Connections Between Relational Event Model and Inverse Reinforcement Learning for Characterizing Group Interaction Sequences

In this article, we explore previously unidentified connections between
relational event model
(REM) from the field of network science and
inverse reinforcement learning
(IRL) from the field of machine learning with respect to their ability to characterize sequences of directed social interaction events in group settings. REM is a conventional approach to tackle such a problem, whereas the application of IRL is a largely unbeaten path. We begin by examining the mathematical components of both REM and IRL and find straightforward analogies between the two methods and the unique characteristics of the IRL approach. We demonstrate the special utility of IRL in characterizing group social interactions with two empirical experiments. In the first experiment, we use IRL to infer personal behavioral preferences based on a sequence of directed communication events from a multiple-team system, specifically a group of virtual-reality game players interacting and cooperating to accomplish a shared goal. In the second experiment, we draw evidence from a machine-coded international event database called the Kansas Event Data System (KEDS) and use IRL to infer the national behavioral inclinations of four major regional and global powers (Iran, Iraq, Russia, and the USA) underlying their interactions with one another during the 1990s Gulf area struggles. Our comparison and experiments introduce fresh perspectives for social behavior analytics and help inspire new research opportunities at the nexus of social network analysis and machine learning.