Hypergraph node representation learning with one-stage message passing.

Area of Science:

• Hypergraph representation learning
• Graph neural networks
• Machine learning

Background:

• Existing hypergraph node representation learning methods often use a two-stage message passing paradigm.
• This paradigm focuses on local information, neglecting global context, leading to suboptimal representations.
• Theoretical analysis reveals limitations of two-stage methods and suggests a unified one-stage approach.

Purpose of the Study:

• To propose a novel one-stage message passing paradigm for hypergraph node representation learning.
• To develop a Transformer-based framework, HGraphormer, integrating global and local information.
• To enhance the effectiveness of hypergraph learning by considering both local structure and global context.

Main Methods:

• Developed a one-stage message passing paradigm for hypergraphs.
• Integrated this paradigm into HGraphormer, a Transformer-based framework.
• Combined attention matrices and hypergraph Laplacian to inject structure information into Transformers.

Main Results:

• HGraphormer demonstrated superior performance on five benchmark datasets for semi-supervised hypernode classification.
• Achieved accuracy improvements ranging from 2.52% to 6.70% compared to recent methods.
• Established new state-of-the-art results in hypergraph learning.

Conclusions:

• The proposed one-stage message passing paradigm effectively models global and local information for hypergraphs.
• HGraphormer offers a powerful framework for hypergraph node representation learning.
• The findings advance the field of hypergraph analysis and its applications.