基于kNCN的标签噪声在线核学习方法  

An online kernel learning method for label noise based on kNCN

作　　者：李世鑫 文益民[1] LI Shixin;WEN Yimin(Guangxi Key Laboratory of Image and Graphic Processing,Guilin University of Electronic Technology,Guilin 541004,China)

机构地区：[1]桂林电子科技大学广西图像与图形智能处理重点实验室,广西桂林541004

出　　处：《桂林电子科技大学学报》2024年第6期551-559,共9页Journal of Guilin University of Electronic Technology

基　　金：广西重点研发计划(桂科AB21220023);国家自然科学基金(61866007);广西图像图形与智能处理重点实验室基金(GIIP2005)。

摘　　要：核方法被开发用于处理在线分类中的非线性分类问题,在计算核函数时,为了避免支持向量的数量随数据流而无限增加,近年来发展了预算维护算法。现有的固定预算的核分类算法在分类性能上会受标签噪声的严重影响。为了解决该问题,提出一种基于kNCN的标签噪声在线核学习方法。当缓冲区达到预算规模时,该方法利用kNCN原则,为缓冲区内每个支持向量找到k个近质心近邻点,通过计算它们之间的局部标签不一致性来构建删除候选集和锚点集,再对删除候选集内所有实例建立试错模型,在锚点集上检验试错模型的分类准确率,从而判断哪个支持向量最应该从缓冲区中删除,实现固定预算的维护。人工合成数据集和真实数据集上的实验结果表明,本方法在固定预算的感知机和被动攻击算法上的应用中,在标签噪声场景下有效提升了分类性能,在6个数据集上的综合排名优于其他对比算法。Kernel methods have been developed to handle nonlinear classification problems in online classification.In recent years,budget maintenance algorithms have been developed to avoid the infinite increase in the number of support vectors with data flow when calculating kernel functions.The existing fixed budget kernel classification algorithms are severely affected by label noise in classification performance.To address this issue,a label noise online kernel learning method based on kNCN is proposed.When the buffer reaches the budget size,this method uses the kNCN principle to find k near centroid nearest neighbor points for each support vector in the buffer.Then,by calculating the local label inconsistency between them,the deletion candidate set and anchor set are constructed.Then,a trial and error model is established for all instances in the deletion candidate set,and the classification accuracy of the trial and error model is tested on the anchor set to determine which support vector should be most removed from the buffer,maintain a fixed budget.Experimental results on synthetic data sets and real data sets show that the application of this method on fixed-budget perceptrons and passive attack algorithms can effectively improve classification performance in label noise scenarios.The comprehensive results on six data sets Ranking is better than other comparison algorithms.

关 键 词：预算维护 近质心近邻 核方法 在线分类 标签噪声 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]