山西大学学报（自然科学版）2024，Vol.47Issue(6)：1164-1177,14.DOI:10.13451/j.sxu.ns.2023087
基于实例的近邻传播偏标签学习算法
Instance-based Nearest Neighbor Propagation Based Partial Label Learning Method
摘要
Abstract
The candidate label set generation method for most partial label learning methods does not make good use of the reliable prior information of samples,consequently,the probably resulted trained models are unsuitable for realistic situations where some samples have confusing candidate labels.In addition,many partial label learning algorithms only use nearest neighbor nodes to con-struct graphs,but neglect the importance of reliable sample information.A new instance-based nearest neighbor propagation partial label learning method(INNPL)is proposed to address the problem of how to generate more realistic candidate label sets and effec-tively utilize the information from highly reliable samples.In terms of candidate label sets generation,INNPL takes advantages of the prior information.For high-reliability samples,INNPL selects them by calculating the similarity between pairwise samples in the same class,and trusts the original labels of typical samples.For other samples,their candidate labels are generated based on the simi-larity between themselves and the typical sample center.INNPL also iteratively builds and updates graph structures based on nearest neighbor samples and highly reliable samples,and progressively propagates labels in layers over the samples to finally obtain reli-able classification results.The validation of INNPL algorithm is tested based on three UCI((University of California Irvine))datas-ets and two image datasets and compared with seven partial label learning methods.The evaluation results of the four classification metrics including accuracy,precision,recall and F1 coefficient show that INNPL achieves the best performance on three datasets and is close to the optimum algorithm on the other two.Regarding the accuracy metric,the results of INNPL compared with any of the other partial label learning algorithms by paired t-test also confirm the superior performance of INNPL.Our method ranks first in the overall metric of the four classification metrics.In conclusion,INNPL confirms its effectiveness from several experiments,which can effectively improve the classification accuracy and also provides a new solution idea for partial label learning.关键词
偏标签学习/图/基于实例的近邻传播/分类Key words
partial label learning/graph/instance-based nearest neighbor propagation/classification分类
信息技术与安全科学引用本文复制引用
李博,熊天龙,杜宇慧..基于实例的近邻传播偏标签学习算法[J].山西大学学报（自然科学版）,2024,47(6):1164-1177,14.基金项目
国家自然科学基金(62076157 （）
61703253) （）
山西省留学人员科技活动择优资助项目(20210033) （20210033）
