A General Framework for Auto-Weighted Feature Selection via Global Redundancy Minimization.

Area of Science:

• Machine Learning
• Data Science
• Computer Science

Background:

• Existing feature selection methods often overlook feature redundancy, leading to correlated features and compromised performance in classification and clustering.
• This redundancy issue necessitates advanced techniques for selecting truly representative and non-redundant features.

Purpose of the Study:

• To propose novel auto-weighted feature selection frameworks, AGRM and C-AGRM, designed for global redundancy minimization.
• To enhance the selection of representative and non-redundant features by addressing the limitations of existing methods.

Main Methods:

• Introduction of the auto-weighted feature selection framework via global redundancy minimization (AGRM).
• Extension of AGRM to a more concise and efficient compact framework (C-AGRM).
• Development of parameter-free, auto-weighted frameworks applicable as post-processing systems for existing feature selection methods.

Main Results:

• AGRM and C-AGRM effectively reduce feature redundancy from a global perspective, enabling the selection of representative features.
• Extensive experiments on nine benchmark datasets demonstrated the effectiveness and superiority of the proposed frameworks.
• The auto-weighted nature makes the frameworks pragmatic for real-world applications.

Conclusions:

• The proposed AGRM and C-AGRM frameworks offer a significant advancement in feature selection by effectively minimizing redundancy.
• These parameter-free methods can refine feature scores, improving performance in both supervised and unsupervised learning tasks.
• The frameworks provide a robust solution for obtaining non-redundant feature subsets, enhancing machine learning model efficacy.