Deep learning resilience inference for complex networked systems.

Area of Science:

• Complex network analysis
• Data-driven modeling
• Machine learning for systems science

Background:

• Resilience is vital for complex networked systems, enabling functionality during failures.
• Traditional analytical methods for resilience are limited by predefined equations and topology assumptions.
• Real-world systems require more adaptable and data-informed resilience assessment.

Purpose of the Study:

• To introduce ResInf, a deep learning framework for inferring system resilience directly from observational data.
• To overcome the limitations of analytical approaches in modeling real-world network dynamics and topologies.
• To provide accurate resilience inference and visualization capabilities.

Main Methods:

• Developed ResInf, a framework combining transformers and graph neural networks.
• Learned representations of node activity dynamics and network topology from data without simplifying assumptions.
• Evaluated performance against established analytical methods and dimension reduction techniques.

Main Results:

• ResInf significantly outperformed analytical methods, achieving up to 41.59% F1-score improvement.
• Demonstrated superior performance compared to spectral dimension reduction (14.32% F1-score improvement).
• Showcased generalization to unseen network topologies and dynamics, with robust performance under disturbances.

Conclusions:

• ResInf effectively addresses the gap in resilience inference for complex real-world systems.
• The framework offers a data-driven approach to network modeling, enhancing resilience assessment.
• Deep learning integration provides a powerful new perspective for understanding and predicting system resilience.