DHMNN: A Hypergraph Motif-Based Framework for Directed Hyperlink Prediction

Area of Science:

• Graph theory
• Network analysis
• Machine learning

Background:

• Directed hypergraphs model complex group interactions but lack effective link prediction methods.
• Existing models struggle with feature initialization, higher-order substructures, and long-range dependencies.

Purpose of the Study:

• To propose a novel directed hypergraph motif-based neural network (DHMNN) for directed hyperlink prediction.
• To address limitations in existing models by capturing higher-order structural and connectivity information.

Main Methods:

• Introduced directed hypergraph motifs (DH-motifs) for higher-order neighborhood analysis and feature generation.
• Utilized hypergraph incidence matrices for local connectivity measurement and feature quantification.
• Employed hypergraph attention for refining vertex features and capturing dependencies.
• Designed a new scoring layer to assess link reliability based on local properties, feature variance, and directionality.

Main Results:

• DHMNN significantly outperforms state-of-the-art models on seven metabolic and three social networks.
• Achieved accuracy improvements ranging from 3.40% to 9.90% consistently across datasets.
• Demonstrated effective capture of higher-order substructures and long-range dependencies.

Conclusions:

• The proposed DHMNN is a powerful and effective method for link prediction in directed hypergraphs.
• DHMNN addresses key challenges, offering improved performance in complex systems analysis.
• The approach enhances understanding of network topology and vertex relationships.