Multilevel Contrastive Graph Masked Autoencoders for Unsupervised Graph-Structure Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Unsupervised graph-structure learning (GSL) aims to discover graph structures from data without labels for downstream tasks.
• Existing methods often struggle to effectively leverage graph masked autoencoders for robust GSL.
• There is a need for improved unsupervised GSL techniques that can extract richer supervisory signals from data.

Purpose of the Study:

• To develop a novel multilevel contrastive graph masked autoencoder (MCGMAE) for unsupervised GSL.
• To enhance the acquisition of supervision information directly from data for improved graph structure learning.
• To increase the robustness and effectiveness of unsupervised GSL in various applications.

Main Methods:

• Introduced a graph masked autoencoder with a dual feature masking strategy for reconstructing graph data.
• Incorporated inter- and intra-class contrastive loss to maximize mutual information at feature and reconstruction levels.
• Applied contrastive losses to the graph encoder module to strengthen feature-level agreement.

Main Results:

• The proposed MCGMAE effectively learns graph structures without relying on labeled data.
• Demonstrated improved training robustness in unsupervised GSL through multilevel supervision signals.
• Achieved superior performance across three graph analytical tasks and eight diverse datasets.

Conclusions:

• MCGMAE offers a powerful framework for unsupervised graph-structure learning.
• The integration of dual masking and contrastive losses significantly enhances GSL effectiveness.
• The method provides a robust and generalizable approach for learning graph structures from unlabeled data.