Multilevel Reliable Guidance for Unpaired Multiview Clustering

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Unpaired multiview clustering (UMC) presents a significant challenge due to the absence of paired samples across different data views.
• Existing incomplete multiview clustering (IMC) methods often fail in UMC scenarios as they rely on paired data to extract cross-view information.
• Mining consistent cluster structures across views is difficult when cluster confidence is low.

Purpose of the Study:

• To propose a novel method, multilevel reliable guidance for UMC (MRG-UMC), for effective joint clustering of unpaired multiview data.
• To enhance the reliability and confidence of cluster structures by integrating multilevel clustering and view guidance.
• To address the limitations of traditional methods in handling unpaired samples in multiview clustering.

Main Methods:

• Inner view multilevel clustering: Exploits high-confidence sample pairs across different levels to reduce the impact of boundary samples and enhance cluster structure confidence.
• Synthesized-view alignment: Utilizes a synthesized view to minimize cross-view discrepancies and promote data consistency.
• Cross-view guidance: Employs a reliable view guidance strategy to boost the clustering confidence of underperforming views.
• Joint optimization: Integrates these three modules across multiple levels for consistent and confident cluster structure learning.

Main Results:

• Theoretical analyses confirm the effectiveness of MRG-UMC in improving clustering confidence.
• Extensive experiments demonstrate that MRG-UMC surpasses current state-of-the-art UMC methods.
• MRG-UMC achieved an average Normalized Mutual Information (NMI) improvement of 12.95% on multiview datasets.

Conclusions:

• MRG-UMC provides a robust solution for unpaired multiview clustering by effectively mining consistent and confident cluster structures.
• The proposed method overcomes the limitations of traditional approaches that require paired data.
• MRG-UMC offers significant performance gains, establishing a new benchmark for UMC tasks.