Complex event analysis over social media

<p>Complex event analysis of social media has attracted much research interest due to its wide application in many domains, from crisis management and decision-making to estimating information propagation scope and emergency prevention. For example, in a natural disaster, complex events are communicated in real time on social media platforms such as Twitter. As the disaster develops, more and more users discuss facets of the disasters with positive or negative attitudes and then deliver their opinions to followers or friends on social media. Finally, many informative messages are posted that frame the disaster from various perspectives, such as government policy, rescue plans, and emergency situations. Obviously, for that kind of complex event it is important to detect its occurrence as soon as possible, to obtain comprehensive information about it, and to predict its scope. However, existing approaches to analyzing events in social media are not effective at handling complex events with complex contexts and evolutions over time. Furthermore, those approaches disregard the demand to summarized complex events in an integrated manner. Therefore, in this thesis three approaches are proposed to handle analyzing complex events in social media: detecting complex events with their evolvement over time, predicting event popularity from influential hashtags, and summarizing complex events in a comprehensive manner with multiple attributes.</p> <p>An event detection method that assesses retweeting behaviour to determine the event’s evolution is first proposed. Specifically, a topic model called RL-latent Dirichlet allocation (RL-LDA) is proposed to capture social media information by hashtag, location, text, and retweeting behaviour. Using RL-LDA, a complex event can be well managed by exploring the correlation between retweeting behaviour and the event. Then, to maintain the RL-LDA in a dynamic environment, a dynamic update algorithm is proposed that incrementally updates events over real-time streams. Then a novel hashtag-influence-based event popularity prediction model is proposed that mines the effect of an influential hashtag set on the event’s propagation. The model selects the influential hashtags from an event hashtag graph built by pairwise hashtag similarities and the topic distribution of event-related hashtags. A novel measurement is proposed to identify the hashtag influence of an event over its content and social impacts. A hashtag correlation-based algorithm is proposed to optimize the seed selection in a greedy manner. Then an event-fitting boosting model predicts the event popularity by embedding the feature importance of events into an XGBoost model. Moreover, an event-structure-based method is proposed that incrementally updates the prediction model over social streams. Finally, the first online complex social event summarization approach is proposed, namely SOMA, that simultaneously summarizes complex social events over multiple attributes including media content and contexts. Specifically, a deep learning model comprehensively summarizes events in regard to the text descriptions and locations that they appear in, by using their hidden connections in posts. Then a summary generator works over time, using text and location to achieve a maximal coverage of the summary over the original social event and a minimal redundancy of the summary. Furthermore, a location estimation method is used to address the location sparsity problem of complex events by mining the correlations between text and location. Extensive experiments were done to evaluate the effectiveness and efficiency of the proposed approaches in complex event detection with temporal evolution, event popularity prediction, and event summarization.</p>