Logistic Regression

Prerequisite-
-Descriptive statistics
-Linear Regression
Objectives-

- To understand the prerequisite terms such as odds, probability, log odds, logit function
and sigmoid function.
- What is logistic regression and the corresponding cost function?
- Learn how to optimize weights using gradient descent.
Odd and Probability-
The theory of chances for success or failure of the events are often expressed as odds and
probabilities. Probability can be defined as to how likely an event can occur. Odds is ratio
of the probability of success to failure of an event.

kumar.ashish2050@gmail.com
21YORICED7 Probability- Probability is defined as the ratio of the number of way that are favourable to
the occurrence of an event A to the total number of outcomes of the experiment.
Probabilities are expressed either as percentage or decimal and value lies between 0 and 1.
𝑐ℎ𝑎𝑛𝑐𝑒𝑠 𝑖𝑛 𝑓𝑎𝑣𝑜𝑟(𝐴)
𝑃𝑟𝑜𝑎𝑏𝑖𝑙𝑖𝑡𝑦 𝑃(𝐴) =
𝑡𝑜𝑡𝑎𝑙 𝑡𝑟𝑖𝑎𝑙𝑠
Example: The probability of getting Ace in a deck of 52 cards is given by:
4
𝑃(𝐴) = = 0.077 𝑜𝑟 7.7%
52

Odd- Odds is defined as the ratio to chances in the favor of event to the chances against
it. The value of odd may lie between 0 to ∞.
𝑐ℎ𝑎𝑛𝑐𝑒𝑠 𝑖𝑛 𝑓𝑎𝑣𝑜𝑟(𝐴)
𝑜𝑑𝑑𝑟𝑎𝑡𝑖𝑜(𝐴) =
𝑐ℎ𝑎𝑛𝑐𝑒𝑠 𝑎𝑔𝑎𝑖𝑛𝑠𝑡(𝐴)

𝑃
Mathematically Odds Ratio = 1−𝑃 , where P denotes the probability of success of the

desired event.

This file is meant for personal use by kumar.ashish2050@gmail.com only.

Proprietary content. © Great Learning. All Rights Reserved. Unauthorized use or distribution prohibited. 1
Sharing or publishing the contents in part or full is liable for legal action.
 Example: The odds of getting Ace in a deck of 52 cards is given by:
(4/52) 4
𝑜𝑑𝑑𝑠 𝑟𝑎𝑡𝑖𝑜(𝐴) = =
(48/52 ) 48
Relationship Between the Probability and the Odds Ratio-
𝑃𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 𝑜𝑓 𝑒𝑣𝑒𝑛𝑡(𝐴) 𝑜𝑐𝑐𝑢𝑟𝑖𝑛𝑔
𝑜𝑑𝑑𝑟𝑎𝑡𝑖𝑜 =
𝑃𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 𝑜𝑓 𝑒𝑣𝑒𝑛𝑡(𝐴) 𝑛𝑜𝑡 𝑜𝑐𝑐𝑢𝑟𝑖𝑛𝑔
𝑃(𝐴)
𝑜𝑑𝑑𝑟𝑎𝑡𝑖𝑜 (𝐴) =
1 − 𝑃(𝐴)
𝑜𝑑𝑑 𝑟𝑎𝑡𝑖𝑜
𝑃𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 𝑃(𝐴) =
1 + 𝑜𝑑𝑑 𝑟𝑎𝑡𝑖𝑜
Log odds and logit-
We now know that the odds ratio is the probability of an event occurring to the probability
of that event not occurring. Taking log of odd ratio is called log odds and defined as:
𝑃(𝐴)
log(𝐴) = log( )
1 − 𝑃(𝐴)

The logit function is the log odds function modelled for a probability ‘p’.
kumar.ashish2050@gmail.com
21YORICED7 𝑃
log(𝑜𝑑𝑑𝑠) = 𝑙𝑜𝑔𝑖𝑡(𝑃) = log(
)
1−𝑃
Logit Function- Logit function is mainly used in working with probabilities. The logit
function is the log of the odds that Y equals one of the categories. The value of logit
function varies between (−∞ , ∞). The value approaches towards ∞ when the probability
value approaches to 1 and it goes to −∞ when the probability value approaches to 0. The
logit function is very important in the field of statistics as it can map the probability values
ranges (0, 1) to a full range value of real numbers.
𝑧
𝑙𝑜𝑔𝑖𝑡(𝑦(𝑧)) = log( )
1−𝑧

This file is meant for personal use by kumar.ashish2050@gmail.com only.

Proprietary content. © Great Learning. All Rights Reserved. Unauthorized use or distribution prohibited. 2
Sharing or publishing the contents in part or full is liable for legal action.
 Figure 1: Logit Function

Sigmoid function- Sigmoid function can be thought of as an inverse to the logit function. This
means for a probability value P we have:
𝑃 = 𝑠𝑖𝑔𝑚𝑜𝑖𝑑(𝑙𝑜𝑔𝑖𝑡(𝑃))
Sigmoid (sometimes
kumar.ashish2050@gmail.com also called the S-curve) performs the inverse of logit which means it maps
21YORICED7
any arbitrary real number into the range of (0, 1). The function is defined as:
𝟏
𝒔𝒊𝒈𝒎𝒐𝒊𝒅(𝒛) =
𝟏 + 𝒆−𝒛

Figure 2 Sigmoid function

This file is meant for personal use by kumar.ashish2050@gmail.com only.

Proprietary content. © Great Learning. All Rights Reserved. Unauthorized use or distribution prohibited. 3
Sharing or publishing the contents in part or full is liable for legal action.
 Logistic Regression-
In the case of linear regression, the target variable ‘y’ is a continuous variable but let us assume
that the ‘y’ is a categorical variable which has two classes then linear regression should not be
used to predict the value of target variable. The output of the Linear Regression is not bound within
[0,1] as it can take any real value from (−∞ , ∞). Logistic regression is used to solve such problem
which gives us the corresponding probability outputs and then we can decide the appropriate cut-
off points to get the target class outputs.
Precisely Logistic Regression is defined as a statistical approach, for calculating the probability
outputs for the target labels. In its basic form it is used to classify binary data. Logistic regression
is very much similar to linear regression where the explanatory variables(X) are combined with
weights to predict a target variable of binary class(y). The main difference between linear
regression and logistic regression is of the type of the target variable. Logistic regression model
can be expressed as:
kumar.ashish2050@gmail.com 1 𝑒 (𝑤0 +𝑤1 𝑥)
21YORICED7 𝑦= =
1 + 𝑒 −(𝑤0 +𝑤1 𝑥) 𝑒 (𝑤0 +𝑤1 𝑥) + 1
Examples-

1. Churn prediction- Churn is the probability of a client to abandon a service or stop paying
to that particular service provider. The ratio of clients that abandon the service during a
particular time interval is called churn rate. Churn prediction is considered as a problem of
binary classification that whether a client or customer is going to churn. For example a
particular client churn the on the basis of monthly charge of service.
1
𝑃(𝑐ℎ𝑢𝑟𝑛 = 1|𝑚𝑜𝑛𝑡ℎ𝑙𝑦 𝑐ℎ𝑎𝑟𝑔𝑒) =
1 + 𝑒 −(𝑤0 +𝑤1 𝑚𝑜𝑛𝑡ℎ𝑙𝑦 𝑐ℎ𝑎𝑟𝑔𝑒)

1 𝑖𝑓 𝑃(𝑐ℎ𝑢𝑟𝑛 = 1|𝑚𝑜𝑛𝑡ℎ𝑙𝑦 𝑐ℎ𝑎𝑟𝑔𝑒) > 0.5

𝑐ℎ𝑢𝑟𝑛 = {
0 𝑖𝑓 𝑃(𝑐ℎ𝑢𝑟𝑛 = 1|𝑚𝑜𝑛𝑡ℎ𝑙𝑦 𝑐ℎ𝑎𝑟𝑔𝑒) ≤ 0.5

Here, we have selected the probability cut-off to 0.5 for predicting the target classes.

2. Spam Detection- Problem to identify that an email is spam or not.

3. Banking- Problem to predict a particular customer default a loan or not.

This file is meant for personal use by kumar.ashish2050@gmail.com only.

Proprietary content. © Great Learning. All Rights Reserved. Unauthorized use or distribution prohibited. 4
Sharing or publishing the contents in part or full is liable for legal action.
 Cost Function-
Linear regression uses mean squared error as its cost function but unfortunately this cannot
be used with logistic regression. Logistic regression uses Cross-Entropy or Log-Loss
function as its cost function defined for two class classification problem.
𝑚
1
𝑐𝑜𝑠𝑡(𝑤0 , 𝑤1 ) = − ∑{𝑦𝑖 𝑙𝑜𝑔(𝑎𝑖 ) + (1 − 𝑦𝑖 )𝑙𝑜𝑔(1 − 𝑎𝑖 )}
𝑚
𝑖=1

Where
𝑎𝑖 = 𝑠𝑖𝑔𝑚(𝑦ℎ𝑎𝑡𝑖 )

1
𝑠𝑖𝑔𝑚(𝑦ℎ𝑎𝑡𝑖 ) =
1 + 𝑒 −𝑦ℎ𝑎𝑡𝑖
The cost function can be divided into two separate functions as:

kumar.ashish2050@gmail.com 𝑚
21YORICED7 1
𝑐𝑜𝑠𝑡(𝑤0 , 𝑤1 ) = − ∑ 𝑙𝑜𝑔(𝑎𝑖 ) 𝒊𝒇 𝑦 = 1
𝑚
𝑖=1

and
𝑚
1
𝑐𝑜𝑠𝑡(𝑤0 , 𝑤1 ) = − ∑ 𝑙𝑜𝑔(1 − 𝑎𝑖 ) 𝒊𝒇 𝑦 = 0
𝑚
𝑖=1

This file is meant for personal use by kumar.ashish2050@gmail.com only.

Proprietary content. © Great Learning. All Rights Reserved. Unauthorized use or distribution prohibited. 5
Sharing or publishing the contents in part or full is liable for legal action.
 Optimization of coefficient or weight parameter- Again gradient descent is used to
optimize the value of weight parameters. The derivative of Loss function is defined as:
𝑚
1
𝐶𝑜𝑠𝑡 = ∑{−𝑦𝑖 log 𝑎𝑖 − (1 − 𝑦𝑖 )log(1 − 𝑎𝑖 ) }
𝑚
𝑖=1
where

1
𝑎𝑖 =
1 + 𝑒 −𝑦ℎ𝑎𝑡𝑖
and
𝑦ℎ𝑎𝑡𝑖 = 𝑤0 + 𝑤1 𝑥𝑖

Now let find the derivative of cost function (By chain rule of partial derivative):

𝜕𝐶𝑜𝑠𝑡 𝜕𝐶𝑜𝑠𝑡 𝜕𝑎𝑖 𝜕𝑦ℎ𝑎𝑡𝑖

= ∗ ∗
𝜕𝑤𝑖 𝜕𝑎𝑖 𝜕𝑦ℎ𝑎𝑡𝑖 𝜕𝑤𝑖

So,
𝜕𝐶𝑜𝑠𝑡 𝑎𝑖 − 𝑦𝑖
=
𝜕𝑎𝑖 𝑎𝑖 (1 − 𝑦𝑖 )
𝜕𝑎𝑖
= 𝑎𝑖 (1 − 𝑎𝑖 )
kumar.ashish2050@gmail.com 𝜕𝑦ℎ𝑎𝑡𝑖
21YORICED7 and
𝜕𝑦ℎ𝑎𝑡𝑖
= 𝑥𝑖
𝜕𝑤𝑖

By using all above the generalize formula expressed as:

𝑚
𝜕𝐶𝑜𝑠𝑡 1
= ∑(𝑎𝑖 − 𝑦𝑖 )𝑥𝑖 𝑤𝑖𝑡ℎ 𝑥0 = 1 𝑓𝑜𝑟 𝑤0
𝜕𝑤𝑖 𝑚
𝑖=1

Parameter update:

𝜕𝐶𝑜𝑠𝑡
𝒘𝒊 = 𝒘𝒊 − 𝒍𝒓𝒂𝒕𝒆
𝜕𝑤𝑖

Example- Consider an example where we are interested to find the effect of studying hours
per day over the result in examination and predict that a student will pass or fail for given
study hours. We have sample data about six students for their grades and total study hours per
day.

This file is meant for personal use by kumar.ashish2050@gmail.com only.

Proprietary content. © Great Learning. All Rights Reserved. Unauthorized use or distribution prohibited. 6
Sharing or publishing the contents in part or full is liable for legal action.
 Table-1 1

Study Hours Grades

per day

Fail/Pass
6 1(Pass)
5 0(Fail)
4 0(Fail)
7 1(Pass) 0
1 3 5 7 9
8 1(Pass) Grades
2 0(Fail) Figure-1: Scatter Graph

To solve the problem using logistic regression let us model the linear equation as:
𝑦(𝐺𝑟𝑎𝑑𝑒𝑠) = 𝑤0 + 𝑤1 𝑥(𝑆𝑡𝑢𝑑𝑦 𝐻𝑜𝑢𝑟𝑠 𝑝𝑒𝑟 𝑑𝑎𝑦)

and predict the result using:

kumar.ashish2050@gmail.com 1
21YORICED7 𝑃(𝑟𝑒𝑠𝑢𝑙𝑡 = 1|𝑠𝑡𝑢𝑑𝑦ℎ𝑜𝑢𝑟𝑠) =
1 + 𝑒 −(𝑤0 +𝑤1 𝑥(𝑆𝑡𝑢𝑑𝑦 𝐻𝑜𝑢𝑟𝑠 𝑝𝑒𝑟 𝑑𝑎𝑦))

𝑦ℎ𝑎𝑡𝑖 = 𝑤0 + 𝑤1 𝑥𝑖
and
1
𝑎𝑖 = 𝑠𝑖𝑔𝑚(𝑦ℎ𝑎𝑡𝑖 ) =
1+ 𝑒 −𝑦ℎ𝑎𝑡𝑖

𝑦ℎ𝑎𝑡𝑖 = 𝑤0 + 𝑤1 𝑥𝑖

Cost Function:
𝑚
1
𝐶𝑜𝑠𝑡(𝑤0 , 𝑤1 ) = ∑{−𝑦𝑖 log 𝑎𝑖 − (1 − 𝑦𝑖 )log(1 − 𝑎𝑖 ) }
𝑚
𝑖=1

Gradients:
𝑚
𝜕𝐶𝑜𝑠𝑡(𝑤0 , 𝑤1 ) 1
= ∑(𝑎𝑖 − 𝑦𝑖 )
𝜕𝑤0 𝑚
𝑖=1

This file is meant for personal use by kumar.ashish2050@gmail.com only.

Proprietary content. © Great Learning. All Rights Reserved. Unauthorized use or distribution prohibited. 7
Sharing or publishing the contents in part or full is liable for legal action.
 and
𝑚
𝜕𝐶𝑜𝑠𝑡(𝑤0 , 𝑤1 ) 1
= ∑(𝑎𝑖 − 𝑦𝑖 )𝑥𝑖
𝜕𝑤1 𝑚
𝑖=1

Parameter updates:

𝜕𝐶𝑜𝑠𝑡(𝑤0 , 𝑤1 )
𝑤0 = 𝑤0 − 𝑙𝑟𝑎𝑡𝑒
𝜕𝑤0

and

𝜕𝐶𝑜𝑠𝑡(𝑤0 , 𝑤1 )
𝑤1 = 𝑤1 − 𝑙𝑟𝑎𝑡𝑒
𝜕𝑤1

We have,
X: 6 5 4 7 8 2
y: 1 0 0 1 1 0

Iteration #1:
kumar.ashish2050@gmail.comLet 𝑤0 = 1 𝑎𝑛𝑑 𝑤1 = 1, 𝑤𝑖𝑡ℎ 𝑙𝑟𝑎𝑡𝑒 = 0.01
21YORICED7
1
𝑦ℎ𝑎𝑡𝑖 = 𝑤0 + 𝑤1 𝑥𝑖 𝑎𝑛𝑑 𝑎𝑖 = 𝑠𝑖𝑔𝑚(𝑦ℎ𝑎𝑡𝑖 ) =
1+ 𝑒 −𝑦ℎ𝑎𝑡𝑖

yhat: 7 6 5 8 9 3
𝒂: 0.999 0.997 0.993 0.999 0.999 0.995
So,
(0.999 − 1) + (0.997 − 0) + (0.993 − 0) +
𝜕𝐶𝑜𝑠𝑡(𝑤0 , 𝑤1 ) (0.999 − 1) + (0.999 − 1) + (0.995 − 0)
=
𝜕𝑤0 6

𝜕𝐶𝑜𝑠𝑡(𝑤0 , 𝑤1 )
= 0.497
𝜕𝑤0
and
(0.999 − 1) ∗ 6 + (0.997 − 0) ∗ 5 + (0.993 − 0) ∗ 4 +
𝜕𝐶𝑜𝑠𝑡(𝑤0 , 𝑤1 ) (0.999 − 1) ∗ 7 + (0.999 − 1) ∗ 8 + (0.995 − 0) ∗ 2
=
𝜕𝑤1 6
𝜕𝐶𝑜𝑠𝑡(𝑤0 , 𝑤1 )
= 1.821
𝜕𝑤1

This file is meant for personal use by kumar.ashish2050@gmail.com only.

Proprietary content. © Great Learning. All Rights Reserved. Unauthorized use or distribution prohibited. 8
Sharing or publishing the contents in part or full is liable for legal action.
 Parameter update:
𝜕𝐶𝑜𝑠𝑡(𝑤0 , 𝑤1 )
𝑤0 = 𝑤0 − 𝑙𝑟𝑎𝑡𝑒 ∗ = 1 − 0.01 ∗ (0.497) = 0.995
𝜕𝑤0
𝜕𝐶𝑜𝑠𝑡(𝑤0 , 𝑤1 )
𝑤1 = 𝑤1 − 𝑙𝑟𝑎𝑡𝑒 ∗ = 1 − 0.01 ∗ (1.821) = 0.982
𝜕𝑤1
Iteration #1:
Let 𝑤0 = 0.995 𝑎𝑛𝑑 𝑤1 = 0.982, 𝑤𝑖𝑡ℎ 𝑙𝑟𝑎𝑡𝑒 = 0.01
yhat: 6.887 5.905 4.923 7.869 8.851 2.959
𝒂: 0.999 0.997 0.993 0.999 0.999 0.950

(0.999 − 1) + (0.997 − 0) + (0.993 − 0) +

𝜕𝐶𝑜𝑠𝑡(𝑤0 , 𝑤1 ) (0.999 − 1) + (0.999 − 1) + (0.950 − 0)
= = 0.489
𝜕𝑤0 6
𝑎𝑛𝑑
(0.999 − 1) ∗ 6 + (0.997 − 0) ∗ 5 + (0.993 − 0) ∗ 4 +
𝜕𝐶𝑜𝑠𝑡(𝑤0 , 𝑤1 ) (0.999 − 1) ∗ 7 + (0.999 − 1) ∗ 8 + (0.950 − 0) ∗ 2
= = 1.806
𝜕𝑤1 6
kumar.ashish2050@gmail.com
21YORICED7 Parameter update:
𝜕𝐶𝑜𝑠𝑡(𝑤0 , 𝑤1 )
𝑤0 = 𝑤0 − 𝑙𝑟𝑎𝑡𝑒 ∗ = 0.995 − 0.01 ∗ (0.489) = 0.990
𝜕𝑤0
𝜕𝐶𝑜𝑠𝑡(𝑤0 , 𝑤1 )
𝑤1 = 𝑤1 − 𝑙𝑟𝑎𝑡𝑒 ∗ = 0.982 − 0.01 ∗ (1.806) = 0.964
𝜕𝑤1
and so on……….

This file is meant for personal use by kumar.ashish2050@gmail.com only.

Proprietary content. © Great Learning. All Rights Reserved. Unauthorized use or distribution prohibited. 9
Sharing or publishing the contents in part or full is liable for legal action.
 Evaluation of Logistic regression model- Performance measurement of classification
algorithms are judge by confusion matrix which comprise the classification count values
of actual and predicted labels. The confusion matrix for binary classification is given by:

Figure 3 Confusion Matrix

Confusion matrix cells are populated by the terms:
True Positive(TP)- The values which are predicted as True and are actually True.
True Negative(TN)- The values which are predicted as False and are actually False.
False Positive(FP)- The
kumar.ashish2050@gmail.com values which are predicted as True but are actually False.
21YORICED7
False Negative(FN)- The values which are predicted as False but are actually True.
Classification performance metrics are based on confusion matrix values. The most
popularly used metrics are;
Precision- measure of correctness achieved in prediction.
𝑇𝑃
𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =
𝑇𝑃 + 𝐹𝑃
Recall (sensitivity)- measure of completeness, actual observations which are predicted
correctly.
𝑇𝑃
𝑟𝑒𝑐𝑎𝑙𝑙 =
𝑇𝑃 + 𝐹𝑁
Specificity- measure of how many observations of false category predicted correctly.
𝑇𝑁
𝑠𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 =
𝑇𝑁 + 𝐹𝑃
F1-Score- a way to combine precision and recall metric in a single term. F1score is defined
as harmonic mean of precision and recall.

This file is meant for personal use by kumar.ashish2050@gmail.com only.

Proprietary content. © Great Learning. All Rights Reserved. Unauthorized use or distribution prohibited. 10
Sharing or publishing the contents in part or full is liable for legal action.
 2 ∗ 𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 ∗ 𝑟𝑒𝑐𝑎𝑙𝑙
𝐹1𝑠𝑐𝑜𝑟𝑒 =
𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 + 𝑟𝑒𝑐𝑎𝑙𝑙
ROC Curve- Receiver Operating Characteristic(ROC) measures the performance of
models by evaluating the trade-offs between sensitivity (true positive rate) and false (1-
specificity) or false positive rate.
AUC - The area under curve (AUC) is another measure for classification models is based
on ROC. It is the measure of accuracy judged by the area under the curve for ROC.

kumar.ashish2050@gmail.com
21YORICED7

Figure 4 ROC Curve

Pros and cons of Linear Regression:

Pros- Logistic regression classification model is simple and easily scalable for multiple
classes.
Cons- Classifier constructs linear boundaries and the interpretation of coefficients value is
difficult.
********

This file is meant for personal use by kumar.ashish2050@gmail.com only.

Proprietary content. © Great Learning. All Rights Reserved. Unauthorized use or distribution prohibited. 11
Sharing or publishing the contents in part or full is liable for legal action.