Learning a bi-directional discriminative representation for deep clustering

Abstract

Nowadays, deep clustering achieves superior performance by jointly performing representation learning and cluster assignment. Although numerous deep clustering algorithms have emerged, most of them have difficulty learning representations that fit the clustering distribution. To address this issue, we propose a bi-directional discriminative representation learning clustering (BDRC) framework in this paper. In our framework, a dual autoencoder network, a bi-directional mutual information maximization module and a self-supervised cluster prediction module are combined into a joint optimization framework. To learn more cluster-friendly representations, the bi-directional mutual information maximization module is executed on both samples and their nearest neighbors to explore the cluster relationships between samples. In order to improve the stability of the model, a self-supervised cluster prediction module is devised to predict clustering assignments to supervise the autoencoder using the KL-divergence. Moreover, the UMAP is used to find the manifold of the latent representations which can better preserve the global structure. Experiments on some benchmark datasets demonstrate the superiority of the proposed BDRC algorithm.