2024 Sixth International Conference on Computational Intelligence and Communication Technologies (CCICT)

2024 Sixth International Conference on Computational Intelligence and Communication Technologies (CCICT) | 979-8-3503-7462-9/24/$31.00 ©2024 IEEE | DOI: 10.1109/CCICT62777.2024.00091

Efficacy of Customer Churn Prediction System

1 2 3
Ritik Raj Siddhant Gupta Ritik Mishra
Dept. of Computer Science and Engineering Dept. of Computer Science and Engineering Dept. of Computer Science and Engineering
ABES Engineering College, Ghaziabad ABES Engineering College, Ghaziabad ABES Engineering College, Ghaziabad
ritik.20b0101053@abes.ac.in siddhantgupta145@gmail.com ritik.20b0101133@abes.ac.in

4
Medha Malik
Assistant Professor,
Department of Computer Science & Engineering,
ABES Engineering College, Ghaziabad
medha.malik@abes.ac.in

Abstract— Customer churn, defined as customer defection is a big customers at risk of churn. [6] By leveraging machine learning
problem for businesses in today's competition. This paper takes an algorithms and delving into the nuances of customer behavior,
in-depth look at the application of machine learning (ML) in businesses can transition from reactive to proactive customer
customer churn prediction to identify the best models for various retention strategies. The insights derived from such predictive
industries. quantitative performance metrics of seven commonly models not only enable timely intervention but also foster a [7]
used machine learning algorithms: Logistic Regression, Decision deeper understanding of customer dynamics, ultimately
Trees, Random Forests, Naive Bayes, K-Nearest Neighbors (KNN), contributing to enhanced customer satisfaction and sustainable
AdaBoost, and XGBoost. The theoretical background, application business growth.
context, and performance evaluation of each model are examined,
As we navigate through the intricacies of customer churn
focusing on precision, accuracy, recall, and F1 score. Evaluation of
the performance of the model on different business data is also done prediction, this research aims to explore the efficacy of various
to ensure that the application is real in the world. This comparison machine learning methodologies, shedding light on their
shows the advantages and disadvantages of each model and guides strengths, limitations, and practical implications. By doing so,
providers in their selection based on the specific needs of the business [9] businesses can harness the power of predictive analytics to
and profile features. Additionally, this paper explores future changes fortify customer relationships, reduce churn rates, and cultivate
in predictions and highlights the changing role of machine learning a loyal customer base in an era where customer experience
technologies, especially deep learning, in improving the accuracy of reigns supreme.
predictions and personalization. This review is very useful for
professionals and experts; It provides a better understanding of the
current machine learning model for customer churn prediction and II. RELATED WORK
provides insight into future research and practical applications.
Keywords— Customer churn, Machine learning, Churn prediction, The suggested [1] study presents a thorough method for
Classification models and Neural Networks customer churn prediction (CP) in the telecom sector that
makes use of Swish Recurrent Neural Networks (S-RNN),
I. INTRODUCTION creative Feature Selection (FS) via the Brownian Movement-
based Butterfly Optimization Algorithm (BM-BOA), and
In today's competitive business environment, where customer effective customer clustering using the CLARA algorithm.
loyalty is paramount to sustained success, organizations face the Data collection, preprocessing, state and area-based customer
formidable challenge of mitigating customer churn the categorization, and a retention procedure are all included in the
phenomenon where customers discontinue their association with research methodology.
a service or product provider. [1] Customer churn not only Comparative studies show that the suggested S-RNN performs
represents a loss of revenue but also poses a threat to the long- better than traditional techniques like Artificial Neural
term viability of businesses. Recognizing the importance of Networks (ANN), Convolutional Neural Networks (CNN),
proactive customer retention strategies, organizations Deep Neural Networks (DNN), and Recurrent Neural
increasingly turn to advanced analytics and machine learning Networks (RNN). When compared to other optimization
techniques to predict and preempt customer churn. methods, the suggested BM-BOA shows efficacy in FS.
Understanding the factors influencing [2] customer churn Additionally, the CLARA clustering algorithm exhibits
requires a comprehensive analysis of diverse data sources, efficiency in forming clusters.
ranging from transactional records and customer interactions to Key findings include a high sensitivity rate of 98.27%,
demographic information. [5] As customer relationships become specificity of 92.31%, and accuracy of 95.99% achieved by the
more intricate, we need advanced models to understand the subtle S-RNN. The BM-BOA facilitates the selection of informative
patterns and connections in the data. This has led to the features with minimal iterations, enhancing computational
emergence of customer churn prediction as a crucial domain efficiency. Overall, the proposed approach demonstrates
within customer relationship management (CRM), empowering reliability and robustness in customer churn prediction. This
businesses to anticipate and address customer attrition before it research contributes valuable insights for advancing predictive
impacts their bottom line. This study delves into the realm of analytics in the telecommunications sector. The [2] presented
customer churn prediction, employing a data-driven approach to research investigates churn prediction in the banking sector,
develop accurate and actionable models for identifying specifically employing Decision Trees and Artificial Neural

979-8-3503-7462-9/24/$31.00 ©2024 IEEE 546

DOI 10.1109/CCICT62777.2024.00091
Authorized licensed use limited to: K K Wagh Inst of Engg Education and Research. Downloaded on August 22,2024 at 16:35:49 UTC from IEEE Xplore. Restrictions apply.
 Networks (ANN) as predictive models. The Decision Tree a strong prediction model is necessary. For improved accuracy
methodology involves feature-based branching to create a tree and performance, this can be accomplished by combining SVM
structure, with pruning to eliminate noisy data. On the other with boosting algorithms, this might be taken into
hand, ANN, inspired by biological neural systems, utilizes consideration for future work on Churn prediction.
interconnected neurons to collectively solve problems based on The study [6] uses well-liked machine-learning techniques to
weighted attributes. tackle the current problem of customer attrition prediction in
The tests involved preprocessing by removing unnecessary the telecom sector. The effectiveness of Naïve Bayes, Support
features from a Bank dataset from Kaggle. They were carried Vector Machines, Decision Trees, Random Forest, XGBoost,
out using Python and Scikit-Learn. Training sets made up 70% CatBoost, AdaBoost, and Extra tree classifiers is compared
of the dataset while testing sets made up 30%. Performance with Logistic Regression. With an AUC score of 84%, the
measures like accuracy, precision, recall, and F1-score were experimental findings demonstrate the higher accuracy of two
used to assess the two prediction models. collaborative learning techniques: the XGBoost classifier and
The results show that the ANN model is better than the the Adaboost classifier. In terms of accuracy, precision, F-
Decision Tree model, with an accuracy of 86.52% as opposed measure, recall, and AUC score, these models perform better
to 79.77%. The ANN model also exhibited higher metrics for than others.
precision and recall. Subsequent examination of the data To efficiently handle telecom churn, the research [7] focuses
revealed gender and country-based categories, which provided on utilizing data analysis and interpretation technologies. The
insights into turnover trends within various market segments. study highlights the necessity for advanced prediction
In conclusion, the research emphasizes the importance of techniques as well as the significance of clarity in the drown of
effective churn prediction in the banking industry to retain data. The most important aspect impacting decision-making in
valuable customers. The ANN approach demonstrated superior churn prediction is the length of months since the last offer
performance, suggesting its suitability for practical modification, according to three crucial tools: Permutation
implementation in predicting bank churners. The findings Importance, PDP, and SHAP.
contribute to the existing body of knowledge on predictive Management and marketing can take immediate action by
analytics in the banking sector, providing valuable insights for identifying risk thresholds, especially the warning signal
future research and industry applications. related to the activation of offers on the client's account.
The study [3] examined how well the approaches of logistic Important metrics that need to be periodically monitored to
regression and logit boost performed in a model that was avoid loss include minutes spent outside of the network,
applied to a database of 3333 instances belonging to Orange, invoice value, age, and time spent in the network.
an American telecom firm. With Logistic Regression reaching The research [8] explores the application of ML to predict the
an accuracy of 85.2385% and Logit Boost coming in close customers who hold credit cards and are churning in the
behind with 85.1785%, both strategies proved to be highly banking sector using a dataset with both categorical and
effective. The similarity of the outcomes demonstrated how continuous variables. Three models, named Models 1, 2, and 3,
equally effective the two stand-alone approaches were. were proposed, incorporating various forms of independent
Considering the fundamental limitation of individual variables. Additionally, 5 machine learning models (Bayesian
approaches in completely capturing relevant data for maximum network, CR-Tree, C5 tree, CHAID, and neural network) were
precision, the research proposes the possible advantages of employed. Results indicated the effectiveness of machine
adopting a hybrid model in later attempts. The goal of this learning models in predicting credit card customer churn, with
hybrid strategy is to improve overall model performance by the C5 tree model outperforming others across Models 1-3.
integrating features from several techniques. Key findings highlighted the significance of total revolving
The [4] research addresses the critical issue of customer churn balance, total transaction count, and change in transaction
in the telecommunications sector, emphasizing its inevitability count are the top three important variables for developing churn
and the need for proactive management. The study employs prediction models. The study emphasized that not all
machine learning algorithms, specifically a decision tree categorical variables are necessary, but a single categorical
classifier, and a random forest algorithm, to develop effective variable encompassing all categories can enhance model
customer churn prediction models. Initial challenges with the performance. Feature selection, implemented through models
decision tree classifier on an unbalanced dataset are overcome or variable combinations, was identified as a crucial step in
through techniques such as up-sampling and ENN (Edited improving churn prediction models.
Nearest Neighbors), with the random forest classifier The research underlines the importance of early churn
ultimately demonstrating superior results, achieving an prediction to help banks retain customers. Clustering with
impressive overall accuracy of 99%. KNN and the C5 tree model demonstrated superior
To capture nonlinear complicated interactions in data variables, performance in various metrics, including R2, precision, and
the research also acknowledges the possibility for future recall. The study suggested future analyses to identify optimal
improvements and suggests investigating advanced algorithms independent variables, enhance model robustness, and assess
like deep learning recurrent neural networks. Furthermore, it is the efficiency of different machine learning models,
recommended that future research on customer churn acknowledging that conventional models may not always yield
prediction in the telecommunications industry take into the best results for diverse datasets.
account data certainty and a rough set methodology. A detailed In this study [9], a multilayer perceptron (MLP) is used to
examination of the techniques employed in the customer churn compare the effects of various hyperparameter configurations
prediction process is given in this [5] paper. The accuracy and on deep neural networks (DNNs) for churn prediction in the
forecast of each of the churn prediction models mentioned banking industry. The impacts of different training techniques,
above are low. Therefore, to prevent the issue of client attrition, batch sizes, and activation functions are investigated in three

547

Authorized licensed use limited to: K K Wagh Inst of Engg Education and Research. Downloaded on August 22,2024 at 16:35:49 UTC from IEEE Xplore. Restrictions apply.
 studies. The results show that when a rectifier function is used established within a wider recommendation platform for cloud
in the hidden layers and a sigmoid in the output layer, the DNN service subscription levels. Customers that have a high chance
performs better than the MLP. DNN performance is strongly of leaving are identified by the system, allowing for proactive
influenced by the batch size, and performance starts to decline actions to be taken, such as maybe modifying subscription
as the batch size gets closer to the test set data. By filling in levels. Based on studies with CBAR-related data, the suggested
theoretical gaps in earlier research on hyperparameter tuning in solution utilizes traditional machine learning, particularly the
banking churn prediction, the paper makes a theoretical random forest technique, and shows rapid model training and
contribution to the field. It also provides useful insights for efficient prediction performance. The study notes that there
effective DNN training and enhanced customer relationship isn't a single, globally ideal churn prediction system and makes
management in the banking industry. suggestions for future modifications to the classification
This [10] paper shows that the entropy of financial activity algorithm based on changing data and new methods.
choices and the analysis of dynamic behavioral patterns in This [14] research underscores the importance of customer
temporal and spatial domains can greatly enhance the churn predictions in e-commerce and advocates for leveraging
prediction of customer churn decisions in the banking industry. machine learning in customer relationship management. Using
The study draws attention to the influence of regularity and SVM and LR models on customer behavior data from a B2C e-
diversity traits on prediction performance as well as the commerce enterprise, the study employs the k-means algorithm
variations in predictability among age groups. The findings for customer segmentation, classifying them into three types
point to the possibility of leveraging extensive financial for predictions. Evaluation metrics include accuracy, recall,
transaction data to model churn behavior. The paper highlights precision, and AUC. The study demonstrates the effectiveness
the complementarity of the suggested qualities with of customer segmentation and the superior accuracy of SVM
conventional churn prediction techniques, while also over LR in predicting customer behavior. Limitations include
acknowledging the need for additional research to prove causal data selection and segmentation methods, urging verification
linkages. Since the suggested spatio-temporal qualities are across multiple datasets and exploration of diverse
thought to apply to different commercial fields, further research segmentation techniques. Future research recommendations
may combine them with more sophisticated traits for involve collecting and comparing customer behavioral data
improvement. The study acknowledges limitations, such as from multiple companies to enhance model generalizability
potential sampling bias and the specific focus on credit card and continuous churn prediction with a focus on customer
holders. value segmentation.
This [11] article focuses on evaluating supervised classifiers The importance of customer churn predictions in e-commerce
for predicting customer churn in a large Brazilian bank, and the application of machine learning for efficient customer
contributing to the existing literature in several ways. The study relationship management are emphasized in this [15]. The
employs a horserace of classification algorithms, with random study uses k-means clustering for customer segmentation and
forests demonstrating superior performance. The research SVM and LR models to forecast consumer behavior in a B2C
emphasizes the temporal nature of churn prediction, e-commerce business. The investigation assesses the two
highlighting the challenging task of data analysis. The models' accuracy, recall, precision, and AUC, demonstrating
identified attributes with high predictive power include the SVM model's superior predictive performance over LR.
customer service frequency, credit volume, and product The findings underscore the significance of client segmentation
ownership, suggesting a focus on relationship strengthening and the possibility of enhancing customer relationship
through product sales for retention. management in business-to-consumer e-commerce businesses.
The random forests model, achieving an 80.2% recall rate, The study does, however, recognize certain shortcomings in the
proves valuable in identifying potential churn, and the study segmentation and data selection processes. Prospective
estimates the model's potential to forecast annual losses of up avenues for investigation encompass broadening the scope of
to R$ 2.12 billion at the customer level. The practical data sources to enhance generalizability, predicting ongoing
importance of the model is underscored by its ability to churn, and incorporating supplementary variables into
preserve revenue, with a conservative projection indicating segmentation and prediction models.
potential savings of approximately R$ 290 million through
targeted retention efforts. The study recognizes the variation in III. PROPOSED SYSTEM
customer margins and highlights the efficiency of a churn 1. Data preparation: The first step is to prepare the data for
prediction model in optimizing cost/benefit outcomes, training the machine learning and deep learning models. This
especially when tailored to individual customer profiles. involves cleaning the data, removing any outliers, and
This [12] focuses on predicting the churn of young customers converting the data into a format that is compatible with the
in the competitive banking sector, emphasizing the significance models. The dataset used in this work, called "Telco Customer
of customer retention. Employing the Extra TreeClassifier, the Churn," was obtained from Kaggle and has 21 columns and 7043
model achieves a satisfactory AUC of 91.88%. Tree-based rows. With columns listing different client attributes, each row
classifiers and ensembles outperform distribution-based represents a single customer. Particularly, it contains data on
classifiers, indicating their suitability for the dataset. The study subscribed services like internet and phone features, account
acknowledges limitations in sample size and geographical specifics like tenure and charges, demographic information like
context, recommending future research to extend the study to gender and age range, and customer churn status. Except for the
comparable geographies with a larger sample size for enhanced goal variable "Churn" and the non-informative "customerID,"
validity and generalizability. the dataset has 19 useful columns. While numerical variables
To help Customer Service Providers (CSPs) with customer like "tenure" and "Monthly charges" offer quantitative insights,
retention, this [13] paper proposes a churn prediction system categorical features like "Partner" and "Dependents" have two or

548

Authorized licensed use limited to: K K Wagh Inst of Engg Education and Research. Downloaded on August 22,2024 at 16:35:49 UTC from IEEE Xplore. Restrictions apply.
 more values. With four or fewer distinct values, most categorical about the mean and standard deviation of every characteristic for
features offer an in-depth knowledge of the characteristics and every class in the dataset during training. To make a prediction,
actions of their customers. it first uses the Gaussian probability density function to
The dataset can be accessed through the following link: determine the likelihood that a given observation would belong
https://www.kaggle.com/datasets/blastchar/telco-customer- to each class. The class with the highest probability is then
churn chosen to be the predicted class.
2. Feature engineering: Once the data is prepared, it is 3.5 K-nearest Neighbor:
important to engineer relevant features that can be used by the Using the simple and natural K-nearest neighbor (KNN)
models to predict churn. This may involve creating new features classification technique, the class of a new data point is
from existing features or combining multiple features into a determined by the majority class among its nearest neighbors in
single feature. the feature space. The approach considers the classes of a new
3. Machine learning model selection and training: A variety data point's closest neighbors in the feature space when
of ML algorithms can be used for customer churn prediction. classifying it using the K-nearest neighbor (KNN) algorithm.
Some popular choices include random forests, logistic The K data points that are closest to the new data point are then
regression, and support vector machines. The best model chosen after calculating the distance between it and the
depends on the specific dataset and the desired performance preceding data point in the training set. The class of the new data
metrics. Once a machine learning algorithm is selected, it needs point is then determined by the majority class among these K
to be trained with the help of the data sets. neighbors, where each neighbor takes an equal part in the
3.1 Logistic Regression: decision-making process.
A statistical technique called logistic regression is used to 3.6 Adaboost Algorithm:
forecast the likelihood of a binary result, such as win/lose, A strong method for ensemble learning is also called Adaptive
pass/fail, or yes/no, based on one or more predictor factors. In Boosting which combines different weak learners to generate a
machine learning, it is frequently applied to classification tasks strong classifier. It functions by increasing the weight assigned
involving categorical dependent variables. In contrast to linear to incorrectly categorized data points throughout each iteration,
regression, logistic regression uses the logistic function to enabling ensuing weak learners to concentrate more on the
estimate the likelihood that an observation falls into a specific challenging cases.
category while ensuring that the predicted probabilities are The AdaBoost approach begins by using the whole dataset to
confined between 0 and 1. By modifying the linear equation's train a weak learner. The data points that the first weak learner
coefficients to reduce the discrepancy between the training data's misclassified is subsequently given greater weights, and a
actual outcomes and the expected probabilities, logistic second weak learner is trained using the updated weights on the
regression learns the relationship between the predictor variables same dataset. Iteratively, this process is carried out, with each
and the binary outcome. weak learner focusing more on the earlier incorrect classified
3.2 Decision Tree Classifier: data points. In the end, a single strong classifier is created by
A machine learning technique that is utilized for both combining all the weak learners by evaluating their predictions
classification and regression problems is the decision tree according to their performance during training. The final
classifier. For it to operate properly, the area of features is classifier is thus capable of making accurate predictions even on
recursively divided into smaller regions according to the values complex datasets by leveraging the collective wisdom of
of predictor variables. At every stage, the most informative multiple weak learners.
attribute is selected. This process creates a tree-like structure 3.7 XGBoost algorithm:
where the branches represent decisions based on feature values, Extreme Gradient Boosting, or XGBoost, is a strong gradient-
and the leaves correspond to the predicted outcome. boosting method that excels at addressing supervised learning
3.3 Random Forest Classifier: problems with speed, accuracy, and efficiency. To fix mistakes
A Random Forest Classifier is a clever way of merging many produced by the earlier models, XGBoost adds weak learners—
decision trees to make predictions. Each tree is trained on a usually decision trees—to an ensemble one after the other in a
random subset of the data and features, and then they all vote on sequential fashion. It reduces the errors between the expected
the final prediction together. This method helps reduce and actual values during training by applying gradient descent to
overfitting and makes the model more accurate and reliable. optimize a differentiable loss function.
During training, each tree makes decisions by splitting the data
into smaller groups based on different features, aiming to create 4. Evaluating & testing the model results:
groups that are as pure as possible. Then, when it's time to make Evaluating customer churn prediction models involves key
a prediction, each tree gives its opinion, and the most popular metrics like accuracy, precision, recall, and F1-score. The
prediction among all the trees is chosen as the final answer. This dataset is divided into training and testing sets, with cross-
approach makes Random Forests very effective for tasks like validation ensuring robust assessments. The confusion matrix
classification, where you want to predict categories based on provides insights into true positives, true negatives, false
input data. positives, and false negatives. ROC and Precision-Recall curves
3.4 Gaussian Naive Bayes Classifier: visually represent model performance, aiding in optimal
Gaussian Naive With an understanding of feature independence, threshold selection. Continuous refinement and validation of
the Bayes classification technique is based on the Bayes theorem diverse datasets are crucial for model reliability. Beyond
and is best suited for continuous data where the likelihood of the standard metrics, considering business-specific implications of
features is considered to be Gaussian (normal distribution). prediction errors enhances strategic alignment. Thorough
Every feature in Gaussian Naive Bayes is thought to have a evaluation ensures confidence in the model's predictive
Gaussian (normal) distribution. The method gains knowledge capabilities, guiding effective customer retention strategies.

549

Authorized licensed use limited to: K K Wagh Inst of Engg Education and Research. Downloaded on August 22,2024 at 16:35:49 UTC from IEEE Xplore. Restrictions apply.
 4.1 Accuracy: divided by the number of true positive predictions, or accurately
When discussing a model, accuracy refers to the percentage of predicted occurrences of the positive class.
correctly classified examples relative to the total number of
occurrences. It's an indicator of how accurately the model
forecasts the results of the training or testing set of data. When
it comes to classification tasks, accuracy is determined by
dividing the total number of predictions the model made by the
number of correct predictions. Its ability to offer a clear-cut and
understandable performance indicator makes it one of the most
widely utilized evaluation metrics for classification algorithms.

4.2 Precision:
In the context of a model, precision refers to the degree of
accuracy of the positive predictions. The ratio of accurately
predicted positive instances, or true positives, to the overall
number of positive predictions, or both true positives and false Figure 1: Recall metric
positives is used to calculate it.
5. Deploy the model:
4.3 f1- score: Deploying a customer churn prediction model involves
One metric used for evaluating a classification model's integrating it into operational systems for real-time use. After
performance, particularly when working with unbalanced rigorous testing and validation, the model is deployed to make
datasets, is the F1 score. It is a balanced metric that takes into predictions on new data. This deployment phase requires
account both false positives and false negatives. It serves as a collaboration between data scientists and IT teams to ensure
harmonic tool for precision and recall. seamless integration with existing infrastructure. Continuous
monitoring and updates are essential to maintain model accuracy
4.4 recall: and relevance, allowing businesses to proactively identify
Recall is a metric used to identify the accuracy of the potential churners and implement targeted retention strategies
classification models and identify all relevant occurrences of a for enhanced customer relationship management. Figure 2
given class. It is also called as sensitivity or true positive rate. It depicts the general model.
is the proportion of all actual positive occurrences in the dataset

Figure 2. General Model for Churn Prediction

IV. REVIEW PROCESS

Question: How is the customer churn prediction in the collection and preprocessing to customer grouping based on
telecommunications industry, and what were the key findings? states and areas, feature engineering (FE), and finally, the
Answer: The research in the telecommunications industry took retention process.
a multifaceted approach to customer churn prediction. It The key findings of the research were compelling. Comparative
introduced a comprehensive methodology incorporating Swish analyses showcased the superior performance of the S-RNN
Recurrent Neural Networks (S-RNN), innovative Feature model compared to traditional methods such as Recurrent
Selection (FS) using the Brownian Movement-based Butterfly Neural Networks (RNN), Convolutional Neural Networks
Optimization Algorithm (BM-BOA), and efficient customer (CNN), Deep Neural Networks (DNN), and Artificial Neural
clustering through the CLARA algorithm. This methodology Networks (ANN). Notably, the S-RNN achieved an impressive
spanned the entire predictive analytics pipeline, from data sensitivity rate of 98.27%, a specificity of 92.31%, and an

550

Authorized licensed use limited to: K K Wagh Inst of Engg Education and Research. Downloaded on August 22,2024 at 16:35:49 UTC from IEEE Xplore. Restrictions apply.
 overall accuracy of 95.99%. These metrics indicated the churners before disengagement.
effectiveness of the proposed approach in identifying potential

Figure 3. Churn Impact on Revenue

Question: Churn prediction in the banking sector using Question: What is the critical issue of customer churn in the
Decision Trees and Artificial Neural Networks (ANN), what telecommunications sector, and what were the proposed future
were the key models employed, and how did they perform? enhancements?
Answer: Artificial Neural Networks (ANN) and Decision Answer: The research on customer churn in the
Trees were used as prediction models in banking research. telecommunications sector employed machine learning
Feature-based branching was used in the Decision Tree algorithms, specifically a decision tree classifier and a random
methodology to create a tree structure, and pruning was done to forest algorithm. Initial challenges with the decision tree
remove noisy data. However, the Artificial Neural Network classifier on an unbalanced dataset were overcome through
(ANN), inspired by biological neural systems, used a network techniques such as up-sampling and Edited Nearest Neighbors
of interconnected neurons to answer problems as a group using (ENN). The random forest classifier ultimately demonstrated
weighted qualities. superior results, achieving an impressive overall accuracy of
An analysis of experimental findings on a Kaggle Bank dataset 92%. The research acknowledged the potential for future
using Scikit-Learn and Python showed that the ANN model enhancements, proposing the exploration of more advanced
performed better than the Decision Tree model. The ANN algorithms such as deep learning recurrent neural networks to
exceeded the Decision Tree with an accuracy of 86.52% as capture non-linear complex relationships in data variables.
opposed to 79.77%. The ANN model also exhibited higher
metrics for precision and recall. Subsequent examination V. LIMITATIONS
divided the data according to gender and nation, offering Customer churn prediction, while a valuable and widely
insightful information about churn trends within several market employed analytical tool for businesses, is not without its
sectors. limitations. Here are some key limitations associated with
Question: How do the Logistic Regression and Logit Boost customer churn prediction:
techniques in a model applied to an American 1. Data Quality and Availability:
telecommunication company's database perform, and what Limited Historical Data: Churn prediction models heavily rely
future enhancements were suggested? on historical customer data. Limited historical data may hinder
Answer: The main objective was to evaluate the effectiveness the accuracy and effectiveness of predictive models, especially
of the Logistic Regression and Logit Boost algorithms when in industries where customer interactions are infrequent.
utilized on a database of 3,333 instances from an American Data Quality Issues: Inaccurate or incomplete data can
telecom business. With Logistic Regression reaching an adversely impact the performance of churn prediction models.
accuracy of 85.2385% and Logit Boost coming in close behind Issues such as missing values, outliers, and inaccuracies in
with 85.1785%, both strategies proved to be highly effective. customer records can introduce noise and compromise the
The similarity of the outcomes demonstrated how equally reliability of predictions.
effective the two independent approaches present. 2. Dynamic Nature of Behavior: Customer behavior is
However acknowledging the fundamental shortcomings of solo dynamic and can change rapidly due to various factors,
methods in fully capturing pertinent data for maximum including external economic conditions, market trends, or
accuracy, the study proposed the advantages of using a hybrid competitive offerings. Predictive models might struggle to
model in subsequent research. The goal of this hybrid strategy adapt quickly to these changes, leading to less accurate
is to improve overall model performance by integrating features predictions.
from several techniques. Future work will focus on bigger, real- 3. Assumption of Stationarity:
world databases to evaluate the model's effectiveness on a more Assumption of Stable Patterns: Churn prediction models often
substantial scale. assume that the patterns observed in historical data will persist

551

Authorized licensed use limited to: K K Wagh Inst of Engg Education and Research. Downloaded on August 22,2024 at 16:35:49 UTC from IEEE Xplore. Restrictions apply.
 into the future. However, customer preferences and market 1. Evolving Machine Learning Techniques and Model
dynamics may evolve, challenging the assumption of Refinement:
stationarity and reducing the predictive power of historical
patterns. x Deep learning: Techniques like Recurrent Neural Networks
(RNNs), specifically the Long Short-Term Memory
4. Limited Causation Understanding: (LSTM) networks, can capture time-based dependencies in
Correlation vs. Causation: Churn prediction models typically customer data, leading to more accurate predictions.
identify correlations between customer attributes and churn x Convolutional Neural Networks (CNNs) can effectively
likelihood. However, they may fall short of providing a deep analyze image and text data like social media posts or
understanding of causation, making it challenging for website clickstream data, improving prediction accuracy.
businesses to implement effective and targeted retention x Ensemble methods like Gradient Boosting Machines and
strategies based on the root causes of churn. Random Forests can further enhance model performance.
5. Human Factors and Unpredictability:
x Explainable AI (XAI) techniques like LIME and SHAP can
Subjectivity and Individual Choices: Predicting human
provide deeper insights into model decisions, enabling
behavior, including customer decisions to churn, is inherently
better understanding and refinement.
challenging due to the subjective and individual nature of
decision-making. Unforeseen personal circumstances or
emotional factors may play a significant role, making it difficult 2. Real-Time Segmentation and Intervention Strategies:
to account for all variables accurately.
x Dynamic segmentation uses real-time data and continuous
VI. CONCLUSION analysis to create customer segments that adapt to their
To allow organizations to implement active retaining measures, evolving behavior, allowing for targeted and personalized
this paper examined the effectiveness of machine learning in interventions.
predicting customer attrition. x Real-time intervention strategies leverage early
According to our investigation, several machine learning identification of at-risk customers through real-time churn
techniques, such as Support Vector Machines, XGBoost, prediction. This enables proactive measures like
Random Forests, and Logistic Regression, identify customers personalized offers, loyalty programs, or targeted customer
who are likely to churn with high accuracy. These models use support, increasing customer retention.
past customer data to identify trends and patterns that indicate
probable churn. Businesses can create targeted measures to re- 3. Challenges and Considerations:
engage these consumers and reduce turnover rates by knowing Implementing these advancements requires acknowledging the
these patterns. associated challenges:
The examination of case studies from the banking, American
telecom, and telecommunications industries brought even more
x Computational resource requirements: Deep learning
attention to the useful uses of machine learning for churn
models can be applied to some businesses.
prediction. These investigations demonstrated the efficacy of
several algorithms, such as Decision Trees, Artificial Neural x Expertise and talent: Handling complex models necessitates
Networks (ANNs), and Swish Recurrent Neural Networks (S- specialized expertise, which can be scarce.
RNN), in achieving high accuracy. Customer churn predicted x Ethical decision-making: It is important to address biases in
does have certain drawbacks, though. Important challenges still data and processes to ensure the implementation of AI-
include data quality, the dynamic nature of customer behavior, driven solutions.
and the difficulties in establishing causality from correlations.
Accurate predictions may be possible with the development of 4. Examples:
deep learning methods like convolutional neural networks
(CNNs) and recurrent neural networks (RNNs). Furthermore, x Vendor X implemented an RNN-based predictive model to
organizations may personalize their approach and increase identify at-risk customers based on purchase history and
client retention with real-time segmentation and management look at character. This allows them to launch promotional
strategies. Hence, we conclude that machine learning-based campaigns and reduce customer churn.
customer churn prediction is an effective tool that enables x Telecom company Y uses CNN-based models to analyze
companies to understand consumer behavior and develop customer interactions, identify potential customers, and
proactive strategies to retain customers. The future of customer improve customer experience based on sentiment analysis.
churn prediction has great opportunities for enhanced customer This increases customer satisfaction and reduces customer
connections and long-term business success as technology churn.
advances and data analytics become more complex. VIII. ACKNOWLEDGMENT
I want to express my sincere gratitude to all who contributed
VII. FUTURE SCOPE to finishing this work on customer churn prediction. First
The future scope of customer churn prediction holds promising and foremost, I extend my deepest appreciation to my
avenues for innovation and improvement. As technology supervisor Ms. Medha Malik, for all her assistance and
evolves and businesses continue to leverage data analytics, encouragement during this research. Their knowledge and
several potential areas of growth and development emerge: support have been crucial in determining the course of our
study. I am also thankful to the ABES Engineering College
faculty members, whose knowledge and support have

552

Authorized licensed use limited to: K K Wagh Inst of Engg Education and Research. Downloaded on August 22,2024 at 16:35:49 UTC from IEEE Xplore. Restrictions apply.
 enhanced my academic career and given me the tools I Prediction in the Banking Sector. Computation 2021, 9, 34.
needed to carry out this project. https://doi.org/10.3390/computation9030034

REFERENCES [10] Kaya, E., Dong, X., Suhara, Y. et al. Behavioral attributes
and financial churn prediction. EPJ Data Sci. 7, 41 (2018).
[1] R. Sudharsan & E. N. Ganesh (2022) A Swish RNN based https://doi.org/10.1140/epjds/s13688-018-0165-5
customer churn prediction for the telecom industry with a
novel feature selection strategy, Connection Science, 34:1, [11]de Lima Lemos, R.A., Silva, T.C. & Tabak, B.M.
1855-1876, DOI: 10.1080/09540091.2022.2083584 Propension to customer churn in a financial institution: a
machine learning approach. Neural Comput & Applic 34,
[2] G. Ravi Kumar, K. Tirupathaiah, B. Krishna Reddy, 11751–11768 (2022). https://doi.org/10.1007/s00521-022-
“Client Churn Prediction of Banking and Fund Industry 07067-x
utilizing Machine Learning Techniques,” International Journal
of Computer Sciences and Engineering, Vol.7, Issue.6, pp.842- [12] Bharathi S, V.; Pramod, D.; Raman, R. An Ensemble
846, 2019. Model for Predicting Retail Banking Churn in the Youth
Segment of Customers. Data 2022, 7, 61.
[3] Hemlata Jain, Ajay Khunteta, Sumit Srivastava Churn https://doi.org/10.3390/data7050061
Prediction in Telecommunication using Logistic Regression
and Logit Boost, Procedia Computer Science, Volume 167, [13] Saias, J.; Rato, L.; Gonçalves, T. An Approach to Churn
2020, Pages 101-112, ISSN 1877-0509, Prediction for Cloud Services Recommendation and User
https://doi.org/10.1016/j.procs.2020.03.187. Retention. Information 2022, 13, 227.
https://doi.org/10.3390/info13050227
[4] Sharmila K. Wagh, Aishwarya A. Andhale, Kishor S.
Wagh, Jayshree R. Pansare, Sarita P. Ambadekar, S.H. [14] Xiahou, X.; Harada, Y. B2C E-Commerce Customer
Gawande, Customer churn prediction in telecom sector using Churn Prediction Based on K-Means and SVM. J. Theor. Appl.
machine learning techniques, Results in Control and Electron. Commer. Res. 2022, 17, 458-475.
Optimization, Volume 14, 2024, 100342, ISSN 2666-7207, https://doi.org/10.3390/jtaer17020024
https://doi.org/10.1016/j.rico.2023.100342.
[15] Machine Learning to Develop Credit Card Customer
[5] Saran Kumar, A., and D. Chandrakala. "A survey on Churn Prediction Najjar, N Al-Rousan, H AL-Najjar - Journal
customer churn prediction using machine learning of Theoretical and …, 2022
techniques." International Journal of Computer Applications
975 (2016): 8887. [16] K. Goyal, K. Kanishka, K. Vasisth, S. Kansal and R.
Srivastava, "Telecom Customer Churn Prediction: A Survey,"
[6] Lalwani, Praveen, et al. "Customer churn prediction 2021 3rd International Conference on Advances in Computing
system: a machine learning approach." Computing (2022): 1- Communication Control and Networking (ICAC3N), 2021, pp.
24. 276-280, doi: 10.1109/ICAC3N53548.2021.9725621.

[7] Dumitrache, Andreea, Alexandra AM Nastu, and Stelian [17] S. De, P. P and J. Paulose, "Effective ML Techniques to
Stancu. "Churn prediction in telecommunication industry: Predict Customer Churn," 2021 Third International Conference
Model interpretability." Journal of Eastern Europe Research in on Inventive Research in Computing Applications (ICIRCA),
Business and Economics 2020 (2020). 2021, pp. 895-902, doi: 10.1109/ICIRCA51532.2021.9544785.

[8]AL-Najjar, Dana, Nadia Al-Rousan, and Hazem AL-Najjar. [18] V. Geetha, A. Punitha, A. Nandhini, T. Nandhini, S.
"Machine learning to develop credit card customer churn Shakila and R. Sushmitha, "Customer Churn Prediction In
prediction." Journal of Theoretical and Applied Electronic Telecommunication Industry Using Random Forest Classifier,"
Commerce Research 17.4 (2022): 1529-1542 2020 International Conference on System, Computation,
. Automation and Networking (ICSCAN), 2020, pp. 1-5,
[9] Domingos, E.; Ojeme, B.; Daramola, O. Experimental doi:10.1109/ICSCAN49426.2020.9262288
Analysis of Hyperparameters for Deep Learning-Based Churn

553

Authorized licensed use limited to: K K Wagh Inst of Engg Education and Research. Downloaded on August 22,2024 at 16:35:49 UTC from IEEE Xplore. Restrictions apply.