Logistic Regression Hands-On

Agenda
01 Overview of business problem

05 Chi-Square Test
02 Exploratory Data Analysis(EDA) 5.1 Information Value Analysis
03 Missing Data Analysis 5.2 Multicolinearity Analysis
04 Outlier Analysis 06 Train and Test Split
Copyright Intellipaat. All rights reserved.

Agenda
07 Model Fitting 8.3 Specificity
08 Model Validation 8.4 ROC Curve
8.1 Accuracy 9 Finding Optimal Cut-off
8.2 Sensitivity

Overview of business
problem

1. Overview of Business Problem
Consider a bank who wants to launch its insurance product to

target the customers with high probability. They are unaware
of which customer would opt for this product.
In this case the bank will retrieve the data of the customer
from their previous interactions and leverage machine learning
algorithms like Logistic regression to predict which customer
will be interested in the insurance and which will not.

2. What is Exploratory Data Analysis ?
It is the process of performing a prior investigation on the data where it is

collected, analysed and presented in an understandable way.

EDA - Code Walk
Step 1: Import dataset
Step 2: Displaying first 20 rows of dataset with

head() function

EDA - Code Walk
describe() helps us to check for mean,

Step 3: Checking all the statistical data. std, quartiles, minimum and maximum
values in our dataset.

EDA In-depth
Visualization is the best way to present the data

in a most understandable way.
Visualization can be done with different types of

plots like bar, line, box, scatter plot and so on.

EDA - Code Walk

3. What is Missing Data Analysis?
It is the process where the missing value pattern is described and

replaced with values using certain regression.
Question:
Now, what do you think we need
to do with these missing values?

Missing Data Analysis - Code Traverse

Missing Data Imputation
Missing values for quantitative or numeric data is replaced

by mean or median of the column.
Missing values for qualitative data is replaced by mode of

the data.

Cheatsheet for Missing Data Imputation

Missing Data Analysis - Code Traverse
Step 6: Replacing missing data.
If it is a categorical or object type

data replace it with mode.
Else if it is a numerical data replace

it with mean.

Mean Vs Median - Quick Look
Let us consider an example, consider below given ages
where we have to predict the missing age with mean and
median.
Mean 31 years
SI Num Age ( years )
1 18
2 20
3 15 Median 18
4 ?
5 70
If we see the above example closely we can see that

median gives us a more realistic value. Due to this nature
of median we prefer median over mean.

4. What is Outlier Analysis?
The process of discovering the abnormal observation in a

dataset is referred to as outlier analysis.
Question:
Now how do you think we can
handle the outliers?

Outlier Analysis - Description
Well, we use a technique called capping.
Values lower than the first percentile value are

substituted with the first percentile value. This is called as
lower capping.
Values larger than the ninety ninth percentile value

are substituted with the ninety ninth percentile value.
This is called upper capping.

Outlier Analysis - Code View
Step 7: Checking for outlier

values.
Here we can see the outlier

values for ‘duration’ column.
Here we can see the outlier

values for ‘age’ column.

Outlier Analysis - Code View
The graph shows the

outliers for age and
duration.

5.1. What is Chi Square Test?
A chi-square test is used to see if observed findings match predicted

outcomes and to eliminate the possibility that observations are
random.
It is used to demonstrate if two category variables have

a link with each other.

Chi Square Test - Study
A p-value less than or equal to your significance

threshold 0.05 in a Chi-square test shows that there is
enough evidence to determine that the observed
distribution is not the same as the expected
distribution.
This means that the independent categorical

variable is having a strong correlation with
dependent categorical variable.
You can deduce that the

categorical variables have a
relationship.

Chi Square Test - Computing
Step 8: Calculating the p-value in order to decide whether null values

needs to be eliminated or not.

Chi Square Test - Computing

5.2. What is Information Value?
When it comes to picking predictor variables for binary

classification problem, the Information Value is a prominent
strategy.
Variable transformation can also be performed with this

technique!

Information Value Cutoffs
Information Value Predictive Power
<0.02 Useless
0.02 – 0.1 Weak predictors
0.1 - 0.3 Medium Predictors
0.3 - 0.5 Strong predictors
>0.5 Suspicious
If IV <= 0.1 we can drop the

variable.

Information Value - Code Snippet



6. What is Multicollinearity Analysis ?
This is the process of occurrence of multiple intercorrelation

among one or more independent variables
Question
What do you think the diagram is trying to tell us?

Multicollinearity Analysis - Sampling
So, in the diagram we can see that Y has a correlation with

X1, X2, X3 and X4.
We can also see that X1 and X2 also have a high correlation.
Think of overlap between these bubbles as correlation in

variables.
This process helps to avoid overlap between these

independent variables.
Multicollinearity Analysis - Instance
Let us consider an example where we show the correlation of ‘Y’

with ‘X1’, ‘X2’ and ‘X3’.
Y X1
Now, if we can see here Y has correlation with X1, X2 and X3.
Y X2
Y X3 Higher correlation between Y and X1 , X2 = Good Model.
Question:
But, what if X1 ~ X2?

Why Collinearity Analysis
In that case we will have Now, how do we check for

duplicate data. that duplicate data?
X1 X2 X3 X4
X1 1 0.2 0.8 0.7
In order to see that let us X2 0.3 1 0.3 0.1

consider an example. X3 0.2 0.8 1 0.2
X4 0.5 0.1 0.9 1
If we look into the table X1 has the

Now the question is how to
highest correlation with X3 and X4
know which to eliminate
which means values of X3 and X4
between X3 and X4.
are similar to X1.

Why Collinearity Analysis - Continuation
Simply putting we first eliminate the highest correlated

variable and then eliminate the next highest correlated
variable.
The variables with high correlation can be removed by

using chi square test.

Multicollinearity Analysis - Pop Quiz
The most important question is why do we need multicollinearity if we have

collinearity?

Why Multicollinearity Analysis - Instance
Consider the below example,
0.1 If it rains, a person will go shopping
If there is an earthquake(mild) a
person may still go shopping.
0.2
If there is rain and an earthquake at the

same time then a person wouldn’t go
shopping due to fear as this not
expected and new to the locals.
0.9 Hence, the probability will be low.

Multicollinearity Analysis - Continuation
From the above example we can say that hidden correlations

can occur.
As collinearity analysis cannot find the hidden correlations, we

use multicollinearity analysis.

Evaluating Multicollinearity Analysis
X1 X2 X3 X4 (0.9)
~
(R2)
X2 X1 X2 X4 (0.1)
~
(R2)
X3 X1 X2 X4 (0.8)
~
(R2)
X4 X1 X2 X3 (0.2)
~
(R2)
Higher the R value higher is the correlation.
Firstly, eliminate X1 value and run the analysis again

We prefer VIF over R2.
This is because VIF helps us to differentiate

between R2 values better.
Formula:
1
1-r2

Multicollinearity Analysis Evaluation
We only drop the highest VIF It is good practice to remove

value not all VIF values. values one by one.

X2 X3 X4 (0.1)
~
R2
X3 X2 X4 (0.1)
~
R2
X4 X2 X3 (0.1)
~
R2
We either use R2 value or VIF value to detect

multicollinearity.
VIF is similar to R2 but with this we can differentiate

between correlations in a better way.
If VIF>0.5 we eliminate the variable.

Multicollinearity Analysis - Code Walk
Step 9.1: Eliminating columns

with high VIF value.
The previously highlighted value

is eliminated but likewise we
have many other values with high
VIF values. Hence we need to
eliminate them all.

Multicollinearity Analysis - Conclusion
As we have seen above by evaluating the hidden correlation of X1 with other variables
helped eliminate correlation of X3 with other variables.
This simply means that sometimes eliminating hidden correlations can result in a stable
correlations among other variables.

Multicollinearity Analysis - Result
All high VIF values The final result has

were eliminated with all VIF lower than
multiple iterations. cutoff of 0.5.

7. What is Train and Test of Data ?

Train and Test of Data - Computing
Step 10: Splitting the data into train and test set.

8. What is Model Training ?
The process of training the model on the required data is called

model training.
We also refer to it as model fitting.

Model Training - Code Snippet
Step 11: Training our model.

9. What is Model Validation ?
The process of testing the model with the test data after
training the model is called model validation.
This gives us the insights on the model

performance.
This gives us the insites on the

model performance.
This is also called model

testing.

What is Model Validation?
Accuracy
Sensitivity
Shown are the steps for model
validation.
Specificity
ROC Curve

Model Validation
Step 12: Scoring our model.

9.1 What is Accuracy?
The process of checking how accurate the model is

during model validation is called as accuracy.
High precision and is required to get high

Formula: accuracy of the model.
TP+TN
TP+TN+FP+FN

Calculating Accuracy
Step 13: Finding the accuracy of the model.

9.2 What is Sensitivity?
The process of determining how the output varies in
response to changes in the input is known as sensitivity.
It could provide insight into the predictability of a model.
It is used to check the model’s ability to predict true

positive data.
Formula:
True Positive
True Positive +False Negative

Calculating Sensitivity
Step 14: Calculating the

sensitivity.

9.3 What is Specificity?

Specificity - Formula
Formula:
True Negative
True Negative + False Positive

Calculating Specificity
Step 15: Calculating specificity..

9.4 What is ROC Curve?

What is ROC Curve?
ROC Curve shows us the relationship between true positive and true
negative.

ROC Curve
Step 16: Plotting an ROC Curve.

10. What is Cut-off?
Cut-off refers to the point where the decision is made.
Eg., Consider a cut-off, if probability of a is more than cut-off

than b then the machine chooses a as a final decision.

Cut-off
For finding optimum cut-off value we will use predict_proba().
This helps us to get an array of list containing class of

probabilities for input points.

Cut-off - Code
Finding optimum cutoff.

Cut-off - Code
Combining the predictions by the model and the

probability of predicting.

Cut-off - Code
Checking for null values with

isna().
y_predict_proba contains the

probability value of getting
either 0 or 1.
Cut-off
If y_predict_proba > 0.1 then

probability = 1.
Else 0
Firstly, we will categories the
data.
Same goes for other cutoffs.

Cut-off - Code
Repeating the process for

all cutoffs.
y_predict_proba > 0.1=1
y_predict_proba <= 0.1=0

Cut-off
From the above cutoffs we will be calculating sensitivity

and specificity of every cutoff.
Which ever has the highest sensitivity + specificity

value we will be considering that as an optimal
cutoff value.

Cut-off - Code
Appending values into

dataframe.
Calculating sensitivity and

specificity.

Cut-off
From above table we have analyzed that highest total value is

2 and for that cutoff value is 0.5.
Therefore, the optimal cutoff value = 0.5.

Cut-off - Code
Choosing the cutoff with
highest Total_val.
Here, 2 is the highest

Total_val. Hence, 0.5 is chosen
as the cutoff.

India: +91-7022374614
US: 1-800-216-8930 (TOLL FREE)
sales@intellipaat.com
support@intellipaat.co
m
24/7 Chat with Our Course

Advisor

Logistic Regression Hands-On

Uploaded by

Document Informationclick to expand document information

Document Informationclick to expand document information

Copyright:

Available Formats

Logistic Regression Hands-On

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Logistic Regression Hands-On

Uploaded by

Copyright:

Available Formats

Agenda

01 Overview of business problem

02 Exploratory Data Analysis(EDA) 5.1 Information Value Analysis

03 Missing Data Analysis 5.2 Multicolinearity Analysis

04 Outlier Analysis 06 Train and Test Split

Copyright Intellipaat. All rights reserved.

08 Model Validation 8.4 ROC Curve

8.1 Accuracy 9 Finding Optimal Cut-off

Copyright Intellipaat. All rights reserved.

Copyright Intellipaat. All rights reserved.

Consider a bank who wants to launch its insurance product to

Copyright Intellipaat. All rights reserved.

It is the process of performing a prior investigation on the data where it is

Copyright Intellipaat. All rights reserved.

Step 2: Displaying first 20 rows of dataset with

Copyright Intellipaat. All rights reserved.

describe() helps us to check for mean,

Copyright Intellipaat. All rights reserved.

Visualization is the best way to present the data

Visualization can be done with different types of

Copyright Intellipaat. All rights reserved.

Copyright Intellipaat. All rights reserved.

It is the process where the missing value pattern is described and

Copyright Intellipaat. All rights reserved.

Copyright Intellipaat. All rights reserved.

Missing values for quantitative or numeric data is replaced

Missing values for qualitative data is replaced by mode of

Copyright Intellipaat. All rights reserved.

Copyright Intellipaat. All rights reserved.

If it is a categorical or object type

Else if it is a numerical data replace

Copyright Intellipaat. All rights reserved.

SI Num Age ( years )

If we see the above example closely we can see that

Copyright Intellipaat. All rights reserved.

The process of discovering the abnormal observation in a

Copyright Intellipaat. All rights reserved.

Well, we use a technique called capping.

Values lower than the first percentile value are

Values larger than the ninety ninth percentile value

Copyright Intellipaat. All rights reserved.

Step 7: Checking for outlier

Here we can see the outlier

Here we can see the outlier

Copyright Intellipaat. All rights reserved.

The graph shows the

Copyright Intellipaat. All rights reserved.

A chi-square test is used to see if observed findings match predicted

It is used to demonstrate if two category variables have

Copyright Intellipaat. All rights reserved.

A p-value less than or equal to your significance

This means that the independent categorical

You can deduce that the

Copyright Intellipaat. All rights reserved.

Step 8: Calculating the p-value in order to decide whether null values

Copyright Intellipaat. All rights reserved.

Copyright Intellipaat. All rights reserved.

When it comes to picking predictor variables for binary