Intro to ML

Lee-1
:

- PCA cond .

-
what is M2 ?

- ML and SDE

-
ML Tasks

learning
-

Types of
-
1 9650204730
(Whatsapp
# :

Email
jitn. Gupta
:
& scelecom -1

notes + cheat
PCA >
- Short note fall Sheet
=> Lec10
Math for ML
d
lecture notes
in today's
as well .
Rap a Weight Y
X -
x
T

the columns
S1) Standardize
·
> X

direction with
S2) find the

manimum variance

>
- Height

Goal
-
>
-
Maximize
projection
* I

angman-

U
IX
N
·
i= 1
lete
2

argman + Eiti)2 St .
Inli oe

- i* =

argman / (FU)" + In e
-
 i
i. T
-
,I .

U ,

Xi =

+T
F
..
.

Uz =

u
Vector ! as :
-
T
U: . .

i: T T
!
·

!
. .

InT Ud int NX1

Wed dx 1

Xull
" "

1 =
(i ,
+
5) + (F2 5) +...+ (UNT)

Kull- (iiT
Xull (xn) (xn) Vall=
11 =

E
T
BTAT
identify (AB) =

Xull"
* +
=> x xi
I =
 Jul 2 -
argman/UTXTX En e
+
= T* xix
covarice
=
=

Jul
[t 1
* UT Si +
=>
i
-

-angmen
-
it to zeno
& equate
Take the gradient

dul-6] = S +
6 Fn
-fuls
+

u5 Si + x(2)
=* *

Ensures
I

TC ] b(2n)
↓x2si +
-..
=
-

. .

i
 2s 6(2)
↓
if + =
0

- si = =
JNT

ded dx1
-
Si = d'u u
X = constant
vo
so
vector
scales
matrina Vector a 5.
2 (4 , 22

when a matrice is multiplied i40 (2, 1)

of what happens >

a vector
with
 7-
-

3
Es .
s =

45
i =

[5]
2x 1
- +2

Sa :

1 ]13T [
%
=
J Est] List
: =

1 =
3, 23) N
M
Si
si du

(2 3)
.

i
-

C N >
f
 /45
Eg
if
-
-
- -

s =
= =

si :

[ ] ( i / kit
"

[=
d
1 Si (6 24) d'u
Si
,
=

= 6
1 i (1 4)
-

Eigenvector of matric
,

* (1 , u) is a

3 is G
q

S and its Eigenvalue

 Eigenvectors and Eigenvalues

Si = 61n

and bi its eigenvalue

eigenvector of
s
zb i is an

have
'd'
(dxd) will
Covenance
matric s
eigenvalue
=>
has
eigenvector
an
and each
eigenvectors -

has
with maximum variance
i The directions

the largest eigenvalue

 and eigenvalue
- we will find all the eigenvector
pains
order of
- We will sort them in the descending
d 3
eigenvalues . >
-

it is is in its--Id

de da da du As ↓d

> decreasing
eigenvalue
 d d' 3
Eg.
=
>
-

directions Un is
-

My
we
will take our final as :

um
d-dimensional
co-ordinates of all the datapoints
Now just
need to find
datapoint along
we
each
-

directions . - take proj of

these 3
along i, Us
My ,

↑ e
Cas
1 -I
----

2 la
Li , is i Us (
i

i, -
E
, H2 ⑨ 3
, -
2
Miz

Us)
"
E2
Kie (iz" i M2 H2 ,

: !
did
[ETNT U , ,
ITNTE2 ,
IN Us)
 1)
i -- 2

Eg . d = 4 i =
2
=
2
Us I
O

I O

O
I
Y
2,
5
2

-
I fe fo ts fu
-
T
u. U 52 -
1

+ 50+211 + ( 1) X 1
"
5 12
-

4 X1 R2 50
i ,Th ,
-

= -

i ,T in = 4 + 10 + 2 +0
= S

-1 = T
8 + 0
ii is +
T = 0

U
i I 5
5 zb
8
2
T
·
I
 PCA
Information retained after
i, is is in --- id
*

d d d
, da bu Id
* >
-
33

di+ but by +- ... Jal

explained
z
variance

↓ ,
d + 62 +
by -... + dd

into retained
d

&
Edi
·

Edj
ja /
 So fas : mathematical implementation
M
and Pandas)
DS libraries (Numpy data manipulation
Seaborn)"
- >
-

"(Matplotlib and
preprocessing
&
& data viz .

& EDA
conditional prob
Prob stats marginal & -

- -

~
prob dist.
.

hypothesis testing
"

& Coordinate Geometry

Linear algebra
mathematically
-

them
Y Build intuitions I implement

- Calculus & optimization

it -I l >
-
Py = 00g
↳
gradients
Y manima & 0:6
[dell
minima
is
+
P1 =

↑
GradientDescent
in naive bayes models
conditional probabilitie
 Visualization
Pta
>
-
(Basic revision

Numerical >
-
histogram ,
KDE ,
bomplot...
>
-
I variable
↳ Categorical >
- count plot barplot
,
.....

< Num-Num >

-

Scatterphof
Pd Crosstab (Contigency table)
Barplots ,
-

I variables car-cet >

-

>

>
- Bouplot
>
- Num-Cat

with have
Scatte plot
3 variables >
->
Num-Num-cat ->
 height
d
weight , height ,
genden
a :
... o
·
nue-gender
..
ou
·
8 - Male
-

Female
O G O o

· +

...
=> u
O

> weight
marks
in a class
Quiz) Students
- bouplot
histogram
-

↳ Numerical >
-
,

Inso 20
s
 ML Classical
programming
vs

-Span
Email
· non-spam (hem)

Classical programming perspective

emails
* Identity commonly occuring/phrases in span

~ store them - spam-words

span-words O
Grade
words
>
-
Spon
of
in
lot
email >
-

if a

- else >
-

Cham d
-
> Decision
 Final
Rules
>
- Decision
scass
llogic

can learn the rules

* Rules are rigid >
-
Spammers
pattern needs to be
* A lot of
hard coding - every new

added separately
"mine"
o can me
create algorithms which can

the data !
the rules and patterns from
ML perspective
I-
Imput Decision
>
-
Model >
-

SSpan/Non-Spam)
( Machine -

learning
the basis
Decisions
learning patterns
>
- on

3 Machine is
learnt
of patterns

Data Decision/labels
* Visualize ML approach
NS
Es
>
model learns the
data
-

training
NS

Input >
-

Training Ez
pattern from
S
e NS data
↓
per S
Trained model
N21000 ↓
email
Predict whether
is spam) non-spam
 is ML ?
what
emails spam or

classification of
as
Task (T) :
eg. non-spam

labelled emails
Experience (E) :
Training data .
eg

how
well
Metric which
quantities
Performance (P)
:

Your
model is
working

eg
. Actual Predicted
NS % Jemails accurately classified
④ S

② S S
I & x100 = 80%
③ NS NS 100

④ NS NS
i (

L
As you provided more experience (training datal
should increase
performance
.

your

TEP model >

-
Yask-experience
-

performance
 Task Experience Performance

difference in the actual

historical prices
① predict stock value
stock
and predicted
prices of the

& categorize Labelled images , % of accurately predicted

images into labolled
with images
images
rabbit
cat ,
dog , as C/D/R
Types tasks
of
* Classification A Regression
data points
Predict a

classifying numerical value

into one of the categories

Recommendation
*
Clustering *
video,
Similar samples
movies
Recommend a

Group
etc
data points product .

or

* Time Series forecasting

stock values , future sales revenue

Predicting
,
 Classification task
ef .
-
using labelled data >
-

Supervised Algo
fron trand or not-fraud
Frand detection >
-
classifying as

mail as spam non-span

-
Classifying

task labelled date -> Supervised Algo

Regression eg > using
.
-

house prices #ufrooms area

of house , locality etc

Predicting
-

,
-

E
T

Spinny
-
+ model , age mileage , odometer
old
,
-

predicting
car
prices
T
 Clustering task eg .
>
-
Unsupervised learning

labelled data
-
does require
not
their
customers based or
-
Amazon is
grouping
and purchasing patterns
.

profile Amount spent

Address Total purchases d

⑱C
T
I
-
Cust I :
-
·:
!:
O

O
:
O

- ↳I

#of times
- -

customel
visited
 transactions data
Clustering
-

on basis

# of time was rejected

a
transaction

⑭te O -
f f
X X

i
x

De
+ ↓ X
x

#y *

of
y
amount
>
Monetary
transaction
random images
- A lot of
Y
group
all images
which contain dog ,
cat
,
rabbit
 car
Dog
·
>
-

OO
1 B is P
B

BB
Q Dog
B is D
D B B

O
>
-
Rabbit
D
i D D

- Rabbit
 & Unsupervised
Impervised Learning
labelled data
Supervised >
-
They learn
using
d target
f f2 As -- fd

I
.

&

2
labels
3

·
y
-
labdled date
unsupervised >
-

training data has no labels

learning >
-
Clustering
fitz . .
-
fd Y Recommendation
④

②
③
&

!
 Show houses with Grooms
Quiz Customers >
- me

all the houses

with rooms
Seller >
-

group
-
clustering
Field 1
Customer
-
Miz surveys
-
field 2

field 3

d
surveys
on
the basis of
grouping
responses

detering >
- unsupervised learning
 stock prices > historical data
Quiz >
-
predicting

volume
indicators of
Price
stock
labeled
no
Supervised learning ↓
future prices
 candidates
Quiz- job as pirants classifying

A month
HR >
-
notice period - <

> < 12(P

Salary expectation
-

:
Favorable

f
Candidate I -

train
candidate 2
- C Unfavorable -

-s ↑
candidate
Doubts
-

Novic
dataset 4
Genze Director Year

·
-
-

Name
g

M1 Matrin Sailfi
08 Mis
Mr
·

M2 Avatar Selfi O O

Comedy ·
M3 Toy story Animation ·
M10
My O
(
&

>
*
encoding

Ill
user
date
Me Ma M3 MuMgr
----
Miob

T
Us · 10
-

Us

-
Nationality Watch time

insus
-
...

I
 Year
Actors genzes
Director

]
-

Safi
Comedy
>
-

-
-

Mo , 00 TI
-
 SDE models
M2 model
+