Prediction of social dilemmas in networked populations via graph neural networks

Abstract

Human behavior presents significant challenges for data-driven approaches and machine learning, particularly in modeling the emergent and complex dynamics observed in social dilemmas. These challenges complicate the accurate prediction of strategic decision-making in structured populations, which is crucial for advancing our understanding of collective behavior. In this work, we introduce a novel approach to predicting high-dimensional collective behavior in structured populations engaged in social dilemmas. We propose a new feature extraction methodology, Topological Marginal Information Feature Extraction (TMIFE), which captures agent-level information over time. Leveraging TMIFE, we employ a graph neural network to encode networked dynamics and predict evolutionary outcomes under various social dilemma scenarios. Our approach is validated through numerical simulations and transfer learning, demonstrating its robustness and predictive accuracy. Furthermore, results from a Prisoner's Dilemma experiment involving human participants confirm that our method reliably predicts the macroscopic fraction of cooperation. These findings underscore the complexity of predicting high-dimensional behavior in structured populations and highlight the potential of graph-based machine learning techniques for this task.