Spectral embedding network for attributed graph clustering

Attributed graph clustering aims to discover node groups by utilizing both graph structure and node features. Recent studies mostly adopt graph neural networks to learn node embeddings, then apply traditional clustering methods to obtain clusters. However, they usually suffer from the following issues: (1) they adopt original graph structure which is unfavorable for clustering due to its noise and sparsity problems; (2) they mainly utilize non-clustering driven losses that cannot well capture the global cluster structure, thus the learned embeddings are not sufficient for the downstream clustering task. In this paper, we propose a spectral embedding network for attributed graph clustering (SENet), which improves graph structure by leveraging the information of shared neighbors, and learns node embeddings with the help of a spectral clustering loss. By combining the original graph structure and shared neighbor based similarity, both the first-order and second-order proximities are encoded into the improved graph structure, thus alleviating the noise and sparsity issues. To make the spectral loss well adapt to attributed graphs, we integrate both structure and feature information into kernel matrix via a higher-order graph convolution. Experiments on benchmark attributed graphs show that SENet achieves superior performance over state-of-the-art methods.