EGRC-Net: Embedding-Induced Graph Refinement Clustering Network.

Area of Science:

• Machine Learning
• Graph Theory
• Data Mining

Background:

• Existing graph clustering methods struggle with fixed initial graphs that may not represent data topology.
• This limitation can lead to suboptimal clustering performance when the initial graph structure is inaccurate.

Purpose of the Study:

• To propose a novel unsupervised graph clustering network, EGRC-Net, that adaptively refines the initial graph using learned embeddings.
• To enhance clustering accuracy and scalability by integrating semantic and topological information.

Main Methods:

• Leveraging an auto-encoder and graph convolution network to learn latent feature representations.
• Dynamically fusing an initial adjacency matrix with a new one derived from the embedding space's local geometric structure.
• Employing Jeffreys divergence for unsupervised training and an improved approximate personalized propagation for scalability.

Main Results:

• EGRC-Net consistently outperforms state-of-the-art approaches on nine benchmark datasets.
• Achieved over 11.99% improvement in Adjusted Rand Index (ARI) on the DBLP dataset.
• Demonstrated scalability with a 10.73% ARI gain, reduced memory usage by 33.73%, and decreased running time by 19.71%.

Conclusions:

• EGRC-Net effectively refines initial graphs using learned embeddings for superior clustering performance.
• The proposed method offers a scalable and efficient solution for unsupervised graph clustering.
• The publicly available code facilitates further research and application.