Spectral Learning Algorithm Reveals Propagation Capability of Complex Networks.

Area of Science:

• Network Science
• Data Mining
• Complex Systems Analysis

Background:

• Predicting node propagation capability is a significant challenge in network science and data mining.
• Existing unsupervised learning algorithms like PageRank (PR) and LeaderRank (LR) have limitations, particularly in degree-uncorrelated networks where they correlate with in-degree, failing to accurately assess propagation potential.

Purpose of the Study:

• To address the limitations of existing algorithms in predicting node propagation capability.
• To introduce a novel iterative algorithm, SpectralRank (SR), and its weighted variant for more accurate prediction.

Main Methods:

• Proposed SpectralRank (SR): An iterative algorithm assuming node propagation capability is proportional to neighbors after adding a ground node.
• Introduced Weighted SR: An extension incorporating a priori node information.
• Established a probabilistic framework as theoretical support.
• Conducted simulations using the susceptible-infected-removed (SIR) model on 32 diverse networks (directed, undirected, binary).

Main Results:

• SpectralRank (SR) and its weighted variant (SR-family) demonstrate superior performance over PageRank (PR) and LeaderRank (LR) in predicting node propagation capability.
• Simulations on 32 networks confirmed the advantages of SR-family methods.
• SR-family methods outperformed 11 other well-known algorithms in predictive accuracy.

Conclusions:

• The proposed SpectralRank algorithms offer a more accurate approach to predicting node propagation capability in complex networks.
• These findings provide valuable insights for controlling spreading behaviors and have significant implications for network management and intervention strategies.