近邻边界Fisher判别分析.pdfVIP

第31 卷第3 期 电 子 与 信 息 学 报 Vol.31No.3 2009 年3 月 Journal of Electronics Information Technology Mar.2009 近邻边界 Fisher 判别分析 魏 莱① 王守觉② 徐菲菲① 王睿智① ①(同济大学计算机科学与技术系 上海 201804) ②(中国科学院半导体所 北京 100083) 摘 要：将数据集进行合理的维数约简对于一些机器学习算法效率的提高起着至关重要的影响。该文提出了一种利 用数据点邻域信息的线性监督降维算法：近邻边界 Fisher 判别分析(Neighborhood Margin Fisher Discriminant Analysis，NMFDA)。NMFDA 尝试将每一数据点邻域内最远的同类数据点和最近的异类数据点之间的边界在投影 子空间内尽可能地扩大，从而提高基于距离的识别算法的准确率。同时为了解决非线性降维问题，提出了 Kernel NMFDA，通过在几个标准人脸数据库上与其它降维算法的对比识别实验，验证了提出算法的有效性。 关键词：维数约简；流形学习；主成份分析；Fisher 判别分析；人脸识别 中图分类号：TP181 文献标识码：A 文章编号：1009-5896(2009)03-0509-05 Neighborhood Margin Fisher Discriminant Analysis ① ② ① ① Wei Lai Wang Shou-jue Xu Fei-fei Wang Rui-zhi ① (Department of Computer Science and Technology, Tongji University, Shanghai 201804, China) ② (Institute of Semiconductor, Chinese Academy of Sciences, Beijing 100083, China) Abstract: The curse of high dimensionality is usually a major cause of limitations of many machine learning algorithms. A novel algorithm called Neighborhood Margin Fisher Discriminant Analysis (NMFDA) is proposed for supervised linear dimensionality reduction. For every point, NMFDA tries to enlarge the margin of the farthest point with the same class label and the nearest point with the different class label. Also the Kernel NMFDA is proposed for nonlinear dimensionality reduction. The contrastive experiments on several benchmark face database show the effectiveness of proposed method. Key words: Dimensionali