Scalable Graph Convolutional Variational Autoencoders

Autoencoders are widely used for self-supervised representation learning. Variational autoencoders (VAEs), a special type of autoencoders, are proven to be effective in estimating the underlying probability distribution of the training data. Even though VAEs are well explored in many application domains, their utilization for graph-structured data is still under extensive research. Graph variational autoencoders achieved competitive results on various graph-related modeling tasks (e.g. link prediction and node clustering) by incorporating node features. However, current graph VAEs are unable to scale efficiently for larger graphs.In this paper, we propose a novel method that adapts the stochastic multiple partitions (SMP) algorithm to improve on scalability. We also introduce novel graph convolutional layers with general graph filters, which significantly improve the predictive performance of the neural network. The proposed method is evaluated on two popular large-graph datasets. According to the results, the proposed filters outperform the baseline filter in link prediction and node clustering for both datasets.