文献类型：期刊文献

中文题名：基于矩阵的AdaBoost多视角学习

英文题名：Matrix-based multi-view learning with AdaBoost

作者：洪思思[1,2];曹辰捷[1];王喆[1,2];李冬冬[1]

机构：[1]华东理工大学信息科学与工程学院,上海200237;[2]江苏省计算机信息处理技术重点实验室,苏州大学,苏州215006

年份：2018

卷号：54

期号：6

起止页码：1152

中文期刊名：南京大学学报（自然科学版）

外文期刊名：Journal of Nanjing University（Natural Science）

收录：CSTPCD;;北大核心:【北大核心2017】；CSCD:【CSCD2017_2018】；

基金：国家自然科学基金(61672227,61806078);上海市教育发展基金会和上海市教育委员会“曙光计划”

语种：中文

中文关键词：多视角;集成学习;AdaBoost;矩阵分类器;正则化

外文关键词：multi-view;ensemble learning;AdaBoost;matrix-based classifier;regularization

摘要：以往基于矩阵的多视角分类(MultiV-MHKS)是从矩阵的单视角模式Ho-Kashyap(MatMHKS)发展而来的,尽管有些不好的视角可能会拉低分类器的总体表现,但它仍将所有视角都默认为相同的.为保证样本视角的有效性和多样性,提出一种名为AdaMultiV-MHKS的新方法,它充分利用了经典的集成学习方法AdaBoost的优势,实现了动态过滤和视角权重计算.所提的多视角框架不受限于任何特殊方法,可以应用于大多数基于矩阵的分类器.该方法加入了来自Alternative Robust Local Embedding(ARLE)的正则化项Rgl,用来增强样本之间的结构信息.集成工作和正则化使得不同视角的附加信息更具竞争力,不仅增强了视角的多样性而且改进了最终的分类结果.在UCI数据集和USTC-NVIE人脸图像数据集,尤其是液体数据集上的实验结果证明了所提方法的有效性.
The previous work about multi-view matrix-based classifier(MultiV-MHKS)has been developed from single-view matrix-pattern-oriented Ho-Kashyap(MatMHKS).It considers all views equal in default,although few poor views may drag down the overall performance of the classifier.Although multi-view can bring abundant information for the learning,treating improperly will have high risk of loss in accuracy.In order to ensure valid views of samples and enhance the diversity among all views,a new method called AdaMultiV-MHKS is proposed.It implements the dynamic filtering and the calculation of views’ weight by taking the advantage of the classic Ensemble Learning method called AdaBoost.The proposed multi-view framework is not restricted to any special method,and can be applied to most matrix-based classifiers.We select the GLMatMHKS as the base classifier.A further regularization term Rglderived from Alternative Robust Local Embedding(ARLE)is added to emphasize structural information among samples.The integration work and regularization make additional information from several views more competitive,which not only enhances the diversity among them but also improves the final classification results.The experimental results on UCI datasets and USTC-NVIE face images prove the effectiveness of integrating AdaBoostand ARLE techniques into the matrix-based multi-view learning,especially in the experiments of liquid datasets.