Interactive data-driven discovery of temporal behavior models from events in media streams

This paper investigates a technique for the discovery of temporal behavior models within multimedia event data. Advancements in both technology and the marketplace present us the opportunity for research in analysis of situated human behavior using video and other sensor data (media streams). By situated analysis, we mean the study of behavior in time as opposed to looking at behavior in the form of aggregated data divorced from how they occur in context. Human and social scientists seek to model behavior captured in media, and these data may be represented in a multi-dimensional event data space derived from media streams. The knowledge of these scientists (experts) is a valuable resource which can be leveraged to search this space. We propose a solution that incorporates the expert in an iteratively, interactive data-driven discovery process to evolve a desired behavior model. We test our solution's accuracy on a multimodal meeting corpus with a progressive three tiered approach.