Please do not remove this page

Telecom customer churn prediction

Sundararajan, Abhijit; Gursoy, Kemal
https://scholarship.libraries.rutgers.edu/esploro/outputs/journalArticle/Telecom-customer-churn-prediction/991031550004604646/filesAndLinks?inde
x=0

Sundararajan, A., & Gursoy, K. (2020). Telecom customer churn prediction.

https://doi.org/10.7282/t3-76xm-de75
Document Version: Version of Record (VoR)

This work is protected by copyright. You are free to use this resource, with proper attribution, for
research and educational purposes. Other uses, such as reproduction or publication, may require the
permission of the copyright holder.
Downloaded On 2024/12/25 23:29:02 -0500
 Telecom Customer Churn Prediction
Abhijit Sundararajan
Department of MSIS, Rutgers University
E-mail: abhijit.sundararajan@rutgers.edu

Kemal Gursoy
Department of MSIS, Rutgers University,
E-mail: kgursoy@business.rutgers.edu

Abstract 1. Introduction
Customer churn is often referred to as customer Customer churn is a common problem across
attrition, or customer defection which is the rate at several business in different industries. There is a
which the customers are lost. Telecom companies substantial amount of financial loss associated with
often use customer churn as a key business metrics churning, as the businesses must invest huge sum
to predict the number of customers that will leave of money in getting new customers. Time and
a telecom service provider. Churn is significant in efforts need to be invested into replacing the
the telecommunication industry because it directly customers that are lost. Businesses need to be able
affects the competitiveness of the service provider. to predict when a client is likely to leave and offer
Churn is the proportion of clients leaving the them lucrative incentives to stay. The most
service provider. The service provider should important essence of using predictive analysis is to
therefore find new clients to preserve profitability. determine the customers that are likely to churn.
This is not possible due to the high and difficult cost
of acquiring new customers. The cost of retaining Companies typically make a difference between
an existing customer is far less than acquiring a new voluntary churn and involuntary churn. Voluntary
one, so telephone service companies, Internet churn occurs because of a selection by customer to
service providers, pay TV companies, insurance replace to any other organisation or provider,
firms and alarm monitoring services use the involuntary churn happens because of instances
customer churn to predict profitability. Businesses inclusive of a client's relocation for an extended-
time period in a care facility, loss of life, or the
basically believe that a customer has churned with
relocation to a distant region. In maximum
the amount that has passed since the customer’s
packages, involuntary reasons for churn are
last interaction with the site or service. The full cost
excluded for analytical purposes. Analysts tend to
of customer churn includes both lost revenue and
concentrate on voluntary churn, as it normally
the marketing costs involved with replacing those
occurs because of factors of the corporation-
customers with new ones. Reducing customer customer courting which agencies manage, which
churn is important because cost of acquiring a new includes how billing interactions are treated or how
customer is higher than retaining an existing one. after-income assistance is provided. Predictive
This case is related to telecom industry where analytics use churn prediction fashions that expect
organizations want to know that for given certain purchaser churn through assessing their propensity
parameters whether a person will churn or not. of chance to churn. Seeing that these fashions
generate a small prioritized list of ability defectors,
they're effective at focusing consumer retention
advertising packages at the subset of the consumer
base who're most liable to churn.
There are 2 types of churners: Voluntary churner
and Involuntary Churner.
Involuntary Churners: This type includes the
customers that are removed by the service

1
providers from the subscription list. The customers Figure 1.
that come under this category are the customers
who cheat or are churned for fraudulent activities,
the customers who do not pay their subscription
charges and the customers who do not use the
services.

Voluntary Churners: This type includes the

customers that voluntarily decide to leave and
terminate their services. Voluntary churn can be
subdivided into 2 categories: Incidental and
Deliberate churn. Incidental churn occurs due to
location change, financial issues and can happen
even due to bad experiences. Deliberate churn
occurs when the customer wants to upgrade to Dataset is from the IBM Watson Repository and
new service providers with better technologies. It available in Kaggle also. The data set consists of
also happens due to price sensitivity, quality of the 7043 rows & 21 columns. It has 17 categorical
services provided, social and psychological factors variables, 3 numerical variables and 1 unique
and convenience factors.
variable i.e. customerID. The row consists of
Customers, columns consists of Customer
Through means of a Kaggle competition, the
Attributes and Target Variable consists Churn
objective is to find an appropriate model to predict
The questions we would like to answer through
whether a customer may churn from a particular
Machine Learning algorithms include:
service provider in the near future. Having this
What are the demographics of people who choose
model in place can ensure that Santander can take
to stay/leave? What kind of services do the
proactive steps to improve a customer’s happiness
existing customers subscribe to? Does loyalty and
before they would take their business elsewhere.
money billed influence the churn? Growing
First the paper will discuss related work done on
companies and those expanding their product
this. Secondly we do an extensive exploratory
range usually segment their customers using
analysis of the data, analysing groups of variables
previously defined and selected features.
and individual features to give us insight in what is
Customers can be divided into subgroups based
relevant. Thirdly several cleaning procedures that
on their lifecycle stage, needs, used solutions,
were employed to lead to better results are
level of engagement, monetary value, or basic
outlined. Fourthly we explain the performance
information. Since every customer category
measure of this competition and the three models
shares common behaviour patterns, it’s possible
to tackle the problem.
to increase prediction accuracy through the use of
ML models trained specifically on datasets
2. Today’s Approach representing each segment.
The data set includes information about:
Customers that churned during the last month – 3. Models
the column is called Churn. 1. Logistic regression
Services that each customer has been using– This is used for classification problem. We can use
phone, multiple lines, internet, online security, logistic regression for this problem. Logistic
online backup, device protection, tech support, Regression is used when the dependent variable is
and streaming TV and movies. categorical. It has a bias towards classes which
Customer account information – how long they’ve have large number of instances. It tends to only
been a customer, contract, payment method, predict the majority class data. The features of the
paperless billing, monthly charges, and total minority class are treated as noise and are often
charges. ignored. Thus, there is a high probability of
Demographic information about the customers – misclassification of the minority class as compared
gender, age range, and if they have partners and to the majority class. Suppose in a dataset there are
dependents
2
 2 categories, then logistic regression models the Paperless billing, then this customer is more likely
probability that Y belongs to a particular category. to churn.

2. Random Forest
Random is used for classification problems. It is a
4. Experiments
combination of tree predictors such that each tree
depends on the value of a vector randomly Figure 2.
sampled and sampled independently and with the
distribution of all the trees in the forest. After
many trees are generated, they vote for the most
popular class. This procedure is called random
forests. There have been significant improvements
in the classification accuracy from growing an
ensemble of trees and letting them vote for the
most popular class.
3. Support Vector Machine
Support Vector Machine is a supervised learning
algorithm which is used for both classification and
regression challenges. it is mostly used in
classification problems. In this algorithm, we plot
each data item as a point in n-dimensional space. We can infer from the above figure that 27%
where n is number of features you have) with the of the customers in the dataset from Kaggle
value of each feature being the value of a have churned.
coordinate. Then, we perform classification by
finding the hyper-plane that differentiate the two
classes
4. Decision Tree
5. Comparison of Methods
Decision trees can be applied to both classification
and regression problems. It is used to predict a Models Accuracy
qualitative response rather than a quantitative Score
response. We predict that each of the observation Random 0.9355
belongs to the most commonly occurring class. It is Forest
a type of supervised learning algorithm with a Classifier
predefined target variable. While mostly used in SVM 0.8192
classification tasks, it can handle numeric data as
well. This algorithm splits a data sample into two or Decision 0.762
Tree
more homogeneous sets based on the most
significant differentiator in input variables to make Logistic 0.7894
a prediction. With each split, a part of a tree is Regression
being generated. As a result, a tree with decision
nodes and leaf nodes (which are decisions or
classifications) is developed. A tree starts from a 6. Results and Discussion
root node – the best predictor. Out of three We observe that the Support Vector Machine
variables we use, Contract is the most important model is a good fit. This is due to the AUC values for
variable to predict customer churn or not churn. If a training and test are high and similar. We may
customer in a one-year or two-year contract, no consider improving the SVM model by performing
matter he (she) has Paperless billing or not, he (she) a grid search for values of C, gamma and degree of
is less likely to churn. On the other hand, if a the kernel that gives higher accuracy using k-fold
customer is in a month-to-month contract, and in cross-validation.
the tenure group of 0–12 month, and using

3
 Conclusion
From the above example, we can see that SVM,
Logistic Regression and Random Forest performed
better than Decision Tree for customer churn
analysis for this dataset. The following things can be
observed from the dataset:
1. Attributes and features such as tenure group,
Contract, Paperless Billing, Monthly Charges and
Internet Service appear to play a role in customer
churn.
2. There seems to be no relationship between the
gender and the churn rate.
3. Customers having a service plan of month-to-
month contract, with Paperless Billing and are
within 12 months tenure, are more likely to churn.
On the other hand, customers with one- or two-
year contract, with longer than 12 months tenure,
that are not using Paperless Billing, are less likely to
churn.

References
1. https://www.datasciencecentral.com/profi
les/blogs/customer-churn-logistic-
regression-with-r
2. https://pdfs.semanticscholar.org/75d3/73f
987be5c2fb5a3cb1830f417c63f09a68d.pdf
3. https://datascienceplus.com/predict-
customer-churn-logistic-regression-
decision-tree-and-random-forest/
4. https://www.kaggle.com/pavanraj159/tele
com-customer-churn-prediction
5. https://towardsdatascience.com/hands-
on-predict-customer-churn-5c2a42806266