Leave-one-out error

 Leave-one-out error can refer to the following:

• Leave-one-out cross-validation Stability (CVloo, for stability of Cross Validation with leave one out): An algorithm f has CVloo stability β with respect to the loss function V if the following holds:

${\displaystyle \forall i\in \{1,...,m\},\mathbb {P} _{S}\{\sup _{z\in Z}|V(f_{S},z_{i})-V(f_{S^{|i}},z_{i})|\leq \beta _{CV}\}\geq 1-\delta _{CV}}$

• Expected-to-leave-one-out error Stability (${\displaystyle Eloo_{err}}$, for Expected error from leaving one out): An algorithm f has ${\displaystyle Eloo_{err}}$ stability if for each n there exists a${\displaystyle \beta _{EL}^{m}}$ and a ${\displaystyle \delta _{EL}^{m}}$ such that:

${\displaystyle \forall i\in \{1,...,m\},\mathbb {P} _{S}\{|I[f_{S}]-{\frac {1}{m}}\sum _{i=1}^{m}V(f_{S^{|i}},z_{i})|\leq \beta _{EL}^{m}\}\geq 1-\delta _{EL}^{m}}$, with ${\displaystyle \beta _{EL}^{m}}$and ${\displaystyle \delta _{EL}^{m}}$ going to zero for ${\displaystyle n\rightarrow \inf }$

Preliminary notationsEdit

With X and Y being a subset of the real numbers R, or X and Y ⊂ R, being respectively an input space X and an output space Y, we consider a training set:

${\displaystyle S=\{z_{1}=(x_{1},\ y_{1})\ ,..,\ z_{m}=(x_{m},\ y_{m})\}}$ of size m in ${\displaystyle Z=X\times Y}$ drawn independently and identically distributed (i.i.d.) from an unknown distribution, here called "D". Then a learning algorithm is a function ${\displaystyle f}$ from ${\displaystyle Z_{m}}$ into ${\displaystyle F\subset YX}$ which maps a learning set S onto a function ${\displaystyle f_{S}}$ from the input space X to the output space Y. To avoid complex notation, we consider only deterministic algorithms. It is also assumed that the algorithm ${\displaystyle f}$ is symmetric with respect to S, i.e. it does not depend on the order of the elements in the training set. Furthermore, we assume that all functions are measurable and all sets are countable which does not limit the interest of the results presented here.

The loss of an hypothesis f with respect to an example ${\displaystyle z=(x,y)}$ is then defined as ${\displaystyle V(f,z)=V(f(x),y)}$. The empirical error of f can then be written as ${\displaystyle I_{S}[f]={\frac {1}{n}}\sum V(f,z_{i})}$.

The true error of f is ${\displaystyle I[f]=\mathbb {E} _{z}V(f,z)}$

Given a training set S of size m, we will build, for all i = 1....,m, modified training sets as follows:

• By removing the i-th element

${\displaystyle S^{|i}=\{z_{1},...,\ z_{i-1},\ z_{i+1},...,\ z_{m}\}}$

• and/or^{[clarification needed]} by replacing the i-th element

${\displaystyle S^{i}=\{z_{1},...,\ z_{i-1},\ z_{i}',\ z_{i+1},...,\ z_{m}\}}$

This article uses material from the Wikipedia article  Metasyntactic variable, which is released under the  Creative Commons Attribution-ShareAlike 3.0 Unported License.