Abstract Search

Primary Submission Category: Methodological approaches to studying public health

Human Mobility and COVID-19: a Link Prediction-based approach and a Case Study for New York State

Authors:  Jie He,

Presenting Author: Jie He*

Human mobility plays a vital role in spreading infectious diseases. Researchers used human mobility data to investigate disease transmission patterns. Lessani showed that there was a strong association between human mobility and spread [1]. Although researchers have already tried integrating network analysis and machine learning in this research area, utilizing link prediction methods on human mobility data for spread prediction has not been fully explored. In this study, I focused on county-to-county interactions and migration flows in the human mobility network (SafeGraph Neighborhood Patterns dataset). I used link prediction scores from modified Weighted Preferential Attachment as indicators to predict county-level COVID-19 cases for New York State (statewide testing data from the government of the state of New York for overall cases) with statistics and machine learning. This has not been done by previous research. I performed two different prediction tasks: 1) using one 62×1 vector involving scores of 62 counties from a single month to predict another 62×1 vector involving COVID-19 cases from the next month; 2) using fifteen 62×62 matrices involving scores representing county-to-county connections from fifteen months (from March 2020 to May 2021) to predict fifteen 62×1 vectors involving COVID-19 cases from the corresponding next months. In the first task, I ran models like XGBoost, Random Forest, Support Vector Regression, and Ordinary Least Squares with three different types of inputs: Personalized PageRank scores (PPR), link prediction scores (aggregated), and the exact numbers of people moving between counties (aggregated). Though all three inputs were strongly positively correlated with the overall cases based on Spearman’s rank correlation, the results showed that link prediction scores and the exact numbers performed similarly and both outperformed PPR according to the best outputs of each. Results were compared based on R-squared and Mean Absolute Percentage Error with 5-fold Cross-Validation. In the second task (spatiotemporal), I applied models like Graph Convolutional Network, hybrid model with CNN and LSTM, Random Forest, Gaussian Process, and Vector Autoregression with two different types of inputs: link prediction scores and the exact numbers of people. The results showed that the former performed better than the latter across different models. Results were compared based on R-squared and Mean Absolute Percentage Error with Walk-Forward Cross-Validation. The other part of this study is to assess Machine Learning-based variant prediction models for New York State (NYS). This has not been fully tested by previous research. I used the testing data from GISAID database for variant-specific cases and a modified network which combined 5 counties into New York City. I employed various types of Zero-Inflated models (Logistic Regression + XGB; Logistic Regression + Random Forest; Logistic Regression + SVR; Logistic Regression + Linear Regression; Zero-Inflated Negative Binomial Regression) for prediction tasks and compared results based on Mean Absolute Error with 5-fold Cross-Validation. In general, my work showed that link prediction scores can be used as indicators for COVID-19 prediction and performed both overall and variant-specific predictions with NYS data. My work used network science and machine learning methods to solve real-world problems in public health. 

[1] M Naser Lessani, Zhenlong Li, Fengrui Jing, Shan Qiao, Jiajia Zhang, Bankole Olatosi, and Xiaoming Li. Human mobility and the infectious disease transmission: a systematic review. Geo-Spatial Information Science, 27(6):1824–1851, 2024.