Stochastic graph recurrent neural network

Representation learning over dynamic graphs has attracted much attention because of its wide applications. Recently, sequential probabilistic generative models have achieved impressive results they can model data distributions. However, modeling the distribution is still extremely challenging. Existing methods usually ignore mutual interference stochastic states and deterministic states. Besides, assumption that latent variables follow Gaussian distributions inappropriate. To address these problems, we propose graph recurrent neural network (SGRNN), a for representation graphs. It separates in iterative process. improve flexibility variables, set prior posterior as semi-implicit DSI-SGRNN. In addition, to alleviate KL-vanishing problem SGRNN, simple interpretable structure proposed based on lower bound KL-divergence. The introduces few extra parameters be implemented with lines code modification. Extensive experiments real-world datasets demonstrate effectiveness model.