Gradual Network Sparsification and Georeferencing for Location-Aware Event Detection in Microblogging Services

Event detection in microblogging services such as Twitter has become a challenging research topic within the fields of social network analysis and natural language processing. Many works focus on the identification of general events with event types ranging from political news and soccer games to entertainment. However, in application contexts like crisis management, traffic planning, or monitoring people's mobility during pandemic scenarios, there is a high need for detecting localisable physical events. To address this need, this paper introduces an extension of an existing event detection framework by combining machine learning-based geo-localisation of tweets and network analysis to reveal events from Twitter distributed in time and space. Gradual network sparsification is introduced to improve the detection events of different granularity and to derive a hierarchical event structure. Results show that the proposed method is able to detect meaningful events including their geo-locations. This constitutes a step towards using social media data to inform, for example, traffic demand models, inform about infection risks in certain places, or the identification of points of interest.