Contrastive graph auto-encoder for graph embedding.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Representation Learning

Background:

• Graph embedding methods often struggle to balance structural and feature information, limiting generalization.
• Graph contrastive learning (GCL) shows promise but faces challenges in learning discriminative node embeddings.
• Existing GCL methods can degrade semantic information through data augmentation and improper negative sampling.

Purpose of the Study:

• To address the limitations of existing GCL methods for node embedding.
• To propose novel approaches that preserve semantic information and improve negative sampling strategies.
• To enhance the discriminative power of node embeddings for downstream tasks.

Main Methods:

• Introduction of Contrastive Graph Auto-Encoder (CGAE) and Contrastive Variational Graph Auto-Encoder (CVGAE).
• Development of distribution-dependent regularizations for parallel encoders to generate contrastive representations.
• Utilization of truncated triplet loss to refine negative sample selection, preventing over-separation of clustered nodes.

Main Results:

• The proposed methods, CGAE and CVGAE, demonstrate improved performance over baseline approaches.
• Experimental results show advancements in link prediction, node clustering, and graph visualization tasks.
• Theoretical analysis supports the effectiveness and robustness of the developed models.

Conclusions:

• CGAE and CVGAE effectively overcome the challenges of semantic information loss and negative sampling in GCL.
• The proposed regularizations and sampling strategy lead to more discriminative node embeddings.
• The models offer significant improvements for various graph-based machine learning tasks.