Feature Selection Based on Neighborhood Self-Information.

Area of Science:

• Data Science
• Artificial Intelligence
• Machine Learning

Background:

• Traditional dependency measures in neighborhood rough set models often overlook crucial information within the upper approximation.
• This limitation can hinder effective feature selection and impact classification performance.

Purpose of the Study:

• To develop novel uncertainty measures for feature selection that incorporate information from both lower and upper approximations.
• To introduce and evaluate 'relative neighborhood self-information' as a superior measure for feature selection.

Main Methods:

• Construction of a class of uncertainty measures, specifically decision self-information, for feature selection.
• Detailed analysis of the relationships and properties of these proposed measures.
• Design of a greedy algorithm for feature selection utilizing the proposed measures.

Main Results:

• Relative neighborhood self-information is identified as a more effective measure due to its consideration of both approximations and sensitivity to feature subset variations.
• The proposed greedy algorithm demonstrates effectiveness in selecting fewer features.
• The algorithm frequently enhances classification accuracy across various datasets.

Conclusions:

• The developed uncertainty measures, particularly relative neighborhood self-information, offer significant improvements for feature selection in neighborhood rough set models.
• The proposed greedy algorithm provides a practical and effective approach for optimizing feature subsets and boosting classification performance.