一类弱监督数据中多视角扰动的特征选择方法  

Feature selection via multi-view perturbation in a typeof weakly supervised data

作　　者：郭启航 王平心 杜亮 杨习贝[1,5] 钱宇华 GUO Qihang;WANG Pingxin;DU Liang;YANG Xibei;QIAN Yuhua(School of Computer,Jiangsu University of Science and Technology,Zhenjiang 212100,China;School of Science,Jiangsu University of Science and Technology,Zhenjiang 212100,China;Institute of Big Data Science and Industry,Shanxi University,Taiyuan 030006,China;School of Computer and Information Technology,Shanxi University,Taiyuan 030006,China;School of Economics and Management,Jiangsu University of Science and Technology,Zhenjiang 212100,China)

机构地区：[1]江苏科技大学计算机学院,镇江212100 [2]江苏科技大学理学院,镇江212100 [3]山西大学大数据科学与产业研究院,太原030006 [4]山西大学计算机与信息技术学院,太原030006 [5]江苏科技大学经济管理学院,镇江212100

出　　处：《江苏科技大学学报（自然科学版）》2024年第2期101-108,共8页Journal of Jiangsu University of Science and Technology:Natural Science Edition

基　　金：国家自然科学基金项目(62076111);国家自然科学基金重点项目(62136005)。

摘　　要：弱标签消歧技术可以用来消除数据中的噪声标签.然而,经由弱标签消歧后的数据中依然可能存在冗余或不相关特征,因此带来了弱监督数据中的特征选择这一实际问题.在弱标签消歧后得到的数据的基础上,提出了一种基于多视角扰动的特征选择框架,其能够分别从样本和特征多个视角出发,构造不同的扰动数据,以便求解出多个不同的特征选择结果,从而为后续的学习任务提供基础性集成工具.此外,所提的多视角扰动特征选择框架适用于不同类型、不同约束下的搜索进程.在12组高维数据上,通过注入5种不同比例的标签噪声和使用3种不同类型的特征度量准则,实验结果表明,所提方法求得的特征选择结果能够从准确率和稳定性的层面极大地提升分类性能.Technique of disambiguation of weak labels can be used to remove noisy labels for samples from data.However,redundant or irrelevant features may also be observed after disambiguation of weak labels,so the problem of feature selection should be paid much attention to in weakly supervised data.On the basis of the data with disambiguation of weak labels,a general feature selection framework via multi-view perturbation is developed,which can construct different perturbed data from both the levels of sample and feature.Consequently,multiple results of feature selection can be obtained,which provide a basic integration tool for the subsequent learning.The proposed framework can be applied to various forms and constraints of searching.On more than 12 sets of high-dimensional data,by injecting 5 ratios of label noise and using 3 criteria of feature evaluation,the experimental results demonstrate that the feature selection results obtained by our proposed method can significantly improve the classification performance from both the aspects of classification accuracy and classification stability.

关 键 词：特征选择 多视角 粗糙集 超集学习 弱监督 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]