In this paper we explore county-to-county interactions and migration flows in a human mobility network (i.e., SafeGraph), but rather than just looking at the flow of people as a proxy to predict cases we use scores from a feature-aware Weighted Preferential Attachment (WPA) model to predict
county-level COVID-19 cases for New York State utilizing various machine learning models (e.g., XGBoost (XGB), Random Forest (RF), Support Vector Regression
(SVR), and Ordinary Least Squares (OLS)). As such our study shows that a network-based feature like WPA outperformed a direct mobility metric like the exact number of people moving between counties in NYS based on the case study at hand.
If this sounds of interest, below you read the abstract to the paper, see our research pipeline and some of our findings. While at the bottom of the post you can find the full reference and a link to the paper. Finally, the the code to for the pipeline can be found at https://github.com/jhbuffalo2022/COVID_Network.Abstract: The COVID-19 pandemic was a global health crisis and, as such,
became an important research topic in epidemiology and population
health sciences. Significant research has been carried out related to
the prediction of COVID-19 cases and the prevention of its spread
utilizing many different methods, including both network analysis
and machine learning. However, combining results from network based
link prediction methods and machine learning-based models
to make COVID-19 case predictions has not been widely explored.
For example, no one has used Weighted Preferential Attachment
results, which provide pairwise scores (like the exact numbers of
people moving around) between nodes (locations in human mobility
networks), as direct inputs to machine learning models to make
predictions. Is this a valid approach? In this study, we used the
scores from a feature-aware Weighted Preferential Attachment
model as a set of indicators to predict county-level COVID-19 cases
for New York State with statistics and machine learning models.
Meanwhile, we fed the same group of models with another set
of indicators, the raw numbers (i.e., the exact numbers of people
moving between counties), and compared the results. As is the
norm when comparing two indicators, we tested them on a fixed
set of models. Results show that the former outperformed the latter
across different models in our spatiotemporal task, and as such,
using the former is a valid approach in this scenario. In general, our
work provides a framework and creates a novel pipeline to predict
new COVID-19 cases.
Keywords: COVID-19, Data Science, Geographic Information Systems, Human Mobility, Network
Science, Spatiotemporal Analysis, Weighted
Preferential Attachment.
 |
| Flowchart showing the entire workflow of this study |
 |
| A flow map showing the mobility patterns inside NYS in a single month |
 |
| A time series plot for Statewide COVID-19 Testing Data (cases vs time). The figure is a stacked bar plot, and different colors represent different counties. |
 |
| 2D Matrix-to-Vector prediction (overall cases) results based on average Mean Absolute Percentage Error (MAPE). |
Full Reference:
He, J. and Crooks A.T. (2026), Human Mobility Networks and COVID-19 Prediction: A Feature-aware Weighted Preferential Attachment approach and a Case Study for New York State, in Kim, J-S., Anderson, T., Su, Y.H., Hao Xue, H. and Züfle, A. (eds), Proceedings of 2nd ACM SIGSPATIAL International Workshop on Geospatial Computing for Public Health, Riverside, CA. pp 35-43. (pdf)
.png) |
| Infographic generated by Copilot |
.png)
