Pseudo-Label Guided Structural Discriminative Subspace Learning for Unsupervised Feature Selection

Area of Science:

• Machine Learning
• Data Science
• Bioinformatics

Background:

• Traditional feature selection methods often perform stages independently, limiting adaptability.
• Existing methods face challenges with sparsity and parameter tuning, particularly when using L1-norm.
• Unsupervised feature selection is crucial for effective data clustering and analysis.

Purpose of the Study:

• To propose a novel unsupervised feature selection method, P উদ্বেগDSL.
• To unify probability graph construction and feature selection into a single framework.
• To enhance feature discrimination and stability for improved downstream clustering tasks.

Main Methods:

• Introduced probability graph construction into feature selection for adaptive learning.
• Developed a pseudo-label guided learning mechanism.
• Combined graph-based methods with maximizing between-class scatter using trace ratio.
• Employed L0-norm constraint for row sparsity and feature stability, addressing L1-norm limitations.

Main Results:

• Demonstrated the effectiveness of P উদ্বেগDSL on nine real-world datasets.
• Validated the method's performance on three biological single-cell RNA sequencing (ScRNA-seq) gene datasets.
• Achieved improved data clustering results compared to existing methods.

Conclusions:

• P উদ্বেগDSL offers a unified and adaptive framework for unsupervised feature selection.
• The method effectively improves feature discrimination and stability.
• P উদ্বেগDSL shows significant promise for enhancing data clustering in various applications, including bioinformatics.