Clustering by Directly Disentangling Latent Space

To overcome the high dimensionality of data, learning latent feature representations for clustering has been widely studied. Recently, ClusterGAN combined GAN with an encoder to learn a mixture of one-hot discrete and continuous latent variables, and achieved remarkable clustering performance. However, the performance of ClusterGAN decreases when it is applied to complex data. In this paper, we analyze the reasons for performance degeneracy in ClusterGAN. We show that minimizing the cycle-consistency loss of continuous latent variables in ClusterGAN trends to generate trivial latent features. Moreover, the objective of ClusterGAN doesn't include a real conditional distribution term, which makes it difficult to be generalized to real data. Therefore, we propose Disentangling Latent Space Clustering (DLS-Clustering), a new clustering mechanism that directly learns cluster assignments from disentangled latent spacing without additional clustering methods. We enforce the inference network (encoder) and the generator of GAN to form an encoder-generator pair in addition to the generator-encoder pair. We train the encoder-generator pair using real data, which can estimate the real conditional distribution. Moreover, the encoder-generator pair competes with the generator-encoder pair during optimization, which can avoid the triviality of continuous latent variables. Furthermore, we utilize a weight-sharing procedure to disentangle the one-hot discrete and the continuous latent variables generated from the encoder. This process enforces the disentangled latent space to match the independence of GAN inputs. Eventually, the one-hot discrete latent variables can be directly expressed as clusters and the continuous latent variables represent remaining unspecified factors. Experiments on benchmark datasets of different types demonstrate that our method outperforms existing state-of-the-art methods.