Self-supervised graph convolutional clustering by preserving latent distribution

Abstract

Node clustering has attracted widespread concern due to the fact that graph structure helps to seek the geometric structure of data. Existing graph convolutional networks (GCNs) based node clustering methods usually contain two steps. First, learning latent nodes representations. Second, inferring the unknown labels. Despite the promising preliminary results, it cannot take full advantage of the information embedded in current pseudo clustering labels, resulting in suboptimal performance. In fact, it is the truth we do not know the true label of the node in clustering, but the pseudo clustering label of the node generated by each iteration is known. Under this condition, although the pseudo labels of some nodes are incorrect, the partial correct label information is useful and significant. Moreover, GCNs based node clustering aims to learn satisfactory low dimensional node representations which are usually used for clustering. However, existing GCNs based clustering methods fail to take into account distribution consistency between the raw data space and the latent space of nodes representations, resulting in insufficient representations. To address the above problems, we propose a novel GCN-based node clustering method. By introducing a self-supervision module, it can employ the good property of pseudo clustering labels to self-supervise the learning of node representations. Moreover, a latent distribution preserving term, which is measured by KL divergence, is employed to help the latent representations of the same sample to further have a consistent distribution in dimension space as well as in the original dimension space. We formulate the above two concerns into a unified optimization framework. Experimental results on several public datasets indicate our method is effective compared with the state-of-the-art node clustering methods.