Machine learning dynamical phase transitions in complex networks.

Area of Science:

• Complex Systems Science
• Machine Learning Applications
• Network Science

Background:

• Growing interest in machine learning for predicting phase transitions in physical systems.
• Existing machine learning applications on complex networks primarily focus on structural properties, not dynamical processes.
• Detecting phase transitions and critical points in dynamical processes on complex networks remains an open challenge.

Purpose of the Study:

• To develop a general machine learning framework for detecting phase transitions and identifying critical points in dynamical processes on complex networks.
• To address the limitations of existing methods, particularly on heterogeneous networks.
• To create a robust, efficient, and universally applicable framework for complex network analysis.

Main Methods:

• Development of a hybrid framework combining supervised and unsupervised learning.
• Incorporation of proper training data set sampling strategies, including hub-node and k-core based methods.
• Utilizing epidemic spreading dynamics on complex networks as a test case.

Main Results:

• The proposed framework successfully detects phase transitions and critical points on homogeneous networks.
• Performance on heterogeneous networks is improved by implementing specific data sampling techniques.
• The framework demonstrates robustness, computational efficiency, and broad applicability across various network types.

Conclusions:

• The developed machine learning framework offers a comprehensive solution for phase transition detection in complex dynamical systems.
• Data sampling is crucial for achieving high performance on heterogeneous networks.
• This work paves the way for leveraging machine learning in understanding, predicting, and controlling complex dynamical systems.