Temporally Coherent Role Topic Models (TCRTM): deinterlacing overlapping activity patterns

In many domains, individual atomic events can be organized into higher-level coherent activities. Often, these activities possess loose, but important, temporal structure. For example, the activity of prepar- ing a PowerPoint presentation may involve atomic events such as opening files, typing text, downloading images from the internet, saving files, and so on. The exact sequence of these atomic events may not be predictable, but they typically occur in close temporal proximity. Similar structure is present in other domains, such as financial services (e. g., credit card cus- tomers often have typical spending patterns, although the exact sequence of purchases in each pattern is unimportant) or healthcare (conditions such as whiplash injury may have a treatment protocol that includes pre- scription pain killers, physical therapy, and periodic check-ups, in some unspecified order). Modeling these activities is challenging, for several reasons. First, the structure is loose, without a rigid sequence of transitions. This makes traditional sequence modeling methods such as HMM difficult to apply. Second, multiple activities often overlap in time (for example, preparing a presentation may overlap with the OSs installing updates activity). Third, activities themselves possess structure; for example, a computer users set of activities is determined by the users job role. Finally, the sheer scale of the data (e. g., each workstation can generate tens of thousands of events per day, and large organizations may have tens of thousands of workstations) makes its interpretation difficult. In this paper, we present a probabilistic graphical model called TCRTM (Temporally Coherent Role-Topic Model) for analyzing loosely struc- tured activities. The proposed model automatically infers an appropriate set of roles and activities, and successfully addresses the challenges men- tioned above. In our experiments, TCRTM improves a perplexity score by a factor of five compared to using a standard model for analysis.