Consensus guided incomplete multi-view clustering via geometric consistency learning

Abstract

Incomplete multi-view clustering (IMC) aims to uncover meaningful cluster structures by leveraging the similarity information within datasets containing multiple, but partially missing, views. While most existing approaches emphasize learning consensus representations to integrate information across views, they often neglect the inherent geometric structure of the data and overlook inter-view correlations among missing samples. Furthermore, such consensus representations may diverge from the true latent structure of the original data. To address these limitations, this study proposes a novel framework known as consensus guided incomplete multi-view clustering via geometric consistency learning (CGIMC). CGIMC seamlessly integrates consensus representation learning and geometric consistency learning into a unified model through connectivity constraints. Specifically, it leverages consensus learning to capture latent data representations, while geometric consistency learning uncovers intrinsic local structures within the high-dimensional data space across views. Additionally, CGIMC adopts a one-step clustering strategy to yield final cluster assignments directly, thereby avoiding suboptimal post-processing steps. Extensive experiments conducted on multiple benchmark datasets demonstrate the superior clustering performance and robustness of the proposed CGIMC method. The source codes and datasets are available at https://github.com/whbdmu/CGIMC.