Abstract
Incomplete multi-view clustering (IMVC) seeks to harness complementary information across multiple views while tackling the challenges inherent in missing data. Modeling inter-sample relationships can mitigate the impact of missing data, but it is computationally expensive. Anchor-graph-based IMC methods offer a more efficient alternative by constructing graphs with a small set of anchors rather than all samples. Nevertheless, existing methods suffer from some limitations that compromise their effectiveness. First, most of them learn a common structure using only available data from observed views, thereby neglecting valuable but unobserved structure hidden in missing samples. Second, their dependence on separate post-processing steps to obtain labels creates a multi-step pipeline that not only constrains the model's capacity to learn latent information but also introduces the disjoint optimization issues. To address these issues, we propose One-step Incomplete Multi-view Clustering via Imputed Anchor Graph Dscretization (OIMC-IAGD). Our method introduces a novel anchor graph imputation mechanism to integrate both observed and inferred structural information, and a one-step framework that simultaneously performs anchor graph imputation and discrete label acquisition. Comprehensive experiments in multiple datasets demonstrate that OIMC-IAGD consistently outperforms state-of-the-art baselines, achieving a superior balance between clustering effectiveness and computational efficiency. The demo code of this work is publicly available at https://github.com/WuShisi/OIMC-IAGD-master
| Original language | English |
|---|---|
| Article number | 114935 |
| Journal | Knowledge-Based Systems |
| Volume | 333 |
| DOIs | |
| State | Published - 30 Jan 2026 |
Keywords
- Anchor graph discretization
- Graph imputation
- Incomplete multi-view clustering
- One-step clustering
Fingerprint
Dive into the research topics of 'One-step incomplete multi-view clustering via imputed anchor graph discretization'. Together they form a unique fingerprint.Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver