Abstract: The existing graph autoencoder ignores the difference between the neighbor nodes of the graph and the potential data distribution of the graph. In order to improve the embedding ability of the graph autoencoder, the graph attention adversarial variational autoencoder (AAVGA-d) is proposed. This method introduced attention to the encoder and used an adversarial mechanism in the embedding training. The graph attention encoder realized the adaptive allocation of the weights of neighbor nodes, and the adversarial regularization made the distribution of the embedding vector generated by the encoder close to the true distribution of the data. In order to deepen the number of graph attention layers, a random edge deletion technology (RDEdge) for attention networks was designed to reduce the loss of over-smooth information caused by excessively deep layers. The experimental results prove that the graph embedding capability of AAVAG-d has a competitive advantage compared with the current popular graph autoencoders.