Multiview Deep Graph Infomax to Achieve Unsupervised Graph Embedding.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Mining

Background:

• Unsupervised graph embedding extracts node representations for analysis.
• Multiview graphs offer richer topological information than single-view graphs.
• Limited research exists on unsupervised representation learning from multiview heterogeneous networks.

Purpose of the Study:

• Propose a novel unsupervised multiview graph embedding method (MVDGI).
• Extract highly discriminative node representations from multiview heterogeneous networks.
• Improve performance in downstream tasks like classification and clustering.

Main Methods:

• Employ a contrastive learning backbone to maximize mutual information.
• Utilize an encoder for view-dependent representation extraction.
• Apply an aggregator to fuse representations and a discriminator for refinement.

Main Results:

• MVDGI outperforms benchmark methods on five real-world datasets.
• The method generates more discriminative node representations.
• Demonstrated superior performance in classification and clustering tasks.

Conclusions:

• MVDGI is an effective unsupervised method for multiview graph embedding.
• The approach successfully extracts discriminative node representations.
• MVDGI offers significant improvements for graph-based machine learning tasks.