Fast and Robust Attribute Reduction Based on the Separability in Fuzzy Decision Systems

Area of Science:

• Machine Learning
• Data Mining
• Pattern Recognition

Background:

• Attribute reduction is crucial for efficient machine learning and data mining.
• Existing attribute evaluation functions are computationally expensive due to object-object relationships.
• There is a need for more efficient attribute evaluation methods.

Purpose of the Study:

• To propose a novel, computationally efficient separability-based attribute evaluation function and reduction method.
• To introduce a new algorithm, Separability-based Sequential Forward Selection (SFSS), for attribute selection.
• To enhance classification performance and reduce computational costs in data preprocessing.

Main Methods:

• Defined Degree of Aggregation (DA) for intraclass objects and Degree of Dispersion (DD) for between-class objects.
• Developed a novel separability measure using DA and DD for attribute subsets in fuzzy decision systems.
• Designed the Sequentially Forward Selection based on Separability (SFSS) algorithm with a postpruning strategy.

Main Results:

• The SFSS algorithm demonstrated significantly lower computational time compared to typical reduction algorithms.
• SFSS achieved higher classification accuracy and compression ratios across public datasets (UCI, ELVIRA).
• The method proved to be fast, robust, and interpretable, as shown on the MNIST dataset.

Conclusions:

• The proposed separability-based evaluation function and SFSS algorithm offer an efficient and effective approach to attribute reduction.
• SFSS overcomes the computational limitations of existing methods by directly using object-category relationships.
• This method holds promise for improving the performance and efficiency of machine learning and data mining tasks.