Unified Contrastive Learning Framework Improves Multi-View Clustering on Incomplete and Noisy Data

Global: Unified Contrastive Learning Framework Improves Multi-View Clustering on Incomplete and Noisy Data

Researchers have introduced a new contrastive learning‑based approach for multi‑view clustering (MVC) that addresses the challenges of rare‑paired and mis‑paired samples in real‑world datasets. The method was detailed in a preprint posted to arXiv in December 2025 and aims to enhance clustering performance without requiring data imputation.

Problem Context

Multi‑view data often suffer from incompleteness or noise, leading to a scarcity of correctly paired samples (rare‑paired issue) and the presence of incorrectly matched pairs (mis‑paired issue). Both conditions can limit the ability of contrastive learning to extract complementary information across views.

Global‑Graph Guided Contrastive Learning

To mitigate the rare‑paired problem, the authors construct a global‑view affinity graph that connects all samples across views. This graph generates additional sample pairs, enabling the model to fully explore complementary information even when original pairings are missing.

Local‑Graph Weighted Contrastive Learning

To counteract the mis‑paired issue, a local‑graph weighted scheme is applied. By examining each sample’s nearest neighbors, the framework assigns pair‑wise weights that strengthen reliable contrasts and weaken potentially erroneous ones, guiding the learning process in the correct direction.

Integrated Framework

The two graph‑based components are combined into a unified global‑local contrastive learning architecture that operates without any data imputation steps, simplifying deployment on noisy or incomplete datasets.

Experimental Validation

Extensive experiments conducted on benchmark multi‑view datasets under both incomplete and noisy conditions demonstrate that the proposed method outperforms existing state‑of‑the‑art approaches, achieving higher clustering accuracy and robustness.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.