Unsupervised Graph Representation Learning With Variable Heat Kernel

Graph representation learning aims to learn a low-dimension latent representation of nodes, and the learned representation is used for downstream graph analysis tasks. However, most of the existing graph embedding models focus on how to aggregate all the neighborhood node features to encode the semantic information into the representation and neglect the global structural features of the node such as community structure and centrality. In the paper, we propose a novel unsupervised graph representation learning method (VHKRep), where a variable heat kernel is designed to better capture implicit global features via heat diffusion with the different time scale and generate the robust node representation. We conduct extensive experiment on three real-world datasets for node classification and link prediction tasks. Compared with the state-of-the-art seven models, the experimental results demonstrate the effectiveness of our proposed method on both node classification and link prediction tasks.