Feature selection with neighborhood entropy-based cooperative game theory

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• High-dimensional datasets pose challenges for traditional feature selection.
• Existing methods may overlook features with strong group classification ability but weak individual performance.

Purpose of the Study:

• To address limitations in current feature selection techniques.
• To propose a novel method that captures group feature contributions.

Main Methods:

• Redefining feature redundancy, interdependence, and independence using neighborhood entropy.
• Developing a neighborhood entropy-based feature contribution metric within a cooperative game framework.
• Formalizing feature evaluation criteria as a product of contribution and classical measures.

Main Results:

• The Neighborhood Entropy-based Cooperative Game Theory (NECGT) model was evaluated on UCI datasets.
• NECGT demonstrated superior performance compared to classical feature selection methods.
• The method effectively identifies features valuable as a collective.

Conclusions:

• The proposed NECGT model offers an effective solution for feature selection in high-dimensional data.
• Neighborhood entropy and cooperative game theory provide a robust framework for evaluating feature contributions.
• NECGT enhances the identification of synergistic feature subsets.