Master-DataScience-Notes/1year/3trimester/Machine Learning, Statistical Learning, Deep Learning and Artificial Intelligence/Machine Learning/main.tex

\title{Statistical Methods for Machine Learning}
\author{
        Andrea Ierardi \\
                Data Science and Economcis\\
        Università degli Studi di Milano\\
        
}
\date{\today}

\documentclass[12pt]{article}
\usepackage{amsmath}
\usepackage{systeme}
\usepackage{amssymb}
\newcommand\barra[1]{\mathbb{#1}}

\begin{document}
\maketitle


\begin{abstract}
This is the paper's abstract \ldots
\end{abstract}

\section{Lecture 1 - 09-03-2020}

\subsection{Introduction}
This is time for all good men to come to the aid of their party!

MACHINE LEARNING
In this course we look at the principle behind design of Machine learning.
Not just coding but have an idea of algorithm that can work with the data.
We have to fix a mathematical framework: some statistic and mathematics.
Work on ML on a higher level
ML is data inference: make prediction about the future using data about the
past
Clustering —> grouping according to similarity
Planning —> (robot to learn to interact in a certain environment)
Classification —> (assign meaning to data) example: Spam filtering
I want to predict the outcome of this individual or i want to predict whether a
person click or not in a certain advertisement.
Examples
Classify data into categories:
Medical diagnosis: data are medical records and • categories are diseases
• Document analysis: data are texts and categories are topics
• Image analysts: data are digital images and for categories name of objects
in the image (but could be different).
• Spam filtering: data are emails, categories are spam vs non spam.
• Advertising prediction: data are features of web site visitors and categories
could be click/non click on banners.
Classification : Different from clustering since we do not have semantically
classification (spam or not spam) —> like meaning of the image.
I have a semantic label.
Clustering: i want to group data with similarity function.
Planning: Learning what to do next
Clustering: Learn similarity function
Classification: Learn semantic labels meaning of data
Planning: Learn actions given state
In classification is an easier than planning task since I’m able to make
prediction telling what is the semantic label that goes with data points.
If i can do classification i can clustering.
If you do planning you probably classify (since you understanding meaning in
your position) and then you can also do clustering probably.
We will focus on classification because many tasks are about classification.
Classify data in categories we can image a set of categories.
For instance the tasks:
‘predict income of a person’
‘Predict tomorrow price for a stock’
The label is a number and not an abstract thing.
We can distinguish two cases:
The label set —> set of possible categories for each data • point. For each of
this could be finite set of abstract symbols (case of document classification,
medical diagnosis). So the task is classification.
• Real number (no bound on how many of them). My prediction will be a real
number and is not a category. In this case we talk about a task of
regression.
Classification: task we want to give a label predefined point in abstract
categories (like YES or NO)
Regression: task we want to give label to data points but this label are
numbers.
When we say prediction task: used both for classification and regression
tasks.
Supervised learning: Label attached to data (classification, regression)
Unsupervised learning: No labels attached to data (clustering)
In unsupervised the mathematical modelling and way algorithm are score and
can learn from mistakes is a little bit harder. Problem of clustering is harder to
model mathematically.
You can cast planning as supervised learning: i can show the robot which is
the right action to do in that state. But that depends on planning task is
formalised.
Planning is higher level of learning since include task of supervised and
unsupervised learning.
Why is this important ?
Algorithm has to know how to given the label.
In ML we want to teach the algorithm to perform prediction correctly. Initially
algorithm will make mistakes in classifying data. We want to tell algorithm that
classification was wrong and just want to perform a score. Like giving a grade
to the algorithm to understand if it did bad or really bad.
So we have mistakes!
Algorithm predicts and something makes a mistake —> we can correct it.
Then algorithm can be more precisely.
We have to define this mistake.
Mistakes in case of classification:
If category is the wrong one (in the simple case). We • have a binary signal
where we know that category is wrong.
How to communicate it?
We can use the loss function: we can tell the algorithm whether is wrong or
not.
Loss function: measure discrepancy between ‘true’ label and predicted
label.
So we may assume that every datapoint has a true label.
If we have a set of topic this is the true topic that document is talking about.
It is typical in supervised learning.
\\\\
How good the algorithm did?
\\

\[\ell(y,\hat{y})\leq0 \]

were $y $ is true label and $\hat{y}$ is predicted label
\\\\
We want to build a spam filter were $0$ is not spam and $1$ is spam and that 
Classification task: 
\\\\
$
\ell(y,\hat{y} = \begin{cases} 0, & \mbox{if } \hat{y} = y 
\\ 1, & 
\mbox{if }\hat{y} \neq y
\end{cases}
$
\\\\
The loss function is the “interface” between algorithm and data.
So algorithm know about the data through the loss function.
If we give a useless loss function the algorithm will not perform good: is
important to have a good loss function.
Spam filtering
We have two main mistakes:
It is the same mistake? No if i have important email and you classify as spam
that’s bad and if you show me a spam than it’s ok.
So we have to assign a different weight.
Even in binary classification, mistakes are not equal.
e Iotf.TFprIuos.uos
True came
razee
Cussler aircN TASK spam ACG FIRM
ftp.y GO
IF F Y n is soon
IF FEY 0 Nor spam
ZERO CNE Cass
n n
Span No Seamy Binary Classification
I 2
FALSE PEENE Mistake Y NON SPAM J Spam
FN Mistake i f SPAM y NO spam
2 IF Fp Meter Airenita
f Y F on positive
y ye en MISTAKE
0 otherwise


\paragraph{Outline}
The remainder of this article is organized as follows.
Section~\ref{previous work} gives account of previous work.
Our new and exciting results are described in Section~\ref{results}.
Finally, Section~\ref{conclusions} gives the conclusions.


\section{Lecture 2 - 07-04-2020}

\subsection{Argomento}
Classification tasks\\
Semantic label space Y\\
Categorization Y finite and\\ small
Regression Y appartiene ad |R\\
How to predict labels?\\
Using the lost function —> ..\\
Binary classification\\
Label space is Y = { -1, +1 }\\
Zero-one loss\\

$
\ell(y,\hat{y} = \begin{cases} 0, & \mbox{if } \hat{y} = y 
\\ 1, & 
\mbox{if }\hat{y} \neq y
\end{cases}
\\\\
FP \quad \hat{y} = 1,\quad y = -1\\
FN \quad \hat{y} = -1, \quad y = 1
$
\\\\
Losses for regression?\\
$y$, and $\hat{y} \in \barra{R}$, \\so they are numbers!\\
One example of loss is the absolute loss: absolute difference between numbers\\
\subsection{Loss}
\subsubsection{Absolute Loss}
$$\ell(y,\hat{y} = | y - \hat{y} |  \Rightarrow absolute \quad loss\\ $$
--- DISEGNO ---\\\\
Some inconvenient properties:

\begin{itemize}
\item ...
\item Derivative only two values (not much informations)
\end{itemize}

\subsubsection{Square Loss}
$$ \ell(y,\hat{y} = ( y - \hat{y} )^2  \Rightarrow \textit{square loss}\\$$
-- DISEGNO ---\\
Derivative :
\begin{itemize}
\item more informative
\item and differentible 
\end{itemize}
Real numbers as label $\rightarrow$ regression.\\
Whenever taking difference between two prediction make sense (value are numbers) then we are talking about regression problem.\\
Classification as categorization when we have small finite set.\\\\

\subsubsection{Example of information of square loss}

$\ell(y,\hat{y}) = ( y - \hat{y} )^2 = F(y) 
\\
F'(\hat(y)) = -2 \cdot (y-\hat{y})
$
\begin{itemize}
\item I'm under sho or over and how much
\item How much far away from the truth
\end{itemize}
$ \ell(y,\hat{y}) = | y- \hat{y}| = F(y') \cdot F'(y) = Sign (y-\hat{y} )\\\\ $
Question about the future\\
Will it rain tomorrow?\\
We have a label and this is a binary classification problem.\\
My label space will be Y = { “rain”, “no rain” }\\
We don’t get a binary prediction, we need another space called prediction space (or decision space). Z = [0,1]\\
$
Z = [0,1]
\hat{y} \in Z \qquad \hat{y} \textit{ is my prediction of rain tomorrow}
\\
\hat{y} = \barra{P} (y = "rain") \quad  \rightarrow \textit{my guess is tomorrow will rain (not sure)}\\\\ 
y \in Y \qquad \hat{y} \in Z \\quad \textit{How can we manage loss?}
\\
\textit{Put numbers in our space}\\
\{1,0\} \quad \textit{where 1 is rain and 0 no rain}\\\\
$
I measure how much I’m far from reality.\\
So loss behave like this and the punishment is gonna go linearly??\\
\[26..\]\\
However is pretty annoying. Sometime I prefer to punish more so i going quadratically instead of linearly.\\
There are other way to punish this.\\
I called \textbf{logarithmic loss}\\ 
We are extending a lot the range of our loss function.\\

$$
\ell(y,\hat{y}) = | y- \hat{y}| \in |0,1| \qquad \ell(y,\hat{y}) = ( y- \hat{y})^2 \in |0,1|
$$
\\
If i want to expand the punishment i use logarithmic loss\\
\\
$ \ell(y,\hat{y} = \begin{cases} ln \dfrac{1}{\hat{y}, & \mbox{if } y = 1 \textit{(rain)} 
\\ ln \frac{1}{1-\hat{y}}, & 
\mbox{if } y = 0 \textit{(no rain} 
\end{cases}
\\\\
F(\hat{y}) \rightarrow can be 0 if i predict with certainty

\mbox{if} \hat{y} = 0.5 \qquad \ell(y, \dfrac{1}{2}) = ln 2 \quad \textit{costnat losses in each prediction}\\\\
\lim_{\hat{y}\to\0^+} \ell(1,\hat{y}) = + \inf

$

\section{Lecture 3 - 07-04-2020}
\section{Lecture 4 - 07-04-2020}
\section{Lecture 5 - 07-04-2020}
\section{Lecture 6 - 07-04-2020}
\section{Lecture 7 - 07-04-2020}
\section{Lecture 8 - 07-04-2020}
\section{Lecture 9 - 07-04-2020}

\section{Lecture 10 - 07-04-2020}

\subsection{TO BE DEFINE}

$|E[z] = |E[|E[z|x]]$
\\\\
$|E[X] = \sum_{t = 1}^{m} |E[x \Pi(A\begin{small}
t \end{small} ) ]$
\\\\
$x \in \mathbb{R}^d
$
\\
$\mathbb{P}(Y_{\Pi(s,x)} = 1) = \\\\ \mathbb{E}[\Pi { Y_{\Pi(s,x)} = 1 } ] = \\\\
= \sum_{t = 1}^{m} \mathbb{E}[\Pi\{Y_t = 1\} \cdot \Pi { Pi(s,x) = t}] = \\\\
= \sum_{t = 1}^{m} \mathbb{E}[\mathbb{E}[\Pi\{Y_t = 1\} \cdot \Pi\{\Pi(s,x) = t\} | X_t]] = \\\\
given the fact that Y_t \sim \eta(X_t) \Rightarrow give me probability \\
Y_t = 1 and \Pi(s,x) = t are independent given X_Y (e. g. \mathbb{E}[Zx] = \mathbb{E}[x] \ast \cdot \mathbb{E}[z]\\\\
= \sum_{t = 1}^{m} \barra{E}[\barra{E}[\Pi\{Y_t = 1\}|X_t] \cdot \barra{E} [ \Pi(s,x) = t | Xt]] = \\\\
= \sum_{t = 1}^{m} \barra{E}[\eta(X_t) \cdot \Pi \cdot \{\Pi (s,x) = t \}] = \\\\
= \barra{E} [ \eta(X_{\Pi(s,x)}]
$

\[ \barra{P} (Y_{\Pi(s,x)}| X=x = \barra{E}[\eta(X_\Pi (s,x))] \]
\\\\

$
\barra{P} (Y_{\Pi(s,x)} = 1, y = -1 ) = \\\\
= \barra{E}[\Pi\{Y_{\Pi(s,x)  }= 1\} \dot \Pi\{Y= -1|X\} ]] =  \\\\ 
= \barra{E}[\Pi \{ Y_{\Pi(s,x)} = 1\} \cdot \Pi \{ y = -1 \} ] = \\\\
= \barra{E}[\barra{E}[\Pi \{ Y_{\Pi(s,x)} = 1\} \cdot \Pi \{ y = -1 | X \} ]] = \\\\
$

\[ Y_{\Pi(s,x)} = 1  \quad \quad y = -1 (1- \eta(x)) \quad when \quad X = x\]

$
\\\\ = \barra{E}[\barra{E}[\Pi \{Y_\Pi(s,x)\} = 1 | X] \cdot \barra{E}[\Pi \{y = -1\} |X ]] = \\\\
= \barra {E}[\eta_{\Pi(s,x)} \cdot (1-\eta(x))] = \\\\
similarly: \quad \barra{P}(Y_{\Pi(s,x)} = -1 ,  y = 1) = \\
\barra{E} [(1- \eta_{\Pi(s,x)}) \cdot \eta(x)]
\\\\
\barra{E} [ \ell_D (\hat{h}_s)] = \barra{P}(Y_{\Pi(s,x)} \neq y ) = 
\\\\
= \barra{P}(Y_{\Pi(s,x)} = 1, y = -1) + \barra{P}(Y_{Pi(s,x)} = -1, y = 1) = 
\\\\
= \barra{E} [\eta_{\Pi(s,x)} \cdot (1-eta(x))] + \barra{E}[( 1- \eta_{\Pi(s,x)})\cdot \eta(x)]$
\\\\
Make assumptions on $D_x \quad and \quad \eta$: \\


MANCAAAAAAA ROBAAA
\\\\

$
\eta(x') <= \eta(x) + c || X-x'|| --> euclidean distance
\\\\
1-\eta(x') <= 1- \eta(x) + c||X-x'||
\\\\
$


$
X' = X_{Pi(s,x)}
\\\\
\eta(X) \cdot (1-\eta(x')) + (1-\eta(x))\cdot \eta(x') <=
\\\\
<= \eta(x) \cdot((1-\eta(x))+\eta(x)\cdot c||X-x'|| + (1-\eta(x))\cdot c||X-x'|| = 
\\\\
= 2 \cdot \eta(x) \cdot (1- \eta(x)) + c||X-x'|| \\\\
\barra{E}[\ell_d \cdot (\hat{h}_s)] <= 2 \cdot \barra{E} [\eta(x) - (1-\eta(x))] + c \cdot \barra(E)[||X-x_{\Pi(s,x)}||]
$
\\ where $<=$ mean at most
\\\\
Compare risk for zero-one loss
\\
$
\barra{E}[min\{\eta(x),1-\eta(x)\}] = \ell_D (f*)
\\\\
\eta(x) \cdot( 1- \eta(X)) <= min\{\eta(x), 1-eta(x) \} \quad \forall x
\\\\
\barra{E}[\eta(x)\cdot(1-\eta(x)] <= \ell_D(f*)
\\\\
\barra{E}[\ell_d(\hat{l}_s)] <= 2 \cdot \ell_D(f*) + c \cdot \barra{E}[||X-X_{\Pi(s,x)}||]
\\\\
\eta(x) \in \{0,1\}
$
\\\\
Depends on dimension: curse of dimensionality
\\\\--DISEGNO--
\\\\
$
\ell_d(f*) = 0 \iff min\{ \eta(x), 1-\eta(x)\} =0 \quad$ with probability = 1
\\
to be true $\eta(x) \in \{0,1\}$

\section{Previous work}\label{previous work}
A much longer \LaTeXe{} example was written by Gil~\cite{Gil:02}.

\section{Results}\label{results}
In this section we describe the results.

\section{Conclusions}\label{conclusions}
We worked hard, and achieved very little.

\bibliographystyle{abbrv}
\bibliography{main}

\end{document}
This is never printed