Learning with Large - Scale Social Media Networks

Social media such as blogs, Facebook, Twitter, YouTube and Flickr enables people of all walks of life to express their thoughts, voice their opinions, and connect to each other more conveniently than ever. The boom of social media opens up a vast range of possibilities to study human interactions and collective behavior on an unprecedented scale. This dissertation presents a framework for learning with large-scale social media networks in order to understand human interactions and to predict collective behavior. Network interactions are typically heterogeneous, representing disparate relations, but most social media sites present only connections with no or limited relation information. Hence, social dimension is introduced to differentiate heterogeneous relations. A learning approach based on social dimensions is proposed, achieving substantial improvement over the state of the art. It is then extended to unify some unsupervised learning methods to handle networks with various types of entities and interactions. As social media networks are often of colossal size, an edge-clustering method is proposed to extract sparse social dimensions in order to address the scalability challenge. In sum, this research provides novel concepts and efficient algorithms to harness the power of social media networks, enables the integration of data in heterogeneous format and information from networks of multiple modes or dimensions, and offers a learning-based solution to social computing.