[{"source":{"position":4,"lines":[],"changePosition":null},"target":{"position":4,"lines":["$$"],"changePosition":null},"type":"INSERT"},{"source":{"position":5,"lines":[],"changePosition":null},"target":{"position":6,"lines":["$$"],"changePosition":null},"type":"INSERT"},{"source":{"position":7,"lines":["1. 原始数据："," $$\begin{gathered}\\vec\\{x\\}=(x\\^\\{(1)\\})...x\\^\\{(n)\\},1)\\^\\{T\\}=(\\vec\\{xw\\}\\^\\{T\\}, 1)\\^\\{T\\}\end{gathered}$$"],"changePosition":null},"target":{"position":9,"lines":["1. 原始数据：","$$\begin{gathered}","\\vec{x}=(x^{(1)})...x^{(n)},1)^{T}=(\\vec{xw}^{T}, 1)^{T}","\end{gathered}$$"],"changePosition":null},"type":"CHANGE"},{"source":{"position":11,"lines":[" $$\begin{gathered}\\vec\\{w\\}=(w\\^\\{(1)\\})...w\\^\\{(n)\\},b)\\^\\{T\\}=(\\vec\\{w\\}\\^\\{T\\}, b)\\^\\{T\\}\end{gathered}$$"],"changePosition":null},"target":{"position":15,"lines":["$$\begin{gathered}","\\vec{w}=(w^{(1)})...w^{(n)},b)^{T}=(\\vec{w}^{T}, b)^{T}","\end{gathered}$$"],"changePosition":null},"type":"CHANGE"},{"source":{"position":14,"lines":[" $$\begin{gathered}\\vec\\{y\\}=(y\\^\\{(1)\\})...y\\^\\{(n)\\})\\^\\{T\\}=(\\vec\\{y\\})\\^\\{T\\}\end{gathered}$$"],"changePosition":null},"target":{"position":20,"lines":[" $$\begin{gathered}\\vec{y}=(y^{(1)})...y^{(n)})^{T}=(\\vec{y})^{T}\end{gathered}$$"],"changePosition":null},"type":"CHANGE"},{"source":{"position":18,"lines":["$$\begin{gathered}\\sum_\\{i=1\\}\\^\\{N\\}(\\vec\\{w\\} \\*\\vec\\{x\\} +b -y_i)\end{gathered}$$"],"changePosition":null},"target":{"position":24,"lines":["$$\begin{gathered}\\sum_{i=1}^{N}(\\vec{w} *\\vec{x} +b -y_i)\end{gathered}$$"],"changePosition":null},"type":"CHANGE"},{"source":{"position":23,"lines":["$$\begin{gathered}\\sum_\\{i=1\\}\\^\\{N\\}\\vec\\{w\\} \\*=arg\\min(\\vec\\{y\\}-\\vec\\{x\\}\\vec\\{w\\})\\^\\{T\\} (\\vec\\{y\\}-\\vec\\{x\\}\\vec\\{w\\})\end{gathered}$$"],"changePosition":null},"target":{"position":29,"lines":["$$\begin{gathered}","\\sum_{i=1}^{N}\\vec{w} *=arg\\min(\\vec{y}-\\vec{x}\\vec{w})^{T} (\\vec{y}-\\vec{x}\\vec{w})","\end{gathered}$$"],"changePosition":null},"type":"CHANGE"},{"source":{"position":29,"lines":["$$\begin{gathered}E(w)=(\\vec\\{y\\}-\\vec\\{xw\\})\\^\\{T\\} \\* (\\vec\\{y\\}-\\vec\\{xw\\})\end{gathered}$$"],"changePosition":null},"target":{"position":37,"lines":["$$\begin{gathered}","E(w)=(\\vec{y}-\\vec{xw})^{T} * (\\vec{y}-\\vec{xw})","\end{gathered}$$"],"changePosition":null},"type":"CHANGE"},{"source":{"position":31,"lines":["$$\begin{gathered} \\frac\\{\t\\partial E(\\vec\\{w\\})\\}\\{\\partial \\vec\\{w\\}\\}=2\\vec\\{x\\}\\^\\{T\\}(\\vec\\{x\\}\\vec\\{w\\} - \\vec\\{y\\})=0\end{gathered}$$ >$$\begin{gathered}\\vec\\{x\\}\\^\\{T\\}\\vec\\{x\\}\\vec\\{w\\}=\\vec\\{x\\}\\^\\{T\\}\\vec\\{y\\}\end{gathered}$$"],"changePosition":null},"target":{"position":41,"lines":["$$\begin{gathered}","\\frac{\t\\partial E(\\vec{w})}{\\partial \\vec{w}}=2\\vec{x}^{T}(\\vec{x}\\vec{w} - \\vec{y})=0","\end{gathered}$$",">","$$\begin{gathered}","\\vec{x}^{T}\\vec{x}\\vec{w}=\\vec{x}^{T}\\vec{y}","\end{gathered}$$"],"changePosition":null},"type":"CHANGE"},{"source":{"position":33,"lines":["\\当xxT是满秩\\,就有如下结论。","$$\begin{gathered} \\vec\\{w\\}\\^\\{\\*\\}=(\\vec\\{x\\}\\^\\{T\\}\\vec\\{x\\})\\^\\{-1\\} \\vec\\{x\\}\\^\\{T\\}\\vec\\{y\\}\end{gathered}$$"],"changePosition":null},"target":{"position":49,"lines":["当xxT是满秩,就有如下结论。","$$\begin{gathered} \\vec{w}^{*}=(\\vec{x}^{T}\\vec{x})^{-1} \\vec{x}^{T}\\vec{y}\end{gathered}$$"],"changePosition":null},"type":"CHANGE"},{"source":{"position":46,"lines":["当\\xxT不是满秩\\，也就是样本数量小于特征数量，也就是方程可能存在多个解，常见的做法就是引入正则化。"],"changePosition":null},"target":{"position":62,"lines":["当xxT不是满秩，也就是样本数量小于特征数量，也就是方程可能存在多个解，常见的做法就是引入正则化。"],"changePosition":null},"type":"CHANGE"},{"source":{"position":50,"lines":["对于问题属性而言，很多维度都是离散的维度呈现，用线性回归很难拟合。这个时候我们就会考虑将输入的特征数据，映射到同一个维度上，最好是连续的空间，然后在一个连续的空间做拟合，这个思路实际上就是我们要说的\\逻辑回归\\。"],"changePosition":null},"target":{"position":66,"lines":["对于问题属性而言，很多维度都是离散的维度呈现，用线性回归很难拟合。这个时候我们就会考虑将输入的特征数据，映射到同一个维度上，最好是连续的空间，然后在一个连续的空间做拟合，这个思路实际上就是我们要说的逻辑回归。"],"changePosition":null},"type":"CHANGE"},{"source":{"position":67,"lines":["$$\begin{gathered}P(Y=1|X)=\\pi(x)=\\frac\\{exp(wx)\\}\\{1+exp(wx)\\}\end{gathered}$$"],"changePosition":null},"target":{"position":83,"lines":["$$\begin{gathered}P(Y=1|X)=\\pi(x)=\\frac{exp(wx)}{1+exp(wx)}\end{gathered}$$"],"changePosition":null},"type":"CHANGE"},{"source":{"position":71,"lines":["$$\begin{gathered}P(Y=0|X)=1-\\pi(x)=1-\\frac\\{exp(wx)\\}\\{1+exp(wx)\\}\end{gathered}$$"],"changePosition":null},"target":{"position":87,"lines":["$$\begin{gathered}P(Y=0|X)=1-\\pi(x)=1-\\frac{exp(wx)}{1+exp(wx)}\end{gathered}$$"],"changePosition":null},"type":"CHANGE"},{"source":{"position":74,"lines":["$$\begin{gathered}L(w)=\\prod_\\{i=1\\}\\^\\{N\\}[\\pi(x_i)]\\^\\{y_i\\}[1-\\pi(\\{x_i)\\}]\\^\\{1-y_i\\}\end{gathered}$$"],"changePosition":null},"target":{"position":90,"lines":["$$\begin{gathered}L(w)=\\prod_{i=1}^{N}[\\pi(x_i)]^{y_i}[1-\\pi({x_i)}]^{1-y_i}\end{gathered}$$"],"changePosition":null},"type":"CHANGE"},{"source":{"position":76,"lines":["那就很容易的知道\\对数似然函数\\为："],"changePosition":null},"target":{"position":92,"lines":["那就很容易的知道对数似然函数为："],"changePosition":null},"type":"CHANGE"},{"source":{"position":78,"lines":["$$\begin{gathered}L(w)=\\sum_\\{i=1\\}\\^\\{N\\}y_iln(\\pi(\\{x_i\\})) + (1-y_i)\\ln(1-\\pi(\\{x_i\\}))\end{gathered}$$"],"changePosition":null},"target":{"position":94,"lines":["$$\begin{gathered}L(w)=\\sum_{i=1}^{N}y_iln(\\pi({x_i})) + (1-y_i)\\ln(1-\\pi({x_i}))\end{gathered}$$"],"changePosition":null},"type":"CHANGE"},{"source":{"position":84,"lines":["$$\begin{gathered}L(w)\\^\\{'\\}=(y_i - \\pi (x))x\end{gathered}$$"],"changePosition":null},"target":{"position":100,"lines":["$$\begin{gathered}L(w)^{'}=(y_i - \\pi (x))x\end{gathered}$$"],"changePosition":null},"type":"CHANGE"},{"source":{"position":89,"lines":["$$\begin{gathered}w=w+\\alpha(L(w)\\^\\{'\\})\end{gathered}$$","目标就是求$L(w)$的极值，从而得到$\\vec\\{w\\}$的估计值。"],"changePosition":null},"target":{"position":105,"lines":["$$\begin{gathered}w=w+\\alpha(L(w)^{'})\end{gathered}$$","目标就是求$L(w)$的极值，从而得到$\\vec{w}$的估计值。"],"changePosition":null},"type":"CHANGE"},{"source":{"position":94,"lines":["以上考虑都是基于二分类的情况，同样可以推广到对分类的情况，设离散变量含有集合Y=\{1，2，3，4，5...K\},则多分类逻辑回归为。"],"changePosition":null},"target":{"position":110,"lines":["以上考虑都是基于二分类的情况，同样可以推广到对分类的情况，设离散变量含有集合Y={1，2，3，4，5...K},则多分类逻辑回归为。"],"changePosition":null},"type":"CHANGE"},{"source":{"position":96,"lines":["$$\begin{gathered}P(Y=k|x)=\\frac\\{exp(wx)\\}\\{1+\\sum_\\{k=1\\}\\^\\{k=K\\}(exp(wx))\\}\end{gathered}$$"],"changePosition":null},"target":{"position":112,"lines":["$$\begin{gathered}P(Y=k|x)=\\frac{exp(wx)}{1+\\sum_{k=1}^{k=K}(exp(wx))}\end{gathered}$$"],"changePosition":null},"type":"CHANGE"},{"source":{"position":98,"lines":["$$\begin{gathered}P(Y=K|x)=\\frac\\{1\\}\\{1+\\sum_\\{k=1\\}\\^\\{k=K\\}(exp(wx)\\}\end{gathered}$$"],"changePosition":null},"target":{"position":114,"lines":["$$\begin{gathered}P(Y=K|x)=\\frac{1}{1+\\sum_{k=1}^{k=K}(exp(wx)}\end{gathered}$$"],"changePosition":null},"type":"CHANGE"},{"source":{"position":102,"lines":[],"changePosition":null},"target":{"position":118,"lines":["## 线性回归的要求","首先线性回归只能满足X和Y符合线性关系的数据，且要求Y服从正态分布。这一点经常被大家忽略。当然这里需要注意的是，底层也不是说Y服从正态分布，而是误差服从一个正态分布。 如果模型预测完，这个偏差不是正太分布，说明模型不是最优解。","真实模型 $y=bx^2+e$, 用 $y=bx+e$ 拟合， 则可能会得到山形的残差分布"],"changePosition":null},"type":"INSERT"},{"source":{"position":103,"lines":["为了不让问题变得复杂，不 要轻易尝试非线性模型，除非你的理由十分充足。"],"changePosition":null},"target":{"position":122,"lines":["为了不让问题变得复杂， 不要轻易尝试非线性模型，除非你的理由十分充足。"],"changePosition":null},"type":"CHANGE"},{"source":{"position":105,"lines":["[^1]: 提到似然函数就要和概率做对比理解，常说的概率是指给定参数后，预测即将发生的事件的可能性。拿硬币这个例子来说，我们已知一枚均匀硬币的正反面概率分别是0.5，要预测抛两次硬币，硬币都朝上的概率：p(HH | pH = 0.5) = 0.5\0.5 = 0.25。而似然概率正好与这个过程相反，我们关注的量不再是事件的发生概率，而是已知发生了某些事件，我们希望知道参数应该是多少。在我们已经抛了两次硬币，并且知道了结果是两次头朝上，这时候，我希望知道这枚硬币抛出去正面朝上的概率为0.5的概率是多少？正面朝上的概率为0.8的概率是多少？如果我们希望知道正面朝上概率为0.5的概率，这个东西就叫做似然函数，可以说成是对某一个参数的猜想（p=0.5）的概率"],"changePosition":null},"target":{"position":124,"lines":["[^1]: 提到似然函数就要和概率做对比理解，常说的概率是指给定参数后，预测即将发生的事件的可能性。拿硬币这个例子来说，我们已知一枚均匀硬币的正反面概率分别是0.5，要预测抛两次硬币，硬币都朝上的概率：p(HH | pH = 0.5) = 0.50.5 = 0.25。而似然概率正好与这个过程相反，我们关注的量不再是事件的发生概率，而是已知发生了某些事件，我们希望知道参数应该是多少。在我们已经抛了两次硬币，并且知道了结果是两次头朝上，这时候，我希望知道这枚硬币抛出去正面朝上的概率为0.5的概率是多少？正面朝上的概率为0.8的概率是多少？如果我们希望知道正面朝上概率为0.5的概率，这个东西就叫做似然函数，可以说成是对某一个参数的猜想（p=0.5）的概率"],"changePosition":null},"type":"CHANGE"}]