Multi-view deep clustering based on autoencoder

Abstract

In recent years, with the development of deep learning, replacing traditional clustering methods with subspaces extracted by deep neural networks will help better clustering performance. However, due to the instability of unsupervised learning, the features extracted each time are different even if the same data is processed. In order to improve the stability and performance of clustering, we propose a novel unsupervised deep embedding clustering multi-view method, which treats multiple different subspaces as different views through some data expansion methods for the same data. Specifically, our method uses a variety of different deep autoencoders to learn the latent representation of the original data and constrain them to learn different features. Our experimental evaluations on several natural image datasets show that this method has a significant improvement compared to existing methods.