Churn Prediction in Over‑The‑Top (OTT) for Customer Retention

Using Machine Learning Algorithms

G. Shabana, M. Jyothi, K. Ali Mahaboob Basha, D. Mounika Reddy and M. Charan Tej

Department of CSM, Srinivasa Ramanujan Institute of Technology, Rotarypuram, Anantapur, Andhra Pradesh, India

Keywords: Over‑The‑Top, Decision Tree, Random Forest, XG Boost, Gradient Boost, over Sampling, SMOTE, Churn

Prediction.

Abstract: In the view of content providers like Over-The- Top (OTT), the ability to predict the amount of churn is a key

part of the organization. With these predictions, the company can make better strategies in order to reduce the

churn rate. This paper presents a comprehensive study on Churn Prediction in Over-The-Top (OTT) using

various Machine Learning Algorithms. These include Decision Tree using both Entropy and Gini as

parameters, Random Forest, XG Boost, Gradient Boost algorithms. In which the class imbalance is found and

treated using Synthetic Minority Over- sampling Technique (SMOTE) and re-performed the machine learning

algorithms, in which the accuracy all algorithms are greater than 74% and better F1-Score, these findings can

be useful to the companies with real time data and to find the reasons behind customer attrition and increase

their customer life value and customer satisfaction.

1 INTRODUCTION

Churn is defined as the how many numbers of

customers are decided to leave the particular

company. Churn Prediction is the process of

identifying consumers who pose a danger of

cancelling their subscriptions or closing accounts

altogether.

It also detects the customers those who are in risk

of rejecting the subscription. Over-The-Top (OTT)

provides content like movies, web series, etc.

customers take subscription in order to get entertained

but due to some reasons they drop the subscriptions

in the middle and leave the platform, these platforms

will predict on what reasons customers are leaving. In

OTT platforms churn prediction is a vital concern.

Churn prediction is basically use machine learning

algorithms to detect the subscribed people leaving the

platform. Churn prediction considers both the "why"

and the "who". Companies may learn a lot about the

factors behind customer attrition by examining the

data used to forecast churn.

Various causes may contribute to this, such as

competitive products, unsatisfactory customer

service, absence of desired features, or cost.

Machine learning is very important for analysing

the customer data for future prediction over churn in

OTT. First, we collect the data on customers. Train

the data by using machine learning algorithms like

Decision Tree, Random Forest, XG Boost, Gradient

Boosting. After completion of training, we need to

evaluate the model to know the accuracy in

predicting churn. Once the model is done then it used

to predict the churn and reasons for the same.

Developing a model that can precisely anticipate

whether a customer will stick with this platform or not

is the aim of utilizing machine learning to forecast

subscriptions for Over-The-Top (OTT) services.

Strategy for client segmentation can be used with

churn prediction. Businesses are able to construct

more individualized customer experiences that meet

a range of wants and preferences by segmenting their

client base according to churn risk and other pertinent

characteristics. OTT firms need this information to

better understand and manage their marketing and

retention campaigns. To train the model, pertinent

data such as watching preferences and consumer

demographics is gathered and examined. In this

procedure, significant characteristics are chosen, the

data is cleaned, and a machine learning model is

trained.

In the following sections, we see each and every

methodology in detail, how the models are

performed, how the performance metrics is

calculated, finding the class imbalance and

224

Shabana, G., Jyothi, M., Basha, K. A. M., Reddy, D. M. and Tej, M. C.

Churn Prediction in Over-The-Top (OTT) for Customer Retention Using Machine Learning Algorithms.

DOI: 10.5220/0013910700004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 4, pages

224-228

ISBN: 978-989-758-777-1

performing over sampling using SMOTE, drawing

recommendations and conclusions.

2 RELATED WORKS

Mohan, M., & Jadhav, A. (2022). uses machine

learning techniques like Hierarchical logistic

regression, decision tree, random forest, Ada Boost.

Factors like multiple subscription, switching

frequency, content satisfaction, price satisfaction

have higher impact on customer churn. These are

found with most effective algorithm among all those

Random Forest, this approach has higher accuracy

compared to others.

Retention strategies in order to reduce churn in

OTT platforms are clearly discussed in Senthil

Kumar, Needhi Devan, 2023. Finding the most

significant attributes for churn of a particular

individual, the OTT platform can take necessary steps

in order to reduce churn, Content satisfaction shows

more effect on churn, so by taking the videos, movies,

that are highly satisfied by the viewers churn can be

reduced. The highly satisfied content can be collected

by viewers or customers feedback, review of a

particular movie or video, etc. Showing the related

content to the viewers is another strategy of OTT.

This can be done by having the data of one viewer,

what kind of movies they are continuously watching,

what genre they are interested. Author used logistic

regression, multi-layer perceptron, random forest,

decision trees, and gradient boosting machines and

also bought the accuracy of 80%. However, they

faced the problems with the data, the model built was

complex, model drift.

Churn of one organization depends on the

competitors. With the increase of technology, OTT

platforms are increasing day-by-day, this can become

a big hurdle for one platform. So, these should build

strategies by keeping competitors in mind. These are

explained in detailed in Manish Mohan, Anil Jadhav

(2022). The availability of competing services, other

platform price, have effect in churn.

Other than machine learning algorithms,

Comprehensive Understanding was done in Srivalli

Leela., et al, 2021. They stated that there will be

increase of paid subscribers by 16.1% by the year

2028 i.e., the subscriber market will increase from

USD104.2 billion to USD293 billion.

Over-The-Top (OTT) providers and Internet

Services Providers (ISPs) joint service management

approach based on Customer Lifetime Value (CLV)

and benefits of joint services management are

discussed in A. Ahmad, et al, 2017. They also stated

that this can improve the customer experiences,

increase customer loyalty which are key factors in

reducing churn. Over-The- Top (OTT) providers and

Internet Services Providers (ISPs) joint service

management approach based on Quality of

Experience (QoE) and benefits of joint services

management A. Ahmad, et al, 2016. This is a measure

of satisfaction got by particular viewer with respect to

the service they received. Regression analysis is

performed between QoE and Churn in order to get a

relationship among those.

Over-The-Top (OTT) Churn is more affected by

content provided by particular OTT platforms and

also the price charged for that, plans and subscription

options provided. This is analyzed by performing

content analysis and economic analysis Priya

Malhotra, Akshay Kumar (2021). Found that

customers are increasing because they feel that OTT

platforms are for providing entertainment, treat as

stress busters. This was found based on various

factors like variety of content, affordability of OTT

subscriptions.

Sachika Luthra, The author stated there are

increase of OTT subscribers during Covid-19. Almost

7.5 million subscribers have been increased from the

year 2019 to 2020. Factors like Covid-19 pandemic,

increasing availability of high-speed internet, growth

popularity of streaming devices showed significant

effect on this particular growth.

3 METHODOLOGY

3.1 About the Dataset

The dataset contains 16 attributes along with the

target variable 'churn' that is binary which states 0-no,

1- yes. The 15 independent variables are Year,

customer_id, phone_no, Gender, Age, multi_screen,

no_of_days_subcribed, mail_subscrided, weekly_mi

ns_watched, maximum_daily_mins,minimum_daily

_mins, weekly_max_night_mins, videos_watched,

maximum_days_inactive, customer_support_calls.

The dataset contains 2000 entries.

3.2 Data Pre-Processing

The initial step of the project includes pre-processing

steps like removing unnecessary attributes, handling

null values, outlier detection, some visualizations,

creation of dummies. Here we removed customer_id,

year, phone_no attributes. While treating with null

values, we found the attributes gender,

maximum_days_inactive, churn have null values.

Churn Prediction in Over-The-Top (OTT) for Customer Retention Using Machine Learning Algorithms

225

Using imputation methods, we treated these, gender

is filled with mode, maximum_days_inactive is filled

with median, churn is filled with mode. No outliers

are detected hence proceed further. When correlation

matrix is plotted, found out that

maximum_days_inactive is highly correlated

minimum_daily_min. Attributes like gender,

multi_screen, mail_subscribed are converted

categorical to binary through dummies.

Figure 1

show the System Architecture for Ott Churn.

Figure 1: System Architecture for OTT Churn.

3.3 Model Selection

Splitting of the dataset into training and testing is

done, 40% of the data is provided for testing and 60%

for training and random state is taken as 78 and built

several machine learning algorithms. Models like

Decision Tree using entropy and gini, Random

Forest, Gradient Boosting, XG Boosting are trained,

tested, and validated. Figure 2 show the Model

Performance Comparison Using Graph

3.4 Performance Evaluation

Performance metrics, Confusion metrics is plotted for

each model and Accuracy, Precision, Recall, F1-

Score are retained, from that the efficiency of the

model is been said. Random Forest has higher

accuracy followed by Decision tree and then XG

Boost. However, we got F1 Score greater than 0.5 for

most of the models.

Accuracy  TP  TN / TP  FP 

TN FN

(1)

Precision  TP / TP  FN

(2)

Recall  TP / TP  FN

(3)

F1  Score  2 ∗ Precision ∗ Recall / Precision 

Recall

(4)

Table 1: Evaluation Metrics Before Smote.

Accur

F1_S

core

Recall Precis

ion

Random

Forest

0.912 0.602 0.486 0.791

Decision_T

ree_Classifi

ID3

0.911 0.590 0.468 0.797

XG_Boosti

0.911 0.603 0.495 0.771

Decision_T

ree_Classifi

CART

0.906 0.540 0.404 0.815

Gradient

Boosting

0.873 0.346 0.248 0.574

Figure 2: Model Performance Comparison Using Graph.

3.5 Treating with Class Imbalance

Less F1-Score is occurred due to class imbalance, this

is treated with Synthetic Minority Over-sampling

Technique (SMOTE), for oversampling the data, after

performing SMOTE, all the above machine learning

models are performed again for the new data and do

the evaluation. Table 1 show the Evaluation Metrics

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

226

Before SMOTE After this we found out that XG

Boost has high accuracy followed by Random Forest

and Decision Tree with entropy.

4 RESULTS & DISCUSSION

All models are compared with accuracy, precision,

recall and F1-score to understand the performance of

various algorithms for churn prediction.

Figure 3

show the Model Performance Comparison Using

Graph In all the proposed models, XG Boost has high

accuracy of 90.9% followed by Decision Tree with

accuracy of 89.7% and Decision Tree with entropy

with accuracy of 89%. Gradient Boosting has less

accuracy of 87% and Random Forest has accuracy of

89.7%. All the performance metrics is as mentioned

in the following table 2.

Table 2: Evaluation Metrics After Smote.

Accuracy F1_Score Recall Precision

Boostin

SMOTE 0.909 0.644 0.606 0.688

Decision_Tree_Classifier_ID3_SMOTE 0.897 0.643 0.679 0.612

Random_Forest_SMOTE 0.897 0.627 0.633 0.622

Decision_Tree_Classifier_CART_SMOTE 0.890 0.614 0.642 0.588

Gradient

Boostin

SMOTE 0.871 0.502 0.477 0.531

Figure 3: Model Performance Comparison Using Graph.

5 CONCLUSIONS

Churn prediction models are useful in identifying

typical problems, including excessive wait times or

terrible customer service, that result in subscriber

churns. Businesses may utilize this data to promptly

handle subscriber problems and enhance their

customer service. Businesses may lower attrition and

boost subscriber retention by rewarding devoted

customers with special offers and incentives. Which

subscribers are most likely to respond to loyalty

programs and what kinds of rewards work best may

be determined with the use of these models.

Through the analysis of subscriber behavior and

subscription history, businesses may enhance their

pricing methods.

In this project, four machine learning algorithms

had been used and we got the highest accuracy to be

0.90 and 0.89 in XG Boost and Decision Tree.

Random forest has higher accuracy when class

imbalance is not treated where as XG Boost has high

accuracy before and after treating the class

imbalance.

We draw the conclusion that ensemble approaches

for churn prediction will yield excellent accuracy as

well as additional performance measures. Future

developments, such as the use of AI chatbots and

gamification, may contribute to this effort.

Chatbots with artificial intelligence (AI) can

engage with subscribers and detect those who are

likely to leave. In order to reduce customer attrition,

chatbots may also be utilized to send subscribers

tailored offers and suggestions.

It is possible to utilize gamification to motivate

users to stick around on the service. One way to

achieve this is by providing incentives for viewing

particular material or urging others to sign up.

REFERENCES

A. Ahmad, A. Floris and L. Atzori, "QoE- aware service

delivery: A joint-venture approach for content and

network providers," 2016 Eighth InternationalConfere

nce on Quality of Multimedia Experience (QoMEX),

Lisbon, Portugal, 2016, pp. 16, doi:10.1109/QoMEX.2

016.7498972.

A. Ahmad, A. Floris and L. Atzori, "OTT- ISP joint

service management: A Customer Lifetime Value

based approach," 2017 IFIP/IEEE Symposium onInteg

rated Network and Service Management (IM),

Lisbon, Portugal, 2017, pp. 10171022, doi:10.23919/I

NM.2017.7987431

Anish Yousaf, Abhishek Mishra 2021- A cross-country

analysis of the determinants of customerrecommendati

on intentions for over- the-top (OTT) platforms.

Churn Prediction in Over-The-Top (OTT) for Customer Retention Using Machine Learning Algorithms

227

E. Liotou, G. Tseliou, K. Samdanis, D. Tsolkas, F.Adelant

ado and C. Verikoukis, "An SDN QoE-service for

dynamically enhancing the performance of OTT

applications," 2015 Seventh International Workshop on

Quality of Multimedia Experience (QoMEX), Pilos,

Greece, 2015, pp. 12, doi:10.1109/QoMEX.2015.7148

106.

Manish Mohan, Anil Jadhav (2022). Predicting Customer

Churn on OTT Platforms: Customers with Subscription

of Multiple Service Providers. Journal of Information

& Organizational Sciences.

Mohan, M., & Jadhav, A. (2022). Predicting customer

churn on OTT platforms: Customers with subscription

of multiple service providers. Journal of the

Association for Information Science and Technology,

73(1), 1- 15.

Priya Malhotra, Akshay Kumar (2021),” Market Research

and Analytics on Rise of OTT Platforms: A study of

Consumer Behaviour” International Journal ofAdvanc

es in Engineering and Management (IJAEM) Volume

3, Issue 7 July 2021, pp: 4005-4012

Sachika Luthra- The Impact of Covid-19 on Consumer

Perception Towards Subscription Based OTTPlatform

Senthil Kumar, Needhi Devan, "Ott Subscriber Churn

Prediction Using Machine Learning" (2023). Electronic

Theses, Projects, and Dissertations. 1660.

Sistla Srivalli Leela Praveena, Dr. Vinay Negi-Over-The-

Top (OTT) 2021Video Market: Rise of PaidSubscripti

on Viewers Study

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

228