mirror of
https://github.com/Andreaierardi/Master-DataScience-Notes.git
synced 2024-12-13 13:57:55 +01:00
226 lines
7.7 KiB
TeX
226 lines
7.7 KiB
TeX
\documentclass[../main.tex]{subfiles}
|
||
\begin{document}
|
||
\chapter{Lecture 9 - 06-04-2020}
|
||
|
||
$\hat{h}$ is ERM predictor
|
||
\\
|
||
$$
|
||
\ell_D\left(\hat{h}\right) \leq min \, \, \ell_D\left( h \right) + \sqrt[]{\frac{2}{m} \, \ln \, \frac{2 \, H}{\delta}} \qquad \textit{ with prob. at least $1-\delta$}
|
||
$$
|
||
\\
|
||
Now we do it with tree predictors\\
|
||
\section{Tree predictors}
|
||
$$
|
||
X = \{ 0,1\}^d \longrightarrow \blue{Binary classification}
|
||
$$
|
||
$$
|
||
h : \{0, 1 \}^d \longrightarrow \blue{Binary classification H}1
|
||
$$
|
||
How big is this class?
|
||
\\Take the size of codomain power the domain $\longrightarrow $ $|H| = 2^{2^d}$\\
|
||
Can we have a tree predictor that predict every H in this class?
|
||
\\
|
||
For every $ h : \{0,1\}^d$ $\longleftrightarrow$ $\{-1,1\} \quad \exists T$\\\\
|
||
We can \bred{build a tree } such that \quad $h_T = h$
|
||
\begin{figure}[h]
|
||
\centering
|
||
\includegraphics[width=0.6\linewidth]{../img/lez9-img1.JPG}
|
||
\caption{Tree building}
|
||
%\label{fig:}
|
||
\end{figure}\\
|
||
$ X = (0,0,1,...,1) \qquad h\left(x\right) = -1$ \\
|
||
$
|
||
\blue{$
|
||
x_1,x_2,x_3,...,x_d$}
|
||
$
|
||
\\\\
|
||
I can apply my analisys to this predictors
|
||
\\
|
||
If I run ERM on $H$
|
||
$$
|
||
\ell_D\left(\hat{h}\right) \, \leq \, min \, \ell_D \left(\hat{h}\right) + \sqrt[]{\frac{2}{m} \, 2^d \, \ln 2 + \ln \frac{2}{\delta}} \qquad \longrightarrow \bred{$\ln|H|+\ln \frac{2}{\delta}$}
|
||
$$
|
||
No sense! What we find about training set that we need?
|
||
\\
|
||
Worst case of overfitting $m >> 2^D = |X|$ $\Rightarrow$ training sample larger
|
||
\\\\
|
||
\textbf{PROBLEM: }cannot learn from a class to big ( $H$ is too big)
|
||
\\
|
||
I can control $H$ just limiting the number of nodes.
|
||
\\\\
|
||
$H_N$ $\longrightarrow$ tree T with at most $N$ node, $N << 2^D$
|
||
\\
|
||
$|H_N| = \, ?$
|
||
\\
|
||
$$
|
||
|H_N| = \left( \textit{\# of trees with } \leq N \, nodes \right)
|
||
\times
|
||
\left( \textit{\# of test on interval nodes } \right)
|
||
\times
|
||
\left(
|
||
\textit{ \# labels on leaves}
|
||
\right)
|
||
$$
|
||
$$
|
||
|H_N| = \red{\bigotimes} \, \times \, d^M \, \times 2^{N-M}
|
||
$$
|
||
$N$ of which $N-M$ are leaves
|
||
\begin{figure}[h]
|
||
\centering
|
||
\includegraphics[width=0.6\linewidth]{../img/lez9-img2.JPG}
|
||
\caption{Tree with at most N node}
|
||
%\label{fig:}
|
||
\end{figure}\\
|
||
$$\red{\bigotimes}
|
||
\textit{\# of binary trees with N nodes, called \bred{Catalan Number}}
|
||
$$
|
||
\subsection{Catalan Number}
|
||
*We are using a binomial *
|
||
$$
|
||
\frac{1}{N} \binom{2 \, N -2}{N-1} \quad \leq \quad \frac{1}{N} \, \left(e \, \frac{\left(2\, N -2 \right)}{N-1} \right)^{N-1} = \frac{1}{N} \, \left( 2 \, e \right)^{N-1}
|
||
$$
|
||
$$
|
||
\binom{N}{K} \quad \leq \quad \left( \frac{e\, n}{k}\right)^k \qquad \textit{ from Stirling approximation}
|
||
$$
|
||
Counting the number of tree structure: a binary tree with exactly N nodes.
|
||
Catalan counts this number. $\longrightarrow$ \blue{but we need a quantity to interpret easily}. So we compute it in another way.
|
||
\\
|
||
Now we can rearrange everything.
|
||
\\
|
||
$$ | H _N |
|
||
\quad \leq \quad \blue{ $
|
||
\frac{1}{N}$} \, \left( 2 \, e \right)^{N-1} \, H^M \, \bred{$2^{N-M} $}
|
||
\quad \leq \quad
|
||
\left( 2 \, e \, d \right)^N
|
||
$$
|
||
\qquad \qquad \qquad \qquad \qquad \qquad \qquad
|
||
\bred{$d \geq 2$} \qquad \bred{$\leq \, d^{N-M}$ }\\
|
||
where \blue{we ignore $
|
||
\frac{1}{N}$ since we are going to use the $\log$}
|
||
\\\\
|
||
ERM on $H_N \quad \hat{h} \quad $
|
||
$$\ell_D \left(\hat{h}\right) \, \leq \, \min_{\mathbf{h \, \in\, H_N}} \, \ell_D \left( h \right) + \sqrt[]{\frac{2}{m} \, \left( \bred{$ N \cdot \left( 1+ \ln \left(2 \cdot d \right) \right)$} + \ln \frac{2}{\delta} \right) }
|
||
$$
|
||
\\
|
||
were \bred{$ N \cdot \left( 1+ \ln \left(2 \cdot d \right) \right)$} \quad $= \quad \ln \left( H_N \right)
|
||
$\\\\
|
||
In order to not overfit $ m >> N \cdot \ln d
|
||
$\\
|
||
$N \cdot \ln d << 2^d$ for reasonable value of $N$
|
||
\\
|
||
We grow the tree and a some point we stop.
|
||
$$
|
||
\ell_D\left(h\right) \, \leq \, \hat{\ell}_S \left(h\right) + \varepsilon
|
||
\qquad \forall h \in H_N \qquad \textit{with probability at least $1-\delta$}
|
||
$$
|
||
\\
|
||
\bred{remove $N$ in $H_N$ and include $h$ on $\varepsilon$}
|
||
\\
|
||
we remove the $N$ index in $H_N$ adding $h$ on $\varepsilon$
|
||
|
||
$$
|
||
\ell_D \left(h\right) \, \leq \, \hat{\ell}_S \left(h\right) + \varepsilon_{\red{h}} \qquad \forall h \in H_{\not{\red{N}}}
|
||
$$
|
||
$$
|
||
W : H \longrightarrow \left[ 0,1 \right] \qquad \sum_{h\in H}{} w\left(h\right) \leq 1
|
||
$$
|
||
\\
|
||
\blue{How to use this to control over risk?}
|
||
$$
|
||
\barra{P} \left( \exists h \in H \, : \, | \, \hat{\ell}_S \left(h \right) - \ell_D \left( h \right) \, | \, > \varepsilon_h \right) \quad \leq
|
||
$$
|
||
\bred{where $\hat{\ell}_S$ is the prob my training set cases is true}
|
||
$$
|
||
\leq \,
|
||
\sum_{h \in H}{}
|
||
\barra{P} \left( \, | \, \hat{\ell}_S
|
||
\left(h \right)
|
||
-
|
||
\ell_D
|
||
\left( h \right)
|
||
\, | \,
|
||
> \varepsilon_h \right) \, \leq \, \sum_{h \in H}{} 2 \, e^{-2 \, m \, \varepsilon \, h^2 } \, \leq \, $$
|
||
$$
|
||
\leq \, \delta \qquad \longrightarrow \textit{since $w(h)$ sum to $1$ $ \left( \, \sum_{h \in H} \, \right) $}
|
||
$$
|
||
I want to choose \bred{$ \quad 2 \, e^{-2 \, m \, \varepsilon \, h^2 } \, =\, \delta \, w(h)$}
|
||
\\
|
||
$$
|
||
2 \, e^{-2 \, m \, \varepsilon \, h^2 } \, =\, \delta \, w(h) \qquad \Leftrightarrow \qquad \textit{--- MANCA PARTEEEE --- }
|
||
$$\\
|
||
therefore:
|
||
$$
|
||
\ell_D \left(h \right) \leq
|
||
\hat{\ell}_S \left(h\right) +
|
||
\sqrt[]{\frac{1}{2 \, m} \cdot \left( \ln \frac{1}{w(h)} + ln \frac{2}{\delta} \right) } \quad \textit{w. p. at least $1-\delta$ \quad $\forall h \in H$}
|
||
$$
|
||
\\
|
||
Now, instead of using ERM we use
|
||
$$
|
||
\hat{h} = arg\min_{h \in H} \left(\hat{\ell}_S\left( h \right)
|
||
+
|
||
\sqrt[]{ \frac{1}{2 \, m} \cdot \left( \ln \frac{1}{w(h)} + ln \frac{2}{\delta} \right) }
|
||
\right)
|
||
$$
|
||
\bred{where $\sqrt[]{...}$ term is the penalisation term}
|
||
\\\\
|
||
Since our class is very large we add this part in order to avoid overfitting. \\
|
||
Instead of minimising training error alone i minimise training error +
|
||
penalisation error.\\\\
|
||
In order to pick w(h) we are going to use \bred{coding theory}\\
|
||
The idea is I have my trees and i want to encode all tree predictors in H using
|
||
strings of bits.
|
||
\\\\
|
||
$\sigma : H $ $\longrightarrow $ $\{ 0,1 \}^* \qquad \bred{coding function for trees}
|
||
$
|
||
\\
|
||
$\forall \, h, h' \in H$ \qquad $\sigma(h)$ not a prefix of $\sigma(h')
|
||
$\\
|
||
$h \neq h'$ \qquad \qquad where $\sigma(h)$ and $\sigma(h')$ are \bred{string of bits}
|
||
\\\\
|
||
$\sigma$ is called \blue{istantaneous coding function}
|
||
\\
|
||
Istantaneous coding function has a property called \bred{kraft inequality}
|
||
$$
|
||
\sum_{h\in H}{} 2^{-|\, \sigma\left(h\right)\, |} \leq 1 \qquad w(h) = 2^{-|\,\sigma(h)\,|}
|
||
$$
|
||
\\
|
||
I can design $\sigma : H \longrightarrow \{0,1\}^* \quad istantaneous \ |\,\sigma(h)\,|$\\
|
||
|
||
$
|
||
\ln |H_N| = O\left(N \cdot \ln d\right)
|
||
$\\
|
||
\bred{number of bits i need \quad $=$ \quad number of node in $h$}
|
||
\\\\
|
||
Even if i insist in istantaneous i do not lose ... -- MANCA PARTE --
|
||
\\
|
||
$$
|
||
| \, \sigma (h) \, | = O \left( N \cdot \ln d\right)
|
||
$$\\
|
||
Using this $\sigma$ and $w(h) = 2 ^{-|\, \sigma(h)\,|}
|
||
$
|
||
$$
|
||
\ell_D\left(h\right) \, \leq \, \hat{\ell}_S \left( h \right) + \sqrt[]{\frac{1}{2 \, m} \cdot \left( \red{c} \cdot N \cdot \ln d + \ln \frac{2}{\delta} \right) } \qquad \textit{w. p. at least $1-\delta$}
|
||
$$
|
||
where \red{$c$} is a constant
|
||
\\
|
||
$$
|
||
\hat{h} = arg\min_{h\in H} \left( \hat{\ell}_S \left( h \right) + \sqrt[]{\frac{1}{2 \, m} \cdot \left( \red{c} \cdot N \cdot \ln d + \ln \frac{2}{\delta} \right) } \, \right)
|
||
$$
|
||
where \bred{$m >> N \cdot h \cdot \ln d$}
|
||
\\
|
||
If training set size is very small then you should not run this algorithm.
|
||
\begin{figure}[h]
|
||
\centering
|
||
\includegraphics[width=0.6\linewidth]{../img/lez9-img3.JPG}
|
||
\caption{Algorithm for tree predictors}
|
||
%\label{fig:}
|
||
\end{figure}\\
|
||
This blue curve is an alternative example. We can use Information criterion.\\\\
|
||
As I increase the number of nodes, $N_h$ decrease so fast. You should take a
|
||
smaller tree because it gives you a better bound. It’s a principle known as
|
||
Occam Razor ( if I have two tree with the same error, if one is smaller than the
|
||
other than i should pick this one).
|
||
\\\\
|
||
Having $N^*$
|
||
\end{document} |