Inferencing User Characteristics and Behaviors from Social Media with Applications in Regulatory Science

Abstract

Regulatory science is the science of developing new tools, standards, and approaches to assess the safety, efficacy, quality and performance of regulated products. Social media data have been widely used in regulatory science studies, but the lack of demographics and health related behaviors in most social media platforms impairs the effective use of social media data for fine-grained analyses. In my dissertation, I develop machine learning models to inference user characteristics and behaviors from unstructured social media and apply them to regulatory science analyses. The first two essays of my dissertation are concerned with model development for user inference. In essay I, I aim to utilize social network structure, which is strongly associated with social media user demographics but not fully utilized in demographics prediction literature. I develop a supervised deep learning model to represent social networks and predict user profiles on social media. In essay II, I observe that consumer-generated content is valuable for relaxing the long-held assumption in consumer preference analysis that all consumers perceive all products in the same way. Then I report an unsupervised hierarchical Bayesian model to identify heterogeneous relations between consumers and products in online product reviews. I apply these inference models to regulatory science analyses in essay III and essay IV. In essay III, I identify the gender differences in e-cigarette use patterns and analyze the change of use patterns. I also explore how the gender differences change as new e-cigarette products are developed and regulatory rules are enacted. Essay IV aims to analyze the consumption behaviors of e-cigarettes in online reviews via market segmentation analysis. Correlated e-cigarette attributes and transition of consumer preferences are revealed. My dissertation develops novel machine learning algorithms for inferencing user characteristics and behaviors with good performance, and demonstrates their capabilities in comprehending user attitudes and consumption patterns on regulated products for regulatory policy making. Beyond the model development and regulatory science insights provided in the essays, this dissertation offers a systematic research paradigm of regulatory science informatics to guide regulatory science studies based on social media data.