Temporally Coherent Role-Topic Models (TCRTM): Deinterlacing Overlapping Activity Patterns

Abstract

The Temporally Coherent Role-Topic Model (TCRTM) is a probabilistic graphical model for analyzing overlapping, loosely temporally structured activities in heterogeneous populations. Such structure appears in many domains where activities have temporal coherence, but no strong ordering. For instance, editing a PowerPoint presentation may involve opening files, typing text, and downloading images. These events occur together in time, but without fixed ordering or duration. Further, several different activities may overlap – the user might check email while editing the presentation. Finally, the user population has subgroups; for example, managers, salespeople and engineers have different activity distributions. TCRTM automatically infers an appropriate set of roles and activity types, and segments users’ event streams into high-level activity instance descriptions. On two real-world datasets involving computer user monitoring and debit card transactions we show that TCRTM extracts semantically meaningful structure and improves hold-out perplexity score by a factor of five compared to standard models.