Robust multi-view clustering in latent low-rank space with discrepancy induction

Due to the fantastic ability to capture consistent and complementary information between views, multi-view graph clustering has attracted extensive research attention. However, multi-view data are mostly high-dimensional, which may contain many redundant and irrelevant features. At the same time, the original data are usually contaminated by noise and outliers, which may destroy the intrinsic structural information of data and reduce the reliability of the affinity matrix learned. Moreover, most models assign different weights to each view to fully consider the relation between views. The intrinsic information of some views cannot be fully utilized due to the small view weights assigned to them. To deal with these problems, in this study, we propose a robust multi-view clustering model by combining low-dimensional and low-rank latent space learning, self-representation learning, and multi-view discrepancy induction fusion into a unified framework. Specifically, the original high-dimensional data is first reconstructed in a low-dimensional and low-rank space. A self-representation learning method is used to learn the reliable affinity matrix for each view. Furthermore, the Hilbert-Schmidt independence criterion is used as a discrepancy induction module for the complementary fusion of views. Finally, to preserve the data’s local geometric structure, a simple adaptive graph regularization term is applied to the affinity matrix for each view. The comprehensive experiments on six benchmark datasets validate that the proposed model outperforms the six state-of-the-art comparison models in robustness and clustering performance.