Group-preserving label-specific feature selection for multi-label learning
Contribution
Preserves label-group structure while selecting label-specific features.
Abstract
In multi-label learning, features that matter for a group of related labels are often systematically different from features that matter for an isolated label. This paper introduces group-preserving label-specific feature selection, which identifies label clusters and selects features jointly tailored to each group while preserving the cluster structure. Experiments across multi-label benchmarks show consistent gains over both global and label-specific baselines, particularly when label relationships are structured rather than independent.
Why This Paper Matters
Targets the realistic case where labels come in groups (genres, disease categories, taxonomic levels) rather than as independent binary targets — directly useful for hierarchical and structured multi-label applications.
BibTeX
@article{zhang2023group,
title={Group-preserving label-specific feature selection for multi-label learning},
author={Zhang, Jia and Wu, Hanrui and Jiang, Min and Liu, Jinghua and Li, Shaozi and Tang, Yong and Long, Jinyi},
journal={Expert Systems with Applications},
volume={213},
pages={118861},
year={2023},
publisher={Elsevier},
}