Data Mining on Customer Churn Classification

Customer churn classification using machine learning techniques and
comparing various classification methods
Adhithya Vattiam Ganesan, Kaushik Rajan, Sindhujan Dhayalan, Sushmit Kajalekar
X18105203, X17165849, X17170265, X18107001
School of computing
National college of Ireland
Abstract—Abstract: The telecommunication sector is rapidly
developing every day with the service providers offering a wide
range of services and benefits to the customers for various
reasons, but they also have a customer churn rate, i.e. the
number of customers choosing to leave the services offered by
one telecom service provider to another service provider. This
directly affects the business of the company leading to huge
loss. CRISP-Data Mining model is used to plan, understand and
implement machine learning methods to predict the impact of the
factors affecting customer retention and if a customer will stay or
leave the company. To do the customer churn prediction, dataset
from Kaggle is taken. Initially, the factors which has the highest
correlation with the customer churn variable is determined. The
customer dataset is taken and trained, then multiple algorithms
such as Random Forest, Decision tree algorithms (CART & C5.0),
Nave Bayes, KNN & ANN are applied to classify if the customer
will leave or stay with the company. All the used models are
evaluated and compared based on accuracy and ROC curve.
the impact of various factors leading to customer churn rate
and compared the models for churn prediction accuracy. C5.0
approach was found to be the best fit for churn prediction with
easy implementation and best fit for the problem.
Index Terms—Random forest, C5.0, KNN, ANN, Naive Bayes,
Decision tree, SVM.
I. INTRODUCTION
Telecommunication sector is at its peak with the
advancement of technologies, they offer a wide range
of services from internet, mobile services and cable
subscriptions all together as a bundle or one can select the
services separately based on preference. This sector has a lot
of players and each of them with the intention of trying to
control the market share provide offers to lure the customers
to change to their service providers for better services and
offers. This movement of customers from one service provider
to another service provider is known as customer churn rate.
This movement of customers affects the service providers in
many ways, first this tarnishes the reputation of the company
in the market creating a bad image of the service provided.
Then secondly this leads to a business and financial loss. The
service providers can cope up with business loss to some
extent but once the reputation of the company is gone then it
becomes irrevocable, which no organisation would want to
face. Most of the telecommunication companies keep losing
customers to competitors from time to time. The customers
that stopped using the company’s product or service during
a certain time frame is known as customer churn [1]. On
multiple instances it is observed that retaining present clients
is a whole lot profitable and cheaper than attracting an entirely
new client. Hence, customer churn prediction is turning into
the top concerns that many organizations dedicate their time
and assets to deal with it.
The service provider analyses the reason for a customer to
discontinue their services to improve their business strategies,
it is estimated that the cost of acquiring a new customer
is 6-7 times more expensive than the cost of retaining
an existing customer. Hence the service providers spend
time analysing the reasons for churn, this may be due to
dissatisfaction, cheaper or more affordable services provided
by the competitors or better sales and marketing strategies
employed by the competitors. Each of which indicating the
areas where the company falls behind its competitor in the
market [2].
In this paper we analyse the most commonly applied
machine learning techniques for customer churn prediction
by applying Decision Tree algorithms, K-Nearest Neighbour
algorithm, Random Forest algorithm and Artificial
neural network to predict the customer churn rate using
telecommunication customer data and then compare the
models based on their prediction accuracy, error rate and the
ease of performing the classification. The objective of the
paper is to find the best model for Telecom customer churn
prediction.
II. LITERATURE REVIEW
The classification problem associated with customer churn
prediction system can be fixed by using decision tree method.
The information theory is used to create information nodes,
branches as per the various values of different aspects[3]. It
examines the statistics of capabilities from large instances,
calculate the entropy of particular feature, and locate the
maximum essential function for category. For over 13000
records in a customer churn data, the decision tree method
showed 88.53% accuracy for prediction[4].
The imbalance between customer churn data and predicting
applicable solutions on it often results in poor performance
of learning and also it is difficult for machine learning
algorithms[5]. The other classifier is c5.0 which gives good

performance over the large data set. It performs constantly
when variety of variables is large. The boosting technique is
carried out in this set of rules to enhance the precision. The
other purpose is that the C5.0 selection tree can overcome the
scatter feature of the telecommunication information. This
algorithm solves few problems such as low rate of churn,
scatter problem. It provides sampling method which removes
noise, redundancy and performs separate clustering. Precision
ratio is 80% and prediction results are satisfactory enough to
retain the potential churn customers[6].
In case of the company which is recently established
or newly adopted some different technology or has lost
the historical information associated with the customers,
the traditional churn prediction systems wont be useful.
In such cases JIT(Just-In-Time) approach is more practical
alternative. This CCP process categories customers as churn
and non-churn according to the probability of a customer
activity in a nearer future. The four important aspect of this
approach are behaviour of customer, the loyalty of customer
with the company, customers personal details and some
external factors. This model follows a Nave Bayes approach
which calculates posteriori probabilities[7].
Another approach is based on information processing mech-
anism. It is primarily based on a large series of different neural
networks. The structure of these network is determined by the
inter connection between them. These networks are trained to
achieve desirable outcome. Their maximum excellent charac-
teristic is their capability to research automatically from the
information so that one can provide a method for predictions.
Hence these models are used to build efficient, reliable and
high accuracy customer churn prediction system[8].
Support Vector Machines are used for minimizing the risks.
It uses the kernel method to draft the data to a 2D or 3D
space. The data is divided in linear manner so that further
classification can be done. In this paper, an approach for churn
prediction by using SVM is discussed. To gain more accurate
results and performing, a grid search manner along with the
SVM was used. The SVM for customer churn prediction
method by using SVM provided 98.7% accuracy[9].
The analysis of the imbalanced data for the churn prediction
based on various predictive model with respect to KNN
classifier. The ratio of non-churn customers compared to churn
clients is only 2 percent, therefore lowering the imbalance
of the instructions before enforcing the models may be very
critical. Another important issue raised due to ratio imbalance
is over sampling resulting in redundancy. This is handled
by the algorithm SMOTE (Synthetic Minority Oversampling
Technique). In the crucial aspect of model selection, the
implementation of various model like AdaBoost, Extra Trees,
kNN, Neural Network and XGBoost[10] is compared. The
study is carried out based on the different resources provided
by one of the internet service providers of leading Telecom
Company. Usage based on numerous statistical assessments;
121 most relevant functions had been decided on to use for
training. these attributes were accumulated over 365 days
and examined the seasonal effect. The non-parametric lazy
learning algorithm used to carry out the classification is KNN
known as k-Nearest Neighbour. The algorithm takes all the
nearer cases in training set into the consideration using mea-
sures like Euclidian distance. This classifier is to understand
statistical analysis of customer churn and pattern recognition.
XGBoost feature importance score is used to understand the
top number of customer variable and best fit model. In the sim-
ilar way, maximum appropriate top functions for AdaBoost,
more trees, KNN and Neural network version are determined
by evaluating their performances. The important aspect of this
study is feature engineering on modelling phase and providing
methodology which can handle imbalanced information in
churn prediction system[11].
Customer churn analysis, is an important factor to be
considered while analysing company profits and revenues.
The only challenge in case of churn analysis is low rate of
churned customers in entire data which causes imbalance data
information and data sets. The paper put forth two approaches
to analyse the imbalanced data. One is sampling approach and
another one is known as cost-sensitive approach. The motive
of this analysis is to address the imbalanced records in the
dataset used in this study, so that it can produce greater per-
formance in churn prediction model. The dataset is processed
by the sampling approach and WRF classifier classification
used on the data produces inadequate results hence sampling
techniques are used to minimize the unbalancing between the
classes churn and non-churn. SMOTE[12] is used to increase
the probability of fetching the data from churn class. The two
methods mentioned earlier to reduce imbalanced data between
the classes, uses many classifiers to solve this issue such as,
logistic regression, linear classification, nave Bayes, decision
tree, multi-layer perception neural networks, support vector
machine, data mining evolutionary algorithm and random
forest. The paper proposes the classification algorithm along
with WRF and sampling which is used to generate multiple
decision trees from the training data and provide the churn
result with precision. The paper tests proposed model and
data with 10 cross validations to obtain result with maximum
precision. Which proves this model provides the churn analysis
with reliable accuracy and can assist in reducing potential
customer churn[5].
III. METHODOLOGY
CRISP DM Model has been used for this Data Mining
project. Cross-Industry Standard Process for Data Mining is a
6-phase constructed approach used for data mining project.
A. Business Understanding
This is the initial stage of the Data Mining project. In-
formation regarding the customer churn and details on the
determined outputs are decided.
Business Questions:
1) Which variables have the highest impact on the customer
retention?

Fig. 1. CRISP DM Model
2) Which model performs the best at classifying if the
customer will leave or stay?
B. Data Understanding
For this project, Customer churn dataset created on 2018-02-
23 is sourced Kaggle1
. The dataset consists of 21 attributes and
details with the datatype of each variable is given in figure2.
Fig. 2. Dataset information
C. Data Preparation
Data Preparation and pre-processing is the most important
stage of any data mining project. In this stage, the dataset is
checked for outliers, missing values and also if the attributes
are in proper datatype. Following steps were done in order to
convert the data to the stage where it is suitable for model
implementation:
1https://www.kaggle.com/blastchar/telco-customer-churn/metadata
1) The datatypes of few attributes were changed. The
categorical variables were converted into factors.
2) The dataset was checked for outliers. Interquartile range
rule was used and the range value is set based on the Q1
and Q3. The IQR rule was used on the tenure attribute.
3) Missing values were replaced with the mean of the
attribute instead of removing the row.
4) CustomerID was removed as it was not used for predic-
tion.
5) The variables were normalized using Z-score normaliza-
tion. The normalization was done using scale() function
in R.
Fig. 3. Boxplot of Tenure
Fig. 4. Boxplot of Totalcharges
D. Modelling and Evaluation of the models
After the data was cleaned and converted as per the re-
quirement, we split the dataset into training and test data in
order to implement the models. The models learn from the
training set and implement on the test set. The models that are
implemented are Random forest, C5.0, K-Nearest Neighbour,
Decision Tree and Artificial neural network.
Random Forest: Random forest is an ensemble of decision
tree which can be used for regression and classification. The

main advantage of random forest is that it does not produce
overfit results and also produces great accuracy. We have used
random forest to classify if the customer will churn or not.
The confusion matrix is given in figure5.
Fig. 5. Random forest model Results- Rcode
The Random forest produced results with 73.08% accuracy
and with the error rate of 0.27. The area under the curve value
is 0.63 which falls in the D category,The ROC of the model
is displayed in figure6.
Fig. 6. ROC of Random forest model
Business question 1:
Random forest to find the importance the variables in
classifying the customer churn variable:
We use random forest to determine the importance of each
of the variable and to find which variable has the highest
correlation with the customer churn variable.
From the above plot we can infer that the variables Tenure,
Total Charges, contract and monthly charges are the most
important variables which are used to classify the customer
churn.
Fig. 7. Parameters contributing the most to accuracy of the model
C5.0: C5.0 is an updated version of C4.5 algorithm which
can produce 2 kinds of models, namely, decision tree and rule
set. The C5.0 can also find the attribute relevance. We have im-
plemented the C5.0 algorithm by using Tenure, Total Charges,
Contract and monthly charges as independent variables. The
model is evaluated based on the value of accuracy, error rate,
precision, recall and AUC value.
Fig. 8. C5.0 Results - Rcode
The accuracy was found to be 77.05% with error rate of
0.22. The values of precision and recall were found to be
0.81, 0.88 and the F1 value is 0.84.AUC value was found to
be 0.79 which falls under C category.Model of C5.0 and ROC
curve is displayed in figure9 and figure10 respectively.
NAIVE BAYES Naive Bayes algorithm is based on the
Bayes theorem of probability distribution. Nave Bayes was
implemented and the accuracy was found to be 69.90. The
confusion matrix is displayed in figure11.
The error rate was 0.30 and precision, recall and F1 values
are 0.88, 0.67 and 0.76 respectively.The value of area under

Fig. 9. C5.0 Model
Fig. 10. C5.0 Model ROC curve
Fig. 11. Naive Bayes Results RCode
the curve was found to be 0.79 which falls under C category.
The ROC curve of the model is displayed in figure12.
Decision Tree: Decision tree is a graphical representation
of all the possible solution to a decision. The main advantage
Fig. 12. ROC Curve of Naive Bayes model
of a decision tree is that it can be easily explained. We have
implemented the Decision tree algorithm and got the following
results. The confusion matrix is given below in the figure13.
The accuracy, error rate, precision, recall and F value were
Fig. 13. Decision Tree Results Rcode
found to be 77.96, 0.22, 0.82, 0.89 and 0.85 respectively.The
area under the curve value was found to be 0.79 which falls
under the C category.The Model of decision tree and ROC
curve is displayed in figure14 and figure15.
K-Nearest Neighbour: It is a supervised machine learning
classification model. The main advantages if KNN are it is
simple and effective and does not have any assumption. The
disadvantages of KNN are that it requires large amount of
memory and does not work well if the dataset has missing
values. We have implemented the KNN algorithm with K
= 1,5,10 and 30. The results obtained show the accuracy
increases with increase in K value. Figure16, Figure17 and
Figure18 show Rcode implementation of KNN model for K
value= 1, 5 and 30 respectively.
We can see that the KNN produces highest accuracy of
78.30% when the K value is 30.Comparison of Performance

Fig. 14. Decision Tree Model
Fig. 15. Decision Tree ROC curve
Fig. 16. KNN Model with K = 1 - Rcode
measure when values of K are 1,5,10 and 30 are displayed in
figure19.
Support vector machine (SVM): SVM is a supervised
method which can be used for classification and regression.
Fig. 19. Comparison of KNN model with different K values
SVM is good at handling non-linear and complex dataset. The
only disadvantage of SVM is that it is a Blackbox method. We
have implemented SVM and obtained the following results.
The Confusion matrix is displayed in figure20.
The accuracy, error rate, precision, recall and F value were
found to be 78.13%, 0.21, 0.83, 0.87 and 0.85 respectively.
Artificial Neural Network: ANN is a supervised algorithm
which is typically used for regression and classification. The
main disadvantages are it is difficult to understand as it is a
black box method and also takes long time for processing.
ANN was implemented and the results displayed in figure21
obtained. Accuracy of ANN was found to be 79%
The Model of ANN and ROC Curve is displayed in figure22
and figure23 respectively.

Fig. 20. SVM model Results - RCode
Fig. 21. ANN - Results
Fig. 22. ANN - Model
Business question 2:
Which model produce better results?
• Accuracy of Random Forest - 73.08%
• Accuracy of C5.0 - 77.05%
Fig. 23. ANN - ROC Curve
• Accuracy of Decision Tree 77.96%
• Accuracy of Naive Bayes - 69.90%
• Accuracy of KNN - 78.30%
• Accuracy of SVM - 78.13%
• Accuracy of ANN - 79.15%
ANN produces better accuracy compared to other
models. Model comparison of Random Forest,Decision
tree(CART),C5.0 and Naive Bayes is displayed in figure24
Fig. 24. ROC Curve comparison of RF,CART,C5.0 and NB
E. Deployment
This is the final phase of a CRISP DM Model where the
models are deployed in real life scenarios.
IV. CONCLUSION AND FUTURE WORK
In this project, we have classified the customer churn by
using machine learning algorithms such as random forest,
C5.0, Decision tree, KNN, ANN and SVM. We have also
found the importance of each variable in contribution to the
classification of customer churn variable. In future, we can
work with a larger dataset and also use various other attributes
to do the same task of classifying the customer churn.

REFERENCES
[1] S. Amaresan, “What is customer churn?” 2019. [Online]. Available:
https://blog.hubspot.com/service/what-is-customer-churn
[2] M. Karanovic, M. Popovac, S. Sladojevic, M. Arsenovic, and D. Ste-
fanovic, “Telecommunication services churn prediction-deep learning
approach,” in 2018 26th Telecommunications Forum (TELFOR). IEEE,
2018, pp. 420–425.
[3] H. Ma, M. Qin, and J. Wang, “Analysis of the business customer churn
based on decision tree method,” in 2009 9th International Conference
on Electronic Measurement & Instruments. IEEE, 2009, pp. 4–818.
[4] F. Guo and H.-L. Qin, “The analysis of customer churns in e-commerce
based on decision tree,” in 2015 International Conference on Computer
Science and Applications (CSA). IEEE, 2015, pp. 199–203.
[5] V. Effendy, Z. A. Baizal et al., “Handling imbalanced data in customer
churn prediction using combined sampling and weighted random forest,”
in 2014 2nd International Conference on Information and Communica-
tion Technology (ICoICT). IEEE, 2014, pp. 325–330.
[6] H. Li, D. Yang, L. Yang, X. Lin et al., “Supervised massive data
analysis for telecommunication customer churn prediction,” in 2016
IEEE International Conferences on Big Data and Cloud Computing
(BDCloud), Social Computing and Networking (SocialCom), Sustainable
Computing and Communications (SustainCom)(BDCloud-SocialCom-
SustainCom). IEEE, 2016, pp. 163–169.
[7] A. Amin, B. Shah, A. M. Khattak, T. Baker, S. Anwar et al., “Just-in-
time customer churn prediction: With and without data transformation,”
in 2018 IEEE Congress on Evolutionary Computation (CEC). IEEE,
2018, pp. 1–6.
[8] S. H. Dolatabadi and F. Keynia, “Designing of customer and employee
churn prediction model based on data mining method and neural
predictor,” in 2017 2nd International Conference on Computer and
Communication Systems (ICCCS). IEEE, 2017, pp. 74–77.
[9] A. Rodan, H. Faris, J. Alsakran, and O. Al-Kadi, “A support vector ma-
chine approach for churn prediction in telecom industry,” International
journal on information, vol. 17, no. 8, pp. 3961–3970, 2014.
[10] T. Chen and C. Guestrin, “Xgboost: A scalable tree boosting system,”
in Proceedings of the 22nd acm sigkdd international conference on
knowledge discovery and data mining. ACM, 2016, pp. 785–794.
[11] D. Do, P. Huynh, P. Vo, and T. Vu, “Customer churn prediction in an
internet service provider,” in 2017 IEEE International Conference on
Big Data (Big Data). IEEE, 2017, pp. 3928–3933.
[12] N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, “Smote:
synthetic minority over-sampling technique,” Journal of artiﬁcial intel-
ligence research, vol. 16, pp. 321–357, 2002.

Data Mining on Customer Churn Classification

Recommended

More Related Content

What's hot (19)

Similar to Data Mining on Customer Churn Classification (20)

More from Kaushik Rajan (8)

Recently uploaded (20)

Data Mining on Customer Churn Classification