Soft label collaborative view consistency enhancement with application to incomplete multi-view clustering

Abstract

Incomplete multi-view clustering (IMVC) is an unsupervised technique for clustering multi-view data when some view information is absent. However, most existing IMVC methods usually suffer from several significant challenges: (1) Inaccurate imputation or padding of missing data degrades clustering performance; (2) The ability to extract view features may decrease due to low-quality views, especially those that are inaccurately imputed. To overcome these challenges, in this paper, we introduce a novel IMVC framework, called soft label collaborative view consistency enhancement (SLC_CE). Firstly, we leverage the encoders of Transformers to construct a soft-label view information interaction module, which fully utilizes soft-labels to enhance view feature embeddings. Secondly, we employ soft labels to collaboratively impute missing features, addressing the incomplete multi-view data problem. Finally, we implement a consistency enhancement strategy across multi-level view features and soft labels to ensure high-quality feature extraction and imputation. Extensive experiments on several benchmark datasets demonstrate that the proposed SLC_CE method outperforms other state-of-the-art methods in real IMVC tasks.