Deep Neural Networks (DNNs) have achieved impressive success in the domain of Euclidean data such as image. However, designing deep neural network to cluster nodes especially in social networks is still a challenging task. Moreover, recent advanced methods for node clustering have focused on learning node embedding, upon which classic clustering algorithms like K-means are applied. Nevertheless, the resulting node embeddings are customarily task-agnostic. This results in the fact that the performance of clustering is difficult to guarantee. To effectively mitigate the problem, in this paper, we propose a novel clustering-oriented node embedding method named Deep Node Clustering (DNC) for non-attributed network data by resorting to deep neural networks. We first present a preprocessing method via adopting a random surfing model to capture graph structural information directly. Subsequently, we propose to learn a deep clustering network, which could jointly learn node embeddings and cluster assignments. Extensive experiments on three real-world network datasets for node clustering are conducted, which demonstrate that the proposed DNC substantially outperforms the state-of-the-art node clustering methods.

深度神经网络（DNN）在欧氏数据（例如图像）领域取得了令人瞩目的成功。但是，设计深度神经网络以群集节点（尤其是在社交网络中）仍然是一项艰巨的任务。此外，最近的用于节点聚类的高级方法集中于学习节点嵌入，在其上应用了诸如K-means之类的经典聚类算法。然而，所得的节点嵌入通常是任务无关的。这导致难以保证群集性能的事实。为了有效缓解该问题，在本文中，我们借助于深度神经网络，提出了一种针对非聚类网络数据的面向聚类的节点嵌入新方法，称为非节点网络聚类（DNC）。我们首先提出一种通过采用随机冲浪模型直接捕获图结构信息的预处理方法。随后，我们建议学习一个深度聚类网络，该网络可以共同学习节点嵌入和聚类分配。在用于节点群集的三个真实世界网络数据集上进行了广泛的实验，这些实验表明，所提出的DNC明显优于最新的节点群集方法。