: An Expectation-Based Poisson (EBP) statistic compares actual counts against the GP-forecasted baseline. A novel ASYM metric was introduced to detect both "hotspots" (excess activity) and "coldspots" (activity deficits), the latter being critical for monitoring lockdown compliance.
The framework was evaluated using semi-synthetic vehicle count data from approximately 900 road sensors in Westminster, London, with simulated activity surges injected to test sensitivity and accuracy.
Key Insights and Results
The transition from planar to network-based scanning represents a shift from a "blunt instrument" to a "scalpel." The NET scan provides significantly higher spatial precision. When an anomaly is flagged, the system identifies specific street segments rather than general neighborhoods.
Precision-Recall Trade-off
Evaluation revealed a distinct trade-off: while the NET scan is far more precise, it generally has lower recall than planar scans. A broad 2D box is more likely to capture the entirety of a regional surge, whereas a single network path might only capture a portion. However, the precision of the Network scan type significantly outperforms the Planar approach following an outbreak.
Significance and Uncertainty
The NET scan consistently yielded higher significance scores (Corrected EBP scores) compared to planar scans, indicating greater confidence in identified true anomalies. Furthermore, the use of Gaussian Processes allowed the incorporation of "metric uncertainty," making the system more robust against data noise and reducing false positives.
Conclusion and Future Work
While the NET-GP approach offers superior precision and robustness, it introduces higher computational complexity. Generating baselines with GPs takes significantly longer than with simpler methods like Holt-Winters. Future work aims to expand this framework to other UK cities and incorporate additional transport modes, such as cycling and walking, to provide a comprehensive view of urban mobility. The team also plans to explore more complex spatio-temporal point-process models to further refine detection capabilities.
Summary updated