Variational co-embedding learning for attributed network clustering

Recent developments in attributed network clustering combine graph neural networks and autoencoders for unsupervised learning. Although effective, these techniques suffer from either (a) clustering-unfriendly embedding spaces or (b) limited utilization of attribute information. To address these issues, we propose a novel model called Variational Co-embedding Learning Model for Attributed Network Clustering (VCLANC), which utilizes much deeper information from the network by reconstructing both the network structure and the node attributes to perform self-supervised learning. Technically, VCLANC consists of dual variational autoencoders that co-embed nodes and attributes into the same latent space, along with a trainable Gaussian mixture prior that simultaneously performs representation learning and node clustering. To optimize the variational autoencoders and infer the latent variables of embeddings and clustering assignments, we derive a new variational lower bound that maximizes the joint likelihood of the observed network structure and node attributes. Furthermore, we also adopt a mutual distance loss on the cluster centers and a clustering assignment hardening loss on the node embeddings to strengthen clustering quality. Our experimental results on four real-world datasets demonstrate the outstanding performance of VCLANC for attributed network clustering.