Robust label propagation on multiple networks

Area of Science:

• Graph-based machine learning
• Network analysis
• Data mining

Background:

• Transductive inference and label propagation on graphs are crucial for many machine learning tasks.
• Integrating information from multiple networks presents a significant challenge in graph analysis.
• Existing methods may not effectively handle irrelevant or noisy network data.

Purpose of the Study:

• To develop a novel algorithm for label propagation across multiple networks.
• To automatically integrate structural information from diverse networks.
• To improve robustness and efficiency in graph-based inference tasks.

Main Methods:

• A new algorithm for multi-network label propagation is proposed.
• The method automatically emphasizes relevant network structures and deemphasizes irrelevant ones.
• The algorithm is shown to be interpretable as an expectation-maximization (EM) algorithm with a student-t prior.

Main Results:

• The proposed method demonstrates robustness by automatically down-weighting irrelevant networks.
• The algorithm achieves superior efficiency compared to existing methods, as shown analytically and experimentally.
• Successful application in protein function prediction and digit classification tasks.

Conclusions:

• The developed algorithm offers a robust and efficient solution for label propagation on multiple networks.
• The method's ability to integrate diverse network structures enhances transductive inference capabilities.
• This approach advances graph-based learning, particularly for complex biological and classification problems.