Bootstrap-based K-means feature selection strategy for fuzzy regression functions

Area of Science:

• Data Science
• Machine Learning
• Statistics

Background:

• Traditional feature selection methods struggle with complex datasets.
• Need for stable and robust feature selection alternatives is growing.
• Existing methods can be sensitive to random initialization and lack generalizability.

Purpose of the Study:

• Propose a novel bootstrap-stabilized feature selection method.
• Combine stable feature selection with Fuzzy Regression Function (FRF) modeling.
• Evaluate the proposed method's effectiveness on real-world datasets.

Main Methods:

• Utilized k-means clustering for grouping features across multiple resamples.
• Identified stable representative features by assigning them to nearest cluster centers.
• Integrated selected features into a Fuzzy Regression Function (FRF) model.
• Determined the number of feature clusters using grid-search, Silhouette Index (SI), and Davies Bouldin Index (DBI).

Main Results:

• The proposed bootstrap-stabilized clustering approach demonstrated superior or competitive performance.
• Effectiveness was validated across ten diverse real-world datasets.
• The hybrid framework combines interpretable feature selection with flexible fuzzy regression.

Conclusions:

• The bootstrap-stabilized clustering method is effective for feature selection in complex data.
• The approach offers improved stability and generalizability over traditional methods.
• This hybrid framework is suitable for complex data environments requiring robust feature selection.