Learning Disentangled Graph Convolutional Networks Locally and Globally.

Area of Science:

• Graph Machine Learning
• Network Science
• Artificial Intelligence

Background:

• Graph Convolutional Networks (GCNs) are effective for graph-structured data but struggle with entangled latent factors, leading to non-explainable representations.
• Current GCNs often overlook global graph structure, focusing primarily on local information, which limits their comprehensive understanding of complex networks.

Purpose of the Study:

• To propose a novel framework, LGD-GCN, that disentangles node representations by effectively utilizing both local and global graph information.
• To enhance the explainability and performance of GCNs in analyzing real-world graph-structured data.

Main Methods:

• LGD-GCN employs a statistical mixture model for a disentangled latent continuous space, using a neighborhood routing mechanism.
• A novel regularizer promotes inter-factor diversity for model expressivity, while global message passing on disentangled graphs ensures intra-factor consistency.

Main Results:

• Extensive evaluations on synthetic and benchmark datasets demonstrate significant performance gains over existing models in both disentangling and node classification.
• LGD-GCN achieves an average improvement of 7.4% over state-of-the-art methods on social network datasets for disentangled representation learning.

Conclusions:

• LGD-GCN effectively addresses the limitations of existing GCNs by integrating local and global information for improved disentangled node representations.
• The proposed framework offers enhanced explainability and superior performance, particularly in complex graph analysis tasks like node classification on social networks.