Modeling Dynamical Influence In Group Interaction

W. Pan, M. Cebrian, W. Dong, T. Kim, Alex (Sandy) Pentland


We present a new way for modeling 1) social influence and 2) the well-observed property of social influence – the influence strength between individuals changes over time (e.g., friendships break and reform). We show that our unsupervised generative switching Bayesian approach can simultaneously captures the system dynamics as the outcome of both (i) the influence between individuals (each modeled as an HMM), and (ii) the changes of influence itself using only individual observations. We describe here a variational Expectation-Maximization (EM) algorithm for inference. In our experiments, we illustrate applications of predicting turn taking by analyzing a real group discussion behavior dataset and understanding flu influence patterns between US states. Results demonstrate that our approach is a strong alternative for modeling complex interacting social systems.

Related Content