Deep tensor completion graph convolutional subspace clustering

Abstract

Graph Convolutional Subspace Clustering (GCSC) aims to integrate the topological information of data with subspace representations by Graph Convolutional Networks (GCNs). However, existing methods are limited by their emphasis on local topological information, which neglects global relationships in data. Also, their adjacency matrices are fixed and predefined, which fail to adjust to the changing features during training and may be easily affected by noise. To address these issues, we propose Deep Tensor Completion Graph Convolutional Subspace Clustering (DTC-GCSC). Firstly, we treat the initialized adjacency matrix as a trainable parameter, enabling its joint optimization with the model through a deep architecture. Based on this framework, we further incorporate global topological information by integrating conventional subspace clustering (CSC) into GCSC, extending local relationships to a global structure. Finally, to enhance the consistency between local and global information, we introduce a Tensor Nuclear Norm (TNN) constraint to enforce high-order correlations across them. Extensive experiments on multiple datasets demonstrate the superiority of our method over state-of-the-art approaches.