Hyperbolic Variational Graph Neural Network for Modeling Dynamic Graphs

Learning representations for graphs plays a critical role in wide spectrum of downstream applications. In this paper, we summarize the limitations prior works three folds: representation space, modeling dynamics and uncertainty. To bridge gap, propose to learn dynamic graph hyperbolic first time, which aims infer stochastic node representations. Working with present novel Hyperbolic Variational Graph Neural Network, referred as HVGNN. particular, model dynamics, introduce Temporal GNN (TGNN) based on theoretically grounded time encoding approach. uncertainty, devise variational autoencoder built upon proposed TGNN generate normal distributions. Furthermore, reparameterisable sampling algorithm distribution enable gradient-based learning Extensive experiments show that HVGNN outperforms state-of-the-art baselines real-world datasets.