Multi-Scale Fuzzy Fusion-Based Heterogeneous Granular-Ball Flexible Representation Learning for Multi-View Feature...

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Representation learning is crucial for machine learning, bridging human cognition and data.
• Existing multi-view feature selection methods lack comprehensive integration and flexibility.
• Current methods are limited by raw-scale representations, hindering depth-breadth integration.

Purpose of the Study:

• To develop a novel flexible representation learning method for multi-view feature selection.
• To enhance feature selection by integrating multi-scale and multi-view data representations.
• To address limitations in depth-breadth integration and adaptability in existing methods.

Main Methods:

• Established a multi-view, multi-scale analytical foundation for representation learning.
• Developed adaptive heterogeneous granular ball-based flexible representations using multi-scale fuzzy fusion.
• Integrated granular-ball representation learning and feature selection into a unified one-step framework.
• Utilized fuzzy approximation operators to extract implicit fuzzy patterns and mitigate uncertainty.

Main Results:

• The proposed MFHGBFR method effectively captures intricate data manifolds and fuzzy patterns across multi-granularity spaces.
• Heterogeneous granular structures improved the adaptability of multi-granularity representations.
• The unified one-step framework achieved adaptive optimization for flexible adaptability.
• The derived optimization algorithm demonstrated proven convergence.

Conclusions:

• MFHGBFR demonstrates superior performance, robustness, and efficiency in multi-view feature selection.
• The method's adaptive nature and unified framework offer significant advantages over conventional approaches.
• Comparative experiments validate the quantitative and qualitative superiority of MFHGBFR against state-of-the-art algorithms.