Event-based social media data mining

color:windowtext;mso-ansi-language:EN-US">Recent years have witnessed the rapid growth of social media collections available over the Internet. The exponential growth of social media data requires scalable, effective and robust technologies to manage and index them. Event is one of the most important cues to recall people's past memory. The reminder value of event makes it extremely helpful in organizing data. With the development of Web 2.0, many event-based information sharing sites are appearing online, and a wide variety of events are scheduled and described by several social online services. The study of the relation between social media and events could leverage the event domain knowledge and ontologies to formulate the raised problems, and it could also exploit multimodal features to mine the patterns deeply, hence gain better performance compared with some other methods. In this thesis, we study the problem of mining relations between events and social media data. There are mainly three problems that are well investigated. The first problem is event enrichment, in which we investigate how to leverage the social media to events illustration. The second problem is event discovery, which focuses on discovering event patterns from social media stream. We propose burst detection and topic model based methods to find events from the spatial and temporal labeled social media. The third problem is visual event modeling, which studies the problem of automatically collecting training samples to model the visualization of events. The solution of collecting both of the positive and negative samples is also derived from the analysis of social media context. Thanks to the approaches proposed in this thesis, the intrinsic relationship between social media and events are deeply investigated, which provides a way to explore and organize online medias effectively.