DEDC-IMAE: A deep evolutionary document clustering model with inherited mixed autoencoder

Abstract

Recently, deep document clustering has been increasingly attracting interest. However, the design of existing deep document clustering methods is limited to processing only static document datasets because applying these methods to dynamic text documents is challenging as the underlying topics of document datasets are not static but evolve. Deep models must learn the cluster-level representations that are inherited. Thus, methods must be explored to implement the inheritance of cluster-level representations. In this study, we overcome the aforementioned challenges and propose a deep evolutionary document clustering model with an inherited mixed autoencoder (DEDC-IMAE), to discover evolutionary document structures with evolutionary topics in dynamic document datasets. An inherited mixed autoencoder (IMAE) is designed to learn an evolutionary semantic representation of each document considering both the inherited historical topic semantics and the document's own characteristics, to reflect the evolved topic semantics. Using the document evolutionary semantics, a deep evolutionary document clustering (DEDC) module is then investigated to learn a clustering partition that reflects the characteristics of document datasets. Experimental results on a real evolutionary text document dataset showed that DEDC-IMAE achieves better evolutionary clustering results than a variety of baseline models.