Abstract
Attributed graph embedding aims to learn low-dimensional representations by fully exploiting both topological structure and rich attributes. The embeddings contain more useful information, thus yielding better performance in subsequent analysis tasks. Therefore, attributed network embedding has become an important research direction in recent years. However, most existing algorithms ignore the high-frequency noise contained in the features and do not normalize the embedding, which limits their performance in downstream tasks. To solve the above problem, we design an Adversarial Attributed Graph Embedding algorithm (AAGE) in conjunction with a Laplacian smoothing filter. In particular, the AAGE employs a variational autoencoder (VAE) to learn the embedding of nodes and attributes. Unlike most variational autoencoder-based algorithms, we first apply a carefully-designed Laplacian filter to smooth the features before performing autoencoder, secondly normalize the obtained embeddings of nodes and attributes during the autoencoder process, and finally devise a generative adversarial training method to enhance the robustness of the learned representations. To verify the potential of the proposed AAGE, we evaluate its performance on the tasks of node classification and node clustering on three real-world attributed networks. The results show that AAGE significantly outperforms state-of-the-art methods.