Mixtures of posterior and prior variational autoencoders for representation learning and cluster analysis in latent space

Cluster analysis aims to identify groups of similar items within an unlabelled dataset. This is particularly challenging in high-dimensional data, necessitating the finding of hidden or latent structure within the data, for which variational methods have proven to be successful. We introduce a generative model based on the variational autoencoder (VAE) that uses a mixture distribution for both the prior and variational posterior components over the latent variables. This pair of distributions means that the algorithm can better capture the underlying structure of the data. We evaluated clustering performance on a set of benchmark datasets. Our proposed model demonstrates superior clustering performance compared with state-of-the-art deep clustering algorithms, as well as demonstrating reasonable reconstruction performance and generation of realistic examples from the latent space.