Relative Fuzzy Rough Approximations for Feature Selection and Classification.

Area of Science:

• Data Science
• Machine Learning
• Artificial Intelligence

Background:

• Classical fuzzy rough set (FRS) theory struggles with data distributions exhibiting significant class density differences.
• Existing methods for measuring data uncertainty are limited in scenarios with imbalanced data.

Purpose of the Study:

• To propose a novel relative fuzzy rough set (FRS) model capable of handling data with varying class densities.
• To develop an effective feature selection algorithm and a classifier based on the enhanced FRS model.

Main Methods:

• A relative distance measure was introduced to accommodate differing data distribution densities.
• A relative FRS model was developed, defining positive, negative, and boundary regions for uncertainty measurement.
• Relative fuzzy dependency was defined for feature importance evaluation, leading to a feature selection algorithm and a maximal positive region-based classifier.

Main Results:

• The relative fuzzy dependency measure proved effective and efficient for feature evaluation.
• The proposed feature selection algorithm demonstrated superior performance compared to classical algorithms.
• The maximal positive region-based classifier achieved performance comparable to, and slightly better than, the KNN classifier.

Conclusions:

• The relative fuzzy rough set model effectively addresses limitations of classical FRS in imbalanced data scenarios.
• The proposed feature selection and classification methods offer a feasible and effective approach for data analysis and machine learning.