Spectral embedding network for attributed graph clustering.

Area of Science:

• Machine Learning
• Data Mining
• Graph Analytics

Background:

• Attributed graph clustering seeks to identify node communities using graph structure and features.
• Current methods often use graph neural networks for embeddings, followed by traditional clustering, but face challenges with noisy, sparse graph structures and non-clustering-aware losses.

Purpose of the Study:

• To propose a Spectral Embedding Network (SENet) for attributed graph clustering.
• To improve graph structure representation and learn embeddings optimized for clustering.

Main Methods:

• SENet enhances graph structure by incorporating shared neighbor information, addressing noise and sparsity.
• A spectral clustering loss is employed, integrating structure and feature information via higher-order graph convolution for better kernel matrix adaptation.
• Combines first-order and second-order proximity information into the graph structure.

Main Results:

• SENet demonstrates superior performance compared to existing state-of-the-art methods on benchmark attributed graphs.
• The proposed method effectively alleviates noise and sparsity issues in graph structures.

Conclusions:

• SENet offers an effective approach for attributed graph clustering by improving graph representation and utilizing a clustering-driven spectral loss.
• The method provides more sufficient node embeddings for downstream clustering tasks.