Multi-masks and Bi-spaces Reconstruction based Single-Layer Auto-encoder for Heterogeneous Graph Representation Learning

Generative self-supervised learning (SSL) for heterogeneous graph representation has drawn increasing attention due to its powerful capability of capturing rich structural and semantic information without relying on labeled information. However, the multi-layer network structure with multiple attention mechanisms and masking strategies adopted by existing generative SSL methods based on Auto-Encoder, brings semantic confusion and creates dependence of the reconstruction of masked attributes on attribute distinguishability issues. To address these issues, we propose Multi-masks and Bi-spaces reconstruction based Single-Layer Heterogeneous graph AutoEncoder(MBS-HAE), which employs a single-layer masked autoencoder with only semantic attention to learn node embeddings and uses re-masking by randomly selecting nodes to reduce the model’s reliance on the distinguishability of input attributes. Meanwhile, MBS-HAE performs bi-spaces reconstruction in both attribute and latent space to improve the certainty of inferring masked data. The extensive experiments on multiple datasets demonstrate the effectiveness and efficiency of the proposed MBS-HAE.