A sampling-guided unsupervised learning method to capture percolation in complex networks.

Area of Science:

• Complex networks analysis
• Machine learning applications
• Statistical physics

Background:

• Machine learning advances phase classification in physical systems.
• Extending ML to dynamical processes in complex networks is an active research area.
• Network percolation measures system resilience and spread dynamics but faces challenges with noisy data.

Purpose of the Study:

• To develop a robust method for identifying percolation clusters in networks with noisy data.
• To leverage network core-periphery structure for microscopic-scale analysis.
• To accurately detect phase transitions and critical points in time-varying networks.

Main Methods:

• Introduced a sampling approach based on onion decomposition (a refined k-core) to analyze network core-periphery structure.
• Utilized unsupervised clustering on selected node subsets to distinguish percolating from non-percolating phases.
• Employed the confusion scheme learning method to predict critical transition points from sampled data.

Main Results:

• The proposed method effectively distinguishes percolating phases from non-percolating ones, even with missing or noisy data.
• Accuracy in initial sampling is crucial for identifying information-rich nodes.
• The method successfully identified phase transitions in synthetic networks and real-world case studies (airport network, COVID-19 outbreaks).

Conclusions:

• The onion decomposition-based sampling strategy offers a robust solution for detecting phase transitions in empirical time-varying networks.
• This approach overcomes limitations posed by noisy or incomplete network data.
• The method has broad applicability in understanding resilience and spread dynamics across various complex systems.