Fast self-supervised discrete graph clustering with ensemble local cluster constraints

Abstract

Spectral clustering (SC) is a graph-based clustering algorithm that has been widely used in the field of data mining and image processing. However, most graph-based clustering methods ignore the utilization of additional prior information. This information can help clustering models further reduce the difference between their clustering results and ground-truth, but is difficult to obtain in unsupervised settings. Moreover, traditional graph-based clustering algorithms require additional hyperparameters and full graph construction to obtain good performance, increasing the tuning pressure and time cost. To address these issues, a simple fast self-supervised discrete graph clustering (FSDGC) is proposed. Specifically, the proposed method has the following features: (1) a novel self-supervised information, based on ensemble local cluster constraints, is used to constrain the sample indicator matrix; (2) the anchor graph technique is introduced for mining the structure between samples and anchors to handle large scale datasets. Meanwhile, a fast coordinate ascent (CA) optimization method, based on self-supervised constraints, is proposed to obtain discrete indicator matrices. Experimental clustering results demonstrate that FSDGC has efficient clustering performance.