A Topology-Enhanced Multi-Viewed Contrastive Approach for Molecular Graph Representation Learning and Classification

Area of Science:

• Graph representation learning
• Deep learning
• Topological data analysis

Background:

• Graph neural networks (GNNs) are dominant in graph representation learning, achieving state-of-the-art results.
• Graph contrastive learning (GCL) enhances multi-view graph representations by comparing embeddings.
• Existing GCL methods often overlook topological insights crucial for graph learning.

Purpose of the Study:

• To propose a novel topology-enhanced, multi-view graph contrastive learning model (TMGCL).
• To capture and utilize multi-scale topological and global structural information from graphs.
• To improve accuracy and robustness in graph-based applications, such as molecular classification.

Main Methods:

• Developed TMGCL, a model integrating topological data analysis (TDA) with GNN-based GCL.
• Designed TMGCL to capture comprehensive multi-scale topological features and global graph structures.
• Evaluated TMGCL on real-world datasets for molecular classification tasks.

Main Results:

• TMGCL demonstrated superior performance compared to state-of-the-art GNN and GCL baselines.
• The model effectively captures and leverages both topological and structural graph information.
• Experiments confirmed the enhanced accuracy and robustness of TMGCL in molecular classification.

Conclusions:

• TMGCL offers a powerful approach to graph representation learning by incorporating topological insights.
• The integration of TDA significantly benefits GCL models.
• TMGCL shows promise for various graph-based applications requiring robust and accurate representations.