Research on Local-Global Spatio-Temporal Topic Model based on Social Media Text Data with Location Information

Abstract. Social media check-in data contains textual, temporal and spatial information, which is of great value for extracting topics and analyzing spatio-temporal changes to capture online public opinion trends. However, because existing topic models rely on predefined functions or distributions, they struggle to handle short text data sets and separate local and global topics. In this study, a Spatio-Temporal Biterm Topic Model (ST-BTM) is proposed, which integrates word pair modeling, spatio-temporal slicing and local-global topic extraction framework. ST-BTM uses spatio-temporal information from social media data to extract global topics while identifying local topics closely related to a specific region in local documents, thereby analyzing the spatio-temporal changes of topics in detail. Experiments on the Weibo COVID-19 topic check-in dataset show that ST-BTM model's UMass average consistency scores under different topic numbers are 50%, 19% and 8% higher than LDA, BTM and ToT models, respectively. Experiment show that this method captures cross-region local topics effectively, proves its ability to process short text data sets.