Abstract: To solve the problem of overfitting in the method of encoding graph structure information based on graph convolutional network (GCN) in depth clustering, a method is proposed to encode graph structure by fusing graph adjacency into traditional depth network through comparison loss. First, in this method, an auto-encoder (AE) was used to learn the deep potential representation of node features. And then, through comparative learning, discriminative node representations were learned from graph relationships, and more detailed inter node influence was designed to provide a strong basis for selecting positive and negative samples for comparative learning. Finally, the network was trained by self-supervision to implement node clustering. A large number of experiments on six benchmark datasets show that the proposed method significantly improves the clustering accuracy.