Machine Learning for Evolutionary Graph Theory

Abstract

The stability of communities - whether biological, social, economic, technological or ecological depends on the balance between cooperation and cheating. While cooperation strengthens communities, selfish individuals, or "cheaters," exploit collective benefits without contributing. If cheaters become too prevalent, they can trigger the collapse of cooperation and of the community, often in an abrupt manner. A key challenge is determining whether the risk of such a collapse can be detected in advance. To address this, we use a combination of evolutionary graph theory and machine learning to examine how one can predict the unravel of cooperation on complex networks. By introducing few cheaters into a structured population, we employ machine learning to detect and anticipate the spreading of cheaters and cooperation collapse. Using temporal and structural data, the presented results show that prediction accuracy improves with stronger selection strength and larger observation windows, with CNN-Seq-LSTM and Seq-LSTM best performing models. Moreover, the accuracy for the predictions depends crucially on the type of game played between cooperators and cheaters (i.e., accuracy improves when it is more advantageous to defect) and on the community structure. Overall, this work introduces a machine learning approach into detecting abrupt shifts in evolutionary graph theory and offer potential strategies for anticipating and preventing cooperation collapse in complex social networks.