Proceedings of the International Conference on Inventive Computing and Informatics (ICICI 2017)

IEEE Xplore Compliant - Part Number: CFP17L34-ART, ISBN: 978-1-5386-4031-9

Evaluation of Machine Learning Models for

Employee Churn Prediction
Dilip Singh Sisodia, Somdutta Vishwakarma, Abinash Pujahari
Department of Computer Science & Engineering
National Institute of Technology Raipur, India

Abstract— Employees are the valuable assets of any or occupation relinquishment. Attrition happens when an
organization. But if they quit jobs unexpectedly, it may incur employee resigns or when the organization eliminates his
huge cost to any organization. Because new hiring will consume occupation. The real contrast between the two is that when
not only money and time but also the freshly hired employees turnover happens, the organization looks for somebody to
take time to make the respective organization profitable. Hence
supplant the employee. In instances of attrition, the business
in this paper we try to build a model which will predict employee
churn rate based on HR analytics dataset obtained from Kaggle leaves the opportunity unfilled or wipes out that employment
website. To show the relation between attributes, the correlation job.
matrix and heatmap is generated. In the experimental part, the This paper is organized as follows. The next section
histogram is generated, which shows the contrast between left describes the related works done in the past and the motivation
employees vs. salary, department, satisfaction level, etc. For regarding this analysis. Section 3 will describe different
prediction purpose, we use five different machine learning machine learning algorithms used in this paper and their
algorithms such as linear support vector machine, C 5.0 Decision significance. Section 4 describes about the data set and also
Tree classifier, Random Forest, k-nearest neighbor and Naïve shows the statistical information using the data set. Section 5
Bayes classifier. This paper proposes the reasons which optimize
contains the detailed experimental results using the machine
the employee attrition in any organization.
learning algorithms using the mentioned data set, which will
Keywords—Turnover;JobSatisfaction;Attrition;Organization; be followed by the conclusion section.
employee retention strategy
II. LITERATURE SURVEY
I. INTRODUCTION Middle level officers are more likely to leave, may be due
Employee Attrition [1] is a reduction in manpower in any to some disagreement with their senior officer as proposed by
organization where employees may voluntarily leave the [3]. They observed major factors that influenced employee
organization or may be retired. Employee turnover is the abandonment from the firm. The two rules are moderately
number of existing employees replaced by new employees for derived by him. Some set of questions are asked with the both
a specific period. A high attrition causes high employee parties and depending upon their answers he concluded some
turnover in any organization. This in turn causes huge facts based on workload, objectives, carrier opportunity and
expenditure on human resource, by contributing towards new firm management. Human resource management [4]
recruitment, training and development of the freshly appointed endeavors on basically termination rates and dismissal rates
employees, also the performance management. Again, attrition but actual content of them are enormously different. The
[2] which are of voluntary is unavoidable. Hence, by previous model shows that, there are several distinct levels of
improving employee morale and providing a desirable attrition and turnover. Some research dictates that the
working environment, we can certainly reduce this problem consequences of dismissal and termination rates are at
significantly. organizational level. Allen & Meyer (1990) [5] described the
three-basic entity for the negative side of the turnover.
The rate of attrition is defined as the recruitment and
termination criteria of the company. An employee can leave Regulating officer will more probable leave from the
the job for various reason. Here, the ‘Turnover’ and ‘Attrition’ organization because of a contention with the higher
are the business terminologies that always conflicts each other. administration than a representative who is in struggle with his
There are various kinds of ‘turnover’ in an organization. prompt director. He recognized the determinant figures that
Lowering in number of employee is mainly considered as the influence employee acceptance [5] without protest from the
‘attrition’. To analyze the manpower data and other organization. Two arrangements of information social
measurements that are necessary for manpower planning these occasion techniques were directed. An equivalent number of
terminologies can be interchangeably used. When an representative and officer respondents were solicited to answer
employee leaves the company both attrition and turnover a set from polls that were ordered by workload, objectives,
happens. Turnover, be that as it may occur because of various identity, professional success, and hierarchical administration.
work activities, for example, release, termination, abdication The after-effects of the two information gathering methods

978-1-5386-4031-9/17/$31.00 ©2017 IEEE 1016

 Proceedings of the International Conference on Inventive Computing and Informatics (ICICI 2017)
IEEE Xplore Compliant - Part Number: CFP17L34-ART, ISBN: 978-1-5386-4031-9

demonstrated that the most noteworthy component that adds to IV. DATA SET ANALYSIS
employee rejection [6] is money related compensation. The ‘HR Analytics’ data set [13], obtained from Kaggle
Website, is used in this paper for the experimental verification.
III. MACHINE LEARNING ALGORITHMS
This data set comprises ten attributes and 15000 tuples.
There are various kind of machine learning techniques
available to learn from the given data which is called train data.
When new or unseen data arises the learned model analyses
and predict desired class. In our experiment we have used the
HR Analytic data set to apply various machine learning
algorithms to predict the chances of employees to quit the job.
The machine learning algorithms for predicting the same are
described below.
A. K-Nearest Neighbour
k-NN classifier [7] is known as lazy learner in machine
learning community. It never learns from the data and do not
build any models. Rather, it finds out the examples from the
train dataset which are closest to the unknown example. Based
on the neighbor examples it will predict the new example. The
value of ‘k’ determines the no. of closest data points or
examples to be selected from the training example.
B. Supprt Vector Machine
A Support Vector Machine [8] is a kind of classification
technique, where the data points are separated by a line in case
of linear SVM, and a hyperplane in case of non-linear SVM. Fig.1. Correlation Matrix
The separation is chosen in such a way that; the two sides of
the hyperplane categorizes the data set in to two classes. When
an unknown data comes it predicts which side/class it belongs
to. The margin between the hyperplane and the support
vectors are as large as possible to reduce the error in
classification.
C. Naïve Bayes Classifier
Naive Bayes [9] is a popular classification technique which
classifies examples based on the probability of chances that are
likely to be occurred. It often performs very well for complex
data set which are very hard to learn using the traditional
learning algorithms.
D. Decision Tree
This is one of the popular learning techniques in machine
learning. C 4.5 [10] is the benchmark learning algorithm in
decision tree which is often compared with the new algorithms
that are being developed. Here C 5.0 [11] learning algorithm is
used which is an advanced version of traditional decision tree
learning algorithms. The nodes along with the edges are the
series of conditions and the leaves are the class labels.
E. Random Forest Fig. 2. Histogram of employee status and satisfaction level
It is one of the ensemble learning technique [12] which The categorical values are converted to numeric values in order
consists of several decision tree rather than a single decision to make the classification algorithm more efficient. For
tree for classification. While classifying all the trees in the example, categorical attribute ‘salary’ contains three values
random forest gives a class to an unknown example and the such as low, medium and high. Hence it is converted to 0, 1
class having maximum votes will be assigned to the unknown and 2 respectively. The misspelled attributes are also corrected.
example. Figure (1) represents the correlation matrix which helps to
identify attributes with the strong or weak correlation.
Figure (2) represents the histogram of employee status and
satisfaction level [14] . It can be seen from the figure that, there
are three segments or behaviors.

978-1-5386-4031-9/17/$31.00 ©2017 IEEE 1017

 Proceedings of the International Conference on Inventive Computing and Informatics (ICICI 2017)
IEEE Xplore Compliant - Part Number: CFP17L34-ART, ISBN: 978-1-5386-4031-9

x First bin in the histogram is empty. Second and third

bins mostly contain people who left the company.
These people must be unhappy for some reason as
their satisfaction level is below 1.15, i.e. (1st Segment)
25.4% of the total employees who left.
x Then there is a peak around 0.4 (2nd Segment) i.e.
43.8% of total employees who left.
x Lastly, there is a chunk 0.7-0.95 (3rd segment) who
left the company, i.e. 26.1 % of total employees who
left. That’s almost 1/4th of employees who left.
Figure (3) represents the histogram of comparison among
the departments with respect to employee attrition and
turnover. From the first graph of this figure tell that, Turnover
rate (percentage of churned employees) of department of ‘R &
D’ and ‘Management’ are the lowest. When it comes to
employee churn, ‘HR’ department is having the highest with
29% and rest of the departments are roughly the same. Again,
from the second and third graph of this figure it can be seen
that highly paid employees are less likely to leave the job
which makes sense.
From Figure (4), only 5 % of the employees who got
promoted churned away, as compared to 24% of employees
who were not promoted and left the company. Hence
‘promotion’ is also an important factor of retaining employees.
Also, it can be seen from Figure (4) that, employees with
projects ‘2’, ‘6’ and ‘7’ left the company most which represents
most of the employees quit if the work load increases.
However, there is an inconsistency where number of projects is
equal to 2, which may be due to the nature of the job in project
or other. Again, employees, who are there in the company for a
long time are less likely to quit as compared to the ones who
have joined recently, which can be seen from the graph ‘time
spent’ from Figure (4).
From Figure (5), employees with lower ‘Last-Evaluation’
left the company. But it is a little ambiguous why employees
with high ‘Last-Evaluation’ left while employees with 0.6-0.8
evaluation score tend to stay. Employees with heavy working
hours also tend to leave the company, which is obvious.
Using all these attributes and precondition basically employee
churn rate can be generalized to the following criteria:
x Employees with no promotion in last 5 years tend to
leave.
x Employees having high salary but no promotion and
high working hours also tend to leave.
x Employees getting promotion but not hike in salary Fig. 3. Histogram of department and salary
also tend to leave.

978-1-5386-4031-9/17/$31.00 ©2017 IEEE 1018

 Proceedings of the International Conference on Inventive Computing and Informatics (ICICI 2017)
IEEE Xplore Compliant - Part Number: CFP17L34-ART, ISBN: 978-1-5386-4031-9

For evaluating the experimental results, ‘Confusion

Matrix’ is used which is a common evaluation criterion for
any classification model. Using this the parameters like
Accuracy, Precision, Recall and F-Measures are used and the
corresponding values obtained through experiment is
displayed in Table 1 with respect to different learning
techniques.
It can be seen from Table 1 that, Random Forest classifier
gives the highest accuracy on the given HR analytics data set
while, LSVM classifier gives the lowest accuracy for the same
data set. The same result is also displayed in Figure 6 in terms
of graphs. From Figure 6 we can conclude that Random Forest
again gives the highest precision i.e. true positive rate.
Likewise, Random forest also performing well for other
measures like Sensitivity or Recall, F-Measure, Specificity,
FPR and FNR as compared to other classifiers. LSVM
classifier gives the lowest sensitivity and Naïve Bayes gives
the lowest F-Measure for the given data set.

TABLE I. RESULTS OF DIFFERENT CLASSIFIERS

Attribute/ Naïve Decision Random
KNN LSVM
models Bayes Tree Forest
Fig. 4. Histogram of accident, promotion, project and time spend
Accuracy 0.9600 0.7821 0.7918 0.9768 0.9897
Precision 0.9674 0.9457 0.8081 0.9796 0.9981
Sensitivity or
Recall or
0.9799 0.8032 0.9085 0.9898 0.9886
True positive
rate
F-measure 0.9736 0.8686 0.8554 0.9846 0.9933
Specificity or
True negative 0.8997 0.5979 0.5463 0.9368 0.9936
rate
False positive
0.1003 0.4021 0.4537 0.0632 0.0064
rate
False
0.0201 0.1968 0.0915 0.0102 0.0114
negative rate

Fig. 5. Histogram of last evaluation and average monthly hours

V. EXPERIMENTAL RESULSTS AND EVALUAITON

In the data set mentioned above, the attributes like
employee’s salary, current position, promotion etc. is given.
Based on these values the learning algorithm [15] will predict Fig. 6. Performance Comparison of Different Classifiers
whether the employee will quit the organization or not. The
predicted value is compared with the actual value in the
database.

978-1-5386-4031-9/17/$31.00 ©2017 IEEE 1019

 Proceedings of the International Conference on Inventive Computing and Informatics (ICICI 2017)
IEEE Xplore Compliant - Part Number: CFP17L34-ART, ISBN: 978-1-5386-4031-9

minimize that. This study requires further exploration to

minimize the prediction error rate.
REFERENCES
[1] R. Punnoose and P. Ajit, “Prediction of employee turnover in
organizations using machine learning algorithms,” Int. J. Adv. Res.
Artif. Intell., vol. 5, no. 9, 2016.
[2] D. Alao and A. B. Adeyemo, “Analyzing employee attrition using
decision tree algorithms,” Comput. Inf. Syst. Dev. Informatics Allied
Res. J., vol. 4, 2013.
[3] O. Ali and N. Z. Munauwarah, “Factors affecting employee
turnover in organization/Nur Zuhan Munauwarah Omar Ali,” 2017.
[4] A. Frederiksen, “Job Satisfaction and Employee Turnover: A firm-
level perspective,” Ger. J. Hum. Resour. Manag., vol. 31, no. 2, pp.
132–161, 2017.
[5] H. Ongori, “A review of the literature on employee turnover,” 2007.
[6] V. V. Saradhi and G. K. Palshikar, “Employee churn prediction,”
Expert Syst. Appl., vol. 38, no. 3, pp. 1999–2006, 2011.
Fig. 7. ROC Curve for different classifiers [7] L. E. Peterson, “K-nearest neighbor,” Scholarpedia, vol. 4, no. 2, p.
1883, 2009.
From Table 1, using false positive rate (FPR) and true [8] C. Cortes and V. Vapnik, “Support vector machine,” Mach. Learn.,
positive rate (TPR) we have plotted the receiver operating vol. 20, no. 3, pp. 273–297, 1995.
characteristic (ROC) curve in Figure 7 for different classifiers. [9] K. P. Murphy, “Naive Bayes classifiers Generative classifiers,”
The higher the area under the curve, greater is the accuracy of Bernoulli, vol. 4701, no. October, pp. 1–8, 2006.
[10] D. Gupta, D. S. Kohli, and R. Jindal, “Taxonomy of tree based
the classifiers. classification algorithm,” in Computer and Communication
Technology (ICCCT), 2011 2nd International Conference on, 2011,
VI. CONCLUSION pp. 33–40.
This paper find out which machine learning algorithm is [11] A. S. Galathiya, A. P. Ganatra, and C. K. Bhensdadia,
“Classification with an improved decision tree algorithm,” Int. J.
performing well in predicting the employees, those are likely Comput. Appl., vol. 46, no. 23, pp. 1–6, 2012.
to quit the respective organization based on their working [12] L. Breiman, “Random Forests,” Mach. Learn., vol. 45, no. 5, pp. 1–
details and environments. From the experimental results, 35, 1999.
Random Forest is clearly outperformed all other classifiers as [13] Kaggle, “HR Analytic Data set.” [Online]. Available:
https://www.kaggle.com/ludobenistant/hr-analytics.
obtained in evaluation criteria. This study might help [14] K. L. Latha, “A study on employee attrition and retention in
organization what are the factors causing the employee manufacturing industries,” BVIMSRs J. Manag. Res., vol. 5, no. 1,
leaving the organization and can take appropriate steps to pp. 1–23, 2013.
[15] A. Rawat and A. Choubey, “A Survey on Classification Techniques
in Internet Environment,” IJSRSET, vol. 2, no. 3, 2016.

978-1-5386-4031-9/17/$31.00 ©2017 IEEE 1020