SensorTree: Bursty Propagation Trees as Sensors for Protest Event Detection

Abstract

Protest event detection is an important task with numerous benefits to many organisations, emergency services, and other stakeholders. Existing research has presented myriad approaches relying on tweet corpus to solve the event detection problem, with notable improvements over time. Despite the plethora of research on event detection, the use of the implicit social links among users in online communities for event detection is rarely observed. In this work, we propose SensorTree, a novel event detection framework that utilizes the network structural connections among users in a community for protest event detection. SensorTree tracks information propagating among communities of Twitter users as propagation trees to detect bursts based on the sudden changes in size of these communities. Once a burst is identified, SensorTree uses a latent event topic model to extract topics from the corpus over the burst period to describe the event that triggered the burst. Extensive experiments performed on real-world Twitter datasets using qualitative and quantitative evaluations show the superiority of SensorTree over existing state-of-the-art methods. We present case studies to further show that SensorTree detects events with fine granularity descriptions.