Advances in Systems Science using Network Theory and Machine Learning

Systems science is widely used for population, public health, traffic, hazard, and other scientific research. New challenges have come up regarding access to big data as well a deeper consideration of systems complexity. The overarching objective of the research herein is to apply modeling and analytic tools to study complex systems, with an emphasis on network theory and machine learning. Specifically, we analyze systems via methods such as agent-based modeling, dimension reduction, classification, and Monte Carlo sampling. We begin by evaluating the effects of social networks in a migration setting. The application is on an NIH-funded study of the rural-to-urban mass migration in China since the 1980s, the largest human migration in history. We use a hierarchical social network by combing four layers of social networks. We study how endogenous social networks help explain accelerating trends in migration patterns. Then we develop this methodology further in an NSF-funded study of the societal impacts of repeated natural hazards. Here clustering methods and the Exponential Random Graph Model (ERGM) are applied to model social network structures that respond to hazardous events. We demonstrate how mortality and behavior change by considering the role of social capital in social networks. A mathematical technique for generating random social networks that satisfies a set of social network properties is introduced. In our next study, we explore novel applications of machine learning, including dimension reduction and classification techniques, in the NIH-funded work of Boston Birth Cohort, a longitudinal study of 8,509 mother-child dyads, to analyze high-dimensional metabolomics data. We compare the performance of several machine learning methods and provide guidance for public health researchers to systematically approach other similar problems. In the final application in this thesis, we develop classification models to identify critical surge-producing storms in the Mid-Atlantic region. Here we demonstrate how expert opinion and analytical forms of domain knowledge can be incorporated with machine learning methods to build an accurate storm surge prediction model