MACHINE LEARNING OVER USER-GENERATED CONTENT: FROM UNSUPERVISED USER BEHAVIORAL MODELS TO EMOTION RECOGNITION VIA DEEP LEARNING
Date
2021
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
We are now a machine-mediated society, with virtually every transaction and decisionmediated, analyzed, and even recommended by algorithms running in the background. An increasing percentage of our interactions and societal discourse now happen over social media platforms. While this has increased our connectivity, transcending geopolitical borders, it has also provided us with unique problems that we would not have anticipated even a decade ago. The propagation of misinformation on social networks is now a societal problem and is prompting an increasing body of research into how to identify misinformation, how to identify spreaders of such information and, perhaps more importantly, how to design mitigation and intervention strategies. In this dissertation we show that advancing research to address these challenges requires foundational research into automated user behavioral profiles. We present an unsupervised learning algorithm that leverages user-generated con- tent to understand and categorize user behavior in near-real time. This line of work rests fundamentally on the ability of machines to understand language. This, in itself, is a multi- faceted challenge that has spawned entire domains in computer science, such as natural language processing. To advance the ability of machines to understand content, we focus on another fundamental problem, emotion recognition from short text in platforms, such as Twitter and Reddit. We significantly advance this line of research by going beyond binary sentiments and present sophisticated deep neural network-based models that can capture fine-grained emotions in the presence of a host of challenges, including data imbalance and noise due to human annotations. Finally, we ask a fundamental question of whether text is sufficient to understand emotions and show that supplementing text with new modes of interaction, such as emojis and emoticons, advances the ability of machines to disambiguate emotions, much like humans. We believe that the research presented in this dissertation lays the groundwork for further advancing machine understanding of human behavior in social media platforms.