Integrating label confidence-based feature selection for partial multi-label learning

doi:10.1016/j.patcog.2024.111281

Integrating label confidence-based feature selection for partial multi-label learning

2025-05-01

0

PRE

AI

DOI：10.1016/j.patcog.2024.111281

Original

Original request for help

Save

Abstract

En 中文

Partial Multi-Label Learning (PML) is an emerging learning paradigm that deals with candidate label sets containing false positive labels, facing the negative impacts of noisy labels and high-dimensional data. Existing methods evaluate label confidence in the original feature space, neglecting the negative impacts of ambiguous and redundant features. To tackle this issue, we propose a novel feature selection method, Label Confidence Feature Selection-Partial Multi-Label (LCFS-PML). This method establishes abetter mapping relationship by simultaneously optimizing both features and labels. First, label confidence is evaluated within the unique feature subspace of each label by combining the average distance between instances sharing the same label and the distance from the instance to the cluster center. Second, the optimized, more reliable labels are used to guide the optimization process of the feature space. During the alternating optimization between the feature and label spaces, LCFS-PML effectively mitigates the negative impacts of noisy labels, ambiguous features, and redundant features, ultimately identifying the optimal feature subset for each label. Comparative experiments on nine benchmark datasets show that the proposed method demonstrates significant superiority.

Keywords:

Partial multi-label learning

Feature selection

Label confidence

Integrating label confidence-based feature selection for partial multi-label learning

Abstract

Journal

Organization