California State University, San Bernardino

CSUSB ScholarWorks

Electronic Theses, Projects, and Dissertations Office of Graduate Studies

5-2023

OTT SUBSCRIBER CHURN PREDICTION USING MACHINE

Follow this and additional works at: https://scholarworks.lib.csusb.edu/etd

Part of the Other Computer Engineering Commons

Recommended Citation
Senthil Kumar, Needhi Devan, "OTT SUBSCRIBER CHURN PREDICTION USING MACHINE LEARNING"
(2023). Electronic Theses, Projects, and Dissertations. 1660.
https://scholarworks.lib.csusb.edu/etd/1660

This Project is brought to you for free and open access by the Office of Graduate Studies at CSUSB ScholarWorks.
It has been accepted for inclusion in Electronic Theses, Projects, and Dissertations by an authorized administrator
of CSUSB ScholarWorks. For more information, please contact [email protected].
OTT SUBSCRIBER CHURN PREDICTION

USING MACHINE LEARNING

A Project

Presented to the

Faculty of

California State University,

San Bernardino

In Partial Fulfillment

of the Requirements for the Degree

Master of Science

in

Information Systems and Technology

by

Needhi Devan Senthil Kumar

May 2023
 OTT SUBSCRIBER CHURN PREDICTION

USING MACHINE LEARNING

A Project

Presented to the

Faculty of

California State University,

San Bernardino

by

Needhi Devan Senthil Kumar

May 2023

Approved by:

Dr. William Butler, Member, Committee Chair

Dr. Conrad Shayo, Member, Reader & Department Chair, Information and

Decision Sciences
© 2023 Needhi Devan Senthil Kumar
 ABSTRACT

Subscriber churn is a critical issue for companies that rely on recurring

revenue from subscription-based services like the OTT platform. Machine

Learning algorithms can be used to predict churn and develop targeted retention

strategies to address the specific needs and concerns of at-risk subscribers. The

research questions are 1) What Machine Learning algorithms are used to

overcome subscriber churn? 2) How to predict subscribers’ churn in the OTT

platform using Machine Learning? 3) How to retain subscribers and improve

customer targeting? The dataset was collected from the Kaggle repository and

implemented it into the various prediction algorithms used in previous research.

Then, evaluate the performance of each algorithm to find out the highest

accuracy model. The findings and conclusion for each question are 1) Logistic

regression, multi-layer perceptron, random forest, decision trees, and gradient

boosting machines were identified as effective algorithms for churn prediction

analysis. 2) By sending the test data to a trained model by their historical dataset,

customers are likely to leave a company (i.e., churn) based on their

characteristics can be predicted. 3) Personalized offers and promotions,

improving customer service, developing loyalty programs, and optimizing pricing

strategies were suggested strategies for retaining subscribers. The gradient

boosting machine model was found to have the highest accuracy and maximum

AUROC, making it a powerful tool in the fight against customer churn. Areas for

further study include incorporating unstructured data sources, deep learning

iii
techniques, and integrating real-time data sources to improve the accuracy and

effectiveness of churn prediction models.

Search Term: Subscribers, Churn Prediction, OTT Platform

iv
 ACKNOWLEDGEMENTS

I would like to thank my parents and my friends for their support and

encouragement throughout the process of this project.

Also, I would like to thank Dr. William Butler and Dr. Conrad Shayo for

guiding me to finish this project.

v
 TABLE OF CONTENTS

ABSTRACT ................................................................................................. iii

ACKNOWLEDGEMENTS ............................................................................ v

LIST OF TABLES ........................................................................................ viii

LIST OF FIGURES ...................................................................................... ix

CHAPTER ONE: INTRODUCTION

Problem Statement ............................................................................ 1

Research Questions ......................................................................... 2

Organization of the Project ............................................................... 2

CHAPTER TWO: LITERATURE REVIEW ................................................... 3

CHAPTER THREE: RESEARCH METHODOLOGY

Logistic Regression ........................................................................... 6

Multi-Layer Perceptron....................................................................... 7

Random Forest .................................................................................. 7

Decision Tree..................................................................................... 8

Gradient Boosting Machine................................................................ 8

Data Collection ................................................................................. 9

CHAPTER FOUR: DATA DESCRIPTION AND ANALYSIS ......................... 12

System Requirements ....................................................................... 13

Implementation ................................................................................. 13

Experimental Results ........................................................................ 13

vi
CHAPTER FIVE: DISCUSSION, CONCLUSION, AND AREAS OF FURTHER

STUDY

Discussion ........................................................................................ 18

Conclusion ........................................................................................ 19

Area for Further Study ....................................................................... 20

REFERENCES ............................................................................................. 21

vii
 LIST OF TABLES

Table 1. Data Description ............................................................................. 12

viii
 LIST OF FIGURES

Figure 1. Classification Report of Logistics Regression ................................ 14

Figure 2. Classification Report of Multi-Layer Perceptron ............................ 14

Figure 3. Classification Report of Random Forest Classifier ........................ 15

Figure 4. Classification Report of Decision Tree Classifier ........................... 15

Figure 5. Classification Report of Gradient Boosting Machine ..................... 16

Figure 6. Values of AUROC.......................................................................... 17

ix
 CHAPTER ONE

INTRODUCTION

OTT (Over the Top) is a term used to describe the delivery of audio, video,

and other media content over the Internet without the need for a traditional cable or

satellite television provider. OTT services provide users with an alternative source for

content that is normally only available through a cable or satellite provider. They also

offer a cost-effective way to access content, often eliminating the need for expensive

contracts and monthly fees. OTT services typically require users to sign up for an

account, either through a website or an app, and then they can access the content

they desire. OTT services can provide a range of options, from streaming live TV

channels to on-demand shows and movies. These services also often provide

additional features, such as the ability to save favorite shows, get recommendations,

and access exclusive content(Fitzgerald, 2019). Let's start by stating the problem

statement, then go on to the research questions and how this culminating experience

project is structured.

Problem Statement

Subscriber churn is a major issue for online streaming services such as Over-

the-top (OTT) platforms (Madden et al., 1999b). Churn is the rate at which

subscribers to a particular OTT service cancel their subscriptions, and it’s a major

part of customer retention for OTT services. Churn is a complex issue, with potential

causes ranging from rising user-acquisition costs to a lack of personalized content. It

is an important metric for OTT services, as it can indicate customer satisfaction, as

well as the effectiveness of marketing and promotional efforts. A high churn rate can

indicate that customers are not satisfied with the service, or that the marketing efforts

1
are not effective. It can also indicate that the service is too expensive, or that there

are better alternatives available (Kuldeep et al., n.d.).

Research Questions

1. What Machine Learning algorithms are used to overcome subscriber churn?

2. How to predict subscribers’ churn in the OTT platform using Machine

Learning?

3. How to retain subscribers and improve customer targeting?

Organization of Project

By building an optimized model that can be used to predict subscribers’ churn,

the goal is to explore how Machine Learning algorithms can be used to decrease

subscribers’ churn in the OTT platform. This culminating experience project is

organized as follows: Chapter 2 will describe the literature review of my research on

the existing churn analysis of various industries. Chapter 3 will provide research

methodologies. Chapter 4 will provide the data collection and implementation of

research methodologies. Based on the experimental result, Chapter 5 is composed

of a discussion, conclusion, and area for further study.

2
 CHAPTER TWO

LITERATURE REVIEW

To learn more about this issue and come up with a workable solution,

research and implementations done in the past by other authors were examined.

Let’s provide our findings in this section related to the research question: Q1. What

Machine Learning algorithms are used to overcome subscriber churn?

Agrawal et al., (2018) discussed the problem of customer churn and analyzes

previous works to identify gaps in the solutions implemented. Agrawal et al., (2018)

proposed a multi-layered Artificial Neural Network model to predict customer churn,

resulting in an accuracy of 80.03%. Ahmad et al., (2019) aimed to build a system to

predict customer churn in the Syriatel telecom company. The XGBOOST algorithm,

which had an AUC value of 93.301% produced the best results. XGBOOST

continued to produce the highest results with an AUC of 89% when tested on a fresh

dataset pertaining to various time periods The outcomes of predicting churn in the

telecom business have been proven to be improved by the usage of Social Network

Analysis elements. Their research is more focused on Machine learning algorithms

used in churn prediction in various industries, which is a part of our project but their

research is less focused on subscriber churn on OTT platforms.

To answer the research questions: Q2. How to predict subscribers’ churn in

the OTT platform using Machine Learning?

De Caigny et al., (2018) aimed to explore the viability of using the Logistic

Linear Model (LLM) as a classification technique in telecom customer churn

prediction. The results showed that LLM performed better than using Logistic

Regression (LR) and Decision Tree (DT) as standalone techniques and at least

equally well as two homogeneous ensemble methods, Random Forest (RF) and

3
Logistic Model Tree (LMT). The LLM provides a comprehensible method with acting

ability, and it can enrich both DT and LR by adding the coefficients of logistic

regression to the leaves in a decision tree and by fitting several logistic regressions

to consider specific group characteristics. Huang et al., (2012) introduced a brand-

new feature set for predicting customer turnover in the telecom sector, which

includes call details, account information, bill information, and other forms of data.

However, the research examined how customer behavioral characteristic data is

used to predict churn of the customers in financial industries, it lacks the methods in

predicting subscriber churn in the OTT platform.

To answer the research questions: Q3. How to retain subscribers and improve

customer targeting?

Ullah et al., (2019) aimed to build a churn prediction model for a telecom

company to improve its CRM and retain valuable customers. They used machine

learning techniques to analyze customer data and identify the main factors

contributing to churn. The results showed that the proposed model performed better

than other techniques and produced a better F-measure result of 88% using

Random Forest and J48. Sung Won Kim et al., (2019) also performed cluster

profiling to better understand the risk of churn for different groups of customers and

provided guidelines for customer retention. Their research is more focused on the

customer retention churn prediction model in telecom, but their research is less

focused on subscriber churn on OTT platforms.

Therefore, the above reviews discuss different approaches to predicting

customer churn. Agrawal et al., (2018) proposed a multi-layered Artificial Neural

Network model that identified attributes related to churn rate and achieved an

accuracy of 80.03%. Ahmad et al., (2019) applied feature engineering,

4
transformation, and selection to prepare data for four tree-based algorithms and

found that the xgboost algorithm produced the best results. De Caigny et al., (2018)

explored the viability of using the Logistic Linear Model as a classification technique

and found that it performed better than other classifiers. Huang et al., (2012)

presented a new feature set for churn prediction and evaluated seven modeling

techniques, finding that the Logistic Regression and DT/SVM were suitable for

predicting true churn rate and false churn rate. Finally, Ullah et al., (2019) used

machine learning techniques to build a churn prediction model and achieved an F-

measure result of 88% using Random Forest and J48. These studies highlight the

importance of feature selection and modeling technique selection for effective churn

prediction in the telecom, financial and social network industries and let’s implement

these methods into the OTT platform industry.

5
 CHAPTER THREE

RESEARCH METHODOLOGY

Here, the research methodologies used to answer the research questions

have been explained. To answer the research question: Q1. What Machine Learning

algorithms are used to overcome subscriber churn? From the literature review the

research come to the solution that Machine learning algorithms used to overcome

churn in other industries are Logistics Regression, Multi-Layer Perceptron, Random

Forest, Decision Trees, and Gradient Boosting Machines, let’s implement these

methods into the OTT platform industry to find subscriber churn. Let’s introduce the

machine learning algorithms that are used to predict subscriber churn analysis and

list some of their benefits and drawbacks.

Logistic Regression

A supervised learning method used in machine learning for binary

classification issues is logistic regression. The sigmoid function is used to predict the

likelihood of a binary outcome by estimating the coefficients of the input variables.

The implementation of logistic regression is straightforward, computationally

effective, and appropriate for small datasets. To comprehend the underlying causes

of the result, the coefficients can be understood. However, it is limited to binary

classification issues, presumes the independence of the data, and presupposes a

linear relationship between the input variables and the result. As a foundational

model for forecasting binary outcomes, logistic regression is frequently employed in

the fields of healthcare, finance, and marketing (Huang et al., 2012).

6
 Multi-Layer Perceptron

Machine learning uses the Multi-Layer Perceptron (MLP), an artificial neural

network, for a variety of tasks like classification, regression, and prediction. It has

many layers of neurons, which combine the inputs in a weighted way and then pass

them via an activation function. Using non-linear activation functions, MLP can

manage relationships between inputs and outputs that are not linear. It is an effective

tool for difficult classification and regression problems, but it can be computationally

expensive and prone to overfitting for big datasets. MLP is widely utilized in

numerous applications, including financial forecasting, speech recognition, image

recognition, and natural language processing (Agrawal et al., 2018).

Random Forest

It is an ensemble learning technique that mixes the outputs of numerous

decision trees to increase accuracy and lessen the chance of overfitting. A group of

decision trees are trained in Random Forest utilizing random subsets of

characteristics and data. The algorithm chooses the appropriate characteristic to

divide on at each node depending on a parameter like information gain or Gini index.

The outcome is then determined by which decision tree received the most votes (De

Caigny et al., 2018).

In order to obtain high accuracy and lower the risk of overfitting, it integrates

numerous decision trees. It is helpful for churn prediction studies where data quality

may be an issue because it can manage missing data and is robust to outliers. Yet,

compared to decision trees, Random Forests might be more computationally

expensive and challenging to understand. Moreover, it can favor categorical

variables with more levels and not be appropriate for small datasets (Ullah et al.,

2019).

7
 Decision Tree

A tree-like structure known as a decision tree algorithm was developed to

depict decisions and their potential outcomes based on the characteristics of the

data. The tree structure is produced by recursively dividing the data into subsets

depending on the most significant predictors, and the algorithm selects the optimal

feature to split on based on a criterion like information gain or Gini index at each

node (Ahmad et al., 2019).

Categorical variables, non-linear correlations, and big datasets with numerous

variables can all be handled via decision trees. They can be unstable with slight

changes in data, are biased toward variables with more levels, and are prone to

overfitting. They are also only able to predict binary outcomes, which may not be

useful for doing so in cases when there are several alternative outcomes, such as

forecasting churn (Ullah et al., 2019).

Gradient Boosting Machine

Gradient Boosting Machine (GBM) is a powerful and popular machine

learning algorithm that can be used for regression, classification, and ranking tasks.

GBM is capable of handling high-dimensional data and effectively reducing bias and

variance in complex models. It works by building an ensemble of weak learners,

such as decision trees, in a stepwise manner. The algorithm starts with a simple

model and then iteratively improves upon it by fitting new models to the residual

errors of the previous models. GBM optimizes a loss function using a gradient

descent approach to minimize the discrepancy between predicted and actual values

by adjusting the weights of the weak learners. The process is repeated until

convergence or a pre-specified number of iterations is reached (Ahmad et al., 2019).

8
 To answer the research questions: Q2. How to predict subscribers’ churn in

the OTT platform using Machine Learning?

The data related to subscribers’ churn, customer behavior, and demographics

will be collected from the OTT platform. The data may also include customer

feedback, preferences, and engagement with the platform. The collected data will be

preprocessed to remove any inconsistencies, missing values, and outliers. Feature

engineering techniques will also be applied to create new features that can be useful

in predicting churn and retention.

Data Collection

To do churn there are several characteristics that are important for subscriber

churn prediction analysis in OTT (over-the-top) services, and they are:

1. User demographics: Age, gender, location, and other demographic data can

be significant predictors of churn.

2. Usage patterns: Information on how frequently and how long a user accesses

the service, which features are used, and when they are used can provide

insights into churn likelihood.

3. Payment information: Whether a user is using a free trial, or a paid

subscription, their billing history, and their payment method can be used to

predict churn.

4. Content preferences: The types of content a user consumes, the ratings and

reviews they provide, and their engagement with content can be used to

predict churn.

5. Customer support interactions: Data on customer support tickets, complaints,

and other interactions can indicate whether a user is at risk of churn.

9
 6. Technical data: Information on the user's device type, network quality, and

other technical data can also be useful in predicting churn.

It is important to note that not all these characteristics may be relevant for

every OTT service or every subscriber churn analysis. The specific characteristics

that are most important may depend on the type of service, the target audience, and

the specific factors that drive churn for that service. Since the research papers

discussed in the literature review were from various industries like telecom, financial,

and social networks, it is not appropriate to use their dataset. So, after searching

OTT datasets in different data repository websites like UCI, Kaggle, Data.gov,

google dataset search, Data.world, Reddit, and OpenML got the churn modeling

dataset from the Kaggle repository which satisfies the requirement to be used to

develop the churn prediction analysis machine learning model.

To answer the research questions: Q3. How to retain subscribers and improve

customer targeting?

Based on the insights gained from the models, customer retention strategies

such as personalized recommendations, targeted marketing, and loyalty programs

will be developed to improve customer engagement and retention. The developed

models and retention strategies will be implemented in the OTT platform and

monitored for effectiveness. Any necessary modifications will be made to further

improve the performance of the models and strategies. The effectiveness of the

implemented models and retention strategies will be evaluated using appropriate

metrics such as customer churn rate, customer lifetime value, and revenue. The

evaluation will help in identifying the strengths and weaknesses of the implemented

models and strategies and making necessary improvements.

10
 So, the summary is choosing the best algorithm for churn analysis depending

on various factors such as data size and quality, problem complexity, desired

accuracy, and interpretability. It is recommended to start with a simple model like

logistic regression, which is computationally efficient and easy to interpret. For larger

datasets, random forests and gradient boosting can handle missing data and outliers

but may be slower. Performance metrics such as accuracy, precision, recall, and F1

score should be evaluated on a holdout set of data, and interpretability should be

considered. Ensemble methods like random forests and gradient boosting can be

used if no single algorithm performs well on the data. Ultimately, the best algorithm is

chosen by comparing the performance of multiple algorithms on the specific

characteristics of the problem and the data. The following chapter will begin with an

introduction to our dataset, followed by a thorough analysis and results.

11
 CHAPTER FOUR

DATA DESCRIPTION AND ANALYSIS

In this chapter, let’s see the detailed data description, and the system

requirements which need to run the machine learning program. Finally, the

experimental results of the models have been discussed below. The dataset, which

contains 16 columns and 2000 entries, is described in the table below.

Table 1. Data Description

S.No Attribute Name Description

1 Year year of the subscription

2 customer_id id of the subscriber

3 phone_no phone number of the subscriber

4 Gender gender of the subscriber

5 Age age of the subscriber

6 no_of_days_subcribed the number of days since the subscription

7 multi_screen single/multiple-screen subscription

8 mail_subscrided subscribers receive mails

9 weekly_mins_watched number of minutes watched weekly

10 minimum_daily_mins minimum minutes watched

11 maximum_daily_mins maximum minutes watched

12 weekly_max_night_mins number of minutes watched at nighttime

13 videos_watched total number of videos watched

14 maximum_days_inactive days since inactive

15 customer_support_calls number of subscriber support calls

16 Churn 1-yes, 0-No

12
 System Requirements

Preprocessing the dataset, implementing the solution, and training the models

were all carried out in a Jupyter notebook. Version 6.1.4 of the Jupyter notebook is

utilized. Python 3.8 is the language used for programming. Data science and

machine learning both employ Python, a well-liked high-level programming

language. For activities like data preprocessing, model training, and evaluation, it

has a sizable and vibrant community that offers considerable help and resources for

machine learning projects. NumPy, Pandas, Matplolib, Scikit-Learn, TensorFlow,

PyTorch, and Keras are a few of these libraries.

Implementation

To build a Machine Learning algorithm below steps should be followed:

1) Load the churned data.

2) Preprocess the data ready for use.

3) Split the preprocessed data into training and testing sets.

4) Fit and train the model of the chosen algorithm.

5) Evaluate the performance of the model on the testing set.

6) Save and load the trained model.

Experimental Results

In this project, generated two variables, "X" and "Y," after uploading our

dataset to the environment, and allocated all of the features of the dataset to "X,"

with the exception of the column "Target," which was assigned to "Y." After dividing

the datasets, the appropriate algorithm was then employed, in this instance, logistic

regression. sklearn.linear model library's LogisticRegression package has been

imported. Utilizing the testing dataset, a prediction was made after the trained

13
datasets had been fitted into the model. Got the following Classification Report in

Fig.1 and an accuracy of 86.50%. The AUROC value of the model is 0.539 as shown

in Fig.6.

Figure 1. Classification Report of Logistics Regression

Secondly, the Multi-Layer Perceptron algorithm is used by importing the

MLPClassifier package from skleran.neural_network library. Set the hyperparameter

hidden_layer, activation, solver, alpha, and max_iter with the following values 100,

‘relu’, ‘adam’, 0.0001, and 1000 respectively. After fitting the trained datasets into the

model, predicted the outcome using the testing dataset. Got the following

Classification Report in Fig.2 and an accuracy of 87.50%. The AUROC value of the

model is 0.552 as shown in Fig.6.

Figure 2. Classification Report of Multi-Layer Perceptron

Then used the Random Forest algorithm by importing the

RandomForestClassifier package from sklearn.ensemble library. After fitting the

the following Classification Report in Fig.3 and an accuracy of 91%. The AUROC

value of the model is 0.676 as shown in Fig.6.

Figure 3. Classification Report of Random Forest Classifier

Then used the Decision Tree algorithm by importing the

DecisionTreeClassifier package from sklearn.tree library. After fitting the trained

datasets into the model, predicted the outcome using the testing dataset. Got the

following Classification Report in Fig.4 and an accuracy of 92.25%. The AUROC

value of the model is 0.779 as shown in Fig.6.

Figure 4. Classification Report of Decision Tree Classifier

Finally, used the Gradient Boosting Machine algorithm by importing the

GradientBoostingClassifier package from sklearn.ensemble library. After fitting the

trained datasets into the model, predicted the outcome using the testing dataset. Got

15
the following Classification Report in Fig.5 and an accuracy of 93.25%. The AUROC

value of the model is 0.793 as shown in Fig.6.

Figure 5. Classification Report of Gradient Boosting Machine

A common performance metric in binary classification issues is AUROC (Area

Under the Receiver Operating Characteristic Curve). At various categorization

criteria, the true positive rate (TPR) is plotted versus the false positive rate (FPR).

The TPR, sometimes referred to as sensitivity, is calculated as the percentage of

true positives (positive samples that were properly predicted to be positive) among

all positive samples. The FPR is defined as the percentage of false positives

(positive samples that were mistakenly predicted to be positive) among all negative

samples.

The classifier's prediction scores are arranged in descending order to

determine AUROC, and the threshold is steadily raised. TPR and FPR are calculated

for each threshold and plotted on the ROC curve. The AUROC score, which goes

from 0 to 1, is the region under this curve. A score of 1 represents flawless

classification, whereas a score of 0.5 shows that the classifier is only slightly more

accurate than random guessing. Unlike other measures like accuracy, precision, and

recall, the AUROC is not affected by class imbalance and threshold selection. This

makes it a valuable metric. It is especially helpful since the threshold can be

16
changed to maximize the statistic that matters most in the particular context when

the cost of false positives and false negatives differ. The AUROC of the used model

is shown below in Fig.6.

Figure 6. Values of AUROC

17
 CHAPTER FIVE

DISCUSSION, CONCLUSION, AND AREAS OF FURTHER STUDY

Discussion

The research questions were:

1) What Machine Learning algorithms are used to overcome subscriber

churn?

2) How to predict subscribers’ churn in the OTT platform using Machine

Learning?

3) How to retain subscribers and improve customer targeting?

After discussing the results and drawing a conclusion, follows suggestion for

areas for further study. Here are some findings and conclusions from previous

questions. In reference to the 1st question, from the literature review and

experimental results, the research come to the solution that Machine learning

algorithms used to overcome subscriber churn are Logistics Regression, Multi-Layer

Perceptron, Random Forest, Decision Trees, and Gradient Boosting Machines.

Moving on to the 2nd question, by sending the test data to a trained model by their

historical dataset, customers are likely to leave a company (i.e., churn) based on

their historical behavior and characteristics can be predicted.

Lastly referring to the 3rd question, here are some strategies that can be used

based on the insights obtained from churn prediction analysis: Identify at-risk

subscribers, personalized offers and promotions, improve customer service, develop

loyalty programs, and optimize pricing strategies.

Talking about Identifying at-risk subscribers’ strategy, churn prediction

models can help identify subscribers who are at risk of churning. Once these

18
subscribers are identified, targeted retention strategies can be developed to address

their specific needs and concerns. Moving on to personalized offers and promotions

strategy, by analyzing subscriber behavior and preferences, churn prediction models

can help companies develop personalized offers, discounted subscription plans, free

trials, or exclusive content to retain subscribers.

Churn prediction models can help identify common issues that lead to

subscriber churns, such as poor customer service or long wait times. Companies can

use this information to improve their customer service and address subscriber

concerns in a timely manner. By rewarding loyal subscribers with exclusive offers and

discounts, companies can increase subscriber retention and reduce churn. The

models can help identify which subscribers are most likely to respond to loyalty

programs and what types of rewards are most effective. Analyzing subscriber behavior

and subscription history can help companies optimize their pricing strategies.

Conclusion

In this project, five machine learning algorithms have been analyzed

and got the highest accuracy and maximum AUROC of 93.25% and 0.793

respectively in the Gradient Boosting Machine Model. And so, in churn prediction,

GBM can be used to identify the most significant predictors of churn and build a

model that predicts which subscribers are at a high risk of leaving. GBM can handle

many features, including both numerical and categorical variables, making it suitable

for churn prediction analysis. The algorithm can also deal with imbalanced datasets

and handle missing values, which are common in real-world churn prediction

scenarios. GBM can provide accurate and robust predictions, making it a powerful

tool in the fight against customer churn.

19
 Area for Further Study

Churn prediction is a complex and dynamic field, and there are several areas

of further study that could help improve the accuracy and effectiveness of churn

prediction models. One area of study is the incorporation of unstructured data

sources, such as social media and customer reviews, into churn prediction models.

Another area of study is the use of deep learning techniques, such as convolutional

neural networks (CNNs) and recurrent neural networks (RNNs), to improve the

accuracy of churn prediction models. Finally, the integration of real-time data

sources, such as clickstream data and mobile app usage data, could enable more

accurate and timely identification of customers at risk of churn. Overall, there are

many exciting areas of further study in churn prediction, and continued research and

innovation in this field have the potential to greatly improve customer retention

strategies across a variety of industries.

20
 REFERENCES

Agrawal, S., Das, A., Gaikwad, A., & Dhage, S. (2018). Customer Churn Prediction

Modelling Based on Behavioural Patterns Analysis using Deep Learning. 2018

International Conference on Smart Computing and Electronic Enterprise,

ICSCEE 2018. https://doi.org/10.1109/ICSCEE.2018.8538420

Ahmad, A. K., Jafar, A., & Aljoumaa, K. (2019). Customer churn prediction in

telecom using machine learning in big data platform. Journal of Big Data, 6(1).

https://doi.org/10.1186/s40537-019-0191-6

De Caigny, A., Coussement, K., & De Bock, K. W. (2018). A new hybrid classification

algorithm for customer churn prediction based on logistic regression and

decision trees. European Journal of Operational Research, 269(2).

https://doi.org/10.1016/j.ejor.2018.02.009

Fitzgerald, S. (2019). Over-the-Top Video Services in India: Media Imperialism after

Globalization. Media Industries, 6.

Huang, B., Kechadi, M. T., & Buckley, B. (2012). Customer churn prediction in

telecommunications. Expert Systems with Applications, 39(1).

https://doi.org/10.1016/j.eswa.2011.08.024

Kuldeep, C., Rojhe, V., Singh, N., & Rao, A. (n.d.). {YICCISS-2021} HJ Emerging

Trends in Management Sciences. Retrieved March 17, 2023, from

www.houseofjournals.com

Madden, G., Savage, S. J., & Coble-Neal, G. (1999a). Subscriber churn in the

Australian ISP market. Information Economics and Policy, 11(2), 195–207.

https://doi.org/10.1016/S0167-6245(99)00015-3

21
Madden, G., Savage, S. J., & Coble-Neal, G. (1999b). Subscriber churn in the

Australian ISP market. Information Economics and Policy, 11(2), 195–207.

https://doi.org/10.1016/S0167-6245(99)00015-3

Ullah, I., Raza, B., Malik, A. K., Imran, M., Islam, S. U., & Kim, S. W. (2019). A Churn

Prediction Model Using Random Forest: Analysis of Machine Learning

Techniques for Churn Prediction and Factor Identification in Telecom Sector.

IEEE Access, 7, 60134–60149. https://doi.org/10.1109/ACCESS.2019.2914999

22