Partial multi-label learning via label-specific feature corrections  

作　　者：Jun-Yi HANG Min-Ling ZHANG 

机构地区：[1]School of Computer Science and Engineering,Southeast University,Nanjing 210096,China [2]Key Laboratory of Computer Network and Information Integration(Southeast University),Ministry of Education,Nanjing 210096,China

出　　处：《Science China(Information Sciences)》2025年第3期91-105,共15页中国科学(信息科学)(英文版)

基　　金：supported by National Natural Science Foundation of China(Grant No.62225602);Big Data Computing Center of Southeast University。

摘　　要：Partial multi-label learning(PML)allows learning from rich-semantic objects with inaccurate annotations,where a set of candidate labels are assigned to each training example but only some of them are valid.Existing approaches rely on disambiguation to tackle the PML problem,which aims to correct noisy candidate labels by recovering the ground-truth labeling information ahead of prediction model induction.However,this dominant strategy might be suboptimal as it usually needs extra assumptions that cannot be fully satisfied in real-world scenarios.Instead of label correction,we investigate another strategy to tackle the PML problem,where the potential ambiguity in PML data is eliminated by correcting instance features in a label-specific manner.Accordingly,a simple yet effective approach named P_(ASE),i.e.,partial multi-label learning via label-specific feature corrections,is proposed.Under a meta-learning framework,P_(ASE)learns to exert label-specific feature corrections so that potential ambiguity specific to each class label can be eliminated and the desired prediction model can be induced on these corrected instance features with the provided candidate labels.Comprehensive experiments on a wide range of synthetic and real-world data sets validate the effectiveness of the proposed approach.

关 键 词：machine learning multi-label learning partial multi-label learning label-specific features feature correction 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]