Spotting Icebergs by the Tips: Rumor and Persuasion Campaign Detection in Social Media

[摘要] Identifying different types of events in social media, i.e., collective online activities or posts, is critical for researchers who study data mining and online communication. However, the online activities of more than one billion social media users from around the world constitute an ocean of data that is hard to study and understand. In this dissertation, we study the problem of event detection with a focus on two important applications---rumor and persuasion campaign detection.Detecting events such as rumors and persuasion campaigns is particularly important for social media users and researchers. Events in social media spread and influence people much more quickly than traditional news media reporting. Viral spreading of specific events, such as rumors and persuasion campaigns, can cause substantial damage in online communities. Automatic detection of these can benefit analysts in many different research domains.In this thesis, we extend the existing research on social media event detection of online events such as rumors and persuasion campaigns. We conducted content analysis and found that the emergence and spreading of certain types of online events often result in similar user reactions. For example, some users will react to the spreading of a rumor by questioning its truth, even though most posts will not explicitly question it. These explicit questions serve as signals for detecting the underlying events. Our approach to detecting a given type of event first identifies the signals from the myriad of posts in the data corpus. We then use these signals to find the rest of the targeted events. Different types of events have different signals. As case studies, we analyze and identify the signals for rumors and persuasion campaigns, and we apply our proposed framework to detect these two types of events.We began by analyzing large-scale online activities in order to understand the relation between events and their signals. We focused on detecting and analyzing users;; question-asking activities. We found that many social media users react to popular and fast-emerging memes by explicitly asking questions. Compared to other user activities, these questions are more likely to be correlated to bursty events and emergent information needs.We use some of our findings to detect trending rumors. We find that in the case of rumors, a common reaction regardless of the content of the rumor is to question the truth of the statement. We use these questioning activities as signals for detecting rumors. Our experimental results show that our rumor detector can effectively and efficiently detect social media rumors at an early stage. As in the case of rumors, the emergence and spreading of persuasion campaigns can result in similar reactions from the online audience. However, the explicit signals for detecting persuasion campaigns are not clearly understood and are difficult to label. We propose an algorithm that automatically learns these signals from data, by maximizing an objective that considers their key properties. We then use the learned signals in our proposed framework for detecting persuasion campaigns in social media. In our evaluation, we find that the learned signals can improve the performance of persuasion campaign detection compared to frameworks that use signals generated by alternative methods as well as those that do not use signals.

[发布机构] University of Michigan

[效力级别] Persuasion Campaign Detection

[关键词] Rumor Detection;Persuasion Campaign Detection;Social Media Search and Mining;Computer Science;Engineering;Computer Science & Engineering