Signed graph embedding via multi-order neighborhood feature fusion and contrastive learning

Area of Science:

• Graph theory
• Network analysis
• Machine learning

Background:

• Signed graphs model complex networks with positive/negative links.
• Existing signed graph embedding methods often lack generality across tasks.
• The generality problem limits performance in multiple downstream applications.

Purpose of the Study:

• To propose a novel signed graph embedding method, MOSGCN, addressing the generality problem.
• To enhance node representations by capturing both local and global structural features.
• To improve the robustness and discriminative power of embeddings for diverse tasks.

Main Methods:

• Developed MOSGCN with a multi-order neighborhood feature fusion strategy based on structural balance theory.
• Employed a signed graph contrastive learning framework for training.
• Evaluated on link sign prediction and community detection tasks using four benchmark datasets.

Main Results:

• MOSGCN demonstrates superior performance in downstream tasks compared to state-of-the-art methods.
• The proposed method shows good generality across different signed graph analysis tasks.
• Multi-order feature fusion and contrastive learning contribute to more informative node representations.

Conclusions:

• MOSGCN effectively addresses the generality problem in signed graph embedding.
• The method achieves state-of-the-art results on link sign prediction and community detection.
• MOSGCN offers a robust and versatile approach for signed graph analysis.