Multilabel Feature Selection With Constrained Latent Structure Shared Term.

Area of Science:

• Machine Learning
• Data Science
• Computer Science

Background:

• High-dimensional multilabel data presents challenges due to numerous features and labels.
• Existing methods often overlook the influence of latent feature structure on label correlations.

Purpose of the Study:

• To develop a novel multilabel feature selection method that incorporates latent feature and label structures.
• To address the limitations of previous methods by considering the impact of latent feature structure.

Main Methods:

• A Latent Structure Shared (LSS) term was designed to preserve both latent feature and label structures.
• Graph regularization was employed to ensure consistency between the original feature space and the latent structure space.
• The Shared Latent Feature and Label Structure Feature Selection (SSFS) method was derived and optimized.

Main Results:

• The proposed SSFS method demonstrated superior performance on benchmark datasets.
• Experimental results showed improvements across multiple evaluation criteria.
• The method effectively leverages shared latent structures for enhanced feature selection.

Conclusions:

• The SSFS method offers an effective approach for feature selection in high-dimensional multilabel learning.
• Incorporating shared latent structures significantly improves the performance of multilabel feature selection.
• The proposed optimization scheme ensures efficient and convergent solutions.