文档详情

大规模线性支持向量机的牛顿法..ppt

发布:2017-05-25约2.97千字共21页下载文档
文本预览下载声明
大规模线性支持向量机的牛顿法 Evaluation only. Created with Aspose.Slides for .NET 3.5 Client Profile 5.2.0.0. Copyright 2004-2011 Aspose Pty Ltd. 大规模线性支持向量机 大规模问题:样本数和/或特征数非常大,数据矩阵常有稀疏性,如文本分类问题 线性支持向量机:线性核 一个核心问题:模型的求解效率 Evaluation only. Created with Aspose.Slides for .NET 3.5 Client Profile 5.2.0.0. Copyright 2004-2011 Aspose Pty Ltd. L2-SVM Evaluation only. Created with Aspose.Slides for .NET 3.5 Client Profile 5.2.0.0. Copyright 2004-2011 Aspose Pty Ltd. 模型(1)的性质 是个无约束优化问题 f(·)是个分段二次函数,且为严格凸的。从而(1)是个严格凸的无约束问题,具有唯一的最优解 与LS-SVM的关系:当I(β)={1,2,…,m}即所有约束指标集时,(1)就是LS-SVM。因此LS-SVM可看成是(1)的一个特殊情况 Evaluation only. Created with Aspose.Slides for .NET 3.5 Client Profile 5.2.0.0. Copyright 2004-2011 Aspose Pty Ltd. 模型(1)的性质 f(·)是一阶连续可微的,其梯度为 Evaluation only. Created with Aspose.Slides for .NET 3.5 Client Profile 5.2.0.0. Copyright 2004-2011 Aspose Pty Ltd. Evaluation only. Created with Aspose.Slides for .NET 3.5 Client Profile 5.2.0.0. Copyright 2004-2011 Aspose Pty Ltd. 模型(1)的求解 有限牛顿法(Mangasarian,2002) 改进的有限牛顿法(Keerthi,2005) Evaluation only. Created with Aspose.Slides for .NET 3.5 Client Profile 5.2.0.0. Copyright 2004-2011 Aspose Pty Ltd. 有限牛顿法 Evaluation only. Created with Aspose.Slides for .NET 3.5 Client Profile 5.2.0.0. Copyright 2004-2011 Aspose Pty Ltd. 广义Hesee矩阵 模型(1)的目标函数仅是一阶连续可微的,但在0点,它的二阶导数不存在。因此Hesse矩阵不存在。但注意到它几乎处处二阶可微(扣除0点都是二阶连续可微的),因此我们可以用它的广义Hesse矩阵作为其Hesse矩阵 Evaluation only. Created with Aspose.Slides for .NET 3.5 Client Profile 5.2.0.0. Copyright 2004-2011 Aspose Pty Ltd. 模型(1)的广义Hesse矩阵 Evaluation only. Created with Aspose.Slides for .NET 3.5 Client Profile 5.2.0.0. Copyright 2004-2011 Aspose Pty Ltd. 有限牛顿法的性质 算法1称为有限牛顿法是因为该算法能在有限步内终止 算法1给出了大规模线性支持向量机的一个高效求解算法,比SMO类型的算法还高效 Evaluation only. Created with Aspose.Slides for .NET 3.5 Client Profile 5.2.0.0. Copyright 2004-2011 Aspose Pty Ltd. 改进的有限牛顿法 Evaluation only. Created with Aspose.Slides for .NET 3.5 Client Profile 5.2.0.0. Copyright 2004-2011 Aspose Pty Ltd. L2-SVM-MFN与L2-SVM-FN 采用共轭梯度法求解 采用精确的一维搜索 采用一些启发式规则提高算法效率 Eva
显示全部
相似文档