Unsupervised graph-level representation learning with hierarchical contrasts.

Area of Science:

• Machine Learning
• Graph Representation Learning
• Data Mining

Background:

• Unsupervised graph-level representation learning is crucial for domains like bioinformatics and social networks.
• Existing graph contrastive learning methods often overlook hierarchical graph structures, limiting semantic information exploration.
• Current methods frequently require numerous negative samples, causing memory issues during optimization.

Purpose of the Study:

• To develop an unsupervised graph-level representation learning framework addressing limitations of existing methods.
• To investigate hierarchical structural semantics at both node and graph levels for enhanced representation.
• To reduce the demand for excessive negative samples in graph contrastive learning.

Main Methods:

• Introduced Hierarchical Graph Contrastive Learning (HGCL), an unsupervised framework for graph representation.
• HGCL integrates node-level, graph-level, and mutual contrastive learning to capture hierarchical graph semantics.
• Employed Siamese networks and momentum updates to mitigate the need for large numbers of negative samples.

Main Results:

• HGCL effectively explores hierarchical graph structures, capturing richer semantic information.
• The framework demonstrates superior performance compared to state-of-the-art baselines on graph classification tasks.
• Significant improvements were observed in transfer learning tasks using large-scale OGB datasets.

Conclusions:

• HGCL offers a novel approach to unsupervised graph representation learning by leveraging hierarchical structures.
• The method effectively addresses the limitations of existing contrastive learning techniques, including reduced memory burden.
• HGCL shows strong potential for various graph-based machine learning applications.