Abstract <p>Deep graph clustering is a fundamental task of graph data analysis, which aims to partition nodes into different clusters based on the node attributes and structural features of the clusters. Related research mainly focuses on unsupervised graph representation learning and clustering algorithms. Existing work is limited by the shallow graph neural networks’ scope of neighborhood information aggregation. The structural features captured by these methods are usually local and noisy, and cannot represent the global cluster structure well, leading to suboptimal clustering performance. To address this issue, this paper proposes a novel deep graph clustering method with improved cluster structure features learning, which enhances the cluster structure features learning by continuously separating the distribution between the enhanced graph and various structure-damaged graphs to guide the model to learn more structural feature information from the whole graph. Moreover, it uses the node embeddings obtained from representation learning to reconstruct and update the graph structure, thereby integrating the learned cluster structure features into the graph structure, and iteratively enhancing the cluster structure features learning. Finally, it uses traditional shallow clustering algorithms to divide nodes into different clusters according to the similarity of node embeddings. Compared with eight state-of-the-art clustering methods, extensive experiments on four benchmark datasets demonstrate the effectiveness of the proposed method for conducting clustering tasks on graph data, especially on the amap dataset, a dense graph where accuracy is improved by 1.5% compared to the current best algorithm. More supplementary materials and the implementation code are available for access at the following link: <a href="https://github.com/BluessSkies/GCIS">https://github.com/BluessSkies/GCIS</a>.</p> Graphical abstract <p></p>