Opinion Mining in Online Social Networks: Patterns, Influences and Anomalies.

摘要

Online social networks have influenced many aspects of people's lifestyle. By participating in online discussions, people exchange ideas, influence group interactional norms and gradually form their own characteristics over time. The popularity of online social networks not only extensively motivates the exchange of opinions and ideas across boarders, but also provides an ideal medium for researchers to understand how opinions emerge, diffuse and influence each other in online environment.;In this dissertation, first, we propose a new algorithm to classify user comments into different opinions. Instead of only focusing on the content, we leverage user behavior information and build graph models to do opinion classification. In addition, we implement this opinion classification algorithm into our SINCERE system as a real-time service. Based on this opinion classification tool, we analyze how others' feedback with different opinions influence continued user participation in online social networks.;Secondly, we develop a general method to identify the factors that influence the formation of user opinions. In this direction, we first propose a dynamic graph model for recognizing user characteristics automatically in online social network. Then we analyze the influence of early discussion context on the formation of user opinions and build a supervised learning model to predict user characteristics. Our results show that having only the first three months of users' interaction data generates an F1 accuracy level of around 70% in predicting user deliberation and bias in online newsgroups.;Finally, we propose a novel framework to detect anomalous user behavior and spam messages in online social networks. In order to reduce the amount of false positive results, we track the structural change and group dynamics of user opinion communities by leveraging community-centric models. All experimental results on both synthetic and real-world datasets show the effectiveness and consistency of our framework in detecting anomalies with reduced false alarm rate.