摘要
多标签学习已成为当前机器学习的研究热点。为了提高分类性能,对训练集中的噪声数据进行预处理,提出一种基于k近邻(kNN)的多标签分类去噪方法:对现有的多标签数据集进行分析后获得近似正态分布的特征,通过将噪声标记改为其k近邻标记的方法,滤去部分噪声信息,从而得到相对高质量的数据集。在MULAN平台上使用多个数据集对6种多标签分类算法进行了噪声去除前后的对比测试,实验结果表明,多标签的预处理方法有效提高了分类器的性能。此方法对于分布特征明显的数据集具有较好的适用性。
Multi-label learning is a new field in machine learning. In order to improve the multi-label classification precision, a new kNN method was used to remove the noise labels. First, a normal distribution is discovered by analyzing the characteristics of multi-label datasets, and then the high quality datasets are generated by changing the value of noisy labels to their k-Nearest Neighbors. In the experiments, six kinds of multi-label classification methods were tested on MULAN with new datasets. Compared to the primal datasets, the classification precision based on new datasets is better. Research results show this method is suitable for the data set which has a regular distribution.
出处
《计算机科学》
CSCD
北大核心
2015年第5期106-108,131,共4页
Computer Science
基金
浙江省教育厅项目(Y201328291)
浙江省自然科学基金项目(LZ14F030001
LY14F020012)资助
关键词
多标签
分类
正态分布
预处理
KNN
Multi-label, Classification, Normal distribution, Pretreatment, kNN
