Simple and Effective Graph Autoencoders with One-Hop Linear Models

Over the last few years, graph autoencoders (AE) and variational (VAE) emerged as powerful node embedding methods, with promising performances on challenging tasks such link prediction clustering. Graph AE, VAE most of their extensions rely multi-layer convolutional networks (GCN) encoders to learn vector space representations nodes. In this paper, we show that GCN are actually unnecessarily complex for many applications. We propose replace them by significantly simpler more interpretable linear models w.r.t. direct neighborhood (one-hop) adjacency matrix graph, involving fewer operations, parameters no activation function. For two aforementioned tasks, approach consistently reaches competitive GCN-based AE numerous real-world graphs, including all benchmark datasets commonly used evaluate VAE. Based these results, also question relevance repeatedly using compare