Adaptive attention encoder for attribute graph embedding

In recent years, unsupervised graph representation learning based on graph encoders has made significant progress. However, the existing methods have three shortcomings. First, the existing method based on graph convolutional network (GCN) cannot distinguish between different neighbor nodes. Second, these methods ignore the actual distribution of graph data. Finally, most existing embedding algorithms focus on restoring the feature matrix or adjacency matrix, which is not necessarily the best choice. Considering the above problems, an adaptive attention adversarial variational graph autoencoder (AAAVGE) is proposed. Its core ideas are (1) Use graph attention to distinguish the weights of different neighbor nodes. (2) Use the adversary strategy to guide the embedding vector to obey the true distribution of data. (3) Use an adaptive learning strategy to train the encoder. Link prediction experiments on four public benchmark data sets show that our proposed AAAVGE can significantly improve the performance of graph embedding compared with the mainstream graph embedding methods.