EMLFCL: An Efficient Multilevel Fusion Contrastive Learning for Multiview Clustering.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Multiview clustering (MVC) using contrastive learning (CL) is a growing research area.
• Existing MVC methods struggle with inter-view coherence and robustness to noisy data.

Purpose of the Study:

• To propose an efficient multilevel fusion CL framework for MVC, named EMLFCL.
• To enhance the coherence between views at both feature and cluster representation levels.
• To improve robustness against noisy data in multiview clustering tasks.

Main Methods:

• Developed the EMLFCL framework integrating a shared multi-layer perceptron network (MNet) and a fusion network (FNet).
• Implemented a multilevel CL strategy comparing different views against an anchor view at feature and cluster levels.
• Eliminated view-specific private information and mitigated noisy view influence through the anchor-based comparison strategy.

Main Results:

• The proposed EMLFCL method significantly outperforms existing advanced MVC approaches.
• Achieved high clustering accuracies on eleven challenging multiview datasets.
• Demonstrated superior performance on Caltech datasets, reaching 66.4%, 74.7%, 82.3%, and 86.4% accuracies.

Conclusions:

• EMLFCL offers an effective solution for multiview clustering by improving inter-view coherence and data robustness.
• The multilevel fusion and anchor-based CL strategy are key to the model's enhanced performance.
• The framework shows strong potential for various real-world applications requiring robust multiview data analysis.