Marginalized Multiview Ensemble Clustering.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Multiview clustering (MVC) seeks shared structures in data from multiple sources.
• Existing MVC methods often overlook higher-level information like basic partitions (BPs).

Purpose of the Study:

• To propose a novel marginalized multiview ensemble clustering (M²VEC) method.
• To effectively utilize complementary information and basic partitions for robust MVC.

Main Methods:

• M²VEC frames MVC as an ensemble clustering (EC) problem, generating and consolidating BPs.
• A marginalized denoising process with a single-layer autoencoder enhances partition robustness.
• Low-rank and sparse decomposition captures consistency and compensates for feature distinctness.
• Spectral consensus graph partitioning unifies the optimization framework.
• A multilayer M²VEC variant handles complex data by incorporating nonlinearity.

Main Results:

• Experimental results on eight real-world datasets demonstrate M²VEC's efficacy.
• The method shows superior performance compared to state-of-the-art MVC and EC techniques.
• M²VEC performs well even with partial multiview data.

Conclusions:

• M²VEC effectively leverages complementary information and BPs for robust multiview clustering.
• The proposed method offers a unified and adaptable framework for complex data analysis.
• M²VEC demonstrates strong performance and robustness in practical scenarios.