Towards exploring the topological structure of data, numerous graph embedding clustering methods have been developed in recent years, none of them takes into account the cluster-specificity distribution of the nodes representations, resulting in suboptimal clustering performance. Moreover, most existing graph embedding clustering methods execute the nodes representations learning and clustering in two separated steps, which increases the instability of its original performance. Additionally, rare of them simultaneously takes node attributes reconstruction and graph structure reconstruction into account, resulting in degrading the capability of graph learning. In this work, we integrate the nodes representations learning and clustering into a unified framework, and propose a new deep graph attention auto-encoder for nodes clustering that attempts to learn more favorable nodes representations by leveraging self-attention mechanism and node attributes reconstruction. Meanwhile, a cluster-specificity distribution constraint, which is measured by ℓ1,2-norm, is employed to make the nodes representations within the same cluster end up with a common distribution in the dimension space while representations with different clusters have different distributions in the intrinsic dimensions. Extensive experiment results reveal that our proposed method is superior to several state-of-the-art methods in terms of performance.
- Cluster-specificity distribution
- Graph neural networks
- Nodes clustering
- Neural Networks, Computer
- Cluster Analysis