International Journal on Recent and Innovation Trends in Computing and Communication

ISSN: 2321-8169 Volume: x Issue: y

DOI: (do not edit)
Article Received: (do not edit) Revised: (do not edit) Accepted: (do not edit) Publication: (do not edit)
____________________________________________________________________________________________________________________

A Survey and Implementation of Machine Learning

Algorithms for Customer Churn Prediction
Dr. Snehal Rathi 1, Atharva Puranik 2 , Vaishnavi Pophale 3, Prajwal Kutwal 4 , Vibhav Kulkarni 5, Shaantanu
Pratham6, Prof. Vikas Maral 7
1Department of Computer Engineering

Vishwakarma Institute of Information Technology, Pune, India

[email protected]
2Department of Computer Engineering

Vishwakarma Institute of Information Technology, Pune, India

[email protected]
3Department of Computer Engineering

Vishwakarma Institute of Information Technology, Pune, India

[email protected]
4Department of Computer Engineering

Vishwakarma Institute of Information Technology, Pune, India

[email protected]
5Department of Computer Engineering

Vishwakarma Institute of Information Technology, Pune, India

[email protected]
6Department of Computer Engineering

Vishwakarma Institute of Information Technology, Pune, India

[email protected]
7Department of Computer Engineering

Vishwakarma Institute of Information Technology, Pune, India

[email protected]

Abstract— Estimating customer traffic is an important task for businesses because it helps them identify customers who are most likely to
leave and take preventative measures to retain them by improving customer satisfaction and further increasing their own reven ue. In this article,
we focus on developing a machine-learning model for predicting customer churn using historical customer data We performed engineering
operations on the data, addressed the missing digits, coded the categorical variables, and preprocessed the data before evalu ating it using a
variety of performance indicators, including accuracy, precision, recall, f1 score, and ROC AUC_Score. Our feature significance analysis
revealed that monthly fees, customer tenure, contract type, and payment method are the factors that have the most impact on f orecasting
customer churn. Finally, we conclude the best-performing model, the Soft Voting Classifier, implemented on the four best-performing
classifiers with a good accuracy of 0.78 and a relatively better ROC AUC_Score of 0.82. Keywords — Customer churn prediction, Machine
learning, Feature importance analysis, Gradient boosting, Business revenue.

I. INTRODUCTION necessary to determine the most appropriate data for the type of
analysis performed. Different datasets provide better metrics for
Customer churn prediction is a critical problem for companies different problems and services. [11] In order to find trends and
across various industries. It refers to the task of identifying signs that assist in identifying high-risk customers, the
customers who are likely to discontinue using a company's organization employs sophisticated analytics as well as machine
products or services. From a business perspective, customer learning approaches.
churn poses significant challenges and can have a substantial
impact on a company's profitability and growth. Retaining Some popular betting methods include:
existing customers is the most important task for the survival of 1. Machine-Learning Algorithms: Gradient boosting, decision
the business, which has become common sense in the business trees, random forests, logistic regression, etc. The likelihood of
world. [15]. Customer acquisition requires substantial additional incoming losses can frequently be predicted by
marketing and promotional efforts while retaining loyal monitoring machine learning systems as such.
customers can lead to repeat business and increased customer 2. Survival Analysis: Survival analysis is a statistical technique
lifetime value. Therefore, accurately predicting customer churn used to analyze event-time data, such as the time before losing
allows companies to proactively address the underlying issues a customer. It considers the varying time periods during which
and take appropriate measures to retain valuable customers. customers remain active and enables the prediction of the
Customer churn prediction relies on analysing historical probability of churn over time.
customer data, such as demographic information, transactional 3. Neural Networks: Deep learning techniques, specifically
records, service usage patterns, and customer interactions. neural networks that can learn complex patterns and
Different data types have different analysis capabilities. It is
1
IJRITCC | Month 20__, Available @ http://www.ijritcc.org
International Journal on Recent and Innovation Trends in Computing and Communication
ISSN: 2321-8169 Volume: x Issue: y
DOI: (do not edit)
Article Received: (do not edit) Revised: (do not edit) Accepted: (do not edit) Publication: (do not edit)
____________________________________________________________________________________________________________________
relationships from large volumes of data, can be used for Tamaddoni et al. [4] found that both the simple predictive
customer churn prediction.[16],[17],[18] model and the cost-sensitive model were better than the
4. Ensemble Methods: The integrated system combines comparison of both the CART model and the multi-model
multiple models using techniques such as packaging, algorithm, and the cost-sensitive learning model obtained the
promotion, and configuration to increase the accuracy of loss model base only in the CART model but not in many models.
estimation. Praveen et al.[5] employed Logistic regression, Support Vector
Machine, Pruning Tree, and Naive Bayes in their comparative
By applying these techniques, companies can gain valuable study of machine-learning techniques for forecasting attrition of
insights into customer behaviour, identify early warning signs customers and looked into how magnification affects accurate
of churn, and implement proactive strategies to mitigate churn classification.
risk. Lalwani et al.[5] found that combining special selection
techniques such as randomization can improve classification
accuracy.
II. MOTIVATION For predicting customer attrition, Horia Beleiu et al.[6] used
With the help of historical customer data, our research aims to three types of machine learning techniques: deep neural
give a thorough and methodical assessment of numerous networks, support vector models, and Bayesian networks.
machine learning models in order to predict customer attrition. Principal component analysis (PCA) was considered
Although there have been previous research works throughout the feature selection process, which reduced
[1],[3],[5],[6] that involved a comparison of the performances residual data. They used optimization techniques to improve the
of a few machine learning models, we aimed at comparing as feature selection process and thereby improved accurate
many models as possible that one can use for classification. We population classification.
applied a wide array of traditional(logistic regression, ridge, The authors of K Coussement et al. [7] used support vector
decision tree, naïve bayes, knn, etc) and ensemble(catboost, machines, logistic regression (LR), and random forests (RF) to
adaboost, xgboost, LGBM, bagging, etc) machine-learning attempt to model the churn prediction problem. SVM initially
models. We wanted to truly explore the potential of Ensemble performed about as well as LR and RF, but when the best
Learning and we did so by employing a Voting Classifier which parameters were chosen, it surpassed LR and RF with regard to
considered the predictions of the best-performing classifiers of PCC and AUC.
our research and therefore provided better results. We were also The decision tree and logistic regression model were used in the
interested in gaining important insights from the dataset by churn prediction data set by K. Dahiya et al. [8]. WEKA tool
thorough visualization of every numerical and categorical was employed during the trial.
column (via kdeplot, boxplot, and histograms) as well as getting Authors Umman et al. [9] used decision tree and logistic
a quantitative value for each feature signifying its importance regression machine learning models to analyze a large data set,
w.r.t the target column (via the chi-square test). Also, we but the accuracy of the results was poor. And proposed that
addressed the issue of an imbalanced target column using development was, therefore, necessary before using additional
SMOTE. We laid greater emphasis on addressing the above machine learning and feature selection techniques.
specific aspects which drove us to do this research. We believe J. Hadden et al. [10] showed that decision trees are better than
this paper will surely provide valuable insights and other systems because of their rules. By using the existing
methodologies that can be applied in real-world scenarios as feature selection strategy, the acquisition accuracy can be
well as prove to be helpful for future researchers. further improved.
J. Hadden et al. [11] reviewed all machine learning models
considered and provided a thorough study of the methods
III. LITERATURE SURVEY currently used for feature selection. They discovered that
This is a concise outline of related research that has been decision trees outperformed the competition in the prediction
proposed by notable scholars as well as churn prediction in the models. The improvement of the prediction algorithms in
telecom business. feature selection is greatly aided by optimization techniques.
On the dataset, Dhangar et al. [1] discovered that SVM and According to Y. Huang et al. [12], the authors used a variety of
Random Forest had the greatest accuracy rates, at 84 and 87 classifiers on the churn prediction dataset, and the findings
percent, respectively. SVM classifiers surpass others with an showed that random forest outperforms the competition with
AUC score of 92.1 percent, while Random Forest earns the regard to AUC and AUC(PR) analysis. However, it is possible
highest AUC score of 94.5 percent. Working on the Customer to raise accuracy even more by employing feature extraction
DNA website, Saad et al [2] highlighted the usage of the re- optimization approaches.
sampling strategy to address the issue of class imbalance. Their Genetic programming (GP) and the Adaboost machine learning
findings show that decision trees are the most accurate model were combined by researchers working under the
classification system when it comes to detecting losses for data direction of A. Idris et al. [13] in order to compare their results
analysis. Authors Adbelrahim et al. [3] Predict users using with those of other classification algorithms. Adaboost and GP's
decision trees, random forests, GBM tree algorithms, and results were more accurate than those of the competition.
XGBoost as tree-based algorithms. Comparative research However, accuracy can be increased even more by utilizing
shows that XGBoost outperforms its competitors in terms of various optimization strategies, like the gravitational search
AUC accuracy. However, feature selection techniques and algorithm, bio-geography-based optimization, and many others.
optimization algorithms can improve accuracy.
2
IJRITCC | Month 20__, Available @ http://www.ijritcc.org
International Journal on Recent and Innovation Trends in Computing and Communication
ISSN: 2321-8169 Volume: x Issue: y
DOI: (do not edit)
Article Received: (do not edit) Revised: (do not edit) Accepted: (do not edit) Publication: (do not edit)
____________________________________________________________________________________________________________________
Authors P. Kisioglu et al. [14] used Bayesian Belief Networks 2. The clients with higher monthly charges are also more likely
(BBN) to estimate client attrition. Correlation analysis and to churn.
multicollinearity tests were carried out during the experimental 3. Both tenure and monthly charges are likely to be important
analysis. BBN was proven to be a viable alternative for the features in predicting churn, as they show significant variations
prediction of churn. They offered suggestions for future study between churned and non-churned customers.
directions as well.
Furthermore, boxplots were generated for the same three
IV. METHODOLOGY numerical columns in Figure 4.3
4.1 SYSTEM ARCHITECTURE:

Fig. 4.1: System Architecture

Description of the system architecture:

4.2.1 Stage 1 - Data Cleaning: The initial stage of our research
Fig 4.3: Boxplot of the numerical columns
involves data cleaning to ensure the integrity and quality of the
Telecom Churn Dataset. Through manual inspection,
The boxplot analysis revealed additional conclusions:
inconsistencies in columns like ‘Total Charges’ were identified 1. The Monthly plan doesn't seem to be such a big driver of
and rectified by eliminating missing values. churn, but we can see that 75% of churners pay between 60 and
100 dollars a month
4.2.2. Stage 2 - Exploratory Data Analysis (EDA): In the 2. One important mark we can see in this plot is that 50% of
second stage of our research, we conducted Exploratory Data Churners leave the company before the first year goes by.
Analysis (EDA) with a strong emphasis on data visualization
techniques. EDA involves a systematic examination of the 4.2.2.2 Visualization of Categorical Columns:
dataset to uncover patterns, relationships, and trends in the data. Count plots were generated for each categorical column to
4.2.2.1. Visualization of Numerical Columns:
understand the churn rate with respect to the values within each
Probability density distribution (or Kernel Density Estimation)
category in Figure 4.4
plots were generated for the three numerical columns: tenure,
monthly charges, and total charges.

Fig. 4.2: Kernel Density Estimation (KDE) plot of the

numerical columns Fig. 4.4: Count plot of ‘Churn’ w.r.t various categories
The conclusions derived from these plots are as follows:
1. The recent clients, indicated by shorter tenure, are more likely From these count plots, several conclusions can be drawn:
to churn. 1. The analysis of the "gender" column revealed that the churn
3
IJRITCC | Month 20__, Available @ http://www.ijritcc.org
International Journal on Recent and Innovation Trends in Computing and Communication
ISSN: 2321-8169 Volume: x Issue: y
DOI: (do not edit)
Article Received: (do not edit) Revised: (do not edit) Accepted: (do not edit) Publication: (do not edit)
____________________________________________________________________________________________________________________
percentage is relatively equal for both males and females. Fig. 4.5: F-value and P-value of every column obtained via
2. The percentage of churn is higher in the case of senior citizens. the Chi-Square test.
3. The customers who have partners and dependents showed a lower To mitigate the issue of multicollinearity and to ensure the
churn rate compared to those who do not have partners or independence of features, pairs of columns with a correlation
dependents. coefficient greater than 0.9 were considered highly correlated
4. Customers with fiber optic internet services have a higher churn and were removed.
rate.
5. Customers who do not have essential services such as online
security, online backup, and tech support tend to leave the platform.
6. Customers with monthly subscription contracts have a higher
churn rate compared to those with one or two-year contracts.
7. Customers who have opted for paperless billing have a higher
churn rate.
8. Customers who use the electronic check payment method tend to
leave the platform more frequently

4.2.3 Stage 3 - Feature Engineering and Outlier Detection

Outlier detection was performed using the Interquartile Range
(IQR) method and no outliers were found. During the feature
engineering phase, it was observed that the "TotalCharges"
column exhibited a high correlation and therefore it was
dropped from the dataset. The contiguous values of the “tenure”
column were grouped into five groups with a fixed interval
length of 12 and a new column "tenure groups” was created.
Several additional columns were deemed not influential on the
"churn" column and were deleted from the dataset. The string Fig 4.6: Heat Map of the Correlations in the Dataset.
values 'yes' and 'no' were replaced with integer values of 1 and
0, respectively. Dummy variables or one-hot encoded 4.2.5 Stage 5 - Minority Oversampling - SMOTE:
representations of the categorical columns were done. Analysis of the ‘churn’ column revealed that 26.6% of the
responses indicated churn as 'yes', while the remaining 73.4%
4.2.4 Stage 4 - Feature Selection indicated churn as 'no'.
In this stage, the training set was subjected to the chi-square test, Synthetic Minority High Sampling Technique (SMOTE) was a
a statistical technique used in feature selection that helps to dopted for solving the class mismatch issue.
identify the categorical variables that are most likely to have an The SMOTE algorithm follows these steps:
impact on the target variable. F-value and p-value arrays were 1) For each minority class instance, find its k nearest
obtained as a consequence. A higher F-value suggests a stronger neighbours.
association between the feature and the target variable. A lower 2) Select a random neighbour and interpolate between the
p-value indicates a higher level of significance of that chosen instance and the neighbour to generate a new synthetic
association. From the results, we can infer that instance.
‘MonthlyCharges’ and ‘tenure_group’ are the columns that 3) Repeat the process for a desired number of synthetic
have the most influence on the target column. instances.

Applying SMOTE resulted in an increased representation of the

minority class in the dataset.

4.2.6 Stage 6 - Implementation of Machine Learning Models:

After data cleaning and feature engineering, we divide the datase
t into training and test sets, use 80% for training and 20% for test
ing, and evaluate the features between 0 and 1 and believe that th
ey are compatible.
1) Logistic Regression: Utilizes the logistic function f(z) = 1 / (1
+ e^(-z)) (where z represents the linear combination of the feature
values and their respective weights) and the logistic loss function.
2) Ridge Classifier: Involves solving the following equation: w
= (X^T * X + alpha * I)^-1 * X^T * y. Here, X is the matrix of
input features, I is the identity matrix, and y is the target variable.
3) SVM: Locates the hyperplane with the greatest margin, defined
as the distance between the hyperplane and the nearest data points
for each class.
4
IJRITCC | Month 20__, Available @ http://www.ijritcc.org
International Journal on Recent and Innovation Trends in Computing and Communication
ISSN: 2321-8169 Volume: x Issue: y
DOI: (do not edit)
Article Received: (do not edit) Revised: (do not edit) Accepted: (do not edit) Publication: (do not edit)
____________________________________________________________________________________________________________________
4) KNN: Identifies the K nearest neighbors based on the chosen V. RESULTS AND DISCUSSION:
distance metric and determines the majority class among them.
5) Naive Bayes: Assumes that the features are conditionally 5.1 Analysis of the ROC curve:
independent of each other, learns the class probabilities (P(y)), It visualizes the model's performance in distinguishing between
and calculates the likelihood probabilities (P(X|y)) based on the positive and negative instances across various threshold
observed features in the training data. settings.
6) Multi-Layer Perceptron: Uses a process called Here’s a comparative analysis of the ROC Curves of different
backpropagation to optimize the weights of the connections models implemented in Figure 5.1:
between neurons iteratively. Uses optimization algorithms like
gradient descent.
7) Decision Tree: The algorithm continues splitting the data at
each node, creating child nodes and branches, until a termination
condition is met. Gini Impurity and Information are the popular
criteria for splitting data.
8) Random Forest: Each decision tree is trained independently
using the random subsets of data and features, and the combined
forecasts of all the decision trees are used to arrive at the final
prediction.
9) Bagging Classifier: Based on a distinct bootstrapped subset of
the training data, each base classifier is independently trained. The
sum of the forecasts from each model is the final prediction.
10) AdaBoost Classifier: A weighted voting system is used to
decide the final prediction after each weak classifier is trained on
a subset of the training data and given a weight depending on
classification accuracy.
11) Gradient Boosting Classifier: Follows a boosting framework
where weak learners are trained sequentially to correct the errors
of previous models by utilizing gradient descent optimization.
12) XGBoost Classifier: Follows the gradient boosting Fig. 5.1: ROC Curves of the models
framework and uses decision trees as weak learners, which are
constructed in a greedy manner. 5.2 Analysis of the Precision-Recall Curve:
13) CatBoost Classifier: Follows the gradient boosting It provides insights into how well the model performs in
framework and handles categorical variables by encoding them correctly classifying positive instances (precision) and
based on the target variable's values within each category. capturing all positive instances (recall).
14) LGBM Classifier: Follows the gradient boosting framework Here’s a comparative analysis of the Precision vs Recall
and provides efficient handling of categorical features by using Curves of various models implemented in Figure 5.2:
'Optimized Exclusive Feature Bundling' (OEFB).
15) Voting Classifier: In hard voting, the majority class is chosen
as the final guess. In soft voting, the class with the highest average
performance or score is chosen as the final estimate.

4.2.7 Stage 7 - Simultaneous implementation of K-fold Cross

Validation and GridSearchCV:
While training every model, we implemented K-Fold Cross
Validation and GridSearchCV simultaneously. K-fold Cross
Validation is primarily used to evaluate a model’s performance
and generalizability, whereas GridSearchCV is used for
hyperparameter tuning. A parameter grid is passed to the
GridSearchCV along with a cross-validation engine like the K-
fold CV. It tests all possible combinations of hyperparameters and
identifies the optimal set of values that yield the best performance.

4.2.8 Stage 8 - Simultaneous performance evaluation:

Right after hyperparameter tuning and training, predictions of the
model were generated over the testing set. We thoroughly
evaluated every model on the basis of various performance
indicators like accuracy, precision, and recall as well as f1,
AUC(ROC), and AUC(PR) scores. Also, plotted necessary graphs Fig. 5.2: Precision vs Recall Curves of the models.
like the ROC and PR curves along with the Confusion Matrix.

5
IJRITCC | Month 20__, Available @ http://www.ijritcc.org
International Journal on Recent and Innovation Trends in Computing and Communication
ISSN: 2321-8169 Volume: x Issue: y
DOI: (do not edit)
Article Received: (do not edit) Revised: (do not edit) Accepted: (do not edit) Publication: (do not edit)
____________________________________________________________________________________________________________________
5.3 Comparison of Every Model: our assessment, we implemented the voting classifier (soft as
Figure 5.3 shows a detailed comparison of all the models well as hard). The Soft Voting Classifier showed the best
implemented in this research work. performance out of all. It showed an accuracy of 0.78 which is
at par with the highest accuracy achieved in our research so far
(i.e. 0.79 of XGBoost) and a decent f1-score of 0.74. The
difference-maker here is the ROC AUC_Score of 0.82 which is
significantly more than the highest ROC AUC_Score we’ve
achieved so far (i.e. 0.76 of Random Forest and Ridge
Classifier). Therefore, we infer that the Soft Voting Classifier is
the best-performing classifier in our research.

VI. CONCLUSION
In this research, we did a comparative analysis of
the effectiveness of various machine learning models
for user loss prediction in the telecom industry. We used
qualitative selection techniques i.e., chi-
square test and correlation analysis to select the best features.
Fig. 5.3: Values of evaluation metrics of all the models We also used SMOTE on the data and created a new synthetic
model with a small number of classes to sample them. We
The graph in Figure 5.4 visualizes how “accuracy”, “f1 score”, employed numerous learning and machine learning techniques,
and “ROC AUC_Score” with respect to every classifier. including Logistic Regression, Support Vector Machine, K
Neighbors, Naive Bayes, Pruning Tree, Random Forest,
Gradient Boosted Classifier, AdaBoost Classifier, XGBoost
Classifier, Light Gradient Accelerator Machine Classifier,
Ridge Classifier, Bagging
Classifier. We evaluate each model using performance indicat
ors such as accuracy, precision, and recall along with f1
and AUC(ROC) scores.

After comparing the performance of each model used, we conc

luded that XGBoost, AdaBoost, Logistic Regression,
and Multilayer Perceptron (MLP) performed best with regard
to accuracy, f1, and ROC(AUC) scores. Based on the top four
performances, we continue to use the survey. When compared
to other classifiers, the voting classifier had a much better
AUC(ROC) score of 0.82 in addition to an accuracy of 0.78 and
an f1-score of 0.74.

Consequently, a loss forecasting model can be used for market

ing communications to identify customers who will switch to o
ther service providers and provide their own incentives. Projec
Fig. 5.4: Comparison of Accuracy, F1, and ROC(AUC) t requests can help businesses stay ahead of the competition an
Scores of all the models d achieve their goals by leading to more information for decisi
From Figure 5.4, we inferred that the following classifiers are on-making.
the best-performing ones:
XGBoost- It boasts the highest accuracy (0.79), a respectable
f1-score (0.78), which is nearly as high as the previous record- VII. REFERENCES
holder (0.79 of AdaBoost), and a respectable ROC AUC_Score 1. Radosavljevik D, van der Putten P, Larsen KK (2010)
(0.71). The impact of experimental setup in prepaidchurn
Multi-Layer Perceptron(MLP) – With a superior ROC prediction for mobile telecommunications: What to
AUC_Score of 0.74, its accuracy and f1 score of 0.78 are both predict, for whom and does the customerexperience
matter? Trans. MLDM 3(2):80–99
quite respectable and close to the greatest values thus far. 2. Qureshi, S.A., Rehman, A.S., Qamar, A.M., Kamal, A.,
AdaBoost - It also has a respectable ROC AUC_Score (0.73), Rehman, A.: Telecommunication subscribers’churn
the highest f1 score (0.79), and the second-best accuracy (0.78). prediction model using machine learning. In: Eighth
Logistic Regression - The ROC AUC_Score achieves a local International Conference on Digital Infor-mation
maximum of 0.75 for logistic regression with an f1-score of Management (ICDIM 2013), pp. 131–136. IEEE
(2013)
0.77, close to the highest value, and a decent accuracy of 0.76. 3. Ahmad AK, Jafar A, Aljoumaa K (2019) Customer
churn prediction in telecom using machine learning in
Based on the above 4 best-performing classifiers according to big data platform. Journal of Big Data 6(1):28
6
IJRITCC | Month 20__, Available @ http://www.ijritcc.org
International Journal on Recent and Innovation Trends in Computing and Communication
ISSN: 2321-8169 Volume: x Issue: y
DOI: (do not edit)
Article Received: (do not edit) Revised: (do not edit) Accepted: (do not edit) Publication: (do not edit)
____________________________________________________________________________________________________________________
4. Tamaddoni Jahromi, A.: Predicting customer churn in
telecommunications service providers (2009)
5. Lalwani, Praveen & Mishra, Manas & Chadha, Jasroop
& Sethi, Pratyush. (2022). Customer churn prediction
system: a machine learning approach. Computing. 104.
1-24. 10.1007/s00607-021-00908-y.
6. Brându¸soiu, I., Toderean, G., Beleiu, H.: Methods for
churn prediction in the pre-paid mobile
telecommunications industry. In: 2016 International
conference on communications (COMM), pp. 97–
100.IEEE (2016)
7. Coussement K, Van den Poel D (2008) Churn
prediction in subscription services: An application of
support vector machines while comparing two
parameter-selection techniques. Expert systems with
applications 34(1):313–327
8. Dahiya, K., Bhatia, S.: Customer churn analysis in
telecom industry. In: 2015 4th International Confer-
ence on Reliability, Infocom Technologies and
Optimization (ICRITO) (Trends and Future
Directions),pp. 1–6 (2015)
9. Gürsoy U ¸S (2010) Customer churn analysis in
telecommunication sector. Istanbul Üniversitesi
˙I¸sletme Fakültesi Dergisi 39(1):35–49
10. Hadden J, Tiwari A, Roy R, Ruta D (2006) Churn
prediction: Does technology matter. International
Journal of Intelligent Technology 1(2):104–110
11. Hadden J, Tiwari A, Roy R, Ruta D (2007) Computer
assisted customer churn management: State-of-the-art
and future trends. Computers & Operations Research
34(10):2902–2917
12. Huang, Y., Zhu, F.,Yuan, M., Deng, K., Li, Y., Ni, B.,
Dai, W., Yang, Q., Zeng, J.: Telco churn prediction
with big data. In: Proceedings of the 2015 ACM
SIGMOD international conference on management of
data, pp. 607–618 (2015)
13. Idris, A., Khan, A., Lee, Y.S.: Genetic programming
and adaboosting based churn prediction for telecom. In:
2012 IEEE International Conference on Systems, Man,
and Cybernetics (SMC), pp. 1328–1332. IEEE (2012)
14. Kisioglu P, Topcu YI (2011) Applying bayesian belief
network approach to customer churn analysis: A case
study on the telecom industry of turkey. Expert
Systems with Applications 38(6):7151–7157
15. Saleh, Sarkaft & Saha, Subrata. (2023). Customer
retention and churn prediction in the
telecommunication industry: a case study on a Danish
university. SN Applied Sciences. 5. 10.1007/s42452-
023-05389-6.
16. Rathi, S., Kant Hiran, K., & Sakhare, S. (2023).
Affective state prediction of E-learner using SS-ROA
based deep LSTM. Array, 19, 100315.
17. Rathi, S.R. and Deshpande, Y.D. (2022), "Course
complexity in engineering education using E-learner's
affective-state prediction", Kybernetes, Vol. ahead-of-
print No. ahead-of-print.
18. S. R. Rathi and Y. D. Deshpande, "Embedding Affect
Awareness into Online Learning Environment using
Deep Neural Network," 2019 5th International
Conference On Computing, Communication, Control
And Automation (ICCUBEA), Pune, India, 2019, pp.
1-6, doi: 10.1109/ICCUBEA47591.2019.9128811

7
IJRITCC | Month 20__, Available @ http://www.ijritcc.org