Dual generative adversarial graph networks: Unsupervised and semi-supervised learning with spherical graph embeddings

Graph-structured data has emerged as a crucial tool for representing complex systems in various domains. The analysis of such data entails addressing two fundamental problems: attributed graph clustering and semi-supervised node classification. This work proposes a dual generative adversarial framework with spherical graph embeddings for attributed graph clustering. The framework integrates a generative adversarial network (GAN) with a probabilistic encoder that incorporates a von Mises-Fisher mixture model (vMFMM) as the prior distribution over the latent variables, resulting in spherical graph embeddings. The vMFMM prior enables explicit control over the trade-off between the generator's capacity and the representativeness of the graph embeddings, facilitating a principled approach to optimization that is more stable and superior in performance. Furthermore, the proposed framework leverages a dual GAN structure to bolster the robustness of graph embeddings by effectively removing unknown noise present in the latent space through adversarial training. Additionally, the proposed clustering framework is extended to a semi-supervised node classification method by integrating it with a mixture density network. This extension allows the utilization of labeled data, enhancing the accuracy of node classification by incorporating it into the clustering process. To assess the effectiveness of the dual generative adversarial attributed graph clustering method and its semi-supervised extension, a comparative analysis against established baseline methods is conducted. Through extensive experimental results, the advantages of the proposed methods are demonstrated, underscoring their superior performance. The code is publicly available at https://github.com/zxj8806/DGAGN-main.