Double-Structured Sparsity Guided Flexible Embedding Learning for Unsupervised Feature Selection.

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Unsupervised feature selection is crucial for dimensionality reduction in high-dimensional data.
• Existing methods often face limitations with sparsity and parameter tuning.
• The l2,0-norm constraint offers potential advantages over the l2,1-norm but presents optimization challenges.

Purpose of the Study:

• To propose a novel unsupervised feature selection model, Double-Structured Sparsity guided Flexible Embedding Learning (DSFEL).
• To address the optimization challenges associated with the l2,0-norm constraint for feature selection.
• To enhance clustering structure learning and distinctive feature selection.

Main Methods:

• DSFEL integrates a block-diagonal structural sparse graph learning module for clustering.
• It employs a completely row-sparse projection matrix with an l2,0-norm constraint for feature selection.
• An efficient optimization strategy is developed to solve the non-convex and non-smooth l2,0-norm problem, yielding a closed-form solution.

Main Results:

• DSFEL effectively learns clustering structures and selects distinctive features.
• The proposed optimization strategy provides an exact solution, overcoming limitations of previous approximate methods.
• Experimental results on nine real-world datasets demonstrate superior performance compared to state-of-the-art methods.

Conclusions:

• DSFEL offers a powerful and efficient approach to unsupervised feature selection.
• The novel optimization strategy for the l2,0-norm constraint is a significant contribution.
• The method shows strong potential for applications requiring effective dimensionality reduction without labeled data.