Prediction of Absenteeism at Work using Data Mining Techniques

Conference Paper · December 2023

DOI:10.1109/ICITR51448.2023.9310913

CITATIONS READS
8 4,106

7 authors, including:

Mikhail Skorikov Muhammad Abrar Hussain

North South University Rakuten Group Inc.
3 PUBLICATIONS 22 CITATIONS 4 PUBLICATIONS 14 CITATIONS

SEE PROFILE SEE PROFILE

Mohammad Kaosain Akbar Sifat Momen

Concordia University Montreal North South University
1 PUBLICATION 8 CITATIONS 63 PUBLICATIONS 745 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Sifat Momen on 06 April 2023.

The user has requested enhancement of the downloaded file.

 Prediction of Absenteeism at Work using Data
Mining Techniques
Mikhail Skorikov∗ , Muhammad Abrar Hussain† , Mahfujur Rhaman Khan∗ , Mohammad Kaosain Akbar‡ ,
Sifat Momen∗ , Nabeel Mohammed∗ and Taniya Nashin §
∗ Department of Electrical and Computer Engineering, North South University, Dhaka, Bangladesh,
† Fujitsu Research Institute, Tokyo, Japan,
‡ Department of Computer Science and Engineering, Daffodil International University, Dhaka, Bangladesh,
§ Department of Business Administration, Victoria University of Bangladesh, Dhaka, Bangladesh

Email: ∗ {mikhail.skorikov, mahfujur.rhaman, sifat.momen, nabeel.mohammed}@northsouth.edu,

† [email protected], ‡ [email protected], § [email protected]

Abstract—High absenteeism among employees can be detri- In this paper, we present a prediction of absenteeism at
mental to an organization as it can result in productivity and work using data mining techniques in a courier company in
economic loss. This paper looks into a case of absenteeism Brazil. The dataset [3] is publicly available in Kaggle and
in a courier company in Brazil. Machine learning techniques
have been employed to understand and predict absenteeism. comprises of a total of three years of data (from July 2007
Understanding this would provide human resource managers an to July 2010) containing details of absences in the company.
excellent decision aid to create policies that can aim to reduce When employees were absent, their details of absence have
absenteeism. Data has been preprocessed, and several machine been recorded. The novelty of the paper lies in the use of
learning classification algorithms (such as zeroR, tree-based J48, data mining techniques (particularly classification) to predict
naive Bayes, and KNN) have been applied. The paper reports
models that can predict absenteeism with an accuracy of over absenteeism. Such a methodology can be applied to under-
92%. Furthermore, from an initial of 20 attributes, disciplinary stand absenteeism behavior in other companies and develop a
failure turns out to be a very prominent feature in predicting decision-aid system that can help managers in controlling it.
absenteeism. The rest of the paper is organized as follows: In section 2,
Index Terms—absenteeism, prediction, data mining, classifica- a critical literature review has been conducted. This follows
tion
our research methodology in section 3. Experimental results
I. I NTRODUCTION are presented in section 4. Finally, the paper is concluded in
section 5 with the highlights of the results obtained.
In Human Resources and Management (HRM), employees
are regarded as valuable assets to their organization. Employee II. BACKGROUND
productivity directly contributes to the work efficiency and Absenteeism is defined as the temporary withdrawal from
the success of a company. The economic sustainability of a work due to personal reasons, including illness and demise of
company lies in its revenue, which is generated only with close relatives [4]. Cuchiella and colleagues, on the other hand,
sufficient support from the employees. If employees under- remarks absenteeism as an employee’s habitual absence from
perform or do not perform well, it results in an incurment work [5]. Despite the differences in opinion, high absenteeism
of high costs to the organization. Consequently, in good is always found to be correlated with high productivity loss
organizations, HRM professionals have to work prudently to and economic loss. Research in understanding why employees
ensure that employee motivation stays high enough to yield stay absent is a very active area partly due to the fact that
target productivity [1]. it often has serious economic consequences. Different factors
One key indicator of low motivation in an organization is have been found to contribute towards absenteeism. Arai and
high employee absenteeism. Absence in work can arise due to Thoursie [6], for instance, found that incentives have a crucial
an array of factors including age, health condition, bad work- role to play in absenteeism. Using industry-region panel data,
life balance, low motivation, organizational culture, policies, they were able to establish a negative correlation between sick
pay scale, poor recognition in the organization, and many rates and shares of temporary contracts. Arai and Thoursie
more. Understanding why employees stay absent and realizing argue that this is because workers on temporary contracts have
the patterns in absenteeism behavior is an invaluable tool for lower job security and hence, a higher possibility of being laid
human resource managers. Understanding this, in turn, would off compared to the employees on time-unlimited contracts.
help organizations to shape the policies and organizational Hence, workers on temporary contracts tend to be more loyal
culture to transform the organization into a more people-driven in the hope of a renewal of contracts. Hesselius [7] found that
organization, which is expected to result in higher productivity absenteeism is negatively correlated to unemployment. This is
[2]. due to the fact that when unemployment is high, the probability
978-0-7381-4403-0/20/$31.00 ©2023 IEEE of an individual to find a new job is low. This acts as an
 TABLE I
ATTRIBUTES AND THEIR DESCRIPTION

Attribute Data Type Description/Remarks Min Max Mean Distribution

ID Integer Identification number of the employee x x x x
Reason for absence Integer The reason for absence (range from 1 - 28) 17, 3, 2 23 x x
Month of absence Integer Month of absence of the employee (range from 1 - 12) 12 3 x x
Day of the Week Integer Range is from 1 - 7 where Sunday = 1 and so on 5 2 x x
Seasons Integer There are four seasons (1 - 4) 1 4 3 x
Travel Expense Integer Transportation cost from home to work 118 388 221.3 negatively skewed
Distance Integer Distance from residence to work 5 52 29.6 positively skewed
Service time Integer Months of service time 1 29 12.6 negatively skewed
Age Integer Age of the employee in years 27 58 36.5 negatively skewed
Workload / day Integer Average workload per day 206 379 271.5 positively skewed
Hit target Integer Target for the employees 81 100 94.6 negatively skewed
Disciplinary failure Boolean =1 for past disciplinary failure, otherwise 0 1 0 x x
Education Integer Range is from 1 - 4 depending on the education level 4 1 x x
Children Integer Number of children of the employee 0 4 1 normally distributed
Social Drinker Boolean =1 for being a social drinker, otherwise 0 1 0 x x
Social Smoker Boolean =1 for being a social smoker, otherwise 0 1 0 x x
Pet Integer Number of pets that the employee has 0 8 0.7 positively skewed
Weight Integer Weight (in nearest Kg) of the employee 56 108 79 negatively skewed
Height Integer Height (in nearest cm) of the employee 163 196 172.1 positively skewed
BMI Integer Body mass index (to the nearest integer) 19 38 26.7 positively skewed
Absenteeism time Integer Absenteeism time in hours 0 120 6.9 positively skewed

incentive for the individual to be less absent in work and thus 3 years of data from July 2007 to July 2010. Information per-
reducing his/her probability of being laid off. Winkelmann [8] taining to health, work, workload, habit, traveling, and details
found that absenteeism is dependent on factors such as wages of absence have been incorporated into the dataset. Table I
and even the firm size. Organizational policies also impact the shows the details of the attributes recorded in the dataset.
level of absenteeism [9]. For example, Halpern and colleagues One particular attribute to note in the dataset is the reason
[10] found that the smoking policy in the workplace affects for absence, which is an important attribute to investigate.
absenteeism and productivity. Their research concludes that Table II describes the various reasons for absence, along with
current smokers tend to have significantly higher absenteeism the percentage of occurrence of each. The dataset size is small
than former smokers and non-smokers. Absenteeism is re- and as a consequence applying deep learning techniques will
peatedly reported to be strongly correlated with employees’ not be effective. Classical data mining techniques are able to
health status. For instance, Tunceli and colleagues [11] found predict with high accuracy. Hence we refrained from using
that for employees with diabetes, the absolute probability of any deep learning techniques. We split the dataset into training
working for male and female employees is 7.1 % and 4.4 and test sets in an 80-20 ratio with a stratified class attribute
% points less compared to the individuals without diabetes. distribution.
Gates and colleagues [12] found that moderately or extremely
obese workers experience a 4.2% loss in productivity, which B. Research Methodology
is tantamount to 1.18% more than all other employees. Shah
Our prime objective is to develop a model that can predict
and colleagues [13] used deep neural networks to predict
absenteeism (with high accuracy) in the courier company for
absenteeism before employees are actually hired. Research in
aid in decision-making by the managers. Since there exists
understanding absenteeism behavior reveals that it depends on
different reasons for absence among employees and many
various factors that act as incentives (directly or indirectly),
factors contribute towards absenteeism, a deductive learning
including organizational culture, policies (both national and
approach is an infeasible option. We instead use the inductive
organizational), size of the organization, and many more.
learning approach, where we use data mining techniques to
III. R ESEARCH W ORK predict absenteeism. Scikit-learn [14], a Python library for data
This paper looks into the factors affecting absenteeism in a science, has been used to carry out the data mining tasks.
courier company in Brazil. Whenever an employee is absent, The research methodology embraced in this work has been
he/she needs to fill up a form detailing reasons for absence. outlined in figure 1.
This, in conjunction with other personal details, has been The raw data (the Absenteeism dataset) is first preprocessed
recorded in a dataset. The dataset was first published by [3] to a form that is suitable for applying machine learning
and is now publicly available in Kaggle. algorithms. The preprocessing phase includes data clean-
ing (removal of attribute and marking missing labels), data
A. Dataset discretization (i.e., converting continuous data into discrete
The dataset comprises of a total of 740 instances recorded categories), and data transformation (converting data from a
over 21 attributes per instance. The dataset is a collection of numeric form to categorical). After the preprocessing step,
 TABLE II
P REPROCESSING FOR THE ATTRIBUTE ” REASON OF ABSENCE ”

Original New Description Percentage

Value Value in dataset
1 IPD Certain infectious and parasitic diseases 2.1%
2 NP Neoplasms 0.1%
3 BOI Blood-forming organ & immune mechanism 0.1%
4 ENM Endocrine, nutritional and metabolic diseasess 0.3%
5 MBD Mental and behavioural disorders 0.4%
6 DNS Diseases of the nervous system 1.1%
7 DEA Diseases of the eye and adnexa 2.0%
8 DEM Diseases of the ear and mastoid process 0.8%
9 DCS Diseases of the circulatory system 0.5%
10 DRS Diseases of the respiratory system 3.4%
11 DDS Diseases of the digestive system 3.5%
12 DSST Diseases of the skin and subcutaneous tissue 1.1%
13 DMSCT Diseases of musculoskeletal system & tissue 7.4%
14 DGS Diseases of the genitourinary system 2.6%
15 PCP Pregnancy, childbirth and the puerperium 0.3%
16 CPP Conditions originating in the perinatal period 0.4%
17 CMDCA Congenital malformations and chromosomal abnormalities 15.1%
18 ACLF Abnormal clinical and laboratory findings 2.8%
19 IPEC Injury, poisoning and consequences of external causes 5.4%
20 ECMM External causes of morbidity and mortality 0%
21 FHSHS Factors to health status and health services 0.8%
22 PFU Patient follow-up 5.1%
23 MC Medical consultation 20.1%
24 BD Blood Donation 0.4%
25 LE Laboratory examination 4.2%
26 UA Unjustified absence 4.5%
27 PTH Physiotherapy 9.3%
28 DC Dental Consultation 15.1%
? null Null values for employees with no absenteeism 5.8%

ID works as a unique identification number for the employee

and is irrelevant, thus has been removed. Other attributes are
converted to categorical form since a numeric or boolean value
is not appropriate for it. Furthermore, it may cause problems
when applying machine learning algorithms. Seven is higher
than one, but there is no quantitative difference if an employee
is absent in July (7) or January (1). Hence, in places where
a numeric or boolean value does not make sense, we have
converted them to categorical values.
There exists three boolean attributes in the dataset: (1)
disciplinary failure, (2) social drinker, and (3) social smoker.
Fig. 1. Research Methodology We converted the initial 0 and 1 values to ’False’ and ’True’
respectively.
Table II shows the conversion of attribute Reason of absence
two subsets of features are found that would make predictions from numeric to categorical. For the attribute seasons, the
better. Machine learning algorithms are then applied to the value from the numeric is also converted to categorical with
new feature subsets and the full feature set to predict the 1, 2, 3, and 4 substituted to A, B, C, and D, respectively.
absenteeism class (discussed later). Values of the attribute ”Education” are also converted from
C. Preprocessing of Predictor Attributes numeric to categorical. The values in the range 1-4 are
Preprocessing is required to transform the raw data into a substituted by ”High School”, ”Graduate”, ”Postgraduate”, and
form that would be suitable for data mining tasks. Missing ”Master and Doctor”, respectively.
data in the dataset is first marked by null value, because the
D. Preprocessing of the class attribute
missing values were originally marked with question marks
(?). After this, several preprocessing techniques are applied to The class attribute is the attribute that is intended to be
the dataset. All the attributes that were not real-valued were predicted. The class attribute that we would like to predict
converted to categorical, including absenteeism time in hours, is absenteeism time. Initially, the values in this attribute were
the class attribute. continuous. However, it would make more sense if we classify
the absenteeism time in terms of categories. We do so because of redundancy between them. Using the CFS algorithm, it has
it would allow the model to predict different degrees of been found that four attributes (the month of absence, age,
absence on test data. disciplinary failure, social drinker) play as the most influential
attributes to predict absenteeism.

TABLE IV
F EATURE SCORES BY R ELIEF ATTRIBUTE E VALUATOR

Attribute Feature Score

Disciplinary failure 0.3629
Reason for absence 0.3303
Month of absence 0.1816
Seasons 0.1685
Social drinker 0.1594
Distance from residence to work 0.137
Children 0.1225
Transportation Expense 0.119
Education 0.1121
Weight 0.1097
Age 0.1005
Body mass index 0.0993
Fig. 2. Frequency Distribution of the class attribute Day of the Week 0.0783
Pet 0.0685
Service time 0.064
Figure 2 shows the frequency distribution of the class Height 0.0617
attribute. The x-axis of the graph indicates absenteeism time Work load / day 0.055
(in hours), whereas the y-axis signifies the corresponding Hit target 0.0434
Social Smoker 0.0386
frequency value. Figure 2 provides a conspicuous display of
the existence of three classes in the class attribute. Another popular feature evaluator is the relief attribute eval-
uator, based on the relief algorithm [16]. The relief attribute
TABLE III
D IFFERENT CLASSES OF CLASS ATTRIBUTE
evaluator evaluates each attribute’s worth in terms of a score
called the feature score. Each attribute’s feature score lies
Absenteeism time (hours) Class between -1 and 1 with values going towards 1 indicating its
0 A
1 - 15 B
prominence level as a feature for predicting the target attribute.
16 - 120 C IV. E XPERIMENTS AND RESULTS
A comparison cannot be made with relevant previous works
due to the class value ranges for each work being different,
E. Class Imbalance
but our model outperforms similar models [17].
It is sensible that we categorize the class attribute as de- This section discusses the experimental methodology and
scribed in table III. However, this also results in an imbalance corresponding results. Three types of experiments, as outlined
of the three classes with class B taking up about 85% of the in table VI, are devised to assess how well the absenteeism
dataset. To combat this problem, we applied an oversampling class can be predicted. Experiment A uses the four prominent
technique called Synthetic Minority Oversampling TEchnique features (the month of absence, age, disciplinary failure, social
(SMOTE) onto the training set. The resulting dataset was drinker) as found using the CFS method to train data. Exper-
almost equally balanced in terms of the class attribute. We iment B uses all the 19 attributes. Since initial experiments
then split the training process into one tested using SMOTE show that the attribute disciplinary failure has the highest
applied, and one without. information gain as well as the highest feature score from
the relief algorithm, experiment C is conducted with only one
F. Feature Selection
attribute - the disciplinary failure.
The selection of prominent features is crucial in data mining Each experiment is run using a 10-fold stratified cross-
tasks for two main reasons: (1) irrelevant attributes act as validation strategy and several different classifiers (including
noise, and this can degrade the predictability of the model. The ZeroR, naive Bayes, KNN, and tree-based J48 classifiers). The
removal of irrelevant attributes improves the predictability of ZeroR classifier does not have any predictability power as
the model. (2) It results in the reduction of the dimension of it merely predicts the majority class for every query input.
the dataset - thus allowing to avoid the curse of dimensionality. However, the ZeroR classifier has been selected as a baseline
Correlation Feature Set (CFS) [15], a well-known feature classifier. The naive Bayes classifier uses the Bayes theorem
selection technique, has been used to find the prominent (equation 1) to predict the absenteeism class.
features that can be used to predict the class attribute. CFS
evaluates the worth of a subset of attributes by considering P (X1 , X2 , ...., Xn |Y )P (Y )
the individual predictive ability of each feature and the degree P (Y |X1 , X2 , ..., Xn ) = (1)
P (X1 , X2 , ..., Xn )
 TABLE V
W EIGHTED AVERAGE OUTPUT OF ORIGINAL DATASET WITHOUT SMOTE

Expt. & Classifier Precision Recall F-measure ROC area Accuracy

A (zeroR) 0.74 0.86 0.79 0.50 85.5 +/- 0.6 %
A (naive Bayes) 0.83 0.91 0.87 0.76 90.1 +/- 2.0 %
A (J48) 0.83 0.91 0.87 0.74 89.7 +/- 2.9 %
A (KNN-Euclidean) 0.83 0.91 0.87 0.74 88.0 +/- 5.3 %
A (KNN-Manhattan) 0.83 0.91 0.87 0.74 88.0 +/- 5.3 %
A (KNN-Chebyshev) 0.78 0.74 0.75 0.69 81.2 +/- 5.9 %
B (zeroR) 0.74 0.86 0.79 0.50 85.5 +/- 0.6 %
B (naive Bayes) 0.85 0.80 0.82 0.80 66.5 +/- 13.8 %
B (J48) 0.86 0.88 0.87 0.80 89.1 +/- 4.3 %
B (KNN-Euclidean) 0.82 0.89 0.84 0.77 85.4 +/- 1.3 %
B (KNN-Manhattan) 0.82 0.89 0.84 0.77 85.4 +/- 1.3 %
B (KNN-Chebyshev) 0.83 0.91 0.87 0.67 86.9 +/- 2.3 %
C (zeroR) 0.74 0.86 0.79 0.50 85.5 +/- 0.6 %
C (naive Bayes) 0.83 0.91 0.87 0.69 90.9 +/- 1.6 %
C (J48) 0.83 0.91 0.87 0.69 90.9 +/- 1.6 %
C (KNN-Euclidean) 0.83 0.91 0.87 0.69 90.9 +/- 1.6 %
C (KNN-Manhattan) 0.83 0.91 0.87 0.69 90.9 +/- 1.6 %
C (KNN-Chebyshev) 0.83 0.91 0.87 0.69 90.9 +/- 1.6 %

TABLE VI
T YPES OF E XPERIMENTS

Experiment Name Features used

A As found using the CFS method
B All 19 attributes
C Disciplinary failure

A lazy classifier, KNN, is used with a K value of 5. In order

to find nearest neighbors, distance is measured using differ-
ent metrics, including Euclidean, Manhattan, and Chebyshev.
Equations 2, 3, and 4 shows the ways Euclidean, Manhattan
and Chebyshev distances are calculated.
v
u n
uX Fig. 3. Experimental results - original dataset
d(x, y) = t (xi − yi )2 (2)
i=1

n
X
d(x, y) = |xi − yi | (3)
i=1

D(x, y) = max(xi − yi ) (4)

Finally, a decision tree (J48) has been used as a classifier.

J48 uses entropy-based mutual information gain to construct
the decision tree.

A. Experimental results without the application of SMOTE

filter
The experimental results without applying the SMOTE
filter are illustrated in figure 3 , with table V detailing the Fig. 4. Experimental results - SMOTE
particulars.

B. Experimental results after applying the SMOTE filter C. Discussion

Figure 4 illustrates the summary of the results with SMOTE Experimental results from tables V and VII indicate that
applied. Table VII denotes the performance scores of each an accuracy of 90.1% is obtained using the naive Bayes
classifier. classifier without oversampling, and it is the highest achieved
 TABLE VII
W EIGHTED AVERAGE OUTPUT WITH SMOTE

Expt. & Classifier Precision Recall F-measure ROC area Accuracy

A (zeroR) 0.00 0.06 0.01 0.50 33.2 +/- 0.1 %
A (naive Bayes) 0.85 0.62 0.70 0.77 75.0 +/- 4.0 %
A (J48) 0.86 0.41 0.51 0.72 75.1 +/- 3.3 %
A (KNN-Euclidean) 0.85 0.64 0.72 0.73 76.9 +/- 5.1 %
A (KNN-Manhattan) 0.84 0.83 0.84 0.70 85.4 +/- 4.7 %
A (KNN-Chebyshev) 0.80 0.76 0.77 0.64 83.0 +/- 4.7 %
B (zeroR) 0.00 0.06 0.01 0.50 33.2 +/- 0.1 %
B (naive Bayes) 0.87 0.75 0.79 0.80 80.5 +/- 6.1 %
B (J48) 0.86 0.89 0.87 0.76 89.3 +/- 6.5 %
B (KNN-Euclidean) 0.85 0.67 0.72 0.81 84.5 +/- 5.0 %
B (KNN-Manhattan) 0.79 0.75 0.77 0.76 92.1 +/- 4.3 %
B (KNN-Chebyshev) 0.87 0.68 0.75 0.70 92.3 +/- 9.5 %
C (zeroR) 0.00 0.06 0.01 0.50 33.2 +/- 0.1 %
C (naive Bayes) 0.83 0.91 0.87 0.69 64.0 +/- 1.1 %
C (J48) 0.83 0.91 0.87 0.69 64.3 +/- 1.0 %
C (KNN-Euclidean) 0.83 0.91 0.87 0.69 64.5 +/- 1.0 %
C (KNN-Manhattan) 0.83 0.91 0.87 0.69 64.5 +/- 1.0 %
C (KNN-Chebyshev) 0.83 0.91 0.87 0.69 64.5 +/- 1.0 %

for experiment A. The KNN, naive Bayes, and J48 classifiers [2] M. Mayfield, J. Mayfield, and K. Q. Ma, “Innovation matters: creative
yield the highest accuracy in experiment C if applied to environment, absenteeism, and job satisfaction,” Journal of Organiza-
tional Change Management, 2020.
the dataset without the SMOTE filter. Applying the SMOTE [3] A. Martiniano, R. Ferreira, R. Sassi, and C. Affonso, “Application of
filter to the data and conducting experiment C results in a a neuro fuzzy network in prediction of absenteeism at work,” in 7th
significant reduction of performance. For experiment B, the Iberian Conference on Information Systems and Technologies (CISTI
2012). IEEE, 2012, pp. 1–4.
performance of the naive Bayes classifier falls with or without [4] R. L. Mathis and J. H. Jackson, Human resource management: Essential
SMOTE. However, the J48 classifier’s performance stays more perspectives. Cengage Learning, 2011.
or less the same regardless of the sampling strategy. The [5] F. Cucchiella, M. Gastaldi, and L. Ranieri, “Managing absenteeism in
the workplace: the case of an italian multiutility company,” Procedia-
highest measure of accuracy overall is 92.3%, achieved by Social and Behavioral Sciences, vol. 150, pp. 1157–1166, 2014.
the KNN classifier with the Chebyshev distance metric in [6] M. Arai and P. S. Thoursie, “Incentives and selection in cyclical
experiment B after applying the SMOTE filter to the train absenteeism,” Labour Economics, vol. 12, no. 2, pp. 269–280, 2005.
[7] P. Hesselius, “Does sickness absence increase the risk of unemploy-
set. For this model, the accuracy of class A is 67%, for ment?” The Journal of Socio-Economics, vol. 36, no. 2, pp. 288–310,
class B it is 92.1%, and for class C is 8.3%. On another 2007.
note, the experiment indicates that disciplinary failure is a [8] R. Winkelmann, “Wages, firm size and absenteeism,” Applied Economics
Letters, vol. 6, no. 6, pp. 337–341, 1999.
very influential attribute for determining absenteeism as all [9] S. A. Ruhle and S. Süß, “Presenteeism and absenteeism at work—an
classifiers other than the baseline result in over 90% accuracy analysis of archetypes of sickness attendance cultures,” Journal of
for it without SMOTE applied. Experiment B, comprising of Business and Psychology, vol. 35, no. 2, pp. 241–255, 2020.
[10] M. T. Halpern, R. Shikiar, A. M. Rentz, and Z. M. Khan, “Impact of
all attributes, is a good indicator of the performance of the smoking status on workplace absenteeism and productivity,” Tobacco
classifiers and reflect a more realistic view due to the class control, vol. 10, no. 3, pp. 233–238, 2001.
imbalance problem causing overestimations in performance [11] K. Tunceli, C. J. Bradley, D. Nerenz, L. K. Williams, M. Pladevall,
and J. E. Lafata, “The impact of diabetes on employment and work
otherwise. An imbalanced dataset can sometimes lead to productivity,” Diabetes care, vol. 28, no. 11, pp. 2662–2667, 2005.
high bias. When the bias is removed owing to SMOTE, the [12] D. M. Gates, P. Succop, B. J. Brehm, G. L. Gillespie, and B. D.
performance naturally decreases. Sommers, “Obesity and presenteeism: the impact of body mass index
on workplace productivity,” Journal of Occupational and Environmental
Medicine, vol. 50, no. 1, pp. 39–45, 2008.
V. C ONCLUSION [13] S. A. Ali Shah, I. Uddin, F. Aziz, S. Ahmad, M. A. Al-Khasawneh, and
M. Sharaf, “An enhanced deep neural network for predicting workplace
This paper looks into the absenteeism dataset, a dataset absenteeism,” Complexity, vol. 2020, 2020.
detailing information about absence records. There were a [14] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion,
O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vander-
total of 740 instances and 21 initial attributes. After careful plas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duch-
preprocessing and feature selection, machine learning algo- esnay, “Scikit-learn: Machine learning in Python,” Journal of Machine
rithms were applied. Three kinds of experiments with different Learning Research, vol. 12, pp. 2825–2830, 2011.
[15] M. A. Hall, “Correlation-based feature selection for machine learning,”
subsets of features were devised. It has been found that the 1999.
model can predict absenteeism with over 92% accuracy. [16] K. Kira and L. A. Rendell, “The feature selection problem: Traditional
methods and a new algorithm,” in Aaai, vol. 2, 1992, pp. 129–134.
R EFERENCES [17] Z. Wahid, A. Z. Satter, A. Al Imran, and T. Bhuiyan, “Predicting
absenteeism at work using tree-based learners,” in Proceedings of the
[1] C. Navarro and C. Bass, “The cost of employee absenteeism,” Compen- 3rd International Conference on Machine Learning and Soft Computing,
sation & Benefits Review, vol. 38, no. 6, pp. 26–30, 2006. 2019, pp. 7–11.

View publication stats