Space-time dynamics of topics in streaming text

Abstract

Human-generated textual data streams from services such as Twitter increasingly become geo-referenced. The spatial resolution of their coverage improves quickly, making them a promising instrument for sensing various aspects of evolution and dynamics of social systems. This work explores spacetime structures of the topical content of short textual messages in a stream available from Twitter in Ireland. It uses a streaming Latent Dirichlet Allocation topic model trained with an incremental variational Bayes method. The posterior probabilities of the discovered topics are post-processed with a spatial kernel density and subjected to comparative analysis. The identified prevailing topics are often found to be spatially contiguous. We apply Markov-modulated non-homogeneous Poisson processes to quantify a proportion of novelty in the observed abnormal patterns. A combined use of these techniques allows for real-time analysis of the temporal evolution and spatial variability of population's response to various stimuli such as large scale sportive, political or cultural events.