Network Comparison with Interpretable Contrastive Network Representation Learning.

Area of Science:

• Network analysis
• Machine learning
• Bioinformatics

Background:

• Network comparison is crucial for identifying unique characteristics, such as differences in protein interactions between normal and cancer tissues.
• Existing contrastive learning methods are unsuitable for network data, necessitating new approaches.

Purpose of the Study:

• To introduce contrastive network representation learning (cNRL) for analyzing network uniqueness.
• To develop an interpretable variant, i-cNRL, for understanding specific network patterns.

Main Methods:

• Integrated network representation learning with contrastive learning to create cNRL.
• Developed i-cNRL to provide interpretable embeddings revealing network distinctions.
• Evaluated i-cNRL on network models and real-world datasets.

Main Results:

• cNRL effectively embeds network nodes, highlighting inter-network uniqueness.
• i-cNRL demonstrated interpretability in identifying network-specific patterns.
• Quantitative and qualitative evaluations confirmed i-cNRL's effectiveness compared to other designs.

Conclusions:

• cNRL offers a novel framework for network comparison and analysis.
• i-cNRL enhances interpretability, allowing deeper insights into network differences.
• The developed methods are effective for diverse network analysis tasks.