Research on signed directed network link prediction based on dual attention mechanism

Area of Science:

• Graph Neural Networks
• Network Analysis
• Machine Learning

Background:

• Traditional analysis of directed signed networks often relies on sociological and structural balance theories.
• These theories inadequately capture the nuanced relationships between nodes in complex networks.

Purpose of the Study:

• To propose a novel model, the Decoupled Attention Dual Signed Graph Neural Network (DADSGNN), for analyzing directed signed networks.
• To enhance the accuracy and interpretability of node embeddings and link sign prediction in these networks.

Main Methods:

• Employing decoupled representation learning to decompose node features into multiple potential factors.
• Implementing a dual attention mechanism (local and structural) to aggregate neighbor information effectively.
• Developing a novel decoder that considers correlations between potential factors for link sign prediction.

Main Results:

• The DADSGNN model demonstrates improved accuracy in representing complex internodal relationships.
• Enhanced prediction accuracy for link signs in directed signed networks.
• Validation of the model's performance on various directed network datasets.

Conclusions:

• DADSGNN offers a more accurate and interpretable approach to analyzing directed signed networks compared to existing methods.
• The decoupled representation learning and dual attention mechanisms are key to the model's effectiveness.
• The model shows strong predictive capabilities on real-world signed network data.