摘要
为了提升标记分布学习在歧义性分类问题上的预测性能,对标记形态上的模糊度进行研究。提出了标记形态模糊度的概念,给出了基于峰度的一种度量方式,探讨了不同模糊程度样本对于分类学习的影响。根据低模糊度数据更有利于学习的研究发现,基于加权低模糊度样本和对齐模糊度损失这两种策略设计了一种新的标记分布学习算法。14个数据集上的十折交叉试验表明,该文提出的标记分布学习模糊度量化标记算法在各种数据集上都能够降低模糊度损失并提升预测精度,具有良好的鲁棒性。
This paper delves into the fuzziness of label morphology to improve the prediction of label distribution learning on classification problems with ambiguity.First,this paper proposes the concept of label morphological ambiguity and gives a measure based on kurtosis.Then this paper explores the effect of different vagueness samples on classification learning through rich experiments.Finally,based on the finding that the low ambiguity samples are more beneficial for classification learning,this paper designs a new algorithm for label distribution learning by weighting low ambiguity samples and aligning ambiguity loss.Ten-fold crossover experiments on 14 datasets show that the proposed algorithm LDL-KQA can reduce ambiguity loss and improve prediction accuracy with good robustness.
作者
许哲源
秦天
Xu Zheyuan;Qin Tian(School of Computer Science and Engineering,Nanjing University of Scienceand Technology,Nanjing 210094,China)
出处
《南京理工大学学报》
CAS
CSCD
北大核心
2022年第1期89-97,共9页
Journal of Nanjing University of Science and Technology
基金
江苏省自然科学基金(BK20191287)。
关键词
标记分布学习
标记形态模糊度
多标记学习
label distribution learning
label morphological ambiguity
multi-label learning
