Machine Learning for Evolutionary Graph Theory

Guoli Yang, Matteo Cavaliere, Mingtao Zhang, Giovanni Masala, Adam Miles, Mengzhu WangarXiv 2025 · DOI · arXiv

Introduction

In biological, social, and economic systems, cooperation serves as the fundamental "glue" that maintains community stability. However, these systems are perpetually vulnerable to "defectors"—individuals who exploit group benefits without contributing. A rapid spread of defectors can lead to a catastrophic tipping point, resulting in the total collapse of cooperation. Predicting these transitions before they occur is a critical challenge. This research proposes a generalized, automated methodology using deep learning to identify early warning signals (EWS) that distinguish between eventual community recovery and total collapse.

Methodology: Predicting the Fate of Cooperation

The study utilizes a "Death-Birth" process on graphs to simulate evolutionary dynamics. Agents engage in a Prisoner’s Dilemma game, where their payoffs determine their "fitness" and likelihood of reproduction. To predict the long-term outcome from early-stage data, the researchers developed a pipeline that analyzes a specific "observation window" ($ws$) of the network's evolution.

The proposed framework uses an observation window of early network dynamics to predict whether cooperation will recover or collapse.

The models are trained on five key temporal features: the number of cooperators (#C), the number of defectors (#D), and the counts of three edge types (C-C, C-D, and D-D). These features capture both the population state and the underlying structural connectivity of the network over time.

Model Architectures

The researchers evaluated five deep learning architectures, ranging from standard LSTMs to Transformers. The most successful models were those capable of capturing both local structural features and long-term temporal dependencies.

Architectural diagrams for (a) Seq-LSTM and (b) CNN-Seq-LSTM, the two top-performing models in the study.

The CNN-Seq-LSTM architecture proved particularly robust. By utilizing a 1D-Convolutional layer to aggregate local features before passing them to LSTM layers, the model effectively identifies the "rhythm" of evolutionary change. This hybrid approach consistently outperformed more traditional architectures across various experimental scenarios.

Key Findings and Results

The models were tested across various network topologies, including Random, Small-world, and Scale-free networks. Several critical insights emerged:

  • High Predictive Accuracy: Under strong selection pressure and with a sufficient observation window, the best models achieved accuracy scores exceeding 0.90.
  • The Role of Network Heterogeneity: Scale-free networks, characterized by highly connected "hubs," demonstrated greater resilience to defector invasion. Interestingly, the ML models also performed more accurately on these heterogeneous networks, suggesting they produce clearer early-stage signals.
  • Selection Strength and Signal Clarity: While stronger selection pressure makes populations more prone to collapse, it also generates "sharper" signals that are easier for machine learning models to classify, provided the observation window is long enough (e.g., $ws > 30$).
  • Temptation as a Clarifier: Increasing the "temptation to defect" ($T$) in the Prisoner's Dilemma game actually improved prediction accuracy, as high-stakes environments produce more distinct structural patterns early on.

Conclusion

This work demonstrates that deep learning, specifically hybrid CNN-LSTM architectures, can effectively provide early warning signals for the collapse of cooperation in networked populations. While the current study relies on synthetic data, it establishes a powerful framework for automated tipping-point prediction. Future research will focus on scaling these models to larger, dynamic networks and testing their generalizability across different real-world social and biological datasets where the underlying rules may be noisier and more complex.

Summary updated

BibTeX
@misc{yang2025machine,
  author = {Guoli Yang and Matteo Cavaliere and Mingtao Zhang and Giovanni Masala and Adam Miles and Mengzhu Wang},
  title = {{Machine Learning for Evolutionary Graph Theory}},
  year = {2025},
  eprint = {2507.08363},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.08363},
}
Figure