Learning interpretable network dynamics via universal neural symbolic regression.

Area of Science:

• Complex Systems Science
• Computational Science
• Network Dynamics

Background:

• Discovering governing equations for complex network dynamics is crucial for understanding system evolution.
• Rich data necessitates advanced tools to uncover hidden patterns and mechanisms.
• Current methods may lack efficiency or accuracy in inferring symbolic equations.

Purpose of the Study:

• To develop a universal computational tool for automatic, efficient, and accurate learning of symbolic patterns in complex system dynamics.
• To combine deep learning's fitting capabilities with pre-trained symbolic regression's inference abilities.
• To provide a solution for uncovering hidden mechanisms in complex phenomena.

Main Methods:

• Developed a hybrid computational tool integrating deep learning and symbolic regression.
• Employed pre-trained symbolic regression for equation inference.
• Combined deep learning for data fitting and pattern recognition.

Main Results:

• Demonstrated remarkable effectiveness and efficiency across over ten diverse scenarios (physics, biochemistry, ecology, epidemiology).
• Outperformed state-of-the-art symbolic regression techniques in network dynamics.
• Validated practical applicability through real-world systems like global epidemic transmission and pedestrian movements.

Conclusions:

• The developed tool serves as a universal solution for deciphering complex phenomena.
• It advances scientific interpretability and facilitates new discoveries.
• The tool has significant implications for decision-making in various scientific domains.