Deep manifold embedding of attributed graphs

Abstract

Unsupervised attribution graph embedding is challenging because the structure and attribute information should be represented in the latent space. Existing reconstruction-based methods may corrupt the manifold of the attributed graph because they indirectly optimize the latent space with the help of decoders. Therefore, we propose a new graph embedding framework called Deep Manifold Embedding of Attribution Graphs (DMEAG). DMEAG directly imposes constraints on the latent space without decoders, thus better preserving the structural information. Further, we propose a node-to-node geodesic similarity metric to measure data relationships. And we design a novel loss function based on Bergman divergence to minimize the difference between embedding and structure/features. Extensive experiments on graphical and image datasets demonstrate the superiority of DMEAG. Furthermore, the proposed DMEAG outperforms state-of-the-art methods in three downstream tasks: clustering, link prediction, and visualization.