Analysis_of_Telecom_Churn_using_Machine_Learning_Techniques

International Conference on Artificial Intelligence and Data Engineering (AIDE)
Analysis of Telecom Churn using Machine Learning

Techniques
Puneeth B. R Preethi Salian K Prathyakshini
NMAM Institute of Technology, NMAM Institute of Technology, NMAM Institute of Technology,
Nitte (Deemed to be University), Nitte (Deemed to be University), Nitte (Deemed to be University),
Udupi, India. Udupi, India. Udupi, India.
puneethbr9@gmail.com preethi.salian@nitte.edu.in prathyakshini@nitte.edu.in
Anantha Murthy Supriya Salian Surabhi

NMAM Institute of Technology, St. Joseph Engineering College, NMAM Institute of Technology,
Nitte (Deemed to be University), Mangaluru, Karnataka, India Nitte (Deemed to be University),
2022 International Conference on Artificial Intelligence and Data Engineering (AIDE) | 978-1-6654-9304-8/22/$31.00 ©2022 IEEE | DOI: 10.1109/AIDE57180.2022.10060222
Udupi, India. supriyas@sjec.ac.in Udupi, India.

anantham2004@nitte.edu.in surabhiachar16@gmail.com
Abstract— Customers are the cornerstone of any company's periods by providing them with specialized services.
success; thus, businesses understand how crucial it is to ensure Following the implementation of customer loyalty, many
their satisfaction. The telecommunications sector is developing businesses start to think about churn reduction as one of the
quickly, and service providers are now more eager to expand business objectives. Customer churn is interpreted as the
their subscriber bases. Retaining existing customers has become movement of a subscriber of the telecom industry from one
a major issue to meet the need to survive in a competing telecom environment to another. In this study, we are going to
economy. In this study, customer turnover is predicted using predict the reason behind the churning of customers.
data from a telecoms business utilizing the survival analysis
Subscriber churn can be in different forms and not just exit
technique. This study will help telecom corporations better
from the base. It includes tariff plan churn like moving from
understand the risk and hazard of customer turnover by
detecting which customers are likely to go and whether they will.
50Rs to 30Rs pack for a month, Service churn like moving
Analysis of the telecommunications industry has revealed that it from weekly subscription to monthly subscription, Product
is significantly more immoderate to obtain a new client than to churn like churning from post-paid to prepaid and usage churn
continue an existing one. Due to growing business competition, is like unsubscribing from telecom.
the importance of marketing techniques, and users' increasingly The system must be able to predict both whether
aware behaviours in recent years, customer turnover is a crucial users are at a severe rate of churning and how quickly these
issue and is examined one of the most important worries high-risk customers will quit assisting telecom enterprises in
between enterprises. To address the churn problems resulting managing churn. As a result, telecom firms are more able to
from the services they provide, the company must implement employ their innovation resources to keep as many clients as
several ways. In the very dynamic and quickly evolving telecom possible. To put it another way, telecom companies may
market, customer attrition strategies are crucial. The process of construct specialized customer communication and
switching telecom service providers happens because of the rehabilitation services quickly and effectively if they are
competitive firm's competitive rates, excellent services, or a aware of the clients who are most likely to switch and how
variety of benefits it offers clients when they sign up. To they will do so. Customer turnover can be accurately predicted
anticipate customer loyalty and whether they would leave the using traditional statistical techniques like logistic regression
company, information from the telecom business can be useful. and decision trees, among others. These techniques hardly
To boost customer engagement and/or treatment resources as ever managed to forecast when clients will go or how long
part of their churn reduction goals, telecom companies will
they will be around.
benefit from the findings of this study.
The dataset is subjected to the top 5 classification models
Keywords—Telecommunication industry, Churn, Mobile, in this proposed system. After applying all of the classification
Prediction models we verify which algorithm gets the best score. An
application is created to anticipate single customer attrition
I. INTRODUCTION using the best score classification algorithm.
In the telecommunications industry, customers can pick
from a wide range of service providers and actively protect II. LITERATURE SURVEY
their right to change service providers. In this crowded
Shrikhande and Verma [1] indicated that analysing data
marketplace, customers desire personalized goods and better
and records acquired from telecom firms might aid in
services at cheaper prices, while service providers' business
determining the causes of customer churn and using the same
objectives are perpetually centered on acquisitions. Consumer
data to retain consumers. The study uses a decision tree
loyalty has surpassed client revenue as one of the most
technique to forecast client attrition. The goal of Arifin and
important aspects, considering that the telecommunications
Samopa [2] was to figure out what factors influence the
industry has a 30% to 35% annual turnover and that acquiring
turnover rate in telecommunication firms. According to
a new user 5 to 10 times more than retaining a current user.
Dahiya K, Bhatia S. [3] the churn prediction model was built
Many telecommunications companies implement retention
using a Decision tree and logistic regression. The test
tactics into their methodology to keep customers for long
evaluates the performance of a classification model by
978-1-6654-9304-8/22/$31.00 ©2022 IEEE

censed use limited to: MINISTERE DE L'ENSEIGNEMENT SUPERIEUR ET DE LA RECHERCHE SCIENTIFIQUE. Downloaded on January 30,2025 at 14:07:53 UTC from IEEE Xplore. Restric
recording counts predicted correctly and inaccurately. For
customer retention analysis, Wei, C. P and Chiu [4] proposed
the churn prediction technique. Upon those details of the
customer call, the research employs a decision tree technique.
By putting up a decision tree algorithm, Yi-Fan wang, Ding-
an chalang, and Mei-Hua Hsu [5] created a technique for void
customer churn. A churn prediction model was developed by
Shaaban [6], and it can forecast the likelihood that customers
will leave in the coming days. N. Kamalraj and A. Malathi
[7] focused their invention on employing data mining
techniques to better understand churn prediction. This
approach to client retention operations can be used by the
telecommunications industry as part of their customer
relationship management initiatives.
Customer turnover has been identified as the
telecom industry’s most challenging issue by Almena et al.
[8]. The current study has combined data mining techniques
with machine learning algorithms to successfully address the
churn prediction difficulty. According to Dahiya[9], the
telecom industry is rapidly developing, which has allowed
service providers to expand their bases. Retaining current
clients has become a challenging endeavor as a result of the
market’s desire for retention. According to Alwin [10],
service providers can improve their offerings to better meet
consumer requests when they are aware of the causes of client
churning. According to Kumar and Chandrakala [11], the
market's explosive growth across all sectors is increasing
service providers' subscriber bases. With every industry
increasingly dependent on customer input, the current study
examines the most popular machine learning algorithms
researchers have used to predict churn. Sjarif, et al.[12]
emphasized the significance of a churn prediction model for
telecommunication companies to prevent their users from
switching to another operator services. The aim of this
research study is to create a model for predicting customer
Fig. 1. Proposed System Architecture
turnover as a outcome. Because of the costs involved,
customer churn has become a serious customer care
A. Data Aquision
management problem for most mobile service providers,
according to Yabas and Chankya [13]. The contribution on The practice of acquiring and analysing data from a
customer turnover inquiry and evaluation for such services is variety of sources is known as data collection. Data collection
described in the paper. Ebrah and Elnasir [14] investigated phase is critical because the quality and quantity of data you
the ML applications and suggested the desirable fixes for the collect will have a direct impact on the accuracy of prediction
telecommunication sector. Customers are more likely to model. In the proposed system, information is collected from
switch service providers in the competitive the patient's various voice recordings.
telecommunication industry, so service providers are
interested in their users and how to retain them. Azeem and
Usman [15] noted that diverse machine learning churn
prediction algorithms have been created in recent years to
deal with consumer attrition.
In the proposed model, different machine learning
classifiers are tested and the best model is used for the
prediction. Also a comparative analysis of the same is shown
which would help telecom analyst to identify the best
classifier that really works well with huge volume of records.
III. METHODOLOGY
For the successful determination of telecom industry churn
analysis, distinctive sorts of classification strategies were
inspected. The proposed system architecture is depicted in
Fig. 1.
Fig. 2. Categorical count for churned or not churned people
The Telecom industry dataset is mainly composed of C. Prediction Model
7043 rows and 21 columns. Each row stands for customer The process of selecting a single machine learning
information, and each column comprise customer attributes. model from a pool of potential models for a training dataset
The main target is the 'Churn' column. That column contains
is known as model selection. Several models may be used
two values that are ‘Yes’ which means the customer shifted
using the process of model selection. The classification
from one telecom industry to another industry and 'No' means
the customer has not changed their telecom industry. Fig. 2. algorithms available can be used based on the benefits and
shows the county for two categorical churned people and not drawbacks of the same [16]. After applying all of the
churned people. The ‘churn’ column in the dataset allows us classification model, verified which algorithm gets the best
to track the number of persons who switch between the score.
telecom industries. Using this, we can observe that total of a) Logistic Regression:
5K users are not leaving the telecom, while 1,869 are moving Logistic regression is one of the most desirable Machine
to other telecom industries. Learning algorithms used in Supervised Learning. It is used
in variety of applications including social networking
advertisement [17]. A discrete class-dependent variable may
be predicted using this technique using a collection of
independent factors. The output of a dependent categorical
variable is predicted via logistic regression. The result must
thus be a discrete or classification value. It can be True or
False, Yes or No, 0 or 1, and so on, but rather than providing
precise values like 0 and 1, it provides probability values that
are in the range of 0 and 1.
b) Decision Tree:
The Supervised algorithm is a learning technique that is
exploited to solve classification and regression problems
[18]. It is viable to use the decision tree approach for both
continuous and categorical output variables. Using the
python sklearn libraries, this method pre-processes the data
before splitting it into train and test sets. Then enumerate the
prediction and enumerate the trained classifier’s accuracy.
c) Support Vector Machine:
In classification complications, SVM is widely utilized
Fig. 3. Correlation between feature of the Telecom industry
[19]. The SVM is frequently employed. Many people prefer
the SVM because it achieves great accuracy while using
Feature importance refers to a set of strategies for minimal computing resources. The fundamental goal of an
assigning scores to features in a predictive model that SVM is to determine a hyperplane in an N-dimensional space
represent the relative value of each feature. The dataset can that categorizes points distinctly. The Support Vector
be analyzed using feature importance analysis. Fig. 3. shows Machine classifier is a frontier that best combines the two
how several features in the Telecom industry dataset were hyperplane/line classifications. The idea of an SVM is to split
interrelated to each other as well.. a dataset into classes so that a maximum marginal feature
space can be determined.
B. Data preparation and Data Pre-processing
The following steps are performed on the dataset in d) XGBoost Algorithm:
the Customer Churn prediction system The XGBOOST method is a widely used one in machine
learning, regardless of whether the subject is a classification
• Dropping Redundant Columns or regression problem. It is also being used in predicting the
• Checking for Duplicated Rows status of power system [20]. The highest performance of any
• Checking for Missing Values machine learning algorithm makes it stand out. The method
The data is well visualized in this step to recognize the is around ten times quicker than earlier gradient booster
relationship between the parameters included in the data to approaches thanks to its excellent predictive capacity, which
take advantage of as well as obtain data imbalances. As a gives it the best accuracy solution.
result, the data is divided into two halves. The first step is to
train the model, 80% of the data is applied for training and e) Naïve Bayes Algorithm:
20% for testing. This stage involves translating the raw data The Bayes Theorem is the foundation of the supervised
collected into a format that can be used in modeling. Data learning method recognized as the Nave Bayes algorithm,
pre-processing is the most crucial stage in predictive models which is used to get around the classification problem [21].
because the data contains ambiguities, errors, and redundancy These techniques support the progress of quick machine
that need to be cleaned in advance. Data collected from learning models with rapid prediction capabilities. As a
various sources is first accumulated and then cleaned because probabilistic classifier, the nave Bayes method estimates the
the entire collected data is not appropriated for modelling likelihood of an item using the laws of probability. The naive
purposes. Records with distinctive values are meaningless Bayes method outperforms the others in multiclass
because they do not provide much to predictive modelling. predictions when compared to other algorithms. The Naive
Fields with many null values should also be eliminated. Bayes Classifier is one of the most direct and efficient
classification algorithms obtainable present day. It helps in False Negatives: These are the individuals whose negative
developing fast machine learning models capable of making outcomes we anticipated, yet they are departing the network
accurate predictions. (churn).
f) Random Forest Algorithm: Accuracy: One of the most important aspects of classification
Random Forest [22] is most famous supervised learning problem accuracy is this. It details the frequency of accurate
technique which could be used for both regression and result predictions made by the model. It is possible to
classification problems. It works on the principle of ensemble calculate the proportion between the classifier's proper
learning where multiple classifiers are combined in order to predictions and all the classifiers' combined predictions.
solve a complex problem. In order to improve the accuracy
of the dataset decision tree is taken from several subsets of The performance metrics such as F1-score,
the dataset. Prediction is done by taking major part of the Precision, Recall and Accuracy were used in order to evaluate
prediction made from each decision tree instead of single the model.
decision tree. Random forest algorithm is used in variety of
applications [23]. Accuracy=(Trp+Trn)/(Trp+Trn+Frp+Frn) (1)
Precision: It may be characterized as either the proportion of
D. Training all received a strong which that model accurately predicted as
Dataset training is one of the most critical jobs in machine true or the number of the model's outputs that are correctly
learning. Utilize the data to gradually improve the chosen generated. It may be calculated using the following formula:
model's ability to forecast better, so the actual result should
Precision=Trp/Trp+Frp (2)
be near to what was anticipated. Utilize the data in this stage
to increase proposed model's capacity to predict whether Recall: It is described as the proportion of positive classes out
customers like to churn or not. of all positive classes that with our very model correctly
predicted. The recall rate must be as high as is practical.
E. Evaluation
Recall=Trp/Trp+Frn (3)
Accuracy, f1-score, Precision, and other metrics were
determined, and they all pointed to the best method out of the F1 Score: The F1 score, which is a statistic that incorporates
collection. The test harness is the data you'll need to train and both accuracy and memory, is equivalent to the chromatic
an algorithm, as well as the performance metric you'll use to mean of precision and recall. Its value falls between [0,1].
evaluate it. The estimation of how a range of algorithms (The higher the value, the higher the F1 score).
performs on the problem versus a given performance metric
will be the result of testing numerous methods against the F1_score=2*Precesion*Recall/Precision+Recall (4)
harness.
F. Decision V. RESULTS
In this phase, we eventually make the model ready to Multiple prediction models for the prediction of Telecom
identify telecom churn prediction using the available dataset. customer churn are evaluated in the proposed system. Six
A set of rules will allow subscribers discover and categorize machine learning models such as Logistic Regression,
into distinct categories churners and non-churners by setting Decision Tree, SVM, XGBoost, Naïve Bayes and Random
a specific threshold value. Forest algorithms are used to evaluate the system. The
projected outcomes of this investigation reveal that the
proposed system’s classification accuracy has improved.
IV. PERFORMANCE EVALUATION Thus, it can be concluded that the Random Forest Classifier
accuracy has performed superior to other classifiers.
The classification model’s performance could be
measured using a variety of approaches. Evaluation of the TABLE I. DIFFERENT MACHINE LEARNING ALGORITHM’S ACCURACY
performance of a machine learning model can be used to IN PERCENTAGE
point the difficulties mentioned above. Typically, the models’ Classification Accuracy Precision Recall F1-
accuracies and performance are tested applying precision, Algorithms (%) (%) (%) score
recall, and f1-score are some of the key criteria to consider. (%)
A confusion matrix is a machine learning classification
performance evaluation technique. By differentiating the Decision Tree 93 91 97 97.2
actual and predicted classifications, the confusion matrix
visualizes the accuracy of the classifier. The squares that Random Forest 96 95 97 97.5
make up the binary confusion matrix are as follows:
Naïve Bayes 90 90 93 70.5
True Positives: These are the individuals whose "churn" we
foresaw, but they are staying on the network (not churn) Logistic Regression 79 64 49 80
True Negatives: These are the individuals whose absence Support Vector Machine 94 95 93 73.4
from the network we foresaw.
XGBoost 96 97 96 91
False Positives: These are the individuals whose departure

from the network we foresaw, but (not churn).
obtained it is evident that Random forest algorithm gets better
accuracy and can be used for prediction.
VI. CONCLUSION
Because of standardization and governmental rules on
mobile communication, customers may switch carriers with
ease in the competitive telecom business. As a result, the
market is competitive. Predicting churn or figuring out which
customers are most likely to discontinue using a service, is a
profitable and significant challenge in the telecom sector.
Because consumers are hesitant to leave if they cannot find
what they are seeking, client turnover is a significant problem
in the telecom sector. Customers mostly seek improved
service standards, lower prices, and better value for their
money. Customer turnover and customer satisfaction are
Fig. 4. Classification Results of each Algorithm directly related. Customer retention is a challenging business
model to calculate since customer acquisition expenses are
higher than client retention costs. There is no traditional
strategy that effectively addresses the churning challenges
faced by international telecom service providers. Churn can
be handled and kept at a tolerable level even if corporate
losses are inevitable. Good practices must be adopted, and
outdated techniques must be upgraded to prevent
impediments from affecting the telecoms industry. In this
work, we investigated several prediction models and
contrasted quality indicators of prediction models including
regression analysis and decision trees. We found that random
forests are a useful strategy since the accuracy they achieve
is significantly higher than that of other models. In future
various feature selection algorithms can be applied on the
model with other classifiers in order to get the better results.
REFERENCES
Fig. 5. Churn by Monthly charges vs. density
[1] Shrikhande, P.A., Verma, A. (2018). Survey on Customer Churn
Analysis in Telecom Sector Using Decision Tree. JASC: Journal of
Applied Science and Computations, 5(9), pp. 330-335.
[2] Arifin, S., & Samopa, F. (2018, November). Analysis of churn rate
significantly factors in telecommunication industry using support
vector machines method. In Journal of Physics: Conference
Series (Vol. 1108, No. 1, p. 012018). IOP Publishing.
[3] Dahiya, K., & Bhatia, S. (2015, September). Customer churn analysis
in telecom industry. In 2015 4th International Conference on
Reliability, Infocom Technologies and Optimization (ICRITO)(Trends
and Future Directions) (pp. 1-6). IEEE.
[4] Wei, C. P., & Chiu, I. T. (2002). Turning telecommunications call
details to churn prediction: a data mining approach. Expert systems
with applications, 23(2), 103-112.
[5] Wang, Y. F., Chiang, D. A., Hsu, M. H., Lin, C. J., & Lin, I. L. (2009).
A recommender system to avoid customer churn: A case study. Expert
Systems with Applications, 36(4), 8071-8075
[6] Shaaban, E., Helmy, Y., Khedr, A., & Nasr, M. (2012). A proposed
churn prediction model. International Journal of Engineering Research
and Applications, 2(4), 693-697.
Fig. 6. Churn by Total charges vs. density [7] Kamalraj, N., & Malathi, A. (2013). A survey on churn prediction
techniques in communication sector. International Journal of Computer
Applications, 64(5), 39-42.
Random Forest has an accuracy of 96%, a precision of
[8] Almana, A. M., Aksoy, M. S., & Alzahrani, R. (2014). A survey on
95%, and a recall of 97% which is better than all other data mining techniques in customer churn analysis for telecom
machine learning techniques when compared with other industry. International Journal of Engineering Research and
models also refer to Table I to check the different evaluation Applications, 4(5), 165-171
scores. Fig. 4. graph shows the accuracy level of each [9] Dahiya, K., & Bhatia, S. (2015, September). Customer churn analysis
algorithm. Fig. 5. and Fig. 6. shows churn by monthly charges in telecom industry. In 2015 4th International Conference on
and total charges. When monthly fees are high, churn is high Reliability, Infocom Technologies and Optimization (ICRITO)(Trends
and Future Directions) (pp. 1-6). IEEE.
and as higher churn at lower Total charges. From the results
[10] Alwin, P.K.D.N.M., Kumara, B.T.G.S., Hapuarachchi, H.A.C.S.
(2018). Customer Churn Analysis and Prediction in [17] Divya Sai Patnana, Gone Hitesh, IpsitaSahu. N. Suresh Kumar,
Telecommunication for Decision Making. 2018 International “Logistic Regression Analysis on Social Networking Advertisement”,
Conference on Business Innovation (ICOBI), pp. 40- 45 Journal Of Critical Reviews Issn- 2394-5125vol 7, Issue 04, 2020.
[11] Saran Kumar, A., & Chandrakala, D. (2016). A survey on customer [18] Arundhati Navada, Aamir Nizam Ansari, Siddharth Patil, Balwant A.
churn prediction using machine learning techniques. International Sonkamble, “Overview of use of decision tree algorithms in machine
Journal of Computer Applications, 975, 8887. learning”, IEEE Control and System Graduate Research Colloquium,
[12] Sjarif, N. N., Rusydi, M., Yusof, M., Hooi, D., Wong, T., Ya’akob, IEEE, 18 August 2011. 10.1109/ICSGRC.2011.5991826
S., ... & Osman, M. Z. (2019). A customer Churn prediction using [19] Durgesh K. Srivastava, 2lekha Bhambhu, “Data Classification Using
telecommunication industry. Int. J. Advance Soft Compu. Appl, 11(2) Support Vector Machine “, Journal of Theoretical and Applied
[13] Yabas, U., & Cankaya, H. C. (2013, December). Churn prediction in Information Technology,2009.
subscriber management for mobile and wireless communications [20] Minghua Chen , Qunying Liu1, Shuheng Chen,Yicen Liu1, Chang-Hua
services. In 2013 IEEE Globecom Workshops (GC Wkshps) (pp. 991- Zhang, And Ruihua Liu, “XGBoost-Based Algorithm Interpretation
995). IEEE. and Application on Post-Fault Transient Stability Status Prediction of
[14] Ebrah, K., & Elnasir, S. (2019). Churn prediction using machine Power System”, IEEE, Volume 7, 2019.
learning and recommendations plans for telecoms. Journal of [21] I. Rish, “An empirical study of the naive bayes classifier,” IJCAI 2001
Computer and Communications, 7(11), 33-53. Workshop on Empirical Methods in Artificial Intelligence, vol. 22, no.
[15] Azeem, M., & Usman, M. (2018). A fuzzy based churn prediction and 3, pp. 41-46, 2001.
retention model for prepaid customers in telecom industry. [22] Leo Breiman , “Random Forests”, publisher Springer, Machine
International Journal of Computational Intelligence Systems, 11(1), Learning , volume 45, 5–32 (2001).
66-78.
[23] Qingxiang Xu1 and Jiesen Yin, “Application of Random Forest
[16] Aized Amin Soofi* and Arshad Awan, “Classification Techniques in Algorithm in Physical Education”, Scientific Programming for Smart
Machine Learning: Applications and Issues”, Journal of Basic & Internet of Things, 2021. https://doi.org/10.1155/2021/1996904
Applied Sciences, 2017, 13, 459-465, ISSN: 1814-8085.

Analysis_of_Telecom_Churn_using_Machine_Learning_Techniques

Uploaded by

Analysis_of_Telecom_Churn_using_Machine_Learning_Techniques

Uploaded by

International Conference on Artificial Intelligence and Data Engineering (AIDE)

Analysis of Telecom Churn using Machine Learning

Anantha Murthy Supriya Salian Surabhi

Udupi, India. supriyas@sjec.ac.in Udupi, India.

978-1-6654-9304-8/22/$31.00 ©2022 IEEE

Fig. 2. Categorical count for churned or not churned people

False Positives: These are the individuals whose departure

You might also like