Source-linked AI summary
Examining COVID-19 Forecasting using Spatio-Temporal Graph Neural Networks
Amol Kapoor, Xue Ben, Luyang Liu, Bryan Perozzi, Matt Barnes, Martin Blais, Shawn O'Banion
TL;DR
COVID-19 forecasting needs information about interactions among regions, not only local historical case sequences. This paper uses a spatio-temporal graph neural network with mobility data for US county-level forecasting, achieving better or comparable performance than baseline models while highlighting important data limitations.
Problem
Common forecasting approaches often assume relatively closed systems, motivating models that incorporate inter-regional interactions and mobility.
Method
The paper represents county mobility and temporal history in a spatio-temporal graph and applies a skip-connected graph neural network to forecast daily new cases.
Results
The GNN achieves either the best or second-best score on every evaluation metric, and additional mobility data improves results for all deep models.
Takeaways & Limitations
The framework can be extended to regression problems involving large-scale spatio-temporal disease reports and human mobility patterns.
Takeaways & Limitations
The mobility data may not represent the full population, is differentially private and aggregated, and omits factors such as mask wearing and weather.
Abstract
from arXiv · showhide
In this work, we examine a novel forecasting approach for COVID-19 case prediction that uses Graph Neural Networks and mobility data. In contrast to existing time series forecasting models, the proposed approach learns from a single large-scale spatio-temporal graph, where nodes represent the region-level human mobility, spatial edges represent the human mobility based inter-region connectivity, and temporal edges represent node features through time. We evaluate this approach on the US county level COVID-19 dataset, and demonstrate that the rich spatial and temporal information leveraged by the graph neural network allows the model to learn complex dynamics. We show a 6% reduction of RMSLE and an absolute Pearson Correlation improvement from 0.9978 to 0.998 compared to the best performing baseline models. This novel source of information combined with graph based deep learning approaches can be a powerful tool to understand the spread and evolution of COVID-19. We encourage others to further develop a novel modeling paradigm for infectious disease based on GNNs and high resolution mobility data.
1 INTRODUCTION
COVID-19 forecasting is important for managing a pandemic, but common modeling approaches often treat locations as relatively closed systems. The paper proposes combining mobility-based inter-region information with a spatio-temporal GNN for county-level case forecasting.
- Accurate caseload forecasts can inform healthcare-resource provisioning, outbreak control, and public safety during a pandemic.The paper situates forecasting within the substantial health and economic consequences of COVID-19.
- Existing epidemiological models commonly use either mechanistic models with predefined transmission dynamics or time-series learning methods applied to historical data.Examples include compartmental and agent-based models, curve fitting, autoregression, and deep learning.
- These approaches often assume that a location’s forecast depends mainly on its own information or observed patterns from other locations.
- The paper argues that inter-regional interactions, mobility, and similar epidemic patterns provide meaningful information for forecasting future regional cases.
- The proposed approach forecasts US county-level daily new cases with a spatio-temporal graph neural network trained using fine-grained mobility data.The authors also analyze mobility data as a signal for tracking disease spread.
2 BACKGROUND
Human mobility data can represent connections among regions that are difficult to capture with ordinary time-series models. The paper builds on epidemiological mobility research and applies graph neural networks to spatio-temporal relationships.
- Fine-grained mobility data can capture inter- and intra-region human activity relevant to disease-spread modeling and evaluating nonpharmaceutical interventions.
- Earlier work used international-flight data and SafeGraph mobility data to study COVID-19 spread and social-distancing effects.
- The paper examines two Google mobility datasets that provide a global and comprehensive view of inter- and intra-region human mobility.The datasets are described in more detail in the data section.
- Spatio-temporal graphs model connections between nodes as functions of both time and space, while GNNs have been applied to diverse graph-learning tasks.
3 METHOD
The method represents counties and their mobility relationships as a time-indexed graph, then combines temporal node features with spatial message passing. Skip-connections preserve seed-node information while a final MLP produces the forecast.
- Graph Neural Networks: GNN message passing propagates information from neighboring nodes and combines it with node transformations to update hidden representations.The computation first propagates messages and then aggregates them into updated node representations.
- COVID-19 Graph: The graph encodes spatial edges from normalized mobility flows and temporal edges connecting each node to observations from previous days.The graph contains 100 daily county-connectivity layers, with each node linked to the seven preceding temporal nodes.
- COVID-19 Graph: Figure 1 distinguishes within-day county connectivity from between-day temporal relationships across graph slices.In practice, every node has direct temporal edges to nodes in d previous days.
- Skip-Connections Model: The skip-connections model concatenates learned temporal embeddings with outputs from successive spatial graph-convolution layers.This design avoids diluting the self-node feature state while aggregating information across spatial hops.
- Skip-Connections Model: The initial embedding comes from an MLP over a node’s temporal features, and a final MLP over s spatial hops produces prediction P.Figure 2 depicts the seed-node embedding being concatenated and propagated through a 2-hop model.
4.1 Data
The study combines New York Times case data with two Google mobility datasets to represent county-level disease history and human movement. The mobility sources provide both flow quantities and category-specific mobility changes, but have important representativeness and coverage limits.
- The model uses New York Times COVID-19 data, Google Aggregated Mobility Research data, and Google Community Mobility Reports.Together, the mobility datasets augment the standard node features from the New York Times data.
- Node features include state, county, day, and normalized vectors of past cases and deaths extending back d days.
- Aggregated Mobility Research data provides weekly user flows at 5km2 resolution that can be aggregated into inter-county and intra-county flows.
- Community Mobility Reports provide county-level trends for grocery and pharmacy, parks, transit stations, workplaces, residential, and retail and recreation locations.Mobility values are normalized relative to a weekday baseline from January 3 to February 6, 2020.
- Limitations: Google mobility data covers opted-in smartphone users, may not represent the broader population consistently across locations, and is released only through differentially private aggregation.
- Limitations: The datasets omit factors such as mask wearing and weather, which can reduce transmission even when mobility is unchanged.The authors encourage future work incorporating these external features.
4.2 Hyperparameters, Architectures, and Splits
The experiments use fixed temporal and spatial horizons to forecast next-day case changes from county-level data, with mobility-flow figures illustrating the modeled connectivity.
- Hyperparameters and architectures: A 7-day time horizon and two spatial hops define the standard GNN experiments.Each hop uses the 32 neighbors with the highest edge weight.
- Hyperparameters and architectures: The two-hop spatial model uses MLPs around four hidden layers: embedding, two spatial aggregations, and prediction.
- Data split: Models predict the change in cases on day t + 1 from previous information, training on days 59–120 and testing on days 120–150.The split roughly corresponds to March–April for training and May 2020 for testing.
- Mobility connectivity: The mobility graph represents reduced inflow to King County from multiple US counties, including physically distant counties connected through its airport.
4.3 Baselines
The study compares the GNN with simple, statistical, and recurrent sequence baselines, with and without mobility information where applicable.
- Baseline design: Baseline comparisons test whether region-level mobility features improve forecasting beyond models that use no mobility information.“No Mob” variants omit mobility data.
- Simple baselines: Previous Delta predicts the next case-count delta equals the previous day’s delta, while Previous Cases predicts zero delta.
- Statistical baseline: ARIMA models daily new cases as a univariate time series using previous observations and estimation errors.Model order was selected with AIC and BIC, using a constant trend with ARIMA(7, 1.
- Sequence baselines: The LSTM baseline stacks two recurrent layers with 32 and 16 units, followed by a dense output layer.
- Sequence baselines: Seq2Seq uses a dense layer and GRU encoder-decoder architecture to learn from sequential input and produce encoded outputs.Bahdanau attention is applied during decoding according to the accompanying description.
- Training and inference: LSTM and Seq2Seq training uses Huber loss, Adam optimization, learning rate 0.02, dropout 0.2, and ten preceding days during inference.
4.4 Case Prediction Performance
The spatio-temporal GNN performs strongly across case and case-delta metrics, while mobility helps deep models but can hurt ARIMA under changing epidemic conditions.
- Evaluation metrics: Performance is evaluated with RMSLE and Pearson Correlation for predicted caseloads and with ΔRMSLE and ΔCorr for case deltas.The case-delta metrics address the high correlation induced by predicting case levels from prior-day cases.
- Overall performance: The GNN achieves the best or second-best score on every reported evaluation metric.
- Mobility effects: Adding mobility data improves results for all deep models but worsens performance for the ARIMA baseline.
- Mobility effects: ARIMA’s fixed linear mobility relationship may help early in the epidemic but underperform as mobility increases later in May.
5 CONCLUSION
The paper presents a spatio-temporal mobility GNN for COVID-19 forecasting and frames it as extensible to broader large-scale regression problems. It also identifies expanded features, horizons, and geographic settings as future work.
- Conclusion: The proposed framework combines graph neural networks with spatio-temporal mobility signals for COVID-19 forecasting.
- Conclusion: Unlike mechanistic or autoregressive approaches, the model can learn from inter-region interactions and region-level features without relying on assumed disease dynamics.
- Conclusion: The modeling framework can extend to regression problems involving large-scale spatio-temporal data across temporal and geographical scales.
- Future work: Future work includes adding features, expanding the prediction horizon, and testing epidemiological mobility data in other parts of the world.
6 APPENDIX
The study uses aggregated, anonymized mobility flows between regions and time intervals, with differential privacy protections applied before publication. The dataset is processed at population level rather than through inspection of individual user data.
- Google’s mobility dataset aggregates anonymized flows of people from region to region among users with Location History enabled.Location History is off by default, and the dataset resembles place-busyness information shown in Google Maps.
- Machine learning segments log data into semantic trips before trips are anonymized and aggregated with a differentially private mechanism.The resulting data contain heavily aggregated flows from large populations, and no individual user data were manually inspected.
- Each origin-destination-time tensor cell is represented as a trip count plus Laplacian noise, providing an (ϵ,δ)-differential privacy guarantee with ϵ = 0.66 and δ = 2.1 × 10−29.The noise is drawn from a zero-mean Laplace distribution.
- For each week and location pair, the process counts unique users, adds zero-mean Laplace noise, and removes noisy metrics below 100 users.The published dataset retains the remaining metrics after this filtering step.