乐观偏差-预测误差的估计

9

《统计学习的元素》（在线提供PDF版本）讨论了最佳偏差（7.21，第229页）。它指出，乐观偏差是训练误差与样本误差（如果我们在每个原始训练点采样新的结果值时观察到的误差）之间的差（见下）。

在此处输入图片说明

接下来，它声明这种乐观偏差（）等于我们的估计y值与实际y值的协方差（公式如下）。我很难理解为什么这个公式表示乐观偏见。天真的，我会认为实际与预测之间的强协方差仅描述准确性，而不是乐观。让我知道是否有人可以帮助您推导公式或分享直觉。 $\omega$ $y$ $y$

在此处输入图片说明

error bias validation

— 用户名
source

非常有帮助，谢谢！我认为其中一个方程有一个较小的错字，应该是：

= \frac{1}{N} \sum_{i = 1}^{N} (E_{y} [y_{i}^{2}] + E_{y} [{\hat{y}}_{i}^{2}] - 2 E_{y} [y_{i}] E_{y} [{\hat{y}}_{i}] - E_{y} [y_{i}^{2}] - E_{y} [{\hat{y}}_{i}^{2}] + 2 E [y_{i} {\hat{y}}_{i}])

$= {1 \over N}\sum_{i=1}^N \left( E_y[y_i^2] + E_y[\hat{y}_i^2] -2 E_y [y_i] E_y[ \hat{y}_i] - E_y[y_i^2] - E_y[\hat{y}_i^2] + 2E[y_i \hat{y}_i] \right)$

— Sleepster

8

让我们从直觉开始。

使用预测并没有错。实际上，不使用它意味着我们将丢弃有价值的信息。但是，我们越依赖于包含的信息来得出我们的预测，我们的估算器将越过乐观。 $y_i$ $\hat{y}_i$ $y_i$

在一个极端情况下，如果仅为，则您将具有完美的样本预测（），但我们可以肯定的是，样本外预测会很糟糕。在这种情况下（您很容易自己检查），自由度将为。 $\hat{y}_i$ $y_i$ $R^2 = 1$ $df(\hat{y}) = n$

另一方面，如果对所有使用的样本均值：，则您的自由度将仅为1。 $y$ $y_i = \hat{y_i} = \bar{y}$ $i$

查看Ryan Tibshirani的这份精美讲义，了解有关此直觉的更多详细信息

现在与其他答案类似，但有更多解释

请记住，根据定义，平均乐观度是：

ω = E_{y} (E r r_{i n} - \bar{e r r})

$\omega = E_y (Err_{in} - \overline{err})$

= E_{y} (\frac{1}{N} \sum_{i = 1}^{N} E_{Y^{0}} [L (Y_{i}^{0}, \hat{f} (x_{i}) | T)] - \frac{1}{N} \sum_{i = 1}^{N} L (y_{i}, \hat{f} (x_{i})))

$= E_y \left( {1 \over N} \sum_{i=1}^N E_{Y^0} \left[ L(Y_i^0, \hat{f} (x_i) \; |\; T) \right] - {1 \over N} \sum_{i=1}^N L(y_i, \hat{f} (x_i) ) \right)$

现在使用二次损失函数并展开平方项：

= E_{y} (\frac{1}{N} \sum_{i = 1}^{N} E_{Y^{0}} [(Y_{i}^{0} - {\hat{y}}_{i})^{2}] - \frac{1}{N} \sum_{i = 1}^{N} (y_{i} - {\hat{y}}_{i})^{2}))

$= E_y \left( {1 \over N} \sum_{i=1}^N E_{Y^0} \left[ (Y_i^0 - \hat{y}_i)^2 \right] - {1 \over N} \sum_{i=1}^N (y_i - \hat{y}_i)^2 ) \right)$

= \frac{1}{N} \sum_{i = 1}^{N} (E_{y} E_{Y^{0}} [(Y_{i}^{0})^{2}] + E_{y} E_{Y^{0}} [{\hat{y}}_{i}^{2}] - 2 E_{y} E_{Y^{0}} [Y_{i}^{0} {\hat{y}}_{i}] - E_{y} [y_{i}^{2}] - E_{y} [{\hat{y}}_{i}^{2}] + 2 E [y_{i} {\hat{y}}_{i}])

$= {1 \over N} \sum_{i=1}^N\left( E_y E_{Y^0}[(Y_i^0)^2] + E_y E_{Y^0} [\hat{y}_i^2] -2 E_y E_{Y^0} [Y_i^0 \hat{y}_i] - E_y[y_i^2] - E_y[\hat{y}_i^2] + 2E[y_i \hat{y}_i] \right)$

使用来替换： $E_y E_{Y^0}[(Y_i^0)^2] = E_y[y_i^2]$

= \frac{1}{N} \sum_{i = 1}^{N} (E_{y} [y_{i}^{2}] + E_{y} [{\hat{y_{i}}}^{2}] - 2 E_{y} [y_{i}] E_{y} [{\hat{y}}_{i}] - E_{y} [y_{i}^{2}] - E_{y} [{\hat{y}}_{i}^{2}] + 2 E [y_{i} {\hat{y}}_{i}])

$= {1 \over N}\sum_{i=1}^N \left( E_y[y_i^2] + E_y[\hat{y_i}^2] -2 E_y [y_i] E_y[ \hat{y}_i] - E_y[y_i^2] - E_y[\hat{y}_i^2] + 2E[y_i \hat{y}_i] \right)$

= \frac{2}{N} \sum_{i = 1}^{N} (E [y_{i} {\hat{y}}_{i}] - E_{y} [y_{i}] E_{y} [{\hat{y}}_{i}])

$= {2 \over N} \sum_{i=1}^N \left( E[y_i \hat{y}_i] - E_y [y_i] E_y[ \hat{y}_i] \right)$

最后，请注意，得出： $Cov(x, w) = E[xw] - E[x]E[w]$

= \frac{2}{N} \sum_{i = 1}^{N} C o v (y_{i}, {\hat{y}}_{i})

$= {2 \over N} \sum_{i=1}^N Cov(y_i, \hat{y}_i)$

— CD98
source

5

我必须指出，他的名字拼写为“ Ryan Tibshirani” Rob Tibshirani

— robert tibshirani

2

欢迎来到我们的网站，Rob，很荣幸能将您带到这里，只是为了纠正错误！如果您看到更多内容，请告知我们：当然，我们很高兴收到您（或您的学生）可能想发布的任何答案。您的工作在此站点上得到了广泛引用，尤其是ESL和Bootstrap简介。

— ub

介意吗？此外，是吗？

E_{y} E_{Y^{0}} [(Y_{i}^{0})^{2}] = E_{y} [y_{i}^{2}]

$E_y E_{Y^0}[(Y_i^0)^2] = E_y[y_i^2]$

2 E_{y} E_{Y^{0}} [Y_{i}^{0} {\hat{y}}_{i}] = 2 E_{y} [E_{Y^{0}} [Y_{i}^{0}] E_{Y^{0}} [{\hat{y}}_{i}]] = 2 E_{y} [y_{i}] E_{y} [{\hat{y}}_{i}]

$2 E_y E_{Y^0} [Y_i^0 \hat{y}_i]=2 E_y [E_{Y^0} [Y_i^0]E_{Y^0}[\hat{y}_i]]=2 E_y [y_i] E_y[ \hat{y}_i]$

— 肖基

7

让，然后 $\hat{f}(x_i)=\hat{y}_i$

\begin{aligned} ω & = E_{y} [o p] \\ = E_{y} [E r r_{i n} - \bar{e r r}] \\ = E_{y} [E r r_{i n}] - E_{y} [\bar{e r r}] \\ = E_{y} [\frac{1}{N} \sum_{i = 1}^{N} E_{Y^{0}} [L (Y_{i}^{0}, \hat{f} (x_{i}))] - E_{y} [\frac{1}{N} \sum_{i = 1}^{N} L (y_{i}, \hat{f} (x_{i}))] \\ = \frac{1}{N} \sum_{i = 1}^{N} E_{y} E_{Y^{0}} [(Y_{i}^{0} - {\hat{y}}_{i})^{2}] - E_{y} [(y_{i} - {\hat{y}}_{i})^{2}] \\ = \frac{1}{N} \sum_{i = 1}^{N} E_{y} E_{Y^{0}} [(Y_{i}^{0})^{2}] + E_{y} E_{Y^{0}} [{\hat{y}}_{i}^{2}] - 2 E_{y} E_{Y^{0}} [Y_{i}^{0} {\hat{y}}_{i}] - E_{y} [y_{i}^{2}] - E_{y} [{\hat{y}}_{i}^{2}] + 2 E_{y} [y_{i} {\hat{y}}_{i}] \\ = \frac{1}{N} \sum_{i = 1}^{N} E_{y} [y_{i}^{2}] + E_{y} [{\hat{y}}_{i}^{2}] - 2 E_{y} [y_{i}] E_{y} [{\hat{y}}_{i}] - E_{y} [y_{i}^{2}] - E_{y} [{\hat{y}}_{i}^{2}] + 2 E_{y} [y_{i} {\hat{y}}_{i}] \\ = \frac{2}{N} \sum_{i = 1}^{N} E_{y} [y_{i} {\hat{y}}_{i}] - E_{y} [y_{i}] E_{y} [{\hat{y}}_{i}] \\ = \frac{2}{N} \sum_{i = 1}^{N} E_{y} [y_{i} {\hat{y}}_{i} - y_{i} E_{y} [{\hat{y}}_{i}] - E_{y} [y_{i}] {\hat{y}}_{i} + E_{y} [y_{i}] E_{y} [{\hat{y}}_{i}]] \\ = \frac{2}{N} \sum_{i = 1}^{N} E_{y} [({\hat{y}}_{i} - E_{y} [{\hat{y}}_{i}]) ([y_{i} - E_{y} [y_{i}])] \\ = \frac{2}{N} \sum_{i = 1}^{N} c o v ({\hat{y}}_{i}, y_{i}) \end{aligned}

$\begin{aligned} \omega &= E_\boldsymbol{y}[op]\\ &=E_\boldsymbol{y}[Err_{in}-\overline{err}]\\ &=E_\boldsymbol{y}[Err_{in}]-E_\boldsymbol{y}[\overline{err}]\\ &=E_\boldsymbol{y}[\frac{1}{N}\sum_{i=1}^{N}E_{Y^0}[L(Y_i^0,\hat{f}(x_i))]-E_\boldsymbol{y}[\frac{1}{N}\sum_{i=1}^{N}L(y_i,\hat{f}(x_i))]\\ &=\frac{1}{N}\sum_{i=1}^{N}E_\boldsymbol{y}E_{Y^0}[(Y_i^0-\hat{y}_i)^2]-E_\boldsymbol{y}[(y_i-\hat{y}_i)^2]\\ &=\frac{1}{N}\sum_{i=1}^{N}E_\boldsymbol{y}E_{Y^0}[({Y_i^0})^2]+E_\boldsymbol{y}E_{Y^0}[{\hat{y}_i}^2]-2E_\boldsymbol{y}E_{Y^0}[Y_i^0\hat{y}_i]-E_\boldsymbol{y}[y_i^2]-E_\boldsymbol{y}[\hat{y}_i^2]+2E_\boldsymbol{y}[y_i\hat{y}_i]\\ &=\frac{1}{N}\sum_{i=1}^{N}E_\boldsymbol{y}[y_i^2]+E_\boldsymbol{y}[\hat{y}_i^2]-2E_\boldsymbol{y}[y_i]E_\boldsymbol{y}[\hat{y}_i]-E_\boldsymbol{y}[y_i^2]-E_\boldsymbol{y}[\hat{y}_i^2]+2E_\boldsymbol{y}[y_i\hat{y}_i]\\ &=\frac{2}{N}\sum_{i=1}^{N}E_\boldsymbol{y}[y_i\hat{y}_i]-E_\boldsymbol{y}[y_i]E_\boldsymbol{y}[\hat{y}_i]\\ &=\frac{2}{N}\sum_{i=1}^{N}E_\boldsymbol{y}[y_i\hat{y}_i-y_iE_\boldsymbol{y}[\hat{y}_i]-E_\boldsymbol{y}[y_i]\hat{y}_i+E_\boldsymbol{y}[y_i]E_\boldsymbol{y}[\hat{y}_i]]\\ &=\frac{2}{N}\sum_{i=1}^{N}E_\boldsymbol{y}[(\hat{y}_i-E_\boldsymbol{y}[\hat{y}_i])([y_i-E_\boldsymbol{y}[y_i])]\\ &=\frac{2}{N}\sum_{i=1}^{N}cov(\hat{y}_i,y_i) \end{aligned}$ 优质教育

— Maciej Lazarewicz
source

1

通过协方差的此属性可以简化最后四个步骤：

E [x w] - E [x] E [w] = C o v (x, w)

$E[x w ] - E[x] E[w] = Cov(x, w)$

— cd98 2016年