面向类不平衡数据的K近邻偏标记学习算法  

K-nearest neighbor based partial label learning algorithm for class imbalanced data

作　　者：王丽 于明仟 刘文鹏 周瑜 郑蕊蕊 贺建军 WANG Li;YU Mingqian;LIU Wenpeng;ZHOU Yu;ZHENG Ruirui;HE Jianjun(College of Information and Communication Engineering,Dalian Minzu University,Dalian 116000,Liaoning,China)

机构地区：[1]大连民族大学信息与通信工程学院,辽宁大连116000

出　　处：《山东大学学报（工学版）》2022年第3期18-24,41,共8页Journal of Shandong University（Engineering Science）

基　　金：国家自然科学基金资助项目(62102062,61972068);辽宁省自然科学基金资助项目(2020-MS-134,2020-MZLH-29,20180550625)。

摘　　要：针对于类不平衡的偏标记学习问题,在PL-KNN算法的基础上,提出一种可以较有效处理类不平衡问题的偏标记K近邻学习算法(K-nearest neighbor algorithm for class imbalanced partial label learning,IM-PLKNN),利用Parzen窗估计法在样本的不同类别的近邻上设置不同的权重,使多数类样本权重降低,让属于少数类样本的近邻具有更高的权重,降低将少数类样本误测为多数类样本的概率,提高对少数类样本的识别精度。试验结果表明,IM-PLKNN算法较PL-KNN算法在不同评价指标上均有显著提高,特别是对少数类样本的识别精度有大幅度提高。IM-PLKNN算法可以有效提高类不平衡的偏标记K近邻学习算法对数据集整体的预测性能。Towarding the class imbalanced partial label learning problem,IM-PLkNN was proposed to deal with the class imbalance problem more effectively.Based on PL-kNN algorithm,Parzen window estimation method was used to set different weights on the nearest neighbors of different classes of samples,so that the weights of the majority class samples were reduced and the nearest neighbours belonging to the minority class samples were given higher weights,reducing the probability of mismeasuring the minority class samples as the majority class samples and thus improving the accuracy of identifying the minority class samples.The experimental results showed that the IM-PLkNN algorithm had significantly improved over the PL-kNN algorithm in different evaluation indexes,especially the recognition accuracy of minority class samples had been substantially improved.IM-PLkNN algorithm could effectively improve the prediction performance of the class imbalanced partial labeled K-nearest neighbor learning algorithm for the dataset as a whole.

关 键 词：偏标记学习 类不平衡 K近邻分类 Parzen窗估计 代价敏感策略 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]