Efficient and Stable Unsupervised Feature Selection Based on Novel Structured Graph and Data Discrepancy Learning.

Area of Science:

• Data Mining
• Machine Learning
• Pattern Recognition

Background:

• Unsupervised feature selection is crucial when data labels are unavailable but class numbers are known.
• Existing structured graph methods for feature selection face challenges in maintaining component count and hyperparameter tuning.
• These limitations hinder the effectiveness of traditional graph-based unsupervised feature selection.

Purpose of the Study:

• To propose an efficient and stable unsupervised feature selection method (ESUFS).
• To address the limitations of existing structured graph learning algorithms in unsupervised feature selection.
• To enhance the accuracy, stability, and speed of feature selection in machine learning.

Main Methods:

• Developed a novel structured graph incorporating a pairwise data similarity matrix and an indicator matrix.
• Employed data discrepancy learning to identify features that maximize data differences.
• Utilized a discrete optimization problem for efficient learning of the structured graph.

Main Results:

• The proposed ESUFS method demonstrates superior performance compared to state-of-the-art techniques.
• ESUFS achieves higher accuracy (ACC) in feature selection tasks.
• Experiments confirm ESUFS's enhanced stability and computational speed.

Conclusions:

• ESUFS offers an effective solution for unsupervised feature selection challenges.
• The novel structured graph and data discrepancy learning contribute to improved feature selection outcomes.
• ESUFS is a promising method for practical applications in data mining and machine learning.