Graph Representation Learning Beyond Node and Homophily

Unsupervised graph representation learning aims to distill various information into a downstream task-agnostic dense vector embedding. However, existing approaches are designed mainly under the node homophily assumption: connected nodes tend have similar labels and optimize performance on node-centric tasks. Their design is apparently against principle generally suffers poor in tasks, e.g., edge classification, that demands feature signals beyond node-view assumption. To condense different embeddings, this paper proposes PairE, novel unsupervised embedding method using two paired as basic unit of retain high-frequency between support node-related edge-related Accordingly, multi-self-supervised autoencoder fulfill pretext tasks: one retains signal better, another enhances commonality. Our extensive experiments diversity benchmark datasets clearly show PairE outperforms state-of-the-art baselines, with up 101.1\% relative improvement classification tasks rely both high low-frequency pair 82.5\% gain