Deep hierarchical clustering with Dirichlet Forest Prior

Abstract

This work proposes to incorporate Dirichlet Forest Priors to Variational Deep Embedding (VaDE), a deep unsupervised generative model for clustering, in order to implement hierarchical clustering and evaluates its hierarchical clustering accuracy. In this process, the method to alleviate the class imbalance problem in clustering by injecting prior knowledge is presented. Furthermore, this paper suggests a method to give guidance to clustering with few labels. Evaluations on the performance gains of these contributions are done through experiments on both image and text datasets.