Feature selection for label distribution learning based on embedding mutual information optimization

doi:10.1007/s13042-026-03221-2

Return

Feature selection for label distribution learning based on embedding mutual information optimization

2026-07-22

PRE

DOI：10.1007/s13042-026-03221-2

Original

Original request for help

Save

Abstract

En 中文

Label distribution learning is an efficient learning paradigm for handling label ambiguity issues, yet it still faces challenges posed by high-dimensional feature spaces. Embedded feature selection algorithms represented by sparse learning are an effective way to address this challenge. However, these methods often overlook the complex nonlinear dependencies between features and labels, as well as the inherent redundancy among features. To address these issues, in this paper, a feature selection method is proposed for label distribution learning based on embedded mutual information optimization (FSEMI). This method employs mutual information as a theoretical tool to systematically assess the correlations between features and between features and labels, using this as prior knowledge to guide the model in maximizing feature-label dependencies while effectively suppressing feature redundancy. Additionally, the method incorporates the $$L_{2,1}$$ -norm to identify key features shared across labels. Extensive experiments conducted on twelve label distribution datasets demonstrate that the proposed method outperforms current state-of-the-art methods for feature selection in label distribution learning.

Keywords:

Label distribution learning

Mutual information

$L_{2,1}$-norm

Feature selection