UMass Machine Learning and Friends Lunch | Main / Learning Curved Exponential Family Models To Infer Face-to-Face Interaction Networks From Situated Speech Data

Human behavior in the real world is a difficult thing to study: it is not possible to have human observers follow someone around all day, and surveys often tend to biased and unreliable. On the other hand, sensor data is easy to collect but inferring human behavior from this data is still a challenging problem. This talk will provide an overview of the probabilistic framework we have developed for inferring the micro-level dynamics of human interactions as well as the macro-level social network structure from local, noisy sensor observations. I will describe our extension to curved exponential random graph models and present empirical results on both synthetic data and a real world dataset of face-to-face conversations collected from 24 individuals using wearable sensors over the course of 6 months. By studying the micro and macro levels simultaneously we are able to explore the relationship between interaction dynamics (local behavior) and network prominence (a global property), and can identify the behavioral correlates of tie strengths within a network. We believe these methods have the potential to allow more quantitative inquiry into human behavior and social dynamics.