Semi-supervised multi-label feature selection with consistent sparse graph learning

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• High-dimensional data often possess multiple semantic labels, posing challenges for traditional single-label feature selection methods.
• Existing multi-label feature selection methods struggle in semi-supervised settings, particularly with limited labeled samples and suboptimal similarity graph construction.

Purpose of the Study:

• To propose a consistent sparse graph learning method (SGMFS) for multi-label semi-supervised feature selection.
• To enhance feature selection by maintaining space consistency and learning label correlations in semi-supervised scenarios.

Main Methods:

• SGMFS learns a low-dimensional label subspace to capture label correlations from projected features.
• It adaptively learns a similarity graph by simultaneously performing sparse reconstruction in both the label space and the learned subspace.
• An efficient optimization solution with fast convergence is employed.

Main Results:

• The proposed SGMFS method effectively addresses the limitations of existing semi-supervised multi-label feature selection techniques.
• It demonstrates superior performance in learning label correlations and constructing reliable similarity graphs.
• Experimental results validate the effectiveness and superiority of SGMFS.

Conclusions:

• SGMFS offers a robust framework for multi-label semi-supervised feature selection.
• The method enhances feature selection by improving the handling of label correlations and similarity graph construction.
• SGMFS shows significant potential for practical applications involving high-dimensional, multi-label data.