Graph embedding clustering: Graph attention auto-encoder with cluster-specificity distribution.

Area of Science:

• Computer Science
• Data Mining
• Machine Learning

Background:

• Existing graph embedding clustering methods often yield suboptimal performance.
• These methods typically separate representation learning and clustering, leading to instability.
• Few methods simultaneously consider node attribute and graph structure reconstruction.

Purpose of the Study:

• To propose a novel deep graph attention auto-encoder for enhanced graph embedding clustering.
• To integrate representation learning and clustering into a unified framework.
• To improve node representation quality by leveraging self-attention and attribute reconstruction.

Main Methods:

• Developed a deep graph attention auto-encoder framework.
• Integrated node representation learning and clustering.
• Employed a cluster-specificity distribution constraint using the ℓ_1,2-norm.
• Incorporated node attribute reconstruction and self-attention mechanisms.

Main Results:

• The proposed method demonstrates superior clustering performance compared to state-of-the-art techniques.
• Achieved more favorable node representations by considering cluster-specific distributions.
• The unified framework enhanced the stability and capability of graph learning.

Conclusions:

• The novel deep graph attention auto-encoder effectively addresses limitations in existing graph embedding clustering methods.
• The integration of cluster-specificity distribution and attribute reconstruction leads to improved clustering outcomes.
• This approach offers a more robust and performant solution for graph data analysis.