0% found this document useful (0 votes)
5 views50 pages

Understanding Support Vector Machines

Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
5 views50 pages

Understanding Support Vector Machines

Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd

Support Vector

Machine
Contents
 Support Vector Machine
Support Vectors

 Hard Margin

 Linear Separability

 SVM for Linear Classification

 SVM for Non-linear Classification


 Kernel SVM
Support Vector Machine
 Supervised ML algorithm

 Mostly used for Classification

 Used for linear classification

 Can also be used for non-linear classification –Kernel SVM

 Application: text classification, image classification, spam detection,


handwriting identification, face detection, anomaly detection, etc.
Linearly Separable Data
Linearly Separable Data

• SVM helps us choose the best line or in general the best hyperplane that
segregates the data points.
Support Vector Machine
𝒙𝟐
• Objective: Find the optimal line in a 2D
space that can separate the data points

• Best hyperplane or line maximizes the


separation margin between the data
points

𝒙𝟏
Best Separating Hyperplane
𝒙𝟐
𝐿1
𝐿2
𝐿3
: Maximum-margin hyperplane/hard margin

The hyperplane whose distance from the


nearest data points on each side is maximum.

𝒙𝟏
Best Separating Hyperplane
𝒙𝟐
Select the hyperplane so that the Hard margin/
margin is maximized. Decision surface

Margin: The separation between the data


points of both classes 𝑑1

𝑑2
𝑑1=𝑑 2=𝑑 ⇒ 𝑑 1+ 𝑑 2=2 𝑑
Margin

𝒙𝟏
Support Vectors
𝒙𝟐
Decision surface
• Data points that lie closest to the
decision surface or hyperplane.

• Most difficult to classify.


𝑑

𝒙𝟏
Support Vectors
𝒙𝟐
New Decision
surface
• Support vectors have direct influence on
the optimum location of the decision
surface
𝑑𝑛𝑒𝑤
𝑑𝑛𝑒𝑤

𝒙𝟏
Support Vectors
𝒙𝟐
New
Decision
• Support vectors have direct Decision surface surface
influence on the optimum
location of the decision surface

𝒙𝟏
Mathematical Formulation:
Classification

Input SVM Target


Feature Output
Class 1 or class 2
( 1) (𝟐 ) (𝒎 )
𝑋=[ 𝐱 , 𝐱 , … , 𝐱 ]
𝐱 (𝒊 )=[𝑥 1 , 𝑥 2 , … , 𝑥𝑛 ] ( 𝑖)
y ∈ {+ 1 ,−1 }
Parameters: Set of weights, and bias
𝒘 =[𝑤 1 , 𝑤2 , …, 𝑤𝑛 ]
Mathematical Formulation:
Classification
𝒙𝟐
( 1) (𝟐 ) ( 𝟏𝟑 )
𝑋=[ 𝐱 , 𝐱 , … , 𝐱 ]

𝐱 (𝒊 )=[ 𝐱 (1𝒊 ) , 𝐱 (2𝒊 ) ]

or

𝒘 ={𝑤 1 , 𝑤2 }

𝒘 𝑇 𝐱 +𝑏=0 𝒙𝟏
Mathematical Formulation:
Classification
The SVM classifier can be defined as 𝒙𝟐

Decision function for binary


classification

𝑇
𝒘 𝐱 +𝑏=0 𝒙𝟏
Mathematical Formulation:
Classification
𝒙𝟐

(𝐢 )
The distance between data pointand the 𝐱
decision boundary

|𝒘 𝑇 𝐱 ( 𝐢) +𝑏| |𝒘 𝑇 𝐱( 𝐢 )+ 𝑏|
𝑑𝑖 = =
𝒘 𝒘
𝑇
‖𝒘‖2

Decision boundary
𝑇
𝒘 𝐱 +𝑏=0 𝒙𝟏
Mathematical Formulation:
Classification
𝒙𝟐

distance between and (𝐢 )


𝐱

𝒘 𝐱
‖𝒘‖2

‖𝒘‖2 𝒘 𝑇 𝒙 𝒊+𝑏 𝒙𝟏
⟹ 𝛼=(𝒘 ¿¿𝑇 𝒙¿¿ 𝒊+𝑏) 𝑇 = ¿¿
𝒘 𝒘 ‖𝒘‖2
Mathematical Formulation:
Classification
𝒙𝟐
𝒘
𝐱 (𝐢 ) − 𝐱 =𝜶 (𝐢 )
‖𝒘‖2 𝐱

(𝐢 )
𝑑𝑖 =¿ 𝐱 − 𝐱 ∨¿ 𝜶
‖ 𝒘

‖𝒘‖2 2
=‖𝜶‖2 𝒘
‖𝒘‖2
𝐱

|𝒘 𝑇 𝐱 (𝐢 ) +𝑏|
⟹ 𝑑𝑖=
‖𝒘‖2

𝒙𝟏
Mathematical Formulation:
Classification
The objective of SVM is to find a decision line so that 𝒙𝟐
the distance of each data point from the decision
boundary will be very high.

Optimization problem.

But instead of considering all data points, the SVM


considers only difficult data points that is the
support vectors.

𝒙𝟏
Mathematical Formulation :
Classification
𝒙𝟐

Assumption: All data points are at a


distance larger than 1 from the hyperplane.

𝒙𝟏
Mathematical Formulation :
Classification
𝒙𝟐

(Hard margin)

= Distance of from the support vector corresponding to the


class

= Distance of from the support vector corresponding to the 𝐻1


class
𝒙𝟏
𝐻0 𝐻2
Mathematical Formulation :
Classification
𝒙𝟐

𝑑+¿ 𝑑 −=𝑑

𝑑+¿ 𝑑
1
𝑑𝑖 = 𝑑 −=d
‖𝒘 ‖2

2
⟹ 𝑀𝑎𝑟𝑔𝑖𝑛= 𝐻1
‖𝒘‖2
𝒙𝟏
𝐻0 𝐻2
Mathematical Formulation :
Classification
𝒙𝟐
2
𝑀𝑎𝑟𝑔𝑖𝑛= 𝒘 𝑇 𝐱 +𝑏=1
‖𝒘‖2

𝑇
Objective: Maximize the margin 𝒘 𝐱 +𝑏=−1
Minimize
𝑑
Minimize

𝑑
Condition: There are no data points between and
when
when
𝑇
𝑦 (𝒘 𝐱 +𝑏) ≥1 𝒙𝟏
𝑦 ( 𝒘 𝑇 𝐱 + 𝑏 ) −1 ≥ 0 𝒘 𝑇 𝐱 +𝑏=0
Mathematical Formulation :
Classification

Optimization problem:

Constrained optimization problem

Can be solved by the Lagrangian multiplier method


Optimal Parameter Calculation
Langrangian

With respect to and .

slack variables
Optimal Parameter Calculation
Solution: Find partial derivative wrt and b and equate to
Optimal Parameter Calculation
We have the Primal problem as

,
We got

∑𝑎
𝑚 (𝑖) (𝑖)
𝒘 =∑ 𝑎 𝑦 𝐱(𝑖) (𝑖) (𝐢)
and 𝑦 =0
𝑖=1 𝑖=1
Optimal Parameter Calculation
Langrange dual problem: Instead of minimizing subject to constraints
involving ’s, we can maximize over ’s, subject to

𝑚
𝒘 =∑ 𝑎 𝑦 𝐱 ( 𝑖) (𝑖 ) ( 𝐢)

𝑖=1

∑𝑎 (𝑖 ) (𝑖 )
𝑦 =0
𝑖 =1
Optimal Parameter Calculation
𝒙𝟐
𝑇
𝒘 𝐱 +𝑏=−1
Karush-Kuhn-Tucker (KKT) condition

𝑇
𝒘 𝐱 +𝑏=1
𝑎
(𝑖 )
( 𝑦 (𝑖 ) ( 𝒘 𝑇 𝐱 ( 𝐢 ) ,+𝑏 ) −1 ) =0

𝒙𝟏
𝒘 𝑇 𝐱 +𝑏=0
Optimal Parameter Calculation
Langrange dual problem:

𝑚 𝑚 𝑚 𝑚
1
max 𝐿 𝐷=¿ ∑ 𝑎 −(𝑖 )
∑ ∑ 𝑎(𝑖 ) 𝑎( 𝑗 ) 𝑦 ( 𝑖 ) 𝑦 ( 𝑗 ) ( 𝐱 (𝐢 ) . 𝐱 ( 𝐣 ) ) +𝑏 ∑ 𝑎( 𝑖 ) 𝑦 (𝑖 ) ¿
𝑖 =1 2 𝑖=1 𝑗 =1 i=1

s uch that ∑ 𝑎 (𝑖 ) 𝑦 ( 𝑖 )=0


𝑗 =1

and ,
Optimal Parameter Calculation
Langrange dual problem:

𝑚 𝑚 𝑚
1
max 𝐿 𝐷 =¿ ∑ (𝑖 )
𝑎 − ∑ ∑ 𝑎(𝑖 ) 𝑎( 𝑗 ) 𝑦 ( 𝑖 ) 𝑦 ( 𝑗 ) ( 𝐱 (𝐢 ) . 𝐱 ( 𝐣 ) ) ¿
𝑖 =1 2 𝑖=1 𝑗 =1

s uch that ∑ 𝑎 (𝑖 ) 𝑦 ( 𝑖 )=0


𝑗 =1

and ,
Optimal Parameter Calculation
Findby taking the derivative of wrt and equate it to zero.

Find the values of’s for all and compute optimal as

𝑚
𝒘 ∗ =∑ 𝑎 (𝑖 ) 𝑦 ( 𝑖 ) 𝐱 (𝐢 )
𝑖=1

𝑚
𝒘 ∗= ∑ 𝑎(𝑖 ) 𝑦 (𝑖 ) 𝐱 ( 𝐢 )
𝑖=1
(𝑖 )
𝑎 >0
Optimal Parameter Calculation
To compute optimal bias:
Computefor each support vector as

Optimal bias: Average ’s over all support vectors

𝑏∗ = avg { 𝑏( 𝑖 ) }
𝑎 ( 𝑖) >0
Inference
For a new data point
𝑚
𝒘 ∗= ∑ 𝑎( 𝑖) 𝑦 (𝑖 ) 𝐱 (𝐢 )
^𝑦 =sign(𝒘 ∗𝑇 𝐳 +𝑏∗ ) ¿𝑖 =1
( 𝑖)
¿𝑎 >0

(∑ )
𝑚
⇒^
𝑦 =sign (𝑖 ) ( 𝑖)
𝑎 𝑦 𝐱 . 𝐳 +𝑏 (𝐢) ∗
𝑏∗ = avg { 𝑏( 𝑖 ) }
𝑖 =1 𝑎 ( 𝑖) >0
Intuition behind the slack
variables
The first optimization problem

The Primal problem


Intuition behind the slack
variables
𝑚 𝑚
1
𝐿𝑝 = ‖𝒘‖ − ∑ 𝑎(𝑖 ) 𝑦 (𝑖 ) ( 𝒘 𝑇 𝒙 𝒊+𝑏 ) + ∑ 𝑎(𝑖 )
2 2 𝑖=1 𝑖=1

𝑚
1
⇒ 𝐿𝑝 = ‖𝒘‖ − ∑ ¿ ¿
2 2 𝑖=1
Intuition behind the slack
variables
𝑚
1
𝐿𝑝 = ‖𝒘‖ − ∑ ¿ ¿
2 2 𝑖 =1

Consider
Intuition behind the slack
variables
Term 1 can be either or a positive value, i.e.,
Support Vector Machine for Not-
linearly Separable Data
𝒙𝟐

Modify the objective function to solve the problem


𝒙𝟏
Support Vector Machine for Not-
linearly Separable Data
Previously

𝑦 ( 𝒘 𝑇 𝐱 + 𝑏 ) −1 ≥ 0

𝑦 𝒘 𝐱 +𝑏 ) ≥1, ∀ 𝑖
( 𝑇 (𝐢 )
(𝑖 )

Maximize the margin and get as many data points classified correctly as
possible.

Modified constraint
𝑦 (𝑖 ) ( 𝒘 𝑇 𝐱 (𝐢 ) +𝑏 ) ≥1 − 𝜉 𝑖 , ∀ 𝑖
Support Vector Machine for Not-
linearly Separable Data
𝒙𝟐
Misclassified data points

𝜉3

𝜉4 𝜉2 𝜉5
Erroneous data points, within the margin

𝜉1

𝒙𝟏
Support Vector Machine for Not-
linearly Separable Data
𝑦 (𝑖 ) ( 𝒘 𝑇 𝐱 (𝐢 ) +𝑏 ) ≥1 − 𝜉 𝑖 , ∀ 𝑖

𝜉𝑖 ≥ 0 , ∀ 𝑖
𝑚

∑ 𝜉 𝑖 ≤ 𝐶𝑜𝑛𝑠𝑡𝑎𝑛𝑡
𝑖=1
Support Vector Machine for Not-
linearly Separable Data
Modified objective function

𝑚
𝒘 ∗= ∑ 𝑎( 𝑖) 𝑦 (𝑖 ) 𝐱 (𝐢 )
𝑏∗ = avg { 𝑏( 𝑖 ) }
¿𝑖 =1
( 𝑖)
¿𝑎 >0 𝑎 ( 𝑖) >0
Support Vector Machine for Not-
linearly Separable Data

Gain linearly separation by mapping the data to a higher dimensional space


Kernel SVM
Dual problem:

𝑚 𝑚 𝑚
1
max 𝐿 𝐷 =¿ ∑ 𝑎 (𝑖 )
− ∑∑ 𝑎(𝑖 ) 𝑎( 𝑗 ) 𝑦 ( 𝑖 ) 𝑦 ( 𝑗 ) ¿ ¿
𝑖 =1 2 𝑖=1 𝑗 =1
Kernel SVM
For a new data point

^
𝑦 = sign ¿
Kernel SVM
Polynomial Kernel
𝑑
1+ ⟨ 𝐱 , 𝐱 ⟩
(𝐢 ) (𝐣)

RBF Kernel
¿
ANN tanh Kernel
𝐾 1 ⟨ 𝐱 (𝐢 ) , 𝐱 ( 𝐣 ) ⟩ + 𝐾 2
Kernel SVM
Kernel SVM
References
1. NPTEL - Introduction to Machine Learning by Prof. Balaraman Ravindran, IIT
Madras [Link]
2. NPTEL - Machine Learning by Prof. Sudeshna Sarkar, IIT Kharagpur
[Link]
3. Machine Learning-McGraw-Hill Education (1997) BY Tom M. Mitchell
4. Burkov, Andriy. The hundred-page machine learning book. Vol. 1. Quebec
City, QC, Canada: Andriy Burkov, 2019.
5. Rogers, Simon, and Mark Girolami. A first course in machine learning.
Chapman and Hall/CRC, 2016.
Thank You

You might also like