Abstract: Graph Auto-Encoder (GAE) is a learning framework derived from graph neural network. It introduces the idea of clustering neighborhood nodes into the encoder, and the decoder decodes the graph structure data and reconstructs the graph structure data. The introduction of supervisory module in the model can improve the integrity of graph structure data embedded in the model and the accuracy of data generation. Encoder and decoder can use different neural networks to take advantage of the advantages different neural networks. In recent years, GAE recommendation has become a hot topic in recommendation system research. The progress of GAE recommendation research is analyzed from the aspects of unsupervised learning and semi-supervised learning. The deficiencies of the existing GAE recommendations, such as cold startup for users, uninterpretability, high model complexity, and multi-source heterogeneity of data, are discussed. In addition, it looks into the future of GAE recommendation from the perspectives of cross-field recommendation, combined with traditional recommendation methods, introduction of attention mechanism, and integration of various scenarios.

Key words: graph auto-encoder, recommendation, unsupervised learning, semi-supervised learning