Latent Space Clustering via Dual Discriminator GAN

Abstract

In recent years, deep generative models have achieved remarkable success in unsupervised learning tasks. Generative Adversarial Network (GAN) is one of the most popular generative models, which learns powerful latent representations, and hence is potential to improve clustering performance. We propose a new method termed CD2GAN for latent space Clustering via dual discriminator GAN (D2GAN) with an inverse network. In the proposed method, the continuous vector sampled from a Gaussian distribution is cascaded with the one-hot vector and then fed into the generator to better capture the categorical information. An inverse network is also introduced to map data into the separable latent space and a semi-supervised strategy is adopted to accelerate and stabilize the training process. What’s more, the final clustering labels can be obtained by the cross-entropy minimization operation rather than by applying the traditional clustering methods like K-means. Extensive experiments are conducted on several real-world datasets. And the results demonstrate that our method outperforms both the GAN-based clustering methods and the traditional clustering methods.