Dynamic Multi-Valued Network Models for Predicting Face-to-Face Conversations

We introduce a new probabilistic framework for collectively modeling people’s social behavior from local sensor observations. Our approach extends curved exponential random graph models to (1) include features that account for multivalued edges, and (2) model the change in edge values over time. We present empirical results on a real world dataset of face-to-face conversations collected from 24 individuals using wearable sensors over the course of 9 months. The results demonstrate that the model is capable of predicting not just whether but also for how long two people will converse and that the ordinality of discretized observations can be exploited to reduce the number of parameters.