用于观察的分类边缘不用于培训G.ydF4y2Ba
返回交叉验证的G.ydF4y2Ba分类的边缘G.ydF4y2Ba由由线性分类模型组成的交叉验证的纠错输出代码(ECOC)模型获得G.ydF4y2BaE.G.ydF4y2Ba
= kfoldedge(G.ydF4y2Bacvmdl.G.ydF4y2Ba
)G.ydF4y2Bacvmdl.G.ydF4y2Ba
.也就是说,对于每一个折叠,G.ydF4y2BaKfoldedgeG.ydF4y2Ba
估计分类边缘,以便在使用所有其他观察时培训它时它坚持下去。G.ydF4y2Ba
E.G.ydF4y2Ba
包含包含的线性分类模型中的每个正则化强度的分类边沿G.ydF4y2Bacvmdl.G.ydF4y2Ba
.G.ydF4y2Ba
使用一个或多个指定的附加选项G.ydF4y2BaE.G.ydF4y2Ba
= kfoldedge(G.ydF4y2Bacvmdl.G.ydF4y2Ba
那G.ydF4y2Ba名称,价值G.ydF4y2Ba
)G.ydF4y2Ba名称,价值G.ydF4y2Ba
对论点。例如,指定解码方案,该方案折叠以用于边缘计算或冗长级别。G.ydF4y2Ba
cvmdl.G.ydF4y2Ba
-G.ydF4y2Ba交叉验证的ECOC模型由线性分类模型组成G.ydF4y2BaClassificationPartitionedLinearECOCG.ydF4y2Ba
模型对象G.ydF4y2Ba交叉验证的ECOC模型由线性分类模型组成,指定为aG.ydF4y2BaClassificationPartitionedLinearECOCG.ydF4y2Ba
模型对象。你可以创建一个G.ydF4y2BaClassificationPartitionedLinearECOCG.ydF4y2Ba
模型使用G.ydF4y2BafitcecocG.ydF4y2Ba
和:G.ydF4y2Ba
例如,指定任何一个交叉验证,名称值对参数,G.ydF4y2Ba横梁G.ydF4y2Ba
设置名称值对参数G.ydF4y2Ba学习者G.ydF4y2Ba
来G.ydF4y2Ba“线性”G.ydF4y2Ba
或返回的线性分类模型模板G.ydF4y2BatemplateLinearG.ydF4y2Ba
为了获得估计,kfoldEdge应用了用于交叉验证ECOC模型的相同数据(G.ydF4y2BaX.G.ydF4y2Ba
和G.ydF4y2BayG.ydF4y2Ba
)。G.ydF4y2Ba
指定可选的逗号分离对G.ydF4y2Ba名称,价值G.ydF4y2Ba
论点。G.ydF4y2Ba的名字G.ydF4y2Ba
是参数名称和G.ydF4y2Ba价值G.ydF4y2Ba
为对应值。G.ydF4y2Ba的名字G.ydF4y2Ba
必须出现在引号内。可以以任意顺序指定多个名称和值对参数G.ydF4y2Baname1,value1,...,namen,valuenG.ydF4y2Ba
.G.ydF4y2Ba
'二元乐'G.ydF4y2Ba
-G.ydF4y2Ba二进制学习损失功能G.ydF4y2Ba'汉明'G.ydF4y2Ba
|G.ydF4y2Ba“线性”G.ydF4y2Ba
|G.ydF4y2Ba'logit'G.ydF4y2Ba
|G.ydF4y2Ba'指数'G.ydF4y2Ba
|G.ydF4y2Ba'binodeviance'G.ydF4y2Ba
|G.ydF4y2Ba“枢纽”G.ydF4y2Ba
|G.ydF4y2Ba'二次'G.ydF4y2Ba
|G.ydF4y2Ba功能手柄G.ydF4y2Ba二进制学习者丢失功能,指定为逗号分隔对组成G.ydF4y2Ba'二元乐'G.ydF4y2Ba
和内置,丢失功能名称或功能句柄。G.ydF4y2Ba
此表包含内置函数的名称和描述,其中G.ydF4y2BayG.ydF4y2BajG.ydF4y2Ba是特定二进制学习者的类标签(集合{-1,1,0}),G.ydF4y2BaS.G.ydF4y2BajG.ydF4y2Ba是观察的分数G.ydF4y2BajG.ydF4y2Ba,G.ydF4y2BaG.G.ydF4y2Ba(G.ydF4y2BayG.ydF4y2BajG.ydF4y2Ba那G.ydF4y2BaS.G.ydF4y2BajG.ydF4y2Ba)是二进制损失公式。G.ydF4y2Ba
价值G.ydF4y2Ba | 描述G.ydF4y2Ba | 分数域G.ydF4y2Ba | G.G.ydF4y2Ba(G.ydF4y2BayG.ydF4y2BajG.ydF4y2Ba那G.ydF4y2BaS.G.ydF4y2BajG.ydF4y2Ba)G.ydF4y2Ba |
---|---|---|---|
'binodeviance'G.ydF4y2Ba |
二项异常G.ydF4y2Ba | ( - ∞,∞)G.ydF4y2Ba | 日志(1 + exp (2G.ydF4y2BayG.ydF4y2BajG.ydF4y2BaS.G.ydF4y2BajG.ydF4y2Ba日志(2)])]/ [2G.ydF4y2Ba |
'指数'G.ydF4y2Ba |
幂数G.ydF4y2Ba | ( - ∞,∞)G.ydF4y2Ba | exp (-G.ydF4y2BayG.ydF4y2BajG.ydF4y2BaS.G.ydF4y2BajG.ydF4y2Ba)/ 2G.ydF4y2Ba |
'汉明'G.ydF4y2Ba |
汉字G.ydF4y2Ba | [0, 1]或(-∞,∞)G.ydF4y2Ba | [1 - 符号(G.ydF4y2BayG.ydF4y2BajG.ydF4y2BaS.G.ydF4y2BajG.ydF4y2Ba) / 2G.ydF4y2Ba |
“枢纽”G.ydF4y2Ba |
铰链G.ydF4y2Ba | ( - ∞,∞)G.ydF4y2Ba | max(0,1 -G.ydF4y2BayG.ydF4y2BajG.ydF4y2BaS.G.ydF4y2BajG.ydF4y2Ba)/ 2G.ydF4y2Ba |
“线性”G.ydF4y2Ba |
线性G.ydF4y2Ba | ( - ∞,∞)G.ydF4y2Ba | (1 -G.ydF4y2BayG.ydF4y2BajG.ydF4y2BaS.G.ydF4y2BajG.ydF4y2Ba)/ 2G.ydF4y2Ba |
'logit'G.ydF4y2Ba |
物流G.ydF4y2Ba | ( - ∞,∞)G.ydF4y2Ba | 日志[1 + exp (-G.ydF4y2BayG.ydF4y2BajG.ydF4y2BaS.G.ydF4y2BajG.ydF4y2Ba日志(2)])]/ [2G.ydF4y2Ba |
'二次'G.ydF4y2Ba |
二次G.ydF4y2Ba | [0,1]G.ydF4y2Ba | (1 -G.ydF4y2BayG.ydF4y2BajG.ydF4y2Ba(2G.ydF4y2BaS.G.ydF4y2BajG.ydF4y2Ba- 1)]G.ydF4y2Ba2G.ydF4y2Ba/ 2G.ydF4y2Ba |
该软件使二进制损失正常化,当损失是0.5G.ydF4y2BayG.ydF4y2BajG.ydF4y2Ba= 0.此外,软件计算每个类的平均二进制损耗。G.ydF4y2Ba
对于自定义二进制损耗函数,例如:G.ydF4y2BaCustomFunction.G.ydF4y2Ba
,指定其函数句柄G.ydF4y2Ba'binaryloss',@ customfunctionG.ydF4y2Ba
.G.ydF4y2Ba
CustomFunction.G.ydF4y2Ba
应该有这个形式G.ydF4y2Ba
bloss = customfunction(m,s)G.ydF4y2Ba
mG.ydF4y2Ba
是G.ydF4y2BaK.G.ydF4y2Ba-G.ydF4y2BaL.G.ydF4y2Ba编码矩阵存储在G.ydF4y2Bamdl.codingmatrix.G.ydF4y2Ba
.G.ydF4y2Ba
S.G.ydF4y2Ba
是1 -G.ydF4y2BaL.G.ydF4y2Ba分类得分的行矢量。G.ydF4y2Ba
blG.ydF4y2Ba
是分类损失。这个标量集合了特定班级中每个学习者的二进制损失。例如,您可以使用平均二进制损失来汇总每个类的学习者的损失。G.ydF4y2Ba
K.G.ydF4y2Ba是课程的数量。G.ydF4y2Ba
L.G.ydF4y2Ba是二元学习者的数量。G.ydF4y2Ba
有关传递自定义二进制丢失功能的示例,请参阅G.ydF4y2Ba使用自定义二进制损耗函数预测Ecoc模型的测试样本标签G.ydF4y2Ba.G.ydF4y2Ba
默认情况下,如果所有二进制学习者都是线性分类模型,则使用:G.ydF4y2Ba
支持向量机,然后G.ydF4y2Ba二进制数G.ydF4y2Ba
是G.ydF4y2Ba“枢纽”G.ydF4y2Ba
Logistic回归,然后G.ydF4y2Ba二进制数G.ydF4y2Ba
是G.ydF4y2Ba'二次'G.ydF4y2Ba
示例:G.ydF4y2Ba'binaryloss','binodeviance'G.ydF4y2Ba
数据类型:G.ydF4y2BacharG.ydF4y2Ba
|G.ydF4y2Ba串G.ydF4y2Ba
|G.ydF4y2Bafunction_handle.G.ydF4y2Ba
'解码'G.ydF4y2Ba
-G.ydF4y2Ba解码方案G.ydF4y2Ba“失去重量”G.ydF4y2Ba
(默认)|G.ydF4y2Ba'失败'G.ydF4y2Ba
汇总二进制损耗的解码方案,指定为逗号分隔的对组成G.ydF4y2Ba'解码'G.ydF4y2Ba
和G.ydF4y2Ba“失去重量”G.ydF4y2Ba
或G.ydF4y2Ba'失败'G.ydF4y2Ba
.有关更多信息,请参见G.ydF4y2Ba二元损失G.ydF4y2Ba.G.ydF4y2Ba
示例:G.ydF4y2Ba“解码”、“lossbased”G.ydF4y2Ba
'折叠'G.ydF4y2Ba
-G.ydF4y2Ba折叠指数用于分类评分预测G.ydF4y2Ba1: CVMdl。KFoldG.ydF4y2Ba
(默认)|G.ydF4y2Ba正整数的数字矢量G.ydF4y2Ba用于分类评分预测的折叠指标,指定为包括的逗号分隔对G.ydF4y2Ba'折叠'G.ydF4y2Ba
和一个正整数的数字矢量。元素G.ydF4y2Ba折叠G.ydF4y2Ba
必须从G.ydF4y2Ba1G.ydF4y2Ba
通过G.ydF4y2Bacvmdl.kfold.G.ydF4y2Ba
.G.ydF4y2Ba
示例:G.ydF4y2Ba'折叠',[1 4 10]G.ydF4y2Ba
数据类型:G.ydF4y2Ba单G.ydF4y2Ba
|G.ydF4y2Ba双人间G.ydF4y2Ba
'模式'G.ydF4y2Ba
-G.ydF4y2Ba边缘聚合级别G.ydF4y2Ba“平均”G.ydF4y2Ba
(默认)|G.ydF4y2Ba“个人”G.ydF4y2Ba
边缘聚合级别,指定为逗号分隔对组成G.ydF4y2Ba'模式'G.ydF4y2Ba
和G.ydF4y2Ba“平均”G.ydF4y2Ba
或G.ydF4y2Ba“个人”G.ydF4y2Ba
.G.ydF4y2Ba
价值G.ydF4y2Ba | 描述G.ydF4y2Ba |
---|---|
“平均”G.ydF4y2Ba |
返回对所有折叠的平均分类边缘G.ydF4y2Ba |
“个人”G.ydF4y2Ba |
返回每个折叠的分类边G.ydF4y2Ba |
示例:G.ydF4y2Ba'模式','个人'G.ydF4y2Ba
'选项'G.ydF4y2Ba
-G.ydF4y2Ba估算选项G.ydF4y2Ba[]G.ydF4y2Ba
(默认)|G.ydF4y2Ba返回的结构数组G.ydF4y2Ba实例化G.ydF4y2Ba
估算选项,指定为逗号分隔对组成G.ydF4y2Ba'选项'G.ydF4y2Ba
和返回的结构阵列G.ydF4y2Ba实例化G.ydF4y2Ba
.G.ydF4y2Ba
要调用并行计算:G.ydF4y2Ba
您需要一个并行计算工具箱™许可证。G.ydF4y2Ba
指定G.ydF4y2Ba'选项',statset('deverypallellel',true)G.ydF4y2Ba
.G.ydF4y2Ba
'verbose'G.ydF4y2Ba
-G.ydF4y2Ba冗长的水平G.ydF4y2Ba0.G.ydF4y2Ba
(默认)|G.ydF4y2Ba1G.ydF4y2Ba
详细级别,指定为逗号分隔对组成G.ydF4y2Ba'verbose'G.ydF4y2Ba
和G.ydF4y2Ba0.G.ydF4y2Ba
或G.ydF4y2Ba1G.ydF4y2Ba
.G.ydF4y2BaverbG.ydF4y2Ba
控制软件在命令窗口中显示的诊断消息的数量。G.ydF4y2Ba
如果G.ydF4y2BaverbG.ydF4y2Ba
是G.ydF4y2Ba0.G.ydF4y2Ba
,则该软件不会显示诊断消息。否则,软件将显示诊断消息。G.ydF4y2Ba
示例:G.ydF4y2Ba“详细”,1G.ydF4y2Ba
数据类型:G.ydF4y2Ba单G.ydF4y2Ba
|G.ydF4y2Ba双人间G.ydF4y2Ba
E.G.ydF4y2Ba
-交叉验证分类边G.ydF4y2Ba交叉验证G.ydF4y2Ba分类的边缘G.ydF4y2Ba,作为数字标量,向量或矩阵返回。G.ydF4y2Ba
让G.ydF4y2BaL.G.ydF4y2Ba
是交叉验证模型中的正则化强度的数量(即,G.ydF4y2BaL.G.ydF4y2Ba是G.ydF4y2Ba元素个数(CVMdl.Trained {1} .BinaryLearners {1} .Lambda)G.ydF4y2Ba
)和G.ydF4y2BaF.G.ydF4y2Ba
为存储的折叠数G.ydF4y2Bacvmdl.kfold.G.ydF4y2Ba
)。G.ydF4y2Ba
如果G.ydF4y2Ba模式G.ydF4y2Ba
是G.ydF4y2Ba“平均”G.ydF4y2Ba
那么G.ydF4y2BaE.G.ydF4y2Ba
是一个1-by-G.ydF4y2BaL.G.ydF4y2Ba
向量。G.ydF4y2BaE(G.ydF4y2Ba
是使用正则化强度的交叉验证模型的所有折叠的平均分类边缘G.ydF4y2BajG.ydF4y2Ba
)G.ydF4y2BajG.ydF4y2Ba
.G.ydF4y2Ba
否则,G.ydF4y2BaE.G.ydF4y2Ba
是A.G.ydF4y2BaF.G.ydF4y2Ba
-G.ydF4y2BaL.G.ydF4y2Ba
矩阵。G.ydF4y2BaE(G.ydF4y2Ba
分类边是用来折叠的吗G.ydF4y2Ba我G.ydF4y2Ba
那G.ydF4y2BajG.ydF4y2Ba
)G.ydF4y2Ba我G.ydF4y2Ba
使用正则化强度的交叉验证模型G.ydF4y2BajG.ydF4y2Ba
.G.ydF4y2Ba
加载NLP数据集。G.ydF4y2Ba
加载G.ydF4y2Banlpdata.G.ydF4y2Ba
X.G.ydF4y2Ba
是预测器数据的稀疏矩阵,以及G.ydF4y2BayG.ydF4y2Ba
是类标签的分类向量。G.ydF4y2Ba
为简单起见,请使用标签'其他人'进行所有观察G.ydF4y2BayG.ydF4y2Ba
不G.ydF4y2Ba“金宝app模型”G.ydF4y2Ba
那G.ydF4y2Ba“dsp”G.ydF4y2Ba
,或G.ydF4y2Ba'comm'G.ydF4y2Ba
.G.ydF4y2Ba
Y (~ (ismember (Y, {G.ydF4y2Ba“金宝app模型”G.ydF4y2Ba那G.ydF4y2Ba“dsp”G.ydF4y2Ba那G.ydF4y2Ba'comm'G.ydF4y2Ba})))=G.ydF4y2Ba“别人”G.ydF4y2Ba;G.ydF4y2Ba
交叉验证多字符,线性分类模型。G.ydF4y2Ba
RNG(1);G.ydF4y2Ba重复性的%G.ydF4y2BaCVMdl = fitcecoc (X, Y,G.ydF4y2Ba“学习者”G.ydF4y2Ba那G.ydF4y2Ba“线性”G.ydF4y2Ba那G.ydF4y2Ba'横向'G.ydF4y2Ba那G.ydF4y2Ba'开'G.ydF4y2Ba);G.ydF4y2Ba
cvmdl.G.ydF4y2Ba
是A.G.ydF4y2BaClassificationPartitionedLinearECOCG.ydF4y2Ba
模型。默认情况下,软件实现10倍交叉验证。您可以使用使用的折叠数G.ydF4y2Ba'kfold'G.ydF4y2Ba
名称值对参数。G.ydF4y2Ba
估计折叠边缘边缘的平均值。G.ydF4y2Ba
e = kfoldedge(cvmdl)G.ydF4y2Ba
E = 0.7232.G.ydF4y2Ba
或者,您可以通过指定名称值对获取每个折叠边缘G.ydF4y2Ba'模式','个人'G.ydF4y2Ba
in.G.ydF4y2BaKfoldedgeG.ydF4y2Ba
.G.ydF4y2Ba
执行特征选择的一种方法是比较G.ydF4y2BaK.G.ydF4y2Ba-从多个模型折边。仅根据这个准则,具有最高边的分类器就是最好的分类器。G.ydF4y2Ba
加载NLP数据集。预处理数据G.ydF4y2Ba估计k折叠交叉验证边缘G.ydF4y2Ba,并定向预测器数据,以便观察对应于列。G.ydF4y2Ba
加载G.ydF4y2Banlpdata.G.ydF4y2BaY (~ (ismember (Y, {G.ydF4y2Ba“金宝app模型”G.ydF4y2Ba那G.ydF4y2Ba“dsp”G.ydF4y2Ba那G.ydF4y2Ba'comm'G.ydF4y2Ba})))=G.ydF4y2Ba“别人”G.ydF4y2Ba;x = x';G.ydF4y2Ba
创建这两个数据集:G.ydF4y2Ba
fullXG.ydF4y2Ba
包含所有预测器。G.ydF4y2Ba
partx.G.ydF4y2Ba
包含随机选择的预测因子的1/2。G.ydF4y2Ba
RNG(1);G.ydF4y2Ba重复性的%G.ydF4y2Bap =尺寸(x,1);G.ydF4y2Ba%预测器数量G.ydF4y2Bahalfpredidx = randsample(p,ceil(0.5 * p));fullx = x;partx = x(halfpredidx,:);G.ydF4y2Ba
创建一个线性分类模型模板,指定使用sparsa优化目标函数。G.ydF4y2Ba
t = templatelinear(G.ydF4y2Ba“规划求解”G.ydF4y2Ba那G.ydF4y2Ba'sparsa'G.ydF4y2Ba);G.ydF4y2Ba
交叉验证由二进制,线性分类模型组成的两个ecoc模型:使用所有预测器的eCOC型号,以及使用一半的预测器。表明观察对应于列。G.ydF4y2Ba
CVMdl = fitcecoc (fullX YG.ydF4y2Ba'学习者'G.ydF4y2Ba,t,G.ydF4y2Ba'横向'G.ydF4y2Ba那G.ydF4y2Ba'开'G.ydF4y2Ba那G.ydF4y2Ba...G.ydF4y2Ba'观察'G.ydF4y2Ba那G.ydF4y2Ba'列'G.ydF4y2Ba);pcvmdl = fitcecoc(partx,y,G.ydF4y2Ba'学习者'G.ydF4y2Ba,t,G.ydF4y2Ba'横向'G.ydF4y2Ba那G.ydF4y2Ba'开'G.ydF4y2Ba那G.ydF4y2Ba...G.ydF4y2Ba'观察'G.ydF4y2Ba那G.ydF4y2Ba'列'G.ydF4y2Ba);G.ydF4y2Ba
cvmdl.G.ydF4y2Ba
和G.ydF4y2BaPCVMdlG.ydF4y2Ba
是G.ydF4y2BaClassificationPartitionedLinearECOCG.ydF4y2Ba
模型。G.ydF4y2Ba
估计G.ydF4y2BaK.G.ydF4y2Ba-折边为每个分类器。G.ydF4y2Ba
FultEdge = KfoldEdge(CVMDL)G.ydF4y2Ba
ullEdge = 0.3090.G.ydF4y2Ba
伙伴= kfoldedge(pcvmdl)G.ydF4y2Ba
伙伴德德= 0.2617.G.ydF4y2Ba
基于G.ydF4y2BaK.G.ydF4y2Ba- 折叠边缘,使用所有预测器的分类器是更好的模型。G.ydF4y2Ba
为了确定使用逻辑回归学习者的线性分类模型的良好租赁强度,比较K折边缘。G.ydF4y2Ba
加载NLP数据集。预处理数据G.ydF4y2Ba使用k折边缘的功能选择G.ydF4y2Ba.G.ydF4y2Ba
加载G.ydF4y2Banlpdata.G.ydF4y2BaY (~ (ismember (Y, {G.ydF4y2Ba“金宝app模型”G.ydF4y2Ba那G.ydF4y2Ba“dsp”G.ydF4y2Ba那G.ydF4y2Ba'comm'G.ydF4y2Ba})))=G.ydF4y2Ba“别人”G.ydF4y2Ba;x = x';G.ydF4y2Ba
创建一组8个对数间隔的正则化强度G.ydF4y2Ba 通过G.ydF4y2Ba .G.ydF4y2Ba
lambda = logspace(-8,1,8);G.ydF4y2Ba
创建一个线性分类模型模板,指定使用Lasso惩罚的Logistic回归,使用每个正则化强度,使用Sparsa优化目标函数,并降低目标函数梯度的容差G.ydF4y2Ba1E-8G.ydF4y2Ba
.G.ydF4y2Ba
t = templatelinear(G.ydF4y2Ba“学习者”G.ydF4y2Ba那G.ydF4y2Ba'逻辑'G.ydF4y2Ba那G.ydF4y2Ba“规划求解”G.ydF4y2Ba那G.ydF4y2Ba'sparsa'G.ydF4y2Ba那G.ydF4y2Ba...G.ydF4y2Ba'正规化'G.ydF4y2Ba那G.ydF4y2Ba'套索'G.ydF4y2Ba那G.ydF4y2Ba'lambda'G.ydF4y2Baλ,G.ydF4y2Ba'gradienttolerance'G.ydF4y2Ba,1E-8);G.ydF4y2Ba
交叉验证由二进制,线性分类模型组成的ECOC模型,使用5倍交叉验证和该模型G.ydF4y2Ba
RNG(10)G.ydF4y2Ba重复性的%G.ydF4y2BaCVMdl = fitcecoc (X, Y,G.ydF4y2Ba'学习者'G.ydF4y2Ba,t,G.ydF4y2Ba'观察'G.ydF4y2Ba那G.ydF4y2Ba'列'G.ydF4y2Ba那G.ydF4y2Ba'kfold'G.ydF4y2Ba5)G.ydF4y2Ba
cvmdl = classificationedlinearecoc crossvalidatedmodel:'linearecoc'racatectename:'y'numobservations:31572 kfold:5分区:[1x1 cvpartition] classnames:[comm dsp simulink其他] scoreTransfo金宝apprm:'无'属性,方法G.ydF4y2Ba
cvmdl.G.ydF4y2Ba
是A.G.ydF4y2BaClassificationPartitionedLinearECOCG.ydF4y2Ba
模型。G.ydF4y2Ba
为每个折叠和正则化强度估计边缘。G.ydF4y2Ba
efolds = kfoldedge(cvmdl,G.ydF4y2Ba'模式'G.ydF4y2Ba那G.ydF4y2Ba“个人”G.ydF4y2Ba)G.ydF4y2Ba
eFolds =G.ydF4y2Ba5×8.G.ydF4y2Ba0.5553 0.5555 0.5556 0.5544 0.4957 0.2938 0.1044 0.0853 0.5301 0.5306 0.5306 0.5310 0.4826 0.2944 0.1049 0.0868 0.5278 0.5284 0.5293 0.5290 0.4764 0.2906 0.1039 0.0867 0.5387 0.5398 0.5408 0.5377 0.4844 0.2897 0.1016 0.0857 0.5509 0.5569 0.5571 0.5568 0.4951 0.2938 0.1032 0.0850G.ydF4y2Ba
efolds.G.ydF4y2Ba
是一个5×8的边缘矩阵。行对应于折叠,列对应于正则化强度G.ydF4y2Balambda.G.ydF4y2Ba
.你可以使用G.ydF4y2Baefolds.G.ydF4y2Ba
用来识别表现不佳的褶皱,也就是异常低的边缘。G.ydF4y2Ba
估计每个正规化强度在所有折痕上的平均边。G.ydF4y2Ba
e = kfoldedge(cvmdl)G.ydF4y2Ba
E =G.ydF4y2Ba1×8G.ydF4y2Ba0.5406 0.5422 0.5427 0.5418 0.4868 0.2924 0.1036 0.0859G.ydF4y2Ba
确定模型通过绘制每个正则化强度的5倍边缘的平均值的概括。确定最大化电网上的5倍边缘的正则化强度。G.ydF4y2Ba
图绘图(log10(lambda),log10(e),G.ydF4y2Ba'-o'G.ydF4y2Ba)[〜,maxeidx] = max(e);maxlambda = lambda(maxeidx);holdG.ydF4y2Ba在G.ydF4y2Ba绘图(log10(maxlambda),log10(e(maxeidx)),G.ydF4y2Ba'ro'G.ydF4y2Ba) ylabel (G.ydF4y2Ba“log_{10} 5倍边缘的G.ydF4y2Ba)包含(G.ydF4y2Ba'log_ {10} lambda'G.ydF4y2Ba)传奇(G.ydF4y2Ba“边缘”G.ydF4y2Ba那G.ydF4y2Ba'max边缘'G.ydF4y2Ba)举行G.ydF4y2Ba从G.ydF4y2Ba
几个值G.ydF4y2Balambda.G.ydF4y2Ba
产生类似的高边缘。更大的正则化强度值导致预测器可变稀疏性,这是一个良好的分类器质量。G.ydF4y2Ba
选择正规化强度发生在边缘开始下降之前。G.ydF4y2Ba
LambdaFinal =λ(4);G.ydF4y2Ba
使用整个数据集培训由线性分类模型组成的ecoc模型,并指定正则化强度G.ydF4y2BaLambdaFinalG.ydF4y2Ba
.G.ydF4y2Ba
t = templatelinear(G.ydF4y2Ba“学习者”G.ydF4y2Ba那G.ydF4y2Ba'逻辑'G.ydF4y2Ba那G.ydF4y2Ba“规划求解”G.ydF4y2Ba那G.ydF4y2Ba'sparsa'G.ydF4y2Ba那G.ydF4y2Ba...G.ydF4y2Ba'正规化'G.ydF4y2Ba那G.ydF4y2Ba'套索'G.ydF4y2Ba那G.ydF4y2Ba'lambda'G.ydF4y2BaLambdaFinal,G.ydF4y2Ba'gradienttolerance'G.ydF4y2Ba,1E-8);mdlfinal = fitcecoc(x,y,G.ydF4y2Ba'学习者'G.ydF4y2Ba,t,G.ydF4y2Ba'观察'G.ydF4y2Ba那G.ydF4y2Ba'列'G.ydF4y2Ba);G.ydF4y2Ba
要估计新观测值的标签,请通过G.ydF4y2BaMdlFinalG.ydF4y2Ba
和新的数据G.ydF4y2Ba预测G.ydF4y2Ba
.G.ydF4y2Ba
A.G.ydF4y2Ba二元损失G.ydF4y2Ba是类和分类分数的函数,它决定二元学习者如何将观察结果分类到类中。G.ydF4y2Ba
假设如下:G.ydF4y2Ba
mG.ydF4y2BakjG.ydF4y2Ba是元素(G.ydF4y2BaK.G.ydF4y2Ba那G.ydF4y2BajG.ydF4y2Ba)编码设计矩阵G.ydF4y2BamG.ydF4y2Ba(即与类对应的代码G.ydF4y2BaK.G.ydF4y2Ba二进制学习者G.ydF4y2BajG.ydF4y2Ba)。G.ydF4y2Ba
S.G.ydF4y2BajG.ydF4y2Ba二元学习者的分数是多少G.ydF4y2BajG.ydF4y2Ba为了观察。G.ydF4y2Ba
G.G.ydF4y2Ba是二进制损失功能。G.ydF4y2Ba
是预测的观察类。G.ydF4y2Ba
in.G.ydF4y2Ba基于损失的解码G.ydF4y2Ba(Escalera等。)G.ydF4y2Ba,制作二进制学习者的二进制损失的最小总和的类决定了观察的预测类,即G.ydF4y2Ba
in.G.ydF4y2Baloss-weighted解码G.ydF4y2Ba(Escalera等。)G.ydF4y2Ba,制作二进制学习者二进制损失的最小平均值的阶级决定了预测的观察类,即G.ydF4y2Ba
allwein等。G.ydF4y2Ba提出损耗加权译码通过将所有类的损耗值保持在相同的动态范围内来提高分类精度。G.ydF4y2Ba
此表总结了支持的损耗功能,其中金宝appG.ydF4y2BayG.ydF4y2BajG.ydF4y2Ba是特定二进制学习者的类标签(集合{-1,1,0}),G.ydF4y2BaS.G.ydF4y2BajG.ydF4y2Ba是观察的分数G.ydF4y2BajG.ydF4y2Ba,G.ydF4y2BaG.G.ydF4y2Ba(G.ydF4y2BayG.ydF4y2BajG.ydF4y2Ba那G.ydF4y2BaS.G.ydF4y2BajG.ydF4y2Ba)。G.ydF4y2Ba
价值G.ydF4y2Ba | 描述G.ydF4y2Ba | 分数域G.ydF4y2Ba | G.G.ydF4y2Ba(G.ydF4y2BayG.ydF4y2BajG.ydF4y2Ba那G.ydF4y2BaS.G.ydF4y2BajG.ydF4y2Ba)G.ydF4y2Ba |
---|---|---|---|
'binodeviance'G.ydF4y2Ba |
二项异常G.ydF4y2Ba | ( - ∞,∞)G.ydF4y2Ba | 日志(1 + exp (2G.ydF4y2BayG.ydF4y2BajG.ydF4y2BaS.G.ydF4y2BajG.ydF4y2Ba日志(2)])]/ [2G.ydF4y2Ba |
'指数'G.ydF4y2Ba |
幂数G.ydF4y2Ba | ( - ∞,∞)G.ydF4y2Ba | exp (-G.ydF4y2BayG.ydF4y2BajG.ydF4y2BaS.G.ydF4y2BajG.ydF4y2Ba)/ 2G.ydF4y2Ba |
'汉明'G.ydF4y2Ba |
汉字G.ydF4y2Ba | [0, 1]或(-∞,∞)G.ydF4y2Ba | [1 - 符号(G.ydF4y2BayG.ydF4y2BajG.ydF4y2BaS.G.ydF4y2BajG.ydF4y2Ba) / 2G.ydF4y2Ba |
“枢纽”G.ydF4y2Ba |
铰链G.ydF4y2Ba | ( - ∞,∞)G.ydF4y2Ba | max(0,1 -G.ydF4y2BayG.ydF4y2BajG.ydF4y2BaS.G.ydF4y2BajG.ydF4y2Ba)/ 2G.ydF4y2Ba |
“线性”G.ydF4y2Ba |
线性G.ydF4y2Ba | ( - ∞,∞)G.ydF4y2Ba | (1 -G.ydF4y2BayG.ydF4y2BajG.ydF4y2BaS.G.ydF4y2BajG.ydF4y2Ba)/ 2G.ydF4y2Ba |
'logit'G.ydF4y2Ba |
物流G.ydF4y2Ba | ( - ∞,∞)G.ydF4y2Ba | 日志[1 + exp (-G.ydF4y2BayG.ydF4y2BajG.ydF4y2BaS.G.ydF4y2BajG.ydF4y2Ba日志(2)])]/ [2G.ydF4y2Ba |
'二次'G.ydF4y2Ba |
二次G.ydF4y2Ba | [0,1]G.ydF4y2Ba | (1 -G.ydF4y2BayG.ydF4y2BajG.ydF4y2Ba(2G.ydF4y2BaS.G.ydF4y2BajG.ydF4y2Ba- 1)]G.ydF4y2Ba2G.ydF4y2Ba/ 2G.ydF4y2Ba |
该软件规范化二进制损耗,使得损失为0.5G.ydF4y2BayG.ydF4y2BajG.ydF4y2Ba= 0,并使用二进制学习者的平均值聚合G.ydF4y2Ba[Allwein等]G.ydF4y2Ba.G.ydF4y2Ba
不要将二进制丢失与整体分类损失混淆(由此指定)G.ydF4y2Ba“LossFun”G.ydF4y2Ba
的名称-值对参数G.ydF4y2Ba损失G.ydF4y2Ba
和G.ydF4y2Ba预测G.ydF4y2Ba
对象函数),测量Ecoc分类器整体执行的良好。G.ydF4y2Ba
当G.ydF4y2Ba分类边缘G.ydF4y2Ba为分类边界的加权平均值。G.ydF4y2Ba
多个分类器中选择一种方法,例如要执行特征选择,是选择产生最大边缘的分类器。G.ydF4y2Ba
当G.ydF4y2Ba分类保证金G.ydF4y2Ba为每次观察真实类的负损失与假类中最大负损失的差值。如果边际值在同一尺度上,则作为分类置信度的衡量标准。在众多分类公司中,那些利润率更高的公司表现更好。G.ydF4y2Ba
[1] Allwein,E.,R. Schapire和Y.歌手。“减少二进制文件的多牌:保证金分类的统一方法。”G.ydF4y2Ba机床学习研究G.ydF4y2Ba.2000年第1卷,113-141页。G.ydF4y2Ba
[2] Escalera,S.,O. pujol和P. Radeva。“在三元纠错输出代码中解码过程。”G.ydF4y2Ba图案分析和机器智能的IEEE交易G.ydF4y2Ba.卷。32,第7号,2010年第70页,第120-134页。G.ydF4y2Ba
Pujol, S. Escalera, S. O. Pujol, P. Radeva。用于纠错输出码稀疏设计的三元码的可分性。G.ydF4y2Ba模式识别G.ydF4y2Ba.卷。30,第3,2009号,第285-297页。G.ydF4y2Ba
要并行运行,请指定G.ydF4y2Ba'选项'G.ydF4y2Ba
对此函数调用中的名称值参数并设置G.ydF4y2Ba'使用指平行'G.ydF4y2Ba
字段的选项结构G.ydF4y2Ba真正的G.ydF4y2Ba
使用G.ydF4y2Ba实例化G.ydF4y2Ba
.G.ydF4y2Ba
例如:G.ydF4y2Ba'选项',statset('deverypallellel',true)G.ydF4y2Ba
有关并行计算的更多信息,请参阅G.ydF4y2Ba运行MATLAB函数与自动并行支持金宝appG.ydF4y2Ba(并行计算工具箱)G.ydF4y2Ba.G.ydF4y2Ba
Classifiedecoc.G.ydF4y2Ba
|G.ydF4y2BaClassificationLinearG.ydF4y2Ba
|G.ydF4y2BaClassificationPartitionedLinearECOCG.ydF4y2Ba
|G.ydF4y2Ba边缘G.ydF4y2Ba
|G.ydF4y2BafitcecocG.ydF4y2Ba
|G.ydF4y2BaKfoldmargin.G.ydF4y2Ba
|G.ydF4y2BaKfoldpredictG.ydF4y2Ba
|G.ydF4y2Ba实例化G.ydF4y2Ba
您单击了与此MATLAB命令对应的链接:G.ydF4y2Ba
在MATLAB命令窗口中输入它来运行命令。Web浏览器不支持MATLAB命令。金宝appG.ydF4y2Ba
选择一个网站,在那里获得翻译的内容,并看到当地的活动和优惠。根据您的位置,我们建议您选择:G.ydF4y2Ba.G.ydF4y2Ba
选择G.ydF4y2Ba网站G.ydF4y2Ba你也可以从以下列表中选择一个网站:G.ydF4y2Ba
选择中国网站(以中文或英文)以获取最佳网站性能。其他MathWorks国家网站未优化您的位置。G.ydF4y2Ba