Discovering Multiple Co-Clusterings With Matrix Factorization.

Area of Science:

• Data Mining
• Machine Learning
• Computational Statistics

Background:

• Traditional clustering methods produce single partitions, often missing alternative meaningful structures in complex data.
• Existing multiple clustering approaches primarily focus on one-way clustering (samples or features).
• There is a need for methods that explore multiple two-way clusterings (co-clusterings) for richer data insights.

Purpose of the Study:

• To introduce a novel approach, multiple co-clusterings (MultiCCs), for generating multiple alternative co-clusterings simultaneously.
• To address the limitation of existing methods by enabling simultaneous exploration of sample and feature cluster structures.
• To develop a method that captures diverse and meaningful co-clustering patterns.

Main Methods:

• Utilizes matrix tri-factorization to identify co-clustering indicator matrices for samples and features.
• Incorporates row and column redundancy quantification terms to ensure diversity among co-clusterings.
• Integrates matrix tri-factorization with nonredundancy terms into a unified objective function, optimized via an alternative procedure.

Main Results:

• MultiCCs significantly outperforms existing multiple clustering methods in experimental evaluations.
• The approach successfully identifies interesting co-clusters that are missed by comparative methods.
• Demonstrates the effectiveness of simultaneous sample and feature co-clustering for data exploration.

Conclusions:

• Multiple co-clusterings (MultiCCs) offer a powerful framework for discovering diverse and meaningful data structures.
• The proposed method advances the field of multiple clustering by enabling simultaneous two-way clustering.
• MultiCCs provides a valuable tool for comprehensive data exploration in complex datasets.