Multilabel Feature Selection via Shared Latent Sublabel Structure and Simultaneous Orthogonal Basis Clustering.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• High-dimensional multilabel data presents challenges for feature selection due to noisy and incomplete labels.
• Existing methods map label spaces to low-dimensional latent spaces but often retain redundant information.
• This redundancy in latent label spaces can negatively impact the accurate capture of label relevance.

Purpose of the Study:

• To propose a novel method, SLOFS, for multilabel feature selection that addresses the issue of redundant information in latent label spaces.
• To develop a technique that effectively extracts latent label correlations by eliminating redundant information.
• To improve the performance of feature selection in high-dimensional multilabel datasets.

Main Methods:

• Introduced a Latent Orthogonal Base Structure Shared (LOBSS) term to create a redundancy-free latent sublabel space.
• Utilized separated latent clustering centers to guide the LOBSS term, retaining both sublabel information and clustering structure.
• Employed graph regularization for structural consistency between data and latent sublabels, and a dynamic sublabel graph for high-quality sublabel space construction.

Main Results:

• The proposed SLOFS method effectively eliminates redundant information from the latent label space.
• The LOBSS term successfully guides the construction of a nonredundant latent sublabel space.
• Experimental results on 18 datasets show SLOFS consistently outperforms existing feature selection methods.

Conclusions:

• SLOFS offers a significant advancement in multilabel feature selection by effectively managing redundant information.
• The method enhances the extraction of latent label correlations, leading to more accurate feature selection.
• SLOFS demonstrates superior and consistent performance compared to previous approaches on diverse datasets.