Employee turnover prediction and retention policies design: a case

study.
Edouard Ribes∗¶ Karim Touahri† Benoı̂t Perthame‡
arXiv:1707.01377v1 [cs.CY] 5 Jul 2017

July 6, 2017

Abstract
This paper illustrates the similarities between the problems of customer churn and employee
turnover. An example of employee turnover prediction model leveraging classical machine learning
techniques is developed. Model outputs are then discussed to design & test employee retention
policies. This type of retention discussion is, to our knowledge, innovative and constitutes the main
value of this paper.

2010 Mathematics Subject Classification.

Keywords and phrases. Churn prediction; Machine learning techniques; Employee Turnover; Classifi-
cation; Retention Policy;Workforce Planning.

1 Introduction
Machine learning algorithms are often showcased in customer churn study. Applications in fields
such as telecommunication or product marketing (gaming, insurance etc..)(see [1],[2] for a recent
review) are multiple. The implementation of these methods in Customer Relationship Management is
becoming the new norm, as improving customer retention yields superior business results. We argue
that this type of techniques can easily be applied to employee turnover. Note that the employee
turnover can actually be subdivided in 3 buckets: involuntary turnover (induced by the company),
voluntary turnover (employee resignation) and retirements. Retirement and an involuntary turnover
are out of the scope of this paper. Retirement is indeed generally legally enforced through specific
local schemes and does not require a prediction. As for firing decisions, they are a company decision
aiming at achieving the right workforce sizing or productivity levels. Therefore they do not require
any anticipation on workers behavior. On the other hand, voluntary turnover is employee dependent.
Mitigating it can be of interest to companies. This is even, in our opinion, one of the key workforce
dynamic to understand in order to manage a company in a sustainable fashion (see [3] for example).
Assume that a company Y has a portion of its workforce in job j that is deemed critical to achieve
business results. Assume the job j participates in a critical part of a product manufacturing process
and is extremely difficult to source because of the required expertise and associated training time.
∗
Ecole Polytechnique,91128 Palaiseau, France,Email: [email protected]
†
Universit 5 Descartes, Paris 13 Sorbonne City,France,Email: [email protected]
‡
Sorbonne Universités, UPMC Univ Paris 06, Laboratoire Jacques-Louis Lions UMR CNRS 7598, Inria, F75005 Paris,
France

1
Delays and/or failure in the manufacturing process can be extremely costly for the company, so that
retention efforts may have to be considered to mitigate production risks.
Retention is part of the standard Human Resources [HR] activity which is structured around the
”5 Bs”, namely : ”Buy” and ”Borrow”(=hire), ”Build”(=train), ”Bind”(=retain), ”Bounce”(=fire).
Retention investments usually factor individual performance, potential to grow in a company and
are conjugated with a leave risk assessment. Let’s assume that performance and potential are already
assessed in a satisfactory fashion but that the leave risk is currently assessed by the employee’s manager
based on personal intuition. This managerial assessment is in reality a classification problem. The
manager will indeed flag each year its employees has having either a ”High Risk” (the employee
will leave the company this year) or a ”Low Risk” (the employee will not leave the company this
year) of departure. Assume that the company Y does not trust the managerial assessments because
historically, they poorly correlate to the observed voluntary turnover fluxes. The company would
therefore be interested in developing a prediction routine to structure its retention efforts.
The use of machine learning techniques for turnover prediction purposes has risen lately (see[4] and
[5] for recent examples). Yet, to our knowledge, nothing has been done to design and test retention
policies. In this paper, we will therefore leverage HR domain knowledge to build a turnover predictor
and test retention policies. This paper will first explain how features can be engineered to feed a
machine learning algorithm. In a second part, several machine learning methods and training sample
techniques will be discussed and analyzed in terms of performance. Finally a sensitivity analysis will
be carried out to design and test specific retention policies.
Important Legal Remark. The findings and opinions expressed in this paper are those of
the authors and do not reflect the positions from any company or institution. Finally, please bear in
mind that for confidentiality reasons numbers have been disguised in a way that preserves the same
conclusions as the actual case study.

2 Feature Engineering.

2.1 Previous Work.

Turnover is not easy to predict because it results from a combination of elements. So far, no con-
sensus has been reached in terms of key elements to use. For example, an early review of voluntary
turnover studies [6] has found that the strongest predictors for voluntary turnover were age, tenure,
pay, overall job satisfaction, and employees perception of fairness. But other studies have also stressed
the importance of job performance [7], job characteristics (role, seniority in role...) [8], enhanced indi-
vidual demographic characteristics (age/experience, gender, ethnicity, education, marital status) ([9],
[10], [11], [12], [13], [14]), structural characteristics (i.e. team size and performance) and geographical
factors ([15] , [16]).Finally some studies have also set an emphasis on salary, working conditions, job
satisfaction, supervision, advancement, recognition, growth potential, burnout etc. [17], [18],[19], [20].
And our list is far from being exhaustive...
To our understanding, the key to successfully engineer features in machine learning exercises revolves
around flexibility. Companies Human Resource Information Systems [HRIS] are famous for data incon-
sistency and quality issues. The absence of data standards thus calls for company specific engineering
tasks.

2
2.2 Turnover Domain Knowledege & Data.
The dataset available for this study was discussed with a pool of peer-trusted experts among our
clients group. The purpose of this discussion was to hand pick features that, in light of the previous
bibliographical work, would be relevant for the given job j and in the current historical context.
Leveraging peer trusted experts was important because it enabled us to infuse domain knowledge to
the project while creating buy-in in our client group for this kind of methodology. This was also
critical as it helped us remove features that were considered non reliable. The selected features were:

• Localization. Individuals from a given country were selected. This country was described by
a categorical feature representing the regional area where the employee was working. This was
used to account for local labor market specific dynamics.

• Knowledge. In order to account for specific domain expertise among job category j, the type of
business unit in which the individual was working was added to the feature mix.

• Individual data. Several features were suggested as far as individuals are concerned. First, time
in the current position and tenure with a given company were added. Demographics in terms
of age and gender were also highlighted. Additional elements specific to the dataset, such as
performance & potential evaluation, were discussed. Hierarchical level according the dataset
grading structure was finally added to the mix.

• Individual managerial data. Managers demographic information such as age and gender were
added, as well as managers time in the current position and tenure in a given company. Manager’s
performance over the last year and average performance over the last 3 years were also integrated.

• Individual team data. Team size and team percentage of high and low performers were added
to the prepared mix.

The population under study in the given job category j accounted for about 1000 employees in a
given country. 2 years (referred to as year 1 and year 2) of data were available. The problem was of
a classification nature as employees were either tagged as ”Terminated” or ”Active”. Note that the
dataset was curated of individuals who left involuntarily or because of retirement to respect the scope
of the study.
The full dataset was heavily unbalanced as the yearly number of people in job j leaving voluntarily
was about 20%.It was randomly split in a training and a test set, which were of equal size. Each of the
two sets contained the same proportion of year 1 and year 2 information. This choice indeed yielded
more robust performances than having the algorithms trained on one year and tested on the other.

2.3 Feature Selection.

After this pre-processing, a selection layer was deployed in order to keep only the most relevant
elements to the current problem out of the ones mentioned by the experts and the literature. Limiting
the number of features used in machine learning problems is indeed crucial to properly manage the
complexity of the learning phase and to avoid over fitting. In this paper, the mutual information (MI)
criterion was used to assess the information content of each individual feature. This type of feature
selection method is indeed well acknowledged and has proven useful for such topics (see [21] for an
early example, [22] for a customer churn example and [23] for a review.).

3
Ranking with Mutual Information (MI). Mutual Information (MI) is an essential metric of
information and has been widely used for quantifying the mutual dependence of random variables (see
[24] for details).Formally the MI between an output variable Y and an input variable X is defined as:

XX P (x, y)
M I(X; Y ) = P (x, y). log (1)
P (x).P (y)
x∈X y∈Y

where P (x) (resp. P (y)) is the probability associated to X (resp. Y ), and P (x, y) is the joint proba-
bility of having both X and Y .

Features Filtering. Several possibilities exist in order to select features. According to a recent
survey [25], they can be summed up in three types of methods: filters, wrappers and embedded ones.
Filter methods preprocess and rank features in order to keep only the best ones. In wrapper meth-
ods, the feature selection criterion is the performance of the predictor itself. The predictor is indeed
wrapped on a search algorithm which will find a subset that gives the best predictor performance.
Finally embedded methods include variable selection as part of the training process without splitting
the data into training and testing sets.

In this case, a filter method was used based on

the previously described mutual information crite-
rion. This choice was motivated by its simplicity
and the empirical efficiency of the approach. For
each feature X, the mutual information M I(X, Y )
was compiled and only the top 60% features were
kept. Mutual information results are displayed be-
low. Among the features of demographic nature,
employee’s gender and age as well as manager’s
gender were removed. Company bonus level infor-
mation and structural hierarchical details proved
useful but required granularity to be of use. Team
size information disappeared. It is interesting to
see that with the current dataset; only employee
behavioral assessments relates to the turnover de-
cision. Further investigation would be needed on
this aspect of the dataset.
Fig 1. Features Mutual Information

3 Machine Learning Methods.

The analytical techniques deployed rely on well-known standards such as SVM, random forests or
KNN that have been proposed to predict turnover (see [26] or [4] for examples). They have been
adjusted to account for class imbalance and potential associated biases.

4
3.1 Class Imbalance Correction.
Class Imbalance is a common theme on customer churn prediction ([27], [28]...), which is a topic similar
to the one developed in this paper. Standard solutions involve subsampling techniques to smooth the
disparities between the observed classes. In this paper, several methods were tested:
• Down-sampling: sample the majority class to make its frequency closer to the rarest class.
• Up-sampling: resample the minority class to increase its frequency.
• Weighting: place a heavier penalty on misclassifying the minority class [29] by assigning a weight
to each class, with the minority class being given a larger weight (i.e., higher misclassification
cost).
• ”SMOTE”: over-sample the minority (rarest) class and under-sample the majority class ([30]).
• ”ROSE”: generate new artificial data from the classes to correct class imbalance according to a
smoothed bootstrap approach [31].

3.2 Selected Classification Methods.

The classification problem was approached with the following techniques:
• Naive Bayes (NB) classifiers (see [32] for an application) calculate the probability that a given
input sample belongs to a certain class. Assuming a number of classes Y = (y1, y2, , yk) and a
given a unknown sample X, a Naive Bayes classifier assigns the sample X to the class with the
highest probability C = ArgM axy∈(y1...yk) (p(y|X))
• Linear Discriminant Analysis (LDA) is a generalization of Fisher’s linear discriminant analysis.
It can be explained as computing a z-score, which is then used to estimate the probabilities that
a particular member or observation belongs to a class.
• Support vector machines (SVMs) achieve class separation by finding (in their linear version) a
hyper plane in a high dimensional space. The intuition is that a good separation is achieved by
the hyper plane that has the largest distance to the nearest training data points of any class. The
larger the margin is, the lower the generalization error of the classifier. For this reason, SVMs
are also referred to as maximum margin classifiers. Multiple versions exists (linear, polynomials,
gaussian...) (see [33])
• Random Forests (RFs) [34] result from the combination of tree classifiers. Each tree depends in
the values of a random vector sampled independently and with the same distribution for all trees
in the forest. The generalization error of a forest of tree classifiers depends on the strength of
the individual trees in the forest and the correlation between them. RF has empirically achieved
very good classification performance in numerous cases and is robust against over-fitting.
Note that the R Caret package [35] was the library used to support the study.

3.3 Evaluation & Results.

Cross-Validation. The dataset was split 80:20 into training and hold out sets in a 10-fold cross
validation scheme. For each algorithm, a grid-search was performed over the tuning parameters,
including regularization or penalty hyper-parameters.

5
Performance Metric. The Area under the receiver operating characteristic curve (ROC-AUC) was
the chosen performance measure. The AUC is a general measure of accuracy. It decouples classifier
assessment from operating conditions i.e., class distributions and misclassification costs [36]. It is
better suited to non-balanced classification problems than standard accuracy or error test as it allows
a graphical representation of specificity (false positive rate) and selectivity (true positive rate) of the
classifiers on the dataset. Note that, in this case, the specificity can be interpreted as the percentage
of workers flagged as leavers that really left, while the selectivity represents the percentage of workers
appropriately flagged that stayed in their company.

Results. We obtained the following results:

Table 1. Classification Performance among
resampling methods.
In terms of imbalance correction, we found that
up sampling and ROSE sampling yielded the best Sampling Best Algorithm ROC Spe. Sel.
performance in this case. As expected, when imbal- None LDA 0.73 0.94 0.19
ance was not corrected, algorithms tended to favor Down LDA 0.74 0.61 0.67
the majority class to achieve the best performances. Up RF 0.96 0.82 0.96
Rose RF 0.95 0.81 0.96
Smote RF 0.87 0.86 0.74

In terms of algorithms, we saw that the tree based

ones performed best. The following performances
were indeed recorded with ROSE resampling:

Table 2. Classification Performance among

algorithms.
Algorithm ROC Specificity Selectivity
LDA 0.75 0.67 0.74
SVM (Radial) 0.80 0.69 0.76
RF 0.95 0.81 0.96
Treebag 0.94 0.78 0.94

The corresponding ROC Curves have been reported

to the left.
Fig 2. ROC Curves with ROSE resampling

6
As far as the variables of importance are con-
cerned, we saw that the employee behavior and
performance (especially when rated low) are im-
portant elements to the turnover behavior predic-
tion. Structural elements in terms of compensation
schemes and business units specificities were also
rated high. The same goes for certain seniority in
company and time in position year bands.
According to the graph ??, new joiners in the com-
pany are at risk. So are people that have been in
their roles for more than 4 years. This last element
may relate to a perceived lack of career opportunity.
This seems to be corroborated by the importance
of managerial time in position. Interestingly man-
agers importance in turnover decisions appears to
be mainly related to their seniority in the company
or their time on the job. Manager performance or
behavior does not rate extremely high.
Another interesting element is that remote jobs do
not seem to be driving more turnover than other of-
fices location. This would hint to local labor market Fig 3. Varibale Importance (RF with Rose
specificities. Resampling)
In this section, it has been showed that it possible to model voluntary turnover with a couple of
standard variables. Even though this may not hold for all types of jobs, the announced performance
proved replicable on the prepared test set. A discussion about the model use to deliver insights on
retention policies is now in order.

4 Retention Discussion.
According to recent reviews in the field of employee retention (see [37], [38]), the results from the
previous section are not surprising. A number of generic actions to bind people to firms have already
been discussed over the years. Compensation topics, including rewards and recognition schemes are
among the most popular (see [39] for an example). Other elements such as opportunity for growth &
development, accountability and leadership structure have also been stressed. Some further elements
regarding the workplace and work life balance have been raised too.
But as stressed by the literature reviews retention efforts should be tied to an in depth turnover
analysis, which will be the aim of this final part of our study. The previous results motivated us
to test five simple possible programs. The first two are about reviewing the job locations: in the
program P 1, remote jobs are reassigned to the location 1, while in P 2 jobs in location 3 are reassigned
to location 1. The next ones are about the managers for the considered labor categories. In the
program P 3, all managers were forced to have a first internal experience prior to their managerial
role. This translated in reallocating all managers with company tenure between 0 and 2 years to a 3
to the 7 years band. In the program P 4, managers were assumed to rotate frequently between team
so that their time in position would be kept below 2 years. Finally we tested an employee program
P 5 that assume that people were bound to the company during their first 2 years in position.

7
 Table 1: Retention Mass Actions Simulations.

Program Type Program Details % of leavers

None No Retention Policy 41.5%
P1.Location Remote Job reassigned to Location 1 40.9%
P2.Location Location 3 reassigned to Location 1 41.3%
P3.Manager People Managers had 1 role internally 39.7%
P4.Manager Manager Rotation Program 32.3%
P5.Employee Bind newly assigned people 30%

4.1 Retention Policy Simulations.

The five programs mentioned (P 1 to P 5) were tested on a prediction set, which was built from the
current year of data where no turnover indication was available. It was therefore different from both
the test and training set mentioned earlier. Size and characteristics of the prediction set were similar.
Certain parameters of the prediction set were altered to perform the sensitivity analysis required to
simulate the retention programs as previously described. Two types of policies were investigated. It
is indeed either possible to apply the program P 1 to P 5 to all the people in scope of the study or to
target only the individuals flagged as leavers and adopt a policy that suits them best.

Mass Actions. As far as mass actions are concerned, the following table summarizes our findings:
The results displayed above show that programs related to internal mobility are among the most
efficient options to reduce turnover. If the program P 5 appears the most efficient, it also is the
most difficult to set in place and offers little upside compared to the managerial rotation program.
Interestingly location change offers little value from a retention standpoint. This could be explained
in this case by the fact that the labor category at stake can work in a remote fashion. This indeed
creates a country wide labor market to compete in.

Targeted Actions. According to the algorithm results, about 40 % of the population is likely
to leave and requires an action. The investigated course is now to test for each of the supposed
leavers which of the programs(P 1 to P 5) is useful. This will limit the amount of effort required while
optimizing the employee retention. This leads to a final % of leavers of 24%, which represents a
significant mitigation of turnover. The results in terms of actions are summarized in the table below:
These results show that to halve the turnover rate, actions only need to be started for 18% of the
total population. In the end, about 25% of the population at risk of departure needs a new manager
and 11% require an incentive to stay over the first two years of their role. This highlights solutions
that the Human Resource organization and notably the talent management centers of expertise could
deploy in in order to successfully contribute to ones organization management.

4.2 Discussion.
Several limitations to this study exist and are listed below:

• Model Limits. First, we did not have a holistic vision on individual parameters. For instance,
detailed compensation parameters were missing. It is completely possible that fed with another

8
 Table 2: Retention Targeted Actions Summary.

Action Type % of the total % of the potential

Population concerned leavers concerned
None 82,42% 57,67%
P1.Location 0,88% 2,12%
P2.Location 0% 0%
P3.Manager 1,54% 3,70%
P4.Manager 10,55% 25,40%
P5.Employee 4,62% 11,11%

set of features, the algorithms would have yielded better performances and different interpre-
tations, for instance in terms of importance ranking. The story developed in this study case
results from insights driven discussion between experts. We argue that if advanced analytics are
important, the key elements here were the discussions that resulted in a general buy in of the
retention policies simulations.

• Emotional Component in the job relationship. The discussions held with our sounding
board raised another issue. Turnover may stem from a mix of rational and emotional components.
The dataset was mainly organized around standard organizational or labor indicators. Little to
no indication regarding people profiles, behaviors or preferences was available.

• Stress Supply vs Demand Approach. The developed approach works well to propose
turnover mitigation measures. However we believe that it should not be the first thing to
do. Instead we recommend leading the turnover and retention discussion with thorough job
level productivity discussion. This indeed helps companies navigate in the supply & demand
type of environment that are labor markets and therefore helps them define what an acceptable
turnover rate is. Productivity will indeed help defining the cost of the turnover related disrup-
tion while the supply demand equilibrium will help finding the right solution mix in terms of
staffing vs retention efforts. We personally believe that the human resource function is a key
operating body of companies. Yet it needs to tie its activity and services to tangible business
outcomes (sales, cost, productivity) to prove its value. This wasn’t stress enough in our study
and a review of productivity definition across functions appears mandatory to standardize this
type of approach.

5 Conclusion & Next Steps.

In this paper, a method to approach employee retention has been proposed using standard machine
learning techniques. In a first part, a literature review has shown how close this problem is from the
customer churn one. Yet, to our knowledge, little has been done regarding employee turnover from a
quantitative standpoint. This generated a discussion with experts on how to tackle this problem and
generatemeaningful features. Those features were then fed to standard machine learning algorithms
under various resampling schemes to account for classes imbalance. The best performing algorithm
was retained to predict turnover and a sensitivity analysis was led to understand the effects of several
retention policies. We found that for the data at hand, the most efficient retention practice was to

9
develop talent mobility across positions.
The principal next step to this study would be to include detailed compensation elements to the feature
mix. The current model is indeed more around the employee perception of its internal environment
than about rational economic tradeoffs. We indeed believe that such models have the potential to
mimic an individual job related utility function.

References
[1] E. Ngai, L. Xiu, and D. Chau, “Application of data mining techniques in customer relationship
management: A literature review and classification,” Expert Systems with Applications, vol. 36,
no. 2, Part 2, pp. 2592 – 2602, 2009.

[2] T. Vafeiadis, K. Diamantaras, G. Sarigiannidis, and K. Chatzisavvas, “A comparison of machine

learning techniques for customer churn prediction,” Simulation Modelling Practice and Theory,
vol. 55, pp. 1 – 9, 2015.

[3] M. Doumic, B. Perthame, E. Ribes, D. Salort, and N. Toubiana, “Toward an integrated workforce
planning framework using structured equations,” European Journal of Operational Research, pp. –
, 2017.

[4] R. Punnoose and P. Ajit, “Prediction of employee turnover in organizations using machine learn-
ing algorithms,” International Journal of Advanced Research in Artificial Intelligence(ijarai),
vol. 5(9), 2016.

[5] H. yun Chang and H. yun Chang, “Employee turnover: a novel prediction solution with effective
feature selection,” pp. 252–256, 2009.

[6] J. L. Cotton and J. M. Tuttle, “Employee turnover: A meta-analysis and review with implications
for research,” The Academy of Management Review, vol. 11, no. 1, pp. 55–70, 1986.

[7] A. Susskind, C. Borchgrevink, K. Kacmar, and R. Brymer, “Customer service employees’ be-
havioral intentions and attitudes: An examination of construct validity and a path model,”
International Journal of Hospitality Management, vol. 19, pp. 53–77, 3 2000.

[8] N. S. K. Hasan Zarei Matin and M. R. A. Anvari, “Do demographic variables moderate the
relationship between job burnout and its consequences?,” Iranian Journal of Management Studies
IJMS, vol. 5, no. 1, pp. 47–62, 2012.

[9] I. GABARRET, “Une analyse exploratoire du departvolontaire des cadres: Le cas de la sortie
entrepreneuriale hors essaimage,” p. 22.

[10] K. M. R. E. K. L. M. Finkelstein, “What do the young (old) people think of me? content and ac-
curacy of age-based metastereotypes,” European Journal of Work and Organizational Psychology,
vol. 22(6), pp. 633–657, 2013.

[11] T. L. B. Holtom, T. Mitchell and M. Eberly, “Turnover and retention research: A glance at the
past, a closer review of the present, and a venture into the future,” Academy of Management
Annals, vol. 2, pp. 231–274, 2008.

10
[12] E. K. K. C. von Hippel and J. D. Henry, “Stereotype threat among older employees: Relationship
with job attitudes and turnover intentions,” Psychology and aging, 2013.

[13] S. L. Peterson, “Toward a theoretical model of employee turnover: A human resource development
perspective,” Human Resource Development Review, vol. 3(3), 2004.

[14] J. M. Sacco and N. Schmitt, “A dynamic multilevel model of demographic diversity and misfit
effects,” Journal of Applied Psychology, vol. 90(2), 2005.

[15] C. Tanova, “Using Job Embeddedness Factors to Explain Voluntary Turnover in Five European
Countries,” IRISS Working Paper Series 2006-04, IRISS at CEPS/INSTEAD, July 2006.

[16] S. M. Carraher, “Turnover prediction using attitudes towards benefits, pay, and pay satisfac-
tion among employees and entrepreneurs in estonia, latvia, and lithuania,” Baltic Journal of
Managementt, vol. 6, pp. 25 – 52, 2001.

[17] D. G. Allen and R. W. Griffeth, “Test of a mediated performance turnover relationship highlight-
ing the moderating roles of visibility and reward contingency,” Journal of Applied Psychology,
vol. 86(5), 2001.

[18] T. W. L. B. C. H. D. Liu, T. R. Mitchell and T. R. Hinkin, “When employees are out of step
with coworkers: How job satisfaction trajectory and dispersion influence individual-and unit-level
voluntary turnover,” Academy of Management Journal, vol. 55(6), 2012.

[19] B. W. Swider and R. D. Zimmerman, “Born to burnout: A metaanalytic path model of person-
ality, job burnout, and work outcomes,” Journal of Vocational Behavior, vol. 76(3), 2010.

[20] T. M. Heckert and A. M. Farabee, “Turnover intentions of the faculty at a teaching-focused

university,” Psychological reports, vol. 99(1), 2006.

[21] R. Battiti, “Using mutual information for selecting features in supervised neural net learning,”
IEEE Transactions on Neural Networks, vol. 5, pp. 537–550, Jul 1994.

[22] S. Moldovan, E. Muller, Y. Richter, and E. Yom-Tov, “Opinion leadership in small groups,”
International Journal of Research in Marketing, pp. –, 2016.

[23] J. R. Vergara and P. A. Estévez, “A review of feature selection methods based on mutual infor-
mation,” Neural Computing and Applications, vol. 24, no. 1, pp. 175–186, 2014.

[24] T. M. Cover and J. A. Thomas, Elements of Information Theory, vol. 2nd Edition. 2006.

[25] G. Chandrashekar and F. Sahin, “A survey on feature selection methods,” Computers and Elec-
trical Engineering, vol. 40, no. 1, pp. 16 – 28, 2014. 40th-year commemorative issue.

[26] A. M. Esmaieeli Sikaroudi, R. Ghousi, and A. Sikaroudi, “A data mining approach to employee
turnover prediction (case study: Arak automotive parts manufacturing),” Journal of Industrial
and Systems Engineering, vol. 8, no. 4, pp. 106–121, 2015.

[27] K. W. Bowyer, N. V. Chawla, L. O. Hall, and W. P. Kegelmeyer, “SMOTE: synthetic minority

over-sampling technique,” CoRR, vol. abs/1106.1813, 2011.

11
[28] A. Amin, F. Rahim, I. Ali, C. Khan, and S. Anwar, A Comparison of Two Oversampling Tech-
niques (SMOTE vs MTDF) for Handling Class Imbalance Problem: A Case Study of Customer
Churn Prediction, pp. 215–225. Cham: Springer International Publishing, 2015.

[29] L. A. . B. L. Chen, C., “Using random forests to learn imbalanced data,” Technical Report 666.
Statistics Department of University of California at Berkeley., 2004.

[30] N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, “Smote: Synthetic minority

over-sampling technique,” Journal of Artificial Intelligence Research, vol. 16, pp. 321–357, 2002.

[31] G. Menardi and N. Torelli, “Training and assessing classification rules with imbalanced data,”
Data Mining and Knowledge Discovery, vol. 28, no. 1, pp. 92–122, 2014.

[32] G.-e. Xia and W.-d. Jin, “Model of customer churn prediction on support vector machine,”
Systems Engineering-Theory & Practice, vol. 28, no. 1, pp. 71–77, 2008.

[33] B-Gent, K. Coussement, and D. V. D. Poel, “Churn prediction in subscription services: an

application of support vector machines while comparing two parameter-selection techniques,”
2006.

[34] A. Liaw and M. Wiener, “Classification and regression by random forest,” R News, vol. 2, no. 3,
pp. 18–22, 2002.

[35] M. K. C. from Jed Wing, S. Weston, A. Williams, C. Keefer, and A. Engelhardt, caret: Classifi-
cation and Regression Training, 2012. R package version 5.15-044.

[36] S. Lessmann and S. Vo, “A reference model for customer-centric data mining with support vector
machines,” European Journal of Operational Research, vol. 199, 2009.

[37] B. L. Das and M. Baruah, “Employee retention: A review of literature,” Journal of Business and
Management, vol. 14, no. 2, pp. 08–16, 2013.

[38] S. Ramlall, “A review of employee motivation theories and their implications for employee reten-
tion within organizations,” Journal of American Academy of Business, vol. 5, no. 1/2, pp. 52–63,
2004.

[39] P. C. Bryant and D. G. Allen, “Compensation, benefits and employee turnover: Hr strategies for
retaining top talent,” Compensation & Benefits Review, vol. 45, no. 3, pp. 171–175, 2013.

12