Transfer cross entropy for fast sociometric inference in longitudinal collections of multi-party conversation

A means is proposed of extracting participant-pair interaction measures from a binary representation of behavior in multi-party conversation, by leveraging an extension of transfer entropy. The technique allows for the inexpensive construction of sociomatrices, requiring only a minimum of detection technology. It is expected that the method will tractably enable the application of social network analysis to conversational behavior mined from very large collections of unannotated audio.