Hypergraph Representation Learning with Weighted- and Clustering-Biased Random Walks.

Area of Science:

• Graph Representation Learning
• Network Science
• Machine Learning

Background:

• Hypergraphs model complex systems with higher-order interactions.
• Existing methods struggle with heterogeneous hypergraphs, impacting structure-sensitive tasks.
• Node classification performance is suboptimal due to limitations in capturing hypergraph structures.

Purpose of the Study:

• To present WCRW-MLP, a new framework for hypergraph representation learning.
• To improve the capture of higher-order structures in heterogeneous hypergraphs.
• To enhance performance on structure-sensitive tasks like node classification.

Main Methods:

• Introduced Weighted- and Clustering-Biased Random Walk (WCRW) extending second-order random walks.
• Incorporated node-pair co-occurrence weights and triadic-closure clustering bias into random walks.
• Utilized Skip-gram for structural embeddings and concatenated them with node attributes for MLP classification.

Main Results:

• WCRW-MLP consistently outperformed state-of-the-art baselines on real-world hypergraph benchmarks.
• Demonstrated the efficacy of the proposed biasing strategy in random walks.
• Validated the overall framework's effectiveness in hypergraph embedding.

Conclusions:

• Explicitly modeling co-occurrence strength and local clustering is crucial for effective hypergraph embedding.
• The WCRW-MLP framework offers a significant advancement in learning from hypergraph data.
• The approach shows promise for various structure-sensitive applications involving complex systems.