Analyzing and modeling human behavioral dynamics in social networks
Loading...
Authors
Flamino, James, Rafael
Issue Date
2021-12
Type
Electronic thesis
Thesis
Thesis
Language
en_US
Keywords
Physics
Alternative Title
Abstract
Understanding, analyzing, and modeling human behavior within the context of social networks is a constantly ongoing study, being one of the central focuses of many scientific disciplines, include social and network science. Human behavioral dynamics are complex and multifaceted, but the advent of digital communications and online social media has made researching some of the driving mechanics and underlying trends more feasible. And while concepts such as value homophily, polarization, and relationship dynamics are integral to how people interact with each other, only recently have we begun to analyze, model, and forecast such behavior in a quantifiable way. And even with the advance of this kind of research, many aspects of human behavior are still not accurately modeled beyond simulations. In this thesis we seek to elucidate some behavioral dynamics through comprehensive analysis and modeling with empirical validation. In particular, we investigate the behavioral aspects mentioned above: value homophily, polarization, and relationship dynamics (in the context of tie strength analysis), with the addition of conversational dynamics. First, we analyze value homophily and how it drives the creation, evolution, and dissolution of groups. Using insights gained from a unique dataset, we formulate a measure of utility that individuals gain from groups based on their alignment with the opinions of other group members. We hypothesize that group membership changes (as well as opinion changes) are driven by a need to maximize this utility. We empirically validate this hypothesis by analyzing how many actual group membership changes occur within the data, and how many of them improve utility of the individuals making the changes, improve the utility of the changing individuals' like-minded peers, and improve the utility of all affected members. We then analyze how membership changes are affected by the popularity of the opinions held by those making the changes, and how utility gains differ as a consequence. Finally, we implement the mechanic of utility maximization in a predictive analytical model, showing how such a model can accurately forecast group membership and opinion changes. Next, we investigate political polarization on a comprehensive scale, using a massive longitudinal Twitter dataset covering the 2016 and 2020 U.S. presidential elections. Enhanced with media classifications allowing us to identify fake news, extremely biased news, and traditional news propagated within our data, we extract Twitter users that are ``super-spreaders'' of these types of content. We analyze these influencers, measuring their shifts in degree of influence and affiliations (or lack thereof) from 2016 to 2020. We then analyze how the user base of our Twitter data polarize themselves with respect to our media classifications, comparing this analysis between our two elections. We find that users have become increasingly polarized over time on this platform. We expand upon this result by also analyzing how influencers group together based on the similarity of the users that propagate their content to see if influencers inherently separate based on this basis, finding that, between the two elections, ever-tightening ``echo-chambers'' are formed around media classification, similarly increasing polarization. We then introduce a novel framework for continuously predicting tie strengths of dyadic connections over long periods of time. With a suite of analytical and machine learning models implemented within this framework for a pair of longitudinal datasets, we explore the upper bounds of prediction for both single-dataset predictions and cross-domain predictions. Then, taking advantage of the uniquely continuous nature of our generated tie strengths, we analyze these values over the length of our data, augmenting them with relationship classifications to investigate long-term relationship dynamics, observing strong trends that both reinforce, and expand upon, previous related literature. Finally, we design an ensemble of unique, text-agnostic measures to characterize conversational dynamics as they occur within forum-like online social media, exploring the relationship between conversational structure and content topics. We use a Support Vector Machine to accurately classify different content based on their genre using only our measures. We then cluster content using our measures, finding that the resultant groupings effectively delineate topical divisions inside genres, with finer clusters even capturing subtle semantic differences within singular topics. We find that the distance between clusters also correlates to difference in content, with farther clusters being less related in content than closer clusters. We conclude our cluster analysis by using an outlier detection algorithm to identify content that is, according to our measures, substantially different from the content of inliers. We investigate these identified outliers with their associated text and discuss the different aspects that contribute to their differences from the majority.
Description
December2021
School of Science
School of Science
Full Citation
Publisher
Rensselaer Polytechnic Institute, Troy, NY