Partition level multiview subspace clustering.

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Multiview clustering leverages complementary information from multiple data sources (views).
• Existing multiview subspace clustering methods integrate information in the data space, making them susceptible to noise and feature inconsistencies.
• A core assumption in multiview clustering is the existence of a shared partition across all views.

Purpose of the Study:

• To propose a novel multiview clustering method that fuses information at the partition level, rather than the data space.
• To address limitations of existing methods, such as noise sensitivity and feature heterogeneity.
• To develop a unified model integrating graph learning, partition generation, and consensus fusion.

Main Methods:

• Proposed a partition-level fusion strategy based on two assumptions: partitions are perturbations of consensus clustering, and closer partitions receive higher weights.
• Developed a unified multiview subspace clustering model.
• Integrated graph learning from individual views, generation of base partitions, and fusion into a consensus partition, optimized iteratively.

Main Results:

• The proposed unified model demonstrated superior performance compared to state-of-the-art techniques.
• Experiments on four benchmark datasets validated the efficacy of the partition-level fusion approach.
• The iterative boosting of integrated components led to an overall optimal solution.

Conclusions:

• Partition-level fusion is an effective strategy for multiview clustering, offering robustness against noise and feature inconsistencies.
• The proposed unified model provides a powerful framework for leveraging multiview data.
• The method shows significant potential for applications requiring robust clustering of heterogeneous data.