Bidirectional Probabilistic Multi-Graph Learning and Decomposition for Multi-View Clustering.

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Graph-based multi-view clustering shows promise but suffers from unidirectional pipelines and insufficient prior information.
• Existing methods often fail to align learned graphs with data structures due to limitations in graph learning and indicator generation.

Purpose of the Study:

• To propose a Bidirectional Probabilistic Multi-graph Learning and Decomposition (BPMLD) method for multi-view clustering.
• To establish an explicit bidirectional pipeline between graph learning and indicator generation for improved clustering accuracy.
• To address the limitations of unidirectional frameworks and inadequate prior information in current graph-based clustering techniques.

Main Methods:

• Developed a confidence term based on clustering probability indicators to drive graph learning.
• Introduced graph tensor learning to capture high-order correlations among refined graphs.
• Proposed a multi-graph probability decomposition module for adaptive cluster indicator generation with probabilistic representation.

Main Results:

• The seamless integration between graph learning and indicator generation allows for mutual enhancement.
• BPMLD effectively aligns learned graphs with underlying data structures through bidirectional interaction.
• Extensive experiments validate the superior performance of BPMLD against state-of-the-art methods.

Conclusions:

• The proposed BPMLD method significantly improves multi-view clustering by enabling bidirectional interaction between graph learning and indicator generation.
• The integration of clustering confidence, graph tensor learning, and probabilistic decomposition offers a robust approach to multi-view clustering.
• The method demonstrates effectiveness and outperforms existing approaches in extensive experimental evaluations.