["DIV",{"class":"direct-editor","contenteditable":"true"},["DIV",{"data-btype":"img"},["IMG",{"src":"http://static.notedeep.com/FiLVkmkNDd85IC-yj0lfb7dNQpD8","style":"width: 600px;"}]],["DIV",{"data-btype":"basic"},"（1）令",["SPAN",{"class":"formula-wrapper","contenteditable":"false","data-tex":"H=\\{h_0,h_1\\},h_0(x)=0,h_1(x)=1,\n\\forall x\\in [0,1]"},["SPAN",{"class":"formula-content"},"H=\\{h_0,h_1\\},h_0(x)=0,h_1(x)=1,\n\\forall x\\in [0,1]"]]," "],["DIV",{"data-btype":"basic"},"（2）使得H的VC维为4的样本集为{(-1,0),(1,0),(0,-1),(0,1)}"],["DIV",{"data-btype":"basic"},"下面说明不存在大小为5的样本集使得H可将其打散："],["DIV",{"data-btype":"basic"},"假设大小为5的样本集为{A_1=(x_1,y_1),A_2=(x_2,y_2),A_3=(x_3,y_3),A_4=(x_4,y_4),A_5=(x_5,y_5)}"],["DIV",{"data-btype":"basic"},"不妨假设x_1",["X_2 \\varepsilon\\wedge\\hat E(h_D)=0\\}"},["SPAN",{"class":"formula-content"},"P\\{E(h_D)\\le \\varepsilon\\}=P\\{E(h_D)\\le \\varepsilon\\wedge\\hat E(h_D)=0\\}=P\\{\\hat E(h_D)=0\\}-P\\{E(h_D)> \\varepsilon\\wedge\\hat E(h_D)=0\\}"]]," "],["DIV",{"data-btype":"basic"},["SPAN",{"class":"formula-wrapper","contenteditable":"false","data-tex":"=1-P\\{E(h_D)> \\varepsilon\\wedge\\hat E(h_D)=0\\}\n"},["SPAN",{"class":"formula-content"},"=1-P\\{E(h_D)> \\varepsilon\\wedge\\hat E(h_D)=0\\}\n"]]," "],["DIV",{"data-btype":"basic"},["SPAN",{"class":"formula-wrapper","contenteditable":"false","data-tex":"\\ge1-|H|(1-P\\{E(h_D)> \\varepsilon\\wedge h_D(x)\\ne y_x\\})^m"},["SPAN",{"class":"formula-content"},"\\ge1-|H|(1-P\\{E(h_D)> \\varepsilon\\wedge h_D(x)\\ne y_x\\})^m"]]," "],["DIV",{"data-btype":"basic"},["SPAN",{"class":"formula-wrapper","contenteditable":"false","data-tex":"\\ge 1-|H|(1-\\varepsilon)^m"},["SPAN",{"class":"formula-content"},"\\ge 1-|H|(1-\\varepsilon)^m"]]," "],["DIV",{"data-btype":"basic"},["SPAN",{"class":"formula-wrapper","contenteditable":"false","data-tex":"\\ge1-|H|e^{-m\\varepsilon}"},["SPAN",{"class":"formula-content"},"\\ge1-|H|e^{-m\\varepsilon}"]]," "],["DIV",{"data-btype":"basic"},"令",["SPAN",{"class":"formula-wrapper","contenteditable":"false","data-tex":"|H|e^{-m\\varepsilon}\\le\\delta"},["SPAN",{"class":"formula-content"},"|H|e^{-m\\varepsilon}\\le\\delta"]]," ，即",["SPAN",{"class":"formula-wrapper","contenteditable":"false","data-tex":"m\\ge\\frac 1\\varepsilon(\\ln|H|+\\ln\\frac 1 \\varepsilon)"},["SPAN",{"class":"formula-content"},"m\\ge\\frac 1\\varepsilon(\\ln|H|+\\ln\\frac 1 \\varepsilon)"]]," ，"],["DIV",{"data-btype":"basic"},"则有",["SPAN",{"class":"formula-wrapper","contenteditable":"false","data-tex":"P\\{E(h_D)\\le\\varepsilon\\}\\ge 1-\\delta"},["SPAN",{"class":"formula-content"},"P\\{E(h_D)\\le\\varepsilon\\}\\ge 1-\\delta"]]," "],["DIV",{"data-btype":"img"},["IMG",{"src":"http://static.notedeep.com/FgV-Uj7f7K1vO_jn9CsnrmGOqjL_","style":"width: 600px;"}]],["DIV",{"data-btype":"basic"},"1）TSVM原理：TSVM是半监督学习算法，采用局部搜索策略进行迭代求解，首先使用有标记的样本训练出初始SVM，接着使用该学习器对未标记样本进行标记，并基于这些有标记的样本重新训练SVM；之后再找出两个标记指派为异类且很可能发生错误的未标记样本，交换标记；增大无标记样本的折中参数和减小有标记样本的折中参数，重新训练SVM；重复训练、交换标记、调整折中参数，指导无标记样本的折中参数等于有标记样本的折中参数。"],["DIV",{"data-btype":"basic"},"2）TSVM与SVM区别：SVM是监督学习算法，其训练数据是样例，样本有明确标记，只需最小化经验误差即可得到目标假设。"],["DIV",{"data-btype":"basic"},"TSVM是半监督学习算法，其训练数据包含已经样例和未标记的样本，需要通过不断重复训练，调整折中参数，才能得到目标假设。"],["DIV",{"data-btype":"img"},["IMG",{"src":"http://static.notedeep.com/FrliXCIwe02CNWeY8ExL3zarG_zD","style":"width: 600px;"}]],["DIV",{"data-btype":"basic"},"过滤式：先进行特征选择，再进行学习器的训练，特征选择与后续学习器的训练无关。"],["DIV",{"data-btype":"basic"},"包裹式：按照学习器的训练误差来评价属性子集的优劣，并不断搜索更好的子集。就最终学习器性能来看，包裹式优于过滤式；但是包裹式需要不断重复选择与训练过程，因而计算开销通常比过滤式大得多。"],["DIV",{"data-btype":"basic"},"嵌入式：特征选择过程在学习器训练过程自动进行。"]]]]

机器学习第二次作业