Modeling multi-scale data via a network of networks.

Area of Science:

• Network science
• Computational biology
• Data science

Background:

• Node and graph label prediction are key network science tasks.
• Related data can be represented as networks at multiple scales, forming a network of networks (NoNs).

Purpose of the Study:

• To investigate if multi-level NoN data integration improves entity label prediction accuracy over single-level methods.
• To develop and evaluate novel approaches for NoN data analysis.

Main Methods:

• Developed the first synthetic NoN generator.
• Constructed a real biological NoN dataset.
• Evaluated prediction accuracy using synthetic and biological NoNs, comparing NoN approaches with traditional single-level methods.

Main Results:

• NoN approaches matched or outperformed single-level methods on synthetic data, depending on NoN properties.
• On biological data, NoN approaches improved protein function prediction accuracy for nearly half of the functions.
• For 30% of functions, only NoN approaches yielded meaningful predictions, while single-level methods achieved random accuracy.

Conclusions:

• Network of networks (NoNs)-based data integration is crucial for enhancing prediction accuracy in systems with multi-level network structures.
• The proposed NoN approaches offer a significant advancement for complex network analysis and prediction tasks.