Reinforcement Learning to Model Churn Propagation Dynamics in Social Networks

Understanding how behavioral patterns propagate through social networks represents a fundamental process across many domains. Traditional approaches have demonstrated success in identifying individual-level outcomes by leveraging personal characteristics and historical behavior, but typically fail at revealing the dynamics of propagation, whereas propagation techniques do not fully integrate additional feature information. Graph Neural Networks have appeared as a powerful alternative, but also lack interpretability with respect to the dynamics of spread. In this paper, we propose an alternative approach to modeling social network propagation dynamics using reinforcement learning which identifies high-risk nodes but also reveals pathways of influence spread. We apply our methodology on an anonymized call detail record (CDR) graph from a Latin-American Telco provider and provide guidance on how to set up an appropriate environment for an RL agent to learn from and present comparative results across several RL algorithms. Our initial results show competitive predictive performance (AUC of 0.742), and are especially promising with regard to the interpretability aspects described above.