Adaptive node-level weighted learning for directed graph neural network.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Network Analysis

Background:

• Directed graph neural networks (DGNNs) are gaining traction, but node-level representation remains under-explored.
• Existing methods often use mean aggregation, potentially losing information from specific node neighborhoods in directed graphs.

Purpose of the Study:

• To develop a novel approach for enhancing node-level representation in directed graphs.
• To address the limitations of mean aggregation in capturing directional neighbor importance.

Main Methods:

• Estimating node homophily to neighbors in different directions using extended Dirichlet energy.
• Assigning weights to neighbors based on directional homophily ratios.
• Incorporating in-degree and out-degree information into weight learning to boost expressive power.

Main Results:

• The proposed method theoretically enhances the expressive ability of directed graphs.
• Experiments on seven real-world datasets show superior performance compared to state-of-the-art methods.
• Demonstrated effectiveness in both node classification and link prediction tasks.

Conclusions:

• The novel weighting strategy effectively captures directional neighbor importance in DGNNs.
• The method offers a significant improvement over traditional aggregation techniques.
• This work provides a more powerful tool for analyzing directed graph data.