Graph generative adversarial networks with evolutionary algorithm

Graph adversarial Networks (GANs) have shown state-of-the-art results in numerous application domains. While GANs are difficult to be trained to generate distribution from data descriptions. In order to solve this problem, GraphGAN is an innovative graph representation learning framework in which generative models and discriminative models are trained against minimax game based on game theory. However, existing GraphGAN is often affected by mode collapse and gradient problem. This paper proposed a novel GANs framework, called graph generative adversarial networks with evolutionary algorithm (EGraphGAN), for enhancing GANs training performance in graph structure learning and improving the generator's competence in generating high-quality data distribution during iterative training process. The generator is regarded as an evolutionary body to continuously mutate and evolve in the environment (discriminator). The discriminator acts as an environment to evaluate the fitness of the individuals generated by generator through the fitness function. During this training process, only individuals with good fitness in each epoch can be retained for the next stage of training. Experiments on multiple challenging datasets showed that EGraphGAN achieves convincing performance and decreases the negative impact of mode collapse and gradient anomalies. The source code is available at https://github.com/codeedit/EGraphGAN.