2024 International Conference on Emerging Research in Computational Science (ICERCS)

Customer Churn Prediction Using Machine

Learning Algorithms
2024 International Conference on Emerging Research in Computational Science (ICERCS) | 979-8-3315-3496-7/24/$31.00 ©2024 IEEE | DOI: 10.1109/ICERCS63125.2024.10895079

1st Garray Balaji 2nd Neralla Aditya Gowtham 3rd Suthapalli Tarun
Department of Data Science and Department of Data Science and Department of Data Science and
Business Systems Business Systems Business Systems
SRM Institute of Science and SRM Institute of Science and SRM Institute of Science and
Technology Technology Technology
Chennai, India Chennai, India Chennai, India
[email protected] [email protected] [email protected]

4th Gaurav Prajapati 5th N.Manikandan

Department of Data Science and Department of Data Science and
Business Systems Business Systems
SRM Institute of Science and SRM Institute of Science and
Technology Technology
Chennai, India Chennai, India
[email protected] [email protected]

Abstract— This study explores the application of machine also understand the root causes. This information will allow
learning algorithms for predicting customer churn within the the organizations to implement pre-emptive actions, whether
telecommunications sector, using a dataset containing key there are some discounts on their next shopping or
customer attributes like service usage indicators, and account redeemable points for walking an extra mile in a fitness app.
details. Customer churn prediction seeks to determine the
Predicting and preventing churn in the digital economy,
likelihood of customers discontinuing service, a crucial insight
for retention strategies. Multiple predictive models were where acquiring a customer at a high cost and retaining that
developed and compared, including logistic regression, decision customer is a fragile business today, has become extremely
trees, random forests, and gradient boosting classifiers, to significant. This paper explores methodological approaches
assess their individual effectiveness and performance. In this to customer churn prediction, specifically recent
analysis, extensive feature engineering was employed to developments within the field of machine learning, data
enhance data representation, followed by targeted feature preprocessing techniques, and practical challenges that
selection to refine model precision and reduce overfitting. Each businesses meet in the implementation of effective churn-
model was evaluated using essential metrics: accuracy, prediction models. To forecast churn based on call duration,
precision, recall (sensitivity), and the area under the ROC
usage frequency, and billing information, early
curve, with comparative analysis highlighting the relative
strengths of each approach. Results indicate that ensemble telecommunications churn prediction models mainly used
methods such as gradient boosting and random forests statistical techniques including logistic regression, decision
perform best in capturing complex churn patterns. trees, and survival analysis. For increased accuracy,
machine learning techniques such as neural networks,
Keywords— Customer Churn Prediction, Machine learning random forests, and gradient boosting have been employed
Algorithms, Churn Analysis recently.
Companies may now identify at-risk clients and provide
I. INTRODUCTION
prompt incentives or solutions by using real-time analytics,
A major problem faced by companies across different which has been integrated to deliver up-to-date churn
industries is customer churn which is the rate at which projections. Real-time models frequently process client
customers stop doing business with an entity, especially in contacts in real time using streaming data analytics systems
competitive market like telecommunications, retail & Credit risk models, which estimate the likelihood of default,
banking. Anticipating customer churn well in advance can have been used by banking and finance. These models have
offer considerable benefits to businesses, as it would give now been extended to analyze loan repayment patterns,
them ample time and room for implementing retention transaction frequency, customer support interactions, and
strategies thereby limiting the expenses of losing valuable churn predictions. Among the methods employed are
customers. With business decisions being more and more ensemble models, logistic regression, and support vector
data-driven, a customer churn prediction problem has machines. In order to help banks customize offers for high-
acquired natural importance among Data Science / Machine risk customers, predictive models like k-means clustering
Learning practitioners. and neural networks are frequently employed in customer
Historical customer data, such as transaction history segmentation and risk assessment.
and behavioural metrics (customer engagement scores), are To increase consumer engagement and lower attrition, e-
used to develop predictive models that identify patterns commerce platforms use recommendation systems that
associated with churn prediction. Through advanced combine collaborative filtering, content-based filtering, and
practices including machine learning algorithms, enterprises hybrid recommendation algorithms. They can make tailored
can not only forecast which consumers are churn-happy but products and offer recommendations by looking at browser

979-8-3315-3496-7/24/$31.00 ©2024 IEEE

Authorized licensed use limited to: BANGALORE INSTITUTE OF TECHNOLOGY. Downloaded on March 05,2025 at 04:03:18 UTC from IEEE Xplore. Restrictions apply.
 history, purchasing trends, and product interactions. To additional tree in particular showed that tenure,
forecast churn, e-commerce platforms also employ models consumption, and referral volume are important
like XGBoost, KNN, and deep learning algorithms. In order determinants of churn.
to comprehend client engagement and loyalty, these models According to the study [7], the best machine learning
concentrate on behavioural segmentation and customer models for predicting customer attrition are Random Forest
lifetime value (CLV) analysis. and Gradient Boosting (AdaBoost and XGBoost). The
Artificial Neural Network (ANN) model had the highest
II. LITERATURE SURVEY accuracy but needed much more time to train. Because of
According to the study [1], which has performed their better performance and quicker training, the authors
Logistic regression which has outperformed decision trees chose Decision Tree, Random Forest, and XGBoost.
in telecom churn prediction, with an accuracy of 94.24% as Nevertheless, the ANN model's long training period is a
opposed to 85.18%. The success of this model can be recognized drawback, leading to recommendations for more
attributed to efficient features of engineering like data study to improve deep learning models' effectiveness and
processing and cleaning. Its predicted reliability was proven performance. For churn prediction, the study [8] shows the
by a number of criteria, including precision, recall, and F1- robust overall accuracy of 96.39% for an ensemble model
score. Notwithstanding its advantages, the study's and the high ROC AUC accuracy of 99.49% for the
dependence on a single logistic regression model might be Histogram-based Gradient Boost (HistGB) model. However,
increased by investigating hybrid models, which could result real-world applications may be limited by the short, fictional
in increased accuracy. According to the study [2], the dataset. To increase model reliability, future studies could
banking industry, the LightGBM and Random Forest improve feature engineering and ensemble approaches.
algorithms had the best customer turnover forecast accuracy, With an accuracy of 83.9% and an AUC of 0.84, the
at 87%. LightGBM's improvements improve accuracy and study [9] concludes that the Stochastic Gradient Booster
efficiency, which is responsible for its performance. The model is the most accurate for forecasting customer
authors suggest that adding deep learning methods for attrition, providing useful information for client retention
handling imbalanced data, such as LSTM and the SMOTE tactics. Gradient boosting models, however, can be difficult
method, could improve model performance even more. to scale, sensitive to outliers, and necessitate a great deal of
Limitations, however, include the unrealized potential of hyperparameter tuning. According to the study [10], the
these cutting-edge methods and the requirement that banks Gradient Boosting model achieved the highest Area Under
create plans to keep older clients especially women who are the Curve (AUC) metrics and performed better than Logistic
more likely to leave. Regression, SVM, and Random Forest in predicting
According to the study [3], it presents a Hybrid Churn customer turnover in the telecom sector. The study
Prediction (HCP) model that makes use of multiple machine emphasizes how important it is to use machine learning to
learning techniques to enhance accuracy in predicting analyze telecom data in order to understand the dynamics of
customer churn within the telecom sector. The HCP model customer attrition. Its exclusive focus on churn prediction,
significantly outperforms standalone algorithms like however, limits research options because it ignores the
Logistic Regression and Random Forest across two datasets, incorrect classification of non-churning consumers and
while also identifying critical churn determinants such as relies on a binary output. To improve prediction accuracy
monthly charges and tenure. However, the study faces and reduce misclassification problems, the authors suggest
limitations due to imbalanced datasets, which can skew an analysis methodology based on profiles.
traditional models, and the lack of interpretability in certain According to the study [11], it provides an overview of
algorithms, limiting insights into churn factors. According churn analysis techniques across industries, highlighting
to the study [4], a voting classifier that combines multiple fragmentation in churn definitions, feature engineering, and
machine learning models achieved an 81.56% prediction predictive models. It offers a valuable guide for researchers
accuracy rate for customer attrition. The Apriori approach to select appropriate churn definitions and models for
was used in the study to identify characteristics, and models specific service domains. The study emphasizes the
were then trained using these essential features, which were potential of deep learning for modeling non-contractual
then improved by ensemble learning strategies. Notably, customer churn in evolving industries. However, the lack of
association rules and lift metrics showed that service type, a common dataset and standardized evaluation metrics
charges, and tenure were the most important factors makes it difficult to compare model performance across
impacting customer churn. industries, limiting the ability to select the most effective
The best F1-score and 95% accuracy in predicting model. These findings call for unified benchmarks and
customer attrition are obtained when Random Forest and tailored approaches to advance churn prediction research.
SMOTE-ENN are combined based on the study [5]. Using According to the study [12], it examines three AdaBoost
SMOTE with ENN sampling, it successfully addresses class algorithms- Real AdaBoost, Gentle AdaBoost, and Modest
imbalance and shows how much better this approach AdaBoost for predicting customer churn in a credit debt
performs than other models like Decision Tree and database. Real AdaBoost and Gentle AdaBoost showed
conventional Random Forest. According to the study [6], the strong predictive performance, identifying key factors like
best features chosen by the decision tree, random forest, and debt amount, customer duty level, and repayment type,
additional tree classifier models substantially overlap, while also adapting well to unbalanced datasets. In contrast,
indicating that they are appropriate for churn prediction. Modest AdaBoost required a balanced sampling scheme to
Because they represent consumer preferences for the improve its performance. The study highlights the
business, the more intricate models random forest and effectiveness of AdaBoost algorithms in churn prediction

Authorized licensed use limited to: BANGALORE INSTITUTE OF TECHNOLOGY. Downloaded on March 05,2025 at 04:03:18 UTC from IEEE Xplore. Restrictions apply.
 and emphasizes the importance of identifying relevant
factors to enhance decision-making and customer retention
strategies. According to the study [13], it explores customer
churn prediction in the telecommunications sector, with the
Random Forest classifier achieving an impressive accuracy
of 98.25%. The model’s success is attributed to its ability to
capture key variables and patterns, helping telecom
operators identify at-risk customers and implement targeted
retention strategies. However, the study did not include deep
learning models or customer feedback, which could provide
a more holistic view of customer behavior. As a
foundational work, it offers opportunities for further
research, particularly in refining models and evaluating their
robustness with varying hyperparameter settings for
improved performance.
According to the study [14], it investigates the Extra
Trees Classifier combined with Principal Component Fig. 1: Workflow Diagram
Analysis (PCA) for predicting customer churn, achieving
the highest accuracy (85.3%) and F1-score (0.514). The Data preprocessing is essential in machine learning to
model effectively handled large datasets and identified key optimize model design and data analysis. In this study,
churn factors like "customer id" and "credit score." This missing values were handled using imputation, converted
approach can enhance customer management and retention relevant data types (e.g., changing 'Total Charges' from
strategies. However, the focus on accuracy overlooks other string to numeric), and scaled numerical features such as
important metrics such as precision and recall. The study 'Monthly Charges' and 'Tenure Months' to ensure balanced
lacks dataset details, limiting reproducibility and general feature contributions. Categorical variables like 'Gender' and
applicability. Additionally, it does not address scalability, 'Internet Service' were encoded using one-hot encoding for
real-time deployment, or challenges like feature importance model interpretability. As shown in Fig. 2, the customer
analysis, data privacy, and model interpretability in real- churn rate visualization highlights the proportion of
world contexts. This study [15] evaluates machine learning customers who have churned versus those who have
models for predicting customer complaints, with AdaBoost remained, providing an initial insight into the dataset’s
combined with a Decision Tree achieving the best balance. Fig. 3 illustrates the status of customers by gender
performance in accuracy, precision, recall, and F1-score. distribution, showcasing the gender breakdown within the
Other models, including Random Forest, Logistic dataset, which is crucial for understanding potential biases
Regression, and Gaussian Naive Bayes, also performed before model training
well, while KNN and SVM showed significantly lower
performance. However, the study's reliance on data from a
single Egyptian mobile network operator limits the findings'
generalizability to other telecom providers or regions. The
authors suggest that future research could explore deep
learning techniques to analyze the physical status of fixed
broadband line boxes for better network performance and
maintenance insights.
III. METHODOLOGY
As shown in Fig. 1, the methodology outlined in the paper
includes data acquisition, data pre-processing, data
preparation, model selection, and model evaluation to
determine the model's performance on the test dataset. In the
case of customer churn prediction, this means collecting Fig. 2: Customer churn rate
data about the customers from various sources which would
include their demographics details and service usage pattern This process posed challenges, including maintaining data
as well as historical behaviour with respect to your product. quality while addressing outliers and missing values.
The dataset was obtained from Kaggle. The dataset from a Preprocessing steps followed statistical guidelines to ensure
telecommunications company in this project have columns data compatibility with machine learning algorithms and
like CustomerID, Gender, Senior Citizen (Yes or No), enhance prediction accuracy. Clean, well-prepared data
Tenure Months and features of the service: Internet Service underpins a stable and generalizable model.
and Phone Service representing customers as rows.

Authorized licensed use limited to: BANGALORE INSTITUTE OF TECHNOLOGY. Downloaded on March 05,2025 at 04:03:18 UTC from IEEE Xplore. Restrictions apply.
 In this study, a variety of machine learning algorithms
including Logistic Regression, Decision Trees, Support
Vector Classifier (SVC), K-Nearest Neighbors (KNN),
AdaBoost, and Gradient Boosting were selected to explore
different approaches in predicting customer churn. Each of
these algorithms offers distinct benefits that align well with
the churn prediction task. Known for its simplicity and ease
of interpretation, Logistic Regression allows to clearly
understand the influence of each feature on churn
probability, making it a valuable tool for stakeholders
seeking actionable insights. As a baseline model, Logistic
Regression provides a useful point of comparison with more
Fig. 3: Status of customers by gender distribution advanced algorithms.
Decision Trees are effective at capturing complex, non-
In the dataset considered, a huge class imbalance was linear relationships in data, making them well-suited for the
observed . This could drastically affect the performance of varied patterns seen in churn datasets. This model’s
the model. It is observed that there is an 8% increase in the structure is also highly interpretable, allowing easy
model performance when class imbalance is addressed [16]. visualization of decision paths. Although Decision Trees can
So, to enhance the model’s ability to accurately predict both overfit, they are highly effective when used in ensemble
classes, some techniques like SMOTE-ENN and methods. SVC is particularly advantageous for high-
Hyperparameter Tuning were implemented which are dimensional data, enabling it to handle datasets with a large
specifically designed to address this imbalance. SMOTE number of features. It classifies by constructing a
generates synthetic examples of the minority class (churned hyperplane or set of hyperplanes in a multi-dimensional
customers) to balance the class distribution. By interpolating space, helping to distinguish between classes. However,
between instances, SMOTE increases the representation of SVC can be computationally demanding and requires
the minority class without simply duplicating existing data careful tuning, especially in large datasets. As a non-
points, thus preserving variety within the dataset. To further parametric, instance-based learning model, KNN is useful
address imbalance, methods like random under sampling for capturing local data patterns by evaluating the
were explored, which reduces the number of majority class neighborhood of each data point.
instances (non-churn customers) to create a more balanced The effectiveness of KNN relies heavily on selecting the
dataset. While this approach reduces the training set size, it optimal number of neighbors (k) and distance metric, which
prevents the model from being overly biased toward the are refined during hyperparameter tuning. AdaBoost creates
majority class. To achieve a comprehensive evaluation, the a stronger overall model by sequentially combining weak
experiment involved combining SMOTE and under classifiers, usually Decision Trees. Each successive model
sampling. This approach synthesizes minority samples while corrects errors made by the previous one, thereby enhancing
reducing the majority class, allowing the model to achieve a overall accuracy. This model’s adaptability to linear and
balanced perspective on both classes without excessive data non-linear patterns, coupled with its reduced risk of
loss. overfitting, makes it a suitable choice for churn prediction.
In customer churn prediction, feature engineering plays a Gradient Boosting is highly effective for structured data,
vital role in enhancing model performance. The XGBoost with a strong ability to capture complex patterns by building
Classifier, a powerful gradient boosting algorithm, is models iteratively. It adapts well to churn prediction by
particularly useful for assessing feature importance. It correcting the errors of prior models at each step. The
provides methods such as Gain (measuring how much a XGBoost variant of Gradient Boosting, in particular,
feature improves model accuracy), Cover (indicating the includes regularization mechanisms that prevent overfitting,
proportion of observations a feature affects), and Frequency improving model robustness in handling diverse customer
(how often a feature is used in decision trees). By data.These algorithms collectively provide a range of
leveraging these important scores, the most impactful modelling approaches, from simple, interpretable methods
features, such as Tenure Months, Churn Score, contract, to sophisticated techniques that excel with complex, non-
total charges while potentially removing less relevant ones. linear data patterns. Each was selected to address different
XGBoost also helps mitigate over-fitting through facets of the customer churn prediction problem effectively.
regularization, making the model more robust. In the churn Model development uses prepared data with machine
project, this analysis streamlined feature selection, improved learning algorithms to create models. For example, this is
interpretability, and enhanced predictive power, ensuring the likely to be selecting algorithms for customer churn
model performs well on unseen data. This combination of predication, training them on historical data and measuring
statistical theory and machine learning boosted the model’s how good the performance of those models are. After that,
ability to detect churn patterns effectively. dataset is divided into two 70% for training and remaining
Splitting the dataset is a crucial step in machine learning 30% as a testing set in order to work on unseen data.
to evaluate how well the model generalizes to unseen data. The following are some of the algorithms that were tried
Typically, the dataset is divided into two main subsets: out: Logistic Regression, Decision Trees, Support Vector
training (70%) and testing set (30%) i.e 70% of the data Classifier (SVC) and K-Nearest Neighbors (KNN).
available is used for training and the remaining 30% data is Additionally, Ensemble methods such as AdaBoost and
used to test the trained model. Gradient Boosting were tested to increase the accuracy. All

Authorized licensed use limited to: BANGALORE INSTITUTE OF TECHNOLOGY. Downloaded on March 05,2025 at 04:03:18 UTC from IEEE Xplore. Restrictions apply.
 the models were evaluated after training and used metrics The performance evaluation highlights the precision of
such as accuracy, precision, recall & F1-score. E.g: SVC has the model in predicting non-churning customers (Class 0)
92 % accuracy; other models give little different results as with a true positive rate of 96%. This precision indicates that
they are more complex from the LR model side of view. To 96% of the customers predicted as non-churning were
maximize model performance, hyperparameter tuning was accurately classified. Additionally, the recall for Class 0 is
done using methods like grid search. By preventing 0.96, meaning 96% of the actual non-churning customers
overfitting, cross-validation made sure the models were correctly identified. The F1-score, which provides a
performed well when applied to fresh data. To improve balance between precision and recall, is also 0.96, reflecting
the strong predictive performance for this class. The support
prediction accuracy and expedite the model development
for Class 0 is 1,551, representing the total number of actual
process, future research could investigate sophisticated
non-churned customers in the dataset. For Class 1 (churning
models like deep learning or Auto ML. customers), the model demonstrates an accuracy of 0.88,
IV. RESULTS AND DISCUSSION correctly predicting 88% of churning customers. The recall
for Class 1 is also 0.88, indicating that 88% of actual
The proposed model was trained and tested with various churners were correctly identified. The F1-score for this class
machine learning algorithms and the accuracies of those is 0.88, balancing precision and recall. The support for Class
algorithms are as follows The bar graph Fig.4 compares the 1 is 559, which is the total number of actual churned
accuracy of six machine learning models used for customer customers.
churn prediction: Gradient Boosting Classifier achieved the
highest accuracy at 94%, making it the most effective model Overall, the model achieved an accuracy of 94%,
in this analysis. AdaBoost Classifier followed closely with an meaning that 94% of all predictions were correct. The macro
accuracy of 93%, indicating strong performance. Support average for precision, recall, and F1-score across both
Vector Classifier (SVC) performed well with an accuracy of classes is 0.92, reflecting the model's average performance
92%. without considering class distribution. The weighted average
for precision, recall, and F1-score is 0.94, which accounts for
the support of each class, further emphasizing the model's
effectiveness in handling both non-churning and churning
customers. The model was very good at accuracy and more
importantly had excellent performance regarding the
identification of non-churning customers. Excellent metrics
were also attained by churn class although almost slightly
lower than that concerning non-churn. In the confusion
matrix Fig. 5, Type 2 errors, also known as false negatives,
are instances in which the model mis-predicted that a
customer would not churn when in truth and in reality, he or
she does churn. This is reflected in the above matrix by the
67 instances where actual churned customers were wrongly
Fig. 4: Comparison of Accuracy Scores for different algorithms tagged as not churned.
Logistic Regression, a simpler and more interpretable • Type 1 errors (False Positives): 68
model, delivered an accuracy of 91%. Decision Tree • Type 2 errors (False Negatives): 67
Classifier achieved an accuracy of 90.62%, slightly lower
than the top models. K-Nearest Neighbors (KNN) had the
lowest accuracy at 88%, likely due to its sensitivity to the
choice of neighbors and distance metrics. Overall, ensemble
methods like Gradient Boosting and AdaBoost outperformed
individual models, highlighting their strength in improving
prediction accuracy through model aggregation.

TABLE I. EVALUATION METRICS

precision recall f1-score support

0 0.96 0.96 0.96 1551

1 0.88 0.88 0.88 599
accuracy 0.94 2110 Fig. 5: Confusion Matrix
macro avg 0.92 0.92 0.92 2110 Type 2 errors are also of importance while predicting churn,
as it is equally important to identify those customers who
weighted avg 0.94 0.94 0.94 2110 are likely to churn so that the firm can retain them and not
lose them without providing an opportunity to act.
Relatively fewer Type 2 errors may be an indicator of good
Table 1 shows the performance metrics of the performance of the model, but further minimization can lead
classification model Gradient Boosting classifier across to even better strategies to prevent people from churning
two classes: "0" (non-churn) and "1" (churn). through focused marketing campaigns or retention
programs.

Authorized licensed use limited to: BANGALORE INSTITUTE OF TECHNOLOGY. Downloaded on March 05,2025 at 04:03:18 UTC from IEEE Xplore. Restrictions apply.
 V. CONCLUSION International Conference on Artificial Intelligence and Robotics
(QICAR), 2024.
A customer churn prediction is critical for those [3] P. Aggarwal and V. Vijayakumar, “Customer Churn Prediction in the
businesses that intent to keep the at-risk customers from Telecom Sector,” 2024 3rd International Conference on Artificial
leaving and maximize revenues. The project built a model to Intelligence For Internet of Things (AIIoT), 2024.
predict the churn probability with tenure, monthly charge [4] D. Azzam, M. Hamed, N. Kasiem, Y. Eid, and W. Medhat,
“Customer Churn Prediction Using Apriori Algorithm and Ensemble
and service type as their key valuables. With the voting Learning,” Novel Intelligent and Leading Emerging Sciences
algorithm enabled, it was a cinch to overrule the lowest- Conference, 2023.
ranked features and effectively replace them with high- [5] R. Srinivasan, D. Rajeswari, and G. Elangovan, “Customer Churn
ranking recommendations that are likely personalized such Predic9 tion Using Machine Learning Approaches,” 2023
as discounts or better support models. Early identification of International Conference on Artificial Intelligence and Knowledge
Discovery in Concurrent Engineering (ICECONF), 2023.
customers at-risk for churn enables preventative measures to
[6] A. Gaur and R. Dubey, “Predicting Customer Churn in Telecom
be taken, increasing customer satisfaction and loyalty. With Sector Using Various Machine Learning Techniques,” 2018
predictions only being able to improve by considering International Conference on Advanced Computation and
previous model iterations, and churn prediction becomes an Telecommunication (ICACAT), 2018.
indispensable tactic for continued prosperity. A number of [7] L. Ou, “Customer Churn Prediction Based on Interpretable Machine
directions for future work can be investigated in order to Learning Algorithms in Telecom Industry,” 2023 International
Conference on Computer Simulation and Modeling, Information
further improve the system's capabilities and address Security (CSMIS), 2023.
possible areas for improvement. Getting a larger and more [8] P. Pulkundwar, K. Rudani, O. Rane, C. Shah, and S. Virnodkar, “A
varied dataset could be taken into consideration to increase Comparison of Machine Learning Algorithms for Customer Churn
the system's accuracy and generalization. Inclusion of Prediction,” International Conference on Awareness Science and
Technology, 2023.
various features of customer data such as usage patterns,
[9] P. A. Raj, “A Novel Approach to Customer Churn Prediction in
service types and payment histories. Exploring additional Telecom,” 2024 International Conference on Advances in Computing,
customer and organization related features that capture Communication and Applied Informatics (ACCAI), 2024.
interactions between existing features and can lead to [10] B. Prabadevi, R. Shalini, and B. R. Kavitha, “Customer Churning
improved accuracy and performance. By combining the two, Analysis Using Machine Learning Algorithms,” International Journal
such as a recommender system and customer churn of Intelligent Networks, 2023.
prediction, you can provide better tailored strategies for [11] J. Ahn, J. Hwang, D. Kim, H. Choi, and S. Kang, “A Survey on
Churn Analysis in Various Business Domains,” IEEE Access, 2020.
each valuable customer by calculating their individual risk
[12] S. Jinbo, L. Xiu, and L. Wenhuang, “The Application of AdaBoost in
of churning. It uses customer data with usage-preference, Customer Churn Prediction,” 2007 International Conference on
service-type & payment adoption pattern to spot at risk Service Systems and Service Management, 2007.
customers and renders personalized actionable insights. [13] M. Afzal, S. Rahman, D. Singh, and A. Imran, “Cross9Sector
Furthermore, providing personalized suggestions can greatly Applica9 tion of Machine Learning in Telecommunications:
Enhancing Customer Retention Through Comparative Analysis of
increase consumer satisfaction and trust that stakes lead to a Ensemble Methods,” IEEE Access, 2024
higher retention percentage. This targeted approach means [14] G. Gayathri, P. Priyanka, B. Manjeera, K. Niharika, V. P. K. S., and
resources are not wasted and efforts can be focused on high- V. K. K. K., “An Extra Tree Classifier for prediction of End to End
risk customers instead of taking a blanket strategy. Customer Churn and Retention,” 2024 Asia Pacific Conference on
Innovation in Technology (APCIT), 2024.
REFERENCES [15] M. Gewaly, M. Saeed, and M. Ammar, “Machine Learning Approach
[1] A. Aggarwal, D. Jain, P. Garg, and A. Gupta, “Analysis and for Complaint Prediction in the Telecom Industry,” 2024 International
Conference on Intelligent Systems and Applications (IMSA), 2024.
Prediction of Churn and Retention Rate of Customers in Telecom
Industry Using Logistic Regression,” 2024 International Conference [16] K. S. Tonmoy Day Sarkar Md Sayeem Rahman, “Analysis of
on Emerging Innovations and Advanced Computing (INNOCOMP), Customer Churn for Telecom Company with SMOTE9ENN and
2024. Hyperparameter Tuning Randomized9SearchCV Technique in
Advanced Machine Learn9 ing Technology,” IEEE, 2024.
[2] A. Mardi and H. Ghorbani, “Customer Churn Prediction: Leveraging
Data Analysis and Machine Learning Approaches,” 2024 10th

Authorized licensed use limited to: BANGALORE INSTITUTE OF TECHNOLOGY. Downloaded on March 05,2025 at 04:03:18 UTC from IEEE Xplore. Restrictions apply.