Sensing And Detecting Small-Scale Events Using Geosocial Media Data

Small-scale events involve interactive human movement in limited space and time. Social media platforms possibly generate large amount of geospatially-referenced information related to small-scale events. It benefits individuals, management departments, and urban systems if small-scale events can be timely detected from social media platforms, where measuring the abnormal patterns of human movement to discover events and analyzing associated texts to interpret the reasons behind abnormal movement are two keys. Through investigating how people move as different events occur and measuring the patterns on social media platforms, small-scale events can be generally classified into two types, namely type I events with abrupt patterns and type II events with random occurrence of key factors, where social events and traffic events are representative correspondingly.

Despite many studies have been conducted to detect social events and traffic events using geosocial media data, there still are some un-answered questions requiring further research. Most existing studies did not identify occurring events from a full coverage of spatial, temporal, and semantic perspectives. Studies concerning social event detection lack efficient semantic analysis summarizing event content to infer the reasons driving the abnormal movement. The typical classification-based method regarding traffic event detection lacks investigation on how the spatiotemporal distribution of traffic relevant posts associate with the occurring traffic events, and simply assigns the detected events with predefined categories, missing events that indicate traffic anomalies but go beyond the predetermined categories.

In this thesis, spatial-temporal-semantic approaches are proposed to measure spatiotemporal patterns of posts and users of social media platforms to capture abnormal human movement, and analyze the content of associated posts to mine the reasons driving the movement. A variety of techniques including machine learning, natural language processing, and spatiotemporal analysis are adopted to realize effective detection. Based on one-year Twitter data collected in Toronto, 2014 Toronto International Film Festival and traffic anomaly detection are selected as two case studies to evaluate the performance of proposed approaches. Through comparing with the ground truth data, the result reveals that more than 80% of the detected events do refer to real-world events, which illustrates the feasibility and efficiency of proposed approaches.

Keywords: Small-scale event, Event detection, Geosocial media data, Traffic event, Social event, Twitter, Spatiotemporal clustering