Multiview Clustering With Propagating Information Bottleneck.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Hierarchical multiview (HMV) data, originating from multiple sources with multiple views, presents unique clustering challenges.
• Existing multiview clustering (MVC) methods often fail to leverage relationships across both sources and views simultaneously.
• Effective clustering of HMV data requires models that can integrate multivariate information and dynamic information flow.

Purpose of the Study:

• To develop a general hierarchical information propagation model for clustering HMV data.
• To propose a novel self-guided method, propagating information bottleneck (PIB), to realize this model.
• To address the limitations of existing MVC methods in handling complex HMV data structures.

Main Methods:

• A hierarchical information propagation model is constructed, encompassing optimal feature subspace learning (OFSL) and clustering structure learning (CSL).
• The propagating information bottleneck (PIB) method is introduced, utilizing a circulating propagation fashion for self-guided learning.
• A two-step alternating optimization strategy is employed for efficient model optimization.

Main Results:

• The PIB method effectively integrates information across multiple sources and views in HMV data.
• Theoretical analysis confirms the relationship between learned cluster structures and propagated information.
• Experimental results demonstrate the superiority of PIB over existing state-of-the-art clustering methods.

Conclusions:

• The proposed hierarchical information propagation model and PIB method offer a powerful solution for HMV data clustering.
• PIB's self-guided approach enhances the utilization of multivariate and dynamic information.
• The method shows significant performance improvements on various benchmark datasets.