Deep Structured Graph Clustering Network

Abstract

Listen

Translate article

AbstractShallow clustering methods adopt linear or simple nonlinear projections to reduce the feature dimensions, which may suffer from the weak representation capability. Contrastively, deep clustering methods have the advantages on representing the sample characteristics. However, most deep clustering models focus on preserving feature information of samples and ignore the important intrinsic structures of samples. Besides, large amounts of neural network parameters should be optimized in deep clustering models, but no proper semantic supervision can be used in the unsupervised clustering process. To alleviate these problems, in this paper, we propose a unified deep structured graph clustering network to guide the unsupervised deep clustering process with a theoretically ideal cluster structure. Specifically, we simultaneously learn the discriminative feature representation of samples, and the similarity graph of samples with well clustering structure by automatically assigning proper neighbors to each sample. Experiments on several public testing datasets demonstrate the effects of the proposed method.KeywordsClusteringDeep learningStructured graphJoint learning