Learning Human Interactions with the Influence Model

We are interested in quantitatively modeling the interactions between humans in conversational settings. While a variety of models are potentially appropriate, such as the coupled HMM, all require a very large number of parameters to describe the interactions between chains. We propose as an alternative the generative model developed in [1], the Influence Model, which parametrizes the hidden state transition probabilities by taking a convex combination of the pairwise transitions with constant “influence” parameters. We develop a learning algorithm for this model and show its abilities to model chain dependencies in comparison to other standard models using synthetic data. We also show early results of applying this model to human interaction data.