Course Title:
Applications of Artificial Intelligence
Assignment No. 2
Course Number
EAI6010
Term and Year: Winter B:
Start and End Dates:
March 1 - April 10
Name: Nikhil Sanjay Thorat
Steps involved in building the Customer Churn Prediction Model:
1. IMPORTING THE DATASET AND NECESSARY LIBRARIES:
2. DATA CLEANING:
REMOVED THE NULL VALUES OF THE DATASET
FIXED THE DUPLICATE VALUES
SEPERATED CATEGORICAL AND CONTINUOUS VALUES
FIXED THE DATA TYPES OF THE VARIABLES IN THE DATASET
THERE WERE NO OUTLIERS PRESENT IN THE DATASET. GENERALLY IF OUTLIERS ARE
PRESENT WE CAN REPLACE IT WITH CAPPING METHODS
3. PERFORMING EDA ON THE DATASET:
Count plot for payment methods.
Count plot of various payment methods in the dataset. We can see that most of the customers prefer
using electronic checks for payments.
Density plot for tenure.
From the above density plot we can see that the tenure mostly ranges from 10 to 70 values.
Important KPI’s for dashboarding:
Average Tenure:
Average Monthly charges
Average Total Charges
Count Plot for target variable
The number of people who stayed are more as compared to number of people who churned.
Density Plot for Total Charges
The density of most of the total charges range from 0 to 4500.
Count plot for all the features with respect to churn
The above mentioned diagram shows the count of various features with respect to the target variables.
4. SPLITTING THE DATASET INTO TRAIN TEST SPLIT USING SKLEARN PACKAGES
The data was split into 80% training and 20% of testing dataset and a various models were fitted on
the same dataset.
5. APPLIED VARIOUS MACHINE LEARNING MODELS WITHOUT SCALING THE FEATURE VARIBALES
The highest accuracy obtained was for Logistic regression which was 80%.
6. PERFORMED FEATURE ENGINEERING TO INCREASE THE ACCURACY OF THE MODEL
Using get_dummies function dummy variables were created for the dataset for all the categorical
columns and that data was later passed into the model.
7. PERFORMED MIN MAX SCALING ON THE DATASET
Using min max scaling we can normalize the features variables of the dataset. In this dataset we had
three continuous variables which are tenure, monthly charges and total charges which were normalized.
8. APPLYING MACHINE LEARNING ALGORITHMS AFTER SCALING THE DATASET
The results obtained were more accurate as compared to the results which were obtained
without scaling the variables.
9. USING HYPERPARAMETERS TO TUNE THE MODELS TO OBTAIN MORE ACCURATE RESULTS
Hyperparameter Tuning and grid-search
Almost all algorithms have hyperparameters that can be tuned to fine-tune their performance,
reduce over-fitting and better capture the patterns in the dataset. Having a good understanding and
intuition of how algorithms work is essential to fully utilize hyperparameter tuning for the purposes
of improving model performance and testing different modeling strategies.
Hyperparameters are significant because they control the conduct of the training algorithm and
have a significant impact on the performance model is being prepared. Productively search the
space of conceivable hyperparameters. Simple to deal with an enormous arrangement of analyses
for hyperparameter tuning.
Grid-search is a hyperparameter tuning algorithm that sequentially goes through every possible
combination of hyperparameter combination it is fed in space. For example, for
hyperparameters parameter 1 and parameter 2 - it would mean testing out all possible
combinations of their values:
Grid-search can be done using the GridSearchCV() function - it takes in as arguments:
The model being used.
The possible parameters to test - inputted as a dictionary.
cv: The number of cross-validation folds.
verbose: More detailed output if 2
RESULTS:
The results obtained after tuning the model were not having a significant raise in the accuracy
and almost all the models showed an accuracy of around 80%.
10. APPLYING ENSEMBLE LEARNING TO THE DATASET
Ensemble modelling is an interaction where various different base models are utilized to
foresee a result. The reason for utilizing ensemble models is to lessen the speculation error of
the prediction. The base models are different and free, the prediction error diminishes when the
ensemble approach is utilized.
The results obtained for ensemble models were:
Decision Tree Classifier : 80%
Adaboost Classifier: 80%
Logistic Regression Classifier: 80%
11. CONCLUSION AND RESULTS:
After trying various methods on the dataset to increase accuracy the maximum accuracy
obtained was around 81% which was using Logistic regression. Whereas other models also gave
a similar accuracy around 80%. Therefore we can conclude that we can use above mentioned
models for classification of the customer chur prediction.
12. TABLEAU DASHBOARD RESULTS:
TABLEAU LINK: Telco Customer Churn Dashboard - Nikhil Sanjay Thorat | Tableau Public
13. WEBAPP FOR USING THE MODEL:
The model was linked to a front end page which was designed using HTML and the webapp was
deployed using Heroku so that the end user can test the model by giving various feature inputs. As
the current model has a lot of variables which cannot be inserted in a webapp, therefore I
performed feature selection on the dataset and took a few top variables and designed a webapp
which predicted customer churn.
Webapp Link: Home Page (churn-customer.herokuapp.com)
