Dual-decoder graph autoencoder for unsupervised graph representation learning

Unsupervised graph representation learning is a challenging task that embeds graph data into a low-dimensional space without label guidance. Recently, graph autoencoders have been proven to be an effective way to solve this problem in some attributed networks. However, most existing graph autoencoder-based embedding algorithms only reconstruct the feature maps of nodes or the affinity matrix but do not fully leverage the latent information encoded in the low-dimensional representation. In this study, we propose a dual-decoder graph autoencoder model for attributed graph embedding. The proposed framework embeds the graph topological structure and node attributes into a compact representation, and then the two decoders are trained to reconstruct the node attributes and graph structures simultaneously. The experimental results on clustering and link prediction tasks strongly support the conclusion that the proposed model outperforms the state-of-the-art approaches.