A Signed Subgraph Encoding Approach via Linear Optimization for Link Sign Prediction.

Area of Science:

• Graph Neural Networks
• Network Science
• Machine Learning

Background:

• Signed networks represent relationships with positive or negative signs.
• Link sign prediction is crucial for understanding network structures.
• Current methods like signed directed graph neural networks (SDGNNs) show strong performance.

Purpose of the Study:

• To propose a novel architecture for link sign prediction in signed directed networks.
• To introduce the subgraph encoding via linear optimization (SELO) model.
• To evaluate SELO's performance against state-of-the-art methods.

Main Methods:

• Developed a subgraph encoding approach to learn edge embeddings.
• Utilized a linear optimization (LO) method to embed subgraphs into likelihood matrices.
• Conducted experiments on five real-world signed networks.

Main Results:

• SELO achieved leading prediction performances compared to SDGNN.
• The model outperformed existing feature-based and embedding-based methods.
• Consistent improvements were observed across all five networks and four evaluation metrics (AUC, F1, micro-F1, macro-F1).

Conclusions:

• SELO offers a significant advancement in link sign prediction for signed directed networks.
• The subgraph encoding via linear optimization approach is effective for learning edge embeddings.
• The proposed model demonstrates robust and superior performance across diverse real-world datasets.