PhD Proposal: Detecting Credible Events in Near Real Time from Social Media Streams

Given social media's explosive popularity and the rapidity with which we share content, numerous organizations have attempted to tap into its power and provide insight into compelling moments, events, and trends as they occur. Typical approaches for this identification are built around a human-generated list of seed keywords whose frequencies can be tracked to identify events of interest (e.g., tracking the keyword ``goal'' during the World Cup to detect when goals are scored). Though straightforward and capable, these mechanisms cannot detect events for which they lack defined keywords, so many interesting but unexpected events go undetected. Another similar method ingests content from a curated set of users whose information should be consistently trustworthy (the Associated Press's Twitter feed for example). Though such an approach avoids issues of missing unexpected events, it implicitly trusts any content published by these curated users, a dangerous assumption as many financial companies found out when the Associated Press's Twitter account was hacked. Additionally, relying on predefined keywords or curated users constrains these systems only to languages represented in the seed keyword or user sets, which is a significant issue for events of international interest like the World Cup or in areas that do not speak these languages.
My proposed research addresses these fundamental deficiencies in social media-based event detection by first relaxing requirements for prior event knowledge and instead using machine learning to discover interesting moments regardless of language. Results from my preliminary work demonstrate the feasibility of this approach by characterizing and detecting bursts in keyword usage. Though this early research focused primarily on language-agnostic discovery in sporting events, it also showed promising results in adapting this work to earthquake detection. My dissertation will extend this research by transferring my learned models to other types of high-impact events, exploring events with different temporal granularities, and finding methods to connect contextually related events into timelines. To ensure applicability of this research, I will also port these event discovery algorithms to stream processing platforms and evaluate their performance in the real-time context.
Secondly, if one is to make critical decisions based on this real-time information extracted from social media, one should have some degree of confidence in the credibility of that information. To address these issues, my dissertation will also include developing algorithms that integrate the vast array of signals in social media to evaluate information credibility in near real time. Such features include structural signatures of information dissemination, the location from which a social media message was posted relative to the location of the event it describes, and metadata from related multimedia (e.g., pictures and video) shared about the event. My preliminary work also suggests methods that could be applied to social networks for proactively stimulating trustworthy behavior and enhancing information quality.
Contributions from my dissertation will primarily be practical algorithms for discovering events from various social media streams and algorithms for evaluating and enhancing the credibility of these events in near real time. In addition, my research into methods for integrating streams and signals from multiple social networks will hopefully inform future research into the value of multimedia in understanding how events unfold. Finally, the stream processing prototypes I plan to develop should prove useful to journalists and first responders wishing to gain insight into events on the ground with confidence in the credibility of that information.
Examining Committee:
Committee Chair: - Dr. Jennifer Golbeck
Dept's Representative - Dr. Donald Perlis
Committee Member(s): - Dr. Jimmy Lin
- Dr. Hector Corrada Bravo