IAES
Inter
national
J
our
nal
of
Articial
Intelligence
(IJ-AI)
V
ol.
14,
No.
1,
February
2025,
pp.
376
∼
384
ISSN:
2252-8938,
DOI:
10.11591/ijai.v14.i1.pp376-384
❒
376
A
systematic
analysis
on
machine
lear
ning
classiers
with
data
pr
e-pr
ocessing
to
detect
anti-patter
n
fr
om
sour
ce
code
Nazneen
Akhter
1
,
Afrina
Khatun
1
,
Md.
Sazzadur
Rahman
2
,
A.
S.
M.
Sanwar
Hosen
3
,
Mohammad
Shahidul
Islam
2
1
Department
of
Computer
Science
and
Engineering,
Bangladesh
Uni
v
ersity
of
Professionals,
Dhaka,
Bangladesh
2
Institute
of
Information
T
echnology
,
Jahangirnag
ar
Uni
v
ersity
,
Dhaka,
Bangladesh
3
Department
of
Articial
Intelligence
and
Big
Data,
W
oosong
Uni
v
ersity
,
Daejeon,
South
K
orea
Article
Inf
o
Recei
v
ed
Jan
16,
2024
Re
vised
Jul
10,
2024
Accepted
Jul
26,
2024
K
eyw
ords:
Anti-pattern
Code
quality
Code
smell
Data
pre-processing
Machine
learning
ABSTRA
CT
Automatic
detection
of
anti-patterns
from
source
code
can
reduce
softw
are
maintenance
costs
massi
v
ely
.
No
w
adays,
machine
learning
approaches
are
v
ery
commonly
use
d
to
ide
ntify
anti-patterns.
Hence,
it
is
v
ery
crucial
to
choose
a
classier
that
can
be
useful
for
detecting
anti
-patterns.
This
w
ork
aims
to
help
practitioners
to
choose
a
suitable
classier
to
detect
anti-patterns.
In
this
paper
,
we
highlight
16
classiers
in
four
dif
ferent
cate
gories
to
detect
anti-patterns.
Furthermore,
the
performance
of
these
classiers
is
identied
with
the
data
pre-processing
(DPP)
to
detect
four
com
monly
occurring
anti-patterns
from
the
three
commonly
used
open-source
Ja
v
a
projects’
source
code.
The
accurac
y
of
Dagging
classiers
is
98.4%.
K
ernel
logistic
re
gression
(KLR)
also
performs
well
i.e.,
97%.
In
the
case
of
time
comple
xity
,
nai
v
e
Bayes
(NB),
decision
trees
(DT),
support
v
ector
machines
(SVM),
library
for
support
v
ector
machines
(LibSVM),
logistic,
and
LightGBM
(LB)
ha
v
e
less
time
comple
xity
to
b
uild
a
model
in
all
the
projects.
This
is
an
open
access
article
under
the
CC
BY
-SA
license
.
Corresponding
A
uthor:
A.
S.
M.
Sanw
ar
Hosen
Department
of
Articial
Intelligence
and
Big
Data,
W
oosong
Uni
v
ersity
Daejeon
34606,
South
K
orea
Email:
sanw
ar@wsu.ac.kr
1.
INTR
ODUCTION
Softw
are
quality
is
v
ery
important
as
it
denes
the
ease
of
maintainability
,
testability
,
re
adability
,
stability
,
speed,
usability
,
size,
cost,
and
security
.
It
also
ensures
the
smoothness,
conciseness,
and
customer
satisf
action
of
softw
are.
Ho
we
v
er
,
due
to
the
hea
vy
w
orkload
and
time
pressure,
softw
are
quality
issues
are
usually
ignored.
Some
reports
sho
w
that
softw
are
e
v
olution,
understandability
,
usability
,
modularity
,
reusabil-
ity
,
analyzability
,
and
changeability
are
ne
glected
during
the
de
v
elopment
process
of
t
he
softw
are
[1].
As
a
result,
design
a
ws
i.e.,
poor
structure
in
the
source
code
increase
the
de
v
elopment
and
maintenance
cost.
These
design
a
ws
are
termed
as
anti-pattern
or
code
smell.
P
alomba
et
al.
[2]
report
that
the
maintenance
cost
of
softw
are
is
2–100
times
greater
than
the
de
v
elopment
cost.
The
term
code
smell
(anti-pattern)
is
rst
introduced
in
[3].
Anti-patterns
from
a
source
code
can
be
detected
both
by
manual
and
automated
processes.
F
or
lar
ge
softw
are
projects,
the
manual
detection
system
is
a
v
er
y
time-consuming
process
as
it
mainly
de-
pends
on
proper
docum
entation,
source
code
structure,
and
de
v
eloper’
s
e
xperience
[1].
Hence,
an
automated
and
ef
fecti
v
e
anti-pattern
detection
technique
is
essential.
Although
fe
w
studies
are
a
v
ailable
in
this
eld.
Sometimes
people
may
become
confused
about
choosing
machine
learning
(ML)
classiers
and
en
vironmental
J
ournal
homepage:
http://ijai.iaescor
e
.com
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Artif
Intell
ISSN:
2252-8938
❒
377
setup.
Se
v
eral
researchers
introduced
three
dif
ferent
support
v
ector
machines
(SVM)-based
approaches
[4]–[6].
Maig
a
et
al.
did
not
adopt
an
y
data
pre-processing
(DPP)
technique,
while
Akhter
et
al.
adopted
a
DPP
technique
namely
synthetic
minority
o
v
er
-sampling
technique
(SMO
TE)
with
SVM
which
performed
better
than
synthetic
minority
under
-sampling
with
random
forest
(SMURF).
Fi
v
e
dif
ferent
ML
classiers:
J48,
random
forest
(RF),
nai
v
e
Bayes
(NB),
SVM,
and
just-in-time
rule
induction
procedure
(JRIP)
to
detect
anti-
pattern
are
used
in
[7].
The
y
reported
NB
performed
the
best.
The
y
did
a
comparison
among
ML
classiers
and
detection
of
code
and
design
smells
(DECOR)
(a
method
of
code
and
design
smell
detection).
The
y
reported
that
in
terms
of
F-measure
and
recall
,
the
performance
of
DECOR
w
as
higher
than
NB
b
ut
precision
w
as
lo
wer
.
The
author
used
16
dif
ferent
classiers
and
their
boosting
techniques
(in
total
32
classiers)
[8].
The
y
reported
that
the
best
algorithm
had
approximately
97%
of
F-Score.
The
aforementioned
researchers
hardly
use
an
y
techniques
to
process
the
data.
In
comparison,
a
DPP
technique
can
inuence
the
performance
of
the
classier
.
As
there
are
man
y
classiers
most
of
the
time
researchers
become
puzzled
to
choose
a
classier
.
Hence,
an
appropriate
classier
selection
is
also
a
crucial
study
for
anti-pattern
detection.
In
this
paper
,
a
comprehensi
v
e
study
is
conducted
so
that
the
practitioners
can
get
a
complete
idea
of
choosing
the
classier
and
their
w
orking
process.
A
data-prepossessing
technique
is
adopted
to
enhance
the
performance
of
the
classiers.
This
is
ho
w
the
remainder
of
the
paper
is
structured.
The
background
research
on
anti-pattern
det
ection
is
co
v
ered
in
section
2.
In
section
3,
an
e
xplanation
of
the
sugges
ted
methodology
is
pro
vided
along
with
an
analysis
of
sixteen
ML
classiers.
Section
4
addresses
the
outcome
and
e
x
ecution.
Section
5
concludes
this
report
by
discussing
future
research.
2.
LITERA
TURE
REVIEW
ON
ANTI-P
A
TTERN
DETECTION
Inadequate
design
raises
the
system’
s
de
v
elopment
and
maintenance
costs
by
lo
wering
perform
ance
and
reusability
.
As
a
result,
scholars
de
v
el
oped
v
arious
theories
to
address
this
issue.
The
three
main
types
of
methodologies
used
in
anti-pattern
detection
research
that
are
currently
in
use
are
rule-based,
ML-based,
and
deep
learning-based.
Detailed
discussion
is
pro
vided
in
the
follo
wing.
The
main
aim
of
the
rule-based
approach
is
to
propose
a
set
of
rules
to
nd
anti-patterns.
T
o
do
that,
se
v
eral
researchers
[9],
[10]
propose
tw
o
dif
ferent
approaches
to
identify
anti-patterns
from
the
source
code.
Blob,
spaghetti
code,
and
functional
decomposition
by
utilizing
three
heuristic
search
algorithms
[9],
[10].
Aras
and
Selcuk
[11]
pre
sent
an
automated
approach
using
metrics
and
rules
to
detect
anti-patterns
in
an
object-oriented
system.
Three
anti-patterns
are
used
namely
Blob,
Swiss
Army
Knife,
and
La
v
a
Flo
w
.
36
Ja
v
a
classes
are
focused
on
conducting
this
e
xperiment.
Ho
we
v
er
,
the
accurac
y
is
not
satisf
actory
.
Due
to
issues
with
rule-based
systems’
accurac
y
,
researchers
introduced
ML
techniques.
The
primary
issue
with
the
rule-based
technique
is
that
each
smell
has
its
o
wn
set
of
rules,
all
of
which
are
manually
specied
by
the
researcher
.
A
met
h
od
for
identifying
design
patterns
utilizing
ML-based
techniques
and
softw
are
metrics
has
been
put
out
by
[12].
F
or
small
scale,
code
recall
is
90%
b
ut
for
lar
ge
scale
code,
it
is
60%.
T
w
o
S
VM-based
methods,
SMURF
and
SVMDetect,
are
introduced
by
Maig
a
et
al.
[5],
[6]
to
identify
four
anti-patterns:
Swiss
Army
Knife,
Blob,
functional
decomposition,
and
spaghetti
code.
The
precision
and
recal
l
for
SMURF
are
97%
and
84.09%
respecti
v
ely
(Blob
Class).
After
computation,
the
y
report
that
SVMDetect
can
detect
143
anti-
patterns
whereas
DETEX
only
detects
102
.
The
recall
and
precision
rates
of
SVMDetect
are
84
.
09
and
97
.
09
%
respecti
v
ely
for
(Blob
class).
In
terms
of
anti-pattern
identication,
both
methods
perform
similarly
and
can
identify
143
instances
of
blobs
within
the
system.
Barbez
et
al.
[13]
present
an
ensemble
method
called
smart
aggre
g
ation
of
anti-patterns
detectors
(SMAD)
to
detect
god
class
and
featur
e
en
vy
.
SMAD
combined
dif
ferent
approaches
depending
on
their
internal
detection
rule.
Y
in
et
al.
[14]
reported
that
accurac
y
is
more
than
85%
for
the
classier
AB-J48-pruned,
RF
,
and
AdaBoost
random
forest
(AB-RF)
among
15
ML
classiers.
Azadi
et
al.
[15]
propose
a
tool
named
W
ekaNose
to
in
v
estig
ate
code
smell
in
a
system
.
The
y
intend
to
characterize
the
instances
that
inuence
the
detection
of
code
smell
or
not.
Cruz
et
al.
[16]
introduce
a
detection
technique
with
se
v
en
dif
ferent
machine-learning
algorithms
for
detecting
four
types
of
bad
smells.
K
umar
and
Sureka
[17]
de
v
elop
a
predicti
v
e
model
to
detect
anti-patterns.
The
y
ha
v
e
mentioned
that
RF
performed
the
best.
Pritam
et
al.
[18]
e
xamine
the
impact
of
code
smell
on
the
change
tendenc
y
of
a
specic
class
in
an
item
system
which
leads
them
to
disco
v
er
mistak
es
in
e
xpectation
of
progress
inclination
utilizing
code
smell.
The
y
use
4100
unique
classes
of
14
softw
are
systems
after
per
-processing.
The
y
ha
v
e
reported
the
sensiti
vity
is
0.70
and
specicity
higher
than
0.67
for
multi-layer
perceptron
(MLP).
Jesudoss
et
al.
[19]
A
systematic
analysis
on
mac
hine
learning
classier
s
with
data
pr
e-pr
ocessing
to
...
(Nazneen
Akhter)
Evaluation Warning : The document was created with Spire.PDF for Python.
378
❒
ISSN:
2252-8938
propose
another
code
smell
detection
technique
utilizing
tw
o
algorithms
such
as
SVM
and
RF
.
SVM
acted
as
a
classier
and
the
RF
is
used
for
predicting
the
range
of
data.
From
the
abo
v
e
discussion,
we
hardly
found
an
y
DPP
adopted.
As
a
result,
the
accurac
y
rate
of
the
classi
ers
decreases.
Most
of
the
researchers
used
the
same
type
of
ML
classiers
lik
e
SVM,
RF
,
and
NB.
It
is
a
matter
of
concern
that
other
classiers
lik
e
ensemble
or
meta
classiers
performance
are
missing
to
detect
anti-pattern.
Kacem
et
al.
[20]
propose
a
h
ybrid
learning
approach
to
identify
four
code
smells
namely
god
class,
data
class,
feature
en
vy
,
and
long
method
from
74
open-source
systems.
The
y
ha
v
e
reported
that
the
performance
of
god
class
w
as
v
ery
good
(god
class
precision
99.28%,
recall
98.58%,
F-measure
98.93%).
Das
et
al.
[21]
come
up
with
a
technique
named
con
v
olution
neural
netw
orks
(CNN)
to
detect
code
smells.
The
y
ha
v
e
achie
v
ed
97%
accurac
y
for
the
Brain
class
and
95%
for
the
Brain
method.
Barbez
et
al.
[22]
propose
a
method
namely
con
v
olutional
analysis
of
code
metrics
e
v
olution
(CAME)
to
detect
anti-patterns.
From
the
background
study
,
it
is
found
that
DPP
techniques
are
missing.
ML
classiers
can
enhance
their
performance
if
the
model
gets
pre-processed
data.
Recent
research
w
orks
are
using
similar
kinds
of
ML
classiers
(the
most
used
classiers
are
SVM,
NB,
and
RF).
Sometimes
it
becomes
v
ery
tough
to
choose
a
classier
for
the
practitioner
to
w
ork
in
this
eld.
3.
ANTI-P
A
TTERN
DETECTION
MODEL
In
this
paper
,
a
constructi
v
e
analysis
of
ML
classiers
is
presented
to
detect
anti-patterns.
Sixteen
dif
ferent
types
of
classiers
ha
v
e
been
adopted
to
conduct
the
study
.
Considering
their
nature,
classiers
are
di
vided
into
four
subgroups:
function
class
iers,
Bayes
classiers,
tree
classiers,
and
meta
classiers
[23].
Figure
1
represents
the
details
of
the
proposed
model.
A
detailed
description
of
those
is
pro
vided
to
understand
the
w
orking
pri
nciple
of
the
classiers
deeply
in
the
case
of
anti-pattern
detection.
The
follo
wing
describes
the
specics.
Figure
1.
The
proposed
methodology
to
detect
anti-pattern
from
source
code
3.1.
Dataset
and
pr
e-pr
ocessing
Three
separate
open-source
Ja
v
a
projects—Azureus
v2.3.0.6,
Xerces
v2.7.0,
and
Ar
goUML
v0.19.8
are
emplo
yed
in
this
study
.
Ar
goUML
is
an
applicat
ion
for
creating
unied
modeling
language
(UML)
dia-
grams.
The
Ar
goUML
dataset
comprises
1,230
classes.
Azureus
with
1,449
classes
uses
the
bit-torrent
protocol
to
trans
mit
data.
The
XML
softw
are
library
is
called
Xerces.
There
are
513
clas
ses
in
t
he
Xereces
d
a
ta
le.
This
paper
uses
data
sets
supplied
in
[6].
T
o
balance
the
data
and
impro
v
e
the
characteristics,
we
e
xamined
SMO
TE
as
a
DPP
technique
in
this
paper
.
This
technique
randomly
chooses
the
minority
class
instance,
and
the
k
closest
minority
class
neighbors
are
found.
Afterw
ard,
a
line
se
gment
in
the
feature
space
is
formed
by
joining
a
and
b,
one
of
the
k
nearest
neighbors
(b),
at
random
to
generate
the
synthetic
instance.
The
tw
o
selected
e
xamples,
a
and
b,
are
used
to
create
the
synthetic
instances
by
using
the
con
v
e
x
combination
[24].
Consequently
,
ne
w
synthetic
instances
are
generated
that
can
be
referred
to
as
feature-boosting
techniques.
3.2.
Critical
analysis
of
classiers
on
anti-patter
n
detection
Functional
classiers
mainly
include
the
idea
of
neural
netw
orks
and
re
gression.
The
main
procedure
of
a
functional
classier
is
to
assign
data
labels
based
on
a
function
that
is
also
kno
wn
as
the
core
function
of
the
classier
.
In
this
w
ork,
v
e
functional
classiers
are
considered.
The
rst
classier
SVM
performs
the
classication
by
calculating
a
boundary
v
alue
kno
wn
as
a
h
yperplane
that
dif
ferentiates
tw
o
data
clas
ses.
SVM
is
broadly
cate
gorized
into
tw
o
groups;
linear
and
non-linear
SVM
[25].
Non-linear
SVM
utilizes
k
ernel
tricks
to
classify
data.
The
k
ernel
function
is
used
to
transform
lo
wer
dimensional
input
space
into
higher
dimensional
space.
K
ernel
logistic
re
gression
(KLR)
solv
es
the
classication
problems
by
transforming
the
real
input
space
into
a
high-dimensional
feature
space
by
utilizing
k
ernel
functions
[26].
LibSVM
is
a
complete
package
of
SVM.
This
classier
can
support
v
arious
SVM
formulations
for
classication
[27].
In
this
project,
support
Int
J
Artif
Intell,
V
ol.
14,
No.
1,
February
2025:
376–384
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Artif
Intell
ISSN:
2252-8938
❒
379
v
ector
classication
(SVC)
(tw
o-class
and
multi-class)
is
used.
Logistic
re
gression
predicts
t
he
odds
ratio
which
is
the
e
v
ent
that
needs
to
be
predicted,
for
e
xample,
ho
w
lik
ely
a
sample
is
an
anti-pattern
[28].
Multi-
layer
perceptron
classier
(MLPCL)
can
train
fully
connected
MLP
netw
orks
with
one
hidden
layer
.
F
or
further
processing,
all
attrib
utes
are
standardized
(it
gi
v
es
all
features
the
same
inuence).
In
this
project,
conjug
ate
gradient
descent
is
used
as
it
can
reduce
the
time
comple
xity
.
The
Bayes
classier
pro
vides
a
probabilistic
model.
that
mak
es
the
most
probable
prediction
for
a
ne
w
e
xample.
The
Bayes
theorem
is
used
to
calculate
the
conditional
probability
.
Nai
v
e
Bayes
(NB)
is
a
classier
that
uses
a
probabilistic
approach
for
the
classi
cation
tasks.
NB
follo
ws
the
follo
wing
equation
to
calculate
the
posterior
probability
for
each
class
[29].
The
classes
that
ha
v
e
the
highest
posterior
probabil
ity
are
the
output
of
prediction.
T
ree
classiers
di
vide
the
training
data
into
subsets
by
utilizing
a
sequence
of
conditional
statement
s.
Nai
v
e-Bayesian
tree
(NBT
ree)
classier
combines
a
decision
tree
(DT)
with
a
NB
classier
[30].
The
tree
is
constructed
recursi
v
ely
to
represent
the
attrib
utes
of
the
dataset.
The
nodes
of
the
tree
utilize
the
DT
classier
.
On
the
other
hand,
the
lea
v
es
contain
nai
v
e-Bayesian
classiers
[31].
DT
utilizes
instances
of
data
to
create
trees.
When
the
tree
construction
is
done,
it
prunes
the
trees
to
impro
v
e
the
model’
s
performance.
In
this
w
ork,
an
adv
anced
decision
tree
algorithm
namely
the
C4.5
algorithm
is
used
(called
J48
in
W
eka)
[32].
Decision
stump
(DS)
is
simply
a
DT
that
utilizes
a
single
attrib
ute
for
splitting
[33].
It
creates
a
single
root
tree
of
the
gi
v
en
v
alues
using
a
top-do
wn
greedy
approach.
RF
creates
a
forest
of
randomly
generated
DT
.
RF
combines
the
output
of
the
DT
to
form
the
nal
output.
The
number
of
trees
denes
the
accurac
y
of
the
classiers.
F
or
e
xample,
the
more
trees,
the
more
accurac
y
is
produced.
In
our
implementation,
100
trees
are
created
[34].
Meta
classi
ers
can
ensembl
e
dif
ferent
cl
assiers
to
enhance
the
perform
ance
of
the
base
classier
s.
Logit
boost
(LB)
is
a
meta
classier
that
is
based
on
the
AdaBoost
classier
.
Among
the
weighted
sample,
higher
weights
will
be
assigned
to
the
unclassied
data
[35].
Real
AdaBoost
(RAB)
is
a
boosting
classier
for
the
binary
classication
task.
Here,
the
weak
learner
algorithm
pro
vides
class
probability
estimation
[36].
Probability
estimation:
P
(
x
)
=
P
w
(
Y
=
1
|
x
)
ϵ
[0
,
1]
(1)
Here,
w
is
the
weight
of
the
training
data.
In
this
study
,
the
DS
is
used
as
the
base
classier
.
The
number
of
iterations
is
considered
10.
It
updates
the
weights
on
e
v
ery
iteration.
Bagging
is
an
ensemble
learner
that
randomly
redistrib
utes
the
original
training
dataset
for
indi
vi
dual
base
models
[37].
V
oting
or
a
v
eraging
their
prediction
probabilities
gi
v
es
a
nal
prediction.
Di
v
erse
ensemble
creation
by
opposi
tional
re-labeling
of
articial
training
e
xamples
(DECORA
TE)
is
designed
to
use
additional
articially
generated
training
data
[37].
Additionally
,
it
achie
v
es
equi
v
alent
performance
on
bigger
training
sets
and
g
ains
higher
accurac
y
than
boosting
on
the
small
training
set
[38].
Dagging
creates
se
v
eral
disjoint,
stratied
folds
of
the
data
to
predict
the
nal
output
by
a
v
eraging
them
[39].
V
oting
is
another
ensemble
algorithm
that
creates
tw
o
or
more
sub-models
with
the
base
cl
assiers
to
classify
data.
The
nal
prediction
is
done
by
aggre
g
ating
the
sub-model
predictions.
In
majority
v
oting,
the
nal
prediction
is
made
by
combining
the
sub-models
output.
The
class
that
gets
the
most
number
of
v
otes
will
be
the
output
of
the
classication
[40].
4.
RESUL
T
AN
AL
YSIS
The
performance
analysis
of
v
arious
ML
classiers
is
presented
in
this
section.
Fi
v
e
e
v
aluation
metri
cs
are
used:
accurac
y
,
F-measure,
precision,
recall,
and
recei
v
er
operating
characteristic
(R
OC).
An
Inte
l
Core
i3
processor
with
4
GB
RAM
and
a
64-bit
operating
system
is
used
for
the
e
xperiment.
WEKA
is
emplo
yed
in
the
implementation
of
the
classier
models.
The
details
are
described
in
the
follo
wing.
4.1.
Resear
ch
questions
and
e
v
aluation
Three
research
issues
are
addressed
in
this
section
to
v
alidate
the
anti-pattern
detection
technique.
In
RQ1,
the
signicance
of
the
DPP
technique
is
briey
discussed.
The
performance
of
classiers
in
terms
of
computing
time
is
sho
wn
in
RQ2.
Each
study
question
is
thoroughly
described
and
e
v
aluated
in
the
analysis
that
follo
ws.
4.1.1.
RQ1:
Ho
w
well
do
classiers
w
ork
with
data
pr
e-pr
ocessing
techniques
to
nd
anti-patter
ns?
In
the
proj
ect
of
Ar
goUML,
the
Dagging
classier
performs
better
in
terms
of
accurac
y
than
other
classiers.
The
accurac
y
of
the
Dagging
is
98.4%
when
DS
i
s
used
as
a
base
classier
.
The
accurac
y
of
DS
as
A
systematic
analysis
on
mac
hine
learning
classier
s
with
data
pr
e-pr
ocessing
to
...
(Nazneen
Akhter)
Evaluation Warning : The document was created with Spire.PDF for Python.
380
❒
ISSN:
2252-8938
an
indi
vidual
classier
is
91.37%.
Dagging
breaks
the
original
training
dataset
into
sm
all
chunks.
W
ith
these
small
chunks
of
data,
the
base
classier
can
train
itself
well.
So,
it
can
classify
data
more
accurately
from
the
test
dataset.
F
or
this
dataset,
the
NB
classier
performs
poorl
y
in
terms
of
accurac
y
,
79.37%.
This
is
because
of
the
high
dependenc
y
on
the
dataset
attrib
utes.
The
Dagging
classier
is
also
used
on
the
dataset
using
the
NB
as
the
base
classier
.
F
or
that
case,
the
accurac
y
of
Dagging
becomes
91.99%.
Hence,
it
can
be
said
that
meta-classiers
lik
e
Dagging
can
enhance
the
performance
of
weak-performed
classiers.
T
able
1
sho
ws
the
detailed
result
of
the
classiers
in
terms
of
accurac
y
for
Ar
goUML,
Azureus,
and
Xerces
project.
In
the
project
of
Azureus,
data
are
non-linearly
separable.
The
accurac
y
of
KLR
is
97%.
This
is
because
KLR
can
perform
well
with
non-linearly
separate
data.
KLR
of
fers
a
natural
es
timate
of
the
class
probability
.
On
the
other
side,
the
Dagging
classier
performed
poorly
for
this
project.
DS
is
used
as
the
base
classier
.
T
able
1.
Accurac
y
of
the
classiers
for
Ar
goUML,
Azureus,
and
Xerces
projects
Classiers
Accurac
y
(Ar
goUML)
(%)
Accurac
y
(Azureus)
(%)
Accurac
y
(Xerces)
(%)
SVM
92.48
90.32
91.49
NB
79.37
76.51
73.71
RF
92.78
92.03
93.35
DT
90.3
91.15
87.76
LibSVM
96.55
90.37
95.82
Logistic
90.78
81.27
88.58
KLR
91.82
96.99
89.90
NBT
ree
88.03
85.31
86.60
MLPCL
91.86
93.93
90.285
RAB
92.31
92.61
93.08
Bagging
91.78
91.8
88.47
Dagging
98.37
68.69
96.04
LB
91.82
96.01
92.59
DECORA
TE
92.92
91.67
92.75
DS
91.37
87.26
89.02
Performance
of
tree
classiers:
i
n
this
w
ork,
four
tree
classiers
ha
v
e
been
used;
DT
,
RF
,
NBT
ree,
and
DS.
Among
those,
the
RF
classier
performs
better
than
other
classiers
for
all
three
projects.
The
accurac
y
of
RF
for
Ar
goUML,
Azureus,
and
Xerces
are
92.8%,
92.03%,
93.4%,
respecti
v
ely
.
This
i
s
because
the
RF
is
an
ensem
ble
learner
that
nds
the
best
features
from
the
dataset
randomly
.
The
nal
prediction
is
done
by
combining
the
output
of
man
y
decision
trees.
NBT
ree
performs
lo
wer
than
other
tree
classiers.
The
accurac
y
of
NBT
ree
for
the
three
projects
is
88.03%,
84.81%,
and
86.61%,
respecti
v
ely
.
NBT
ree
utilizes
NB
in
the
lea
v
es
of
the
tree.
NB
tree
sho
ws
minimum
performance
for
the
dependent
dataset
which
is
the
main
reason
behind
performing
lo
w
for
NBT
ree.
Performance
of
ensemble
classiers:
for
the
project
of
Ar
goUML,
Dagging
classier
performs
the
best
whose
accurac
y
is
98.48%.
Figure
2
sho
ws
the
detail
ed
result
of
bagging,
boosting,
dagging,
and
DECORA
TE
classiers
according
to
their
accurac
y
.
Here,
DS
is
used
as
a
base
classier
.
Figure
2.
Performance
of
meta
classiers
and
v
oting
classier
for
Ar
goUML
project
Int
J
Artif
Intell,
V
ol.
14,
No.
1,
February
2025:
376–384
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Artif
Intell
ISSN:
2252-8938
❒
381
Performance
of
v
oting
classier:
in
this
w
ork,
tw
o
combination
rules
of
v
oting
are
used;
a
v
erage
v
oting
and
majority
v
oting.
The
sub-models
of
v
oting
classiers
are
decided
on
the
indi
vidual
performance
of
the
classiers.
F
or
e
xample,
in
the
Ar
goUML
project,
NB,
DT
,
logistic,
KLR,
NBT
ree,
and
MLPCL
classier’
s
performance
is
w
orse
than
other
classiers.
So,
the
y
are
used
as
the
base
classier
of
v
oting
for
the
Ar
goUML
project.
In
this
w
ork,
some
combinations
of
base
classie
rs
are
used
to
e
v
aluate
the
result.
First,
the
result
of
NB
and
DT
is
noted
do
wn.
Then
classiers
are
added
to
check
the
performance
of
the
V
oting
classier
one
by
one.
Figure
2
reports
the
detailed
result
of
the
v
oting
classier
according
to
their
base
classiers.
The
accurac
y
of
NB
and
DT
are
79.37%
and
90.3%
respecti
v
ely
.
When
it
is
used
as
the
sub-models
of
the
v
oting
classier
,
the
accurac
y
has
been
increased
to
91.68%
(both
for
a
v
erage
and
majority
v
oting).
In
the
Xerces
project,
NB
performs
poorly
than
other
classiers
in
terms
of
accurac
y
.
Other
classiers
lik
e
DT
,
logistic,
KLR,
and
bagging
cannot
pro
vide
satisf
actory
results
either
.
So,
the
dif
ferent
combinations
of
classiers
are
tak
en
as
a
base
classier
for
the
v
oting
meta
classier
to
e
v
aluate
the
result.
When
NB
and
DT
are
tak
en
as
a
combination
of
base
classiers,
the
accurac
y
has
been
increased
i.e.,
90.61%.
From
this
analysis,
it
can
be
said
that
v
oting
classiers
can
increase
the
performance
(proper
combination
of
base
classiers)
than
indi
vidual
classiers.
T
able
2
sho
ws
the
result
of
the
classiers
in
terms
of
recall
and
accurac
y
.
T
able
2.
Recall
and
accurac
y
of
classiers
with
and
without
DPP
for
Ar
goUML
project
Classiers
Recall
(without
DPP)
Recall
(with
DPP)
Accurac
y
(without
DPP)
Accurac
y
(with
DPP)
SVM
0.925
0.948
0.921
0.925
NB
0.739
0.833
0.739
0.793
RF
0.926
0.949
0.915
0.927
DT
0.903
0.935
0.901
0.903
LibSVM
0.954
0.964
0.953
0.965
Logistic
0.899
0.932
0.899
0.907
KLR
0.918
0.943
0.912
0.918
NBT
ree
0.931
0.941
0.875
0.88
MLP
CL
0.916
0.943
0.916
0.918
RAB
0.919
0.945
0.918
0.923
Bagging
0.916
0.943
0.915
0.917
Dagging
0.981
0.983
0.974
0.983
LB
0.918
0.944
0.908
0.918
DECORA
TE
0.925
0.949
0.925
0.929
DS
0.913
0.942
0.913
0.913
In
comparison
wit
h
the
DPP
,
we
found
most
of
the
classiers
perform
well.
Dagging
classier
per
-
forms
better
than
other
classiers
in
terms
of
accurac
y
i.e.,
97.4%
whereas
the
accurac
y
of
dagging
is
98.3%
with
the
DPP
.
Maig
a
et
al.
[5],
[6]
proposed
tw
o
SVM-based
w
orks
SMURF
and
SVMDetect.
The
recall
of
SMURF
and
SVMDetect
is
84.09%.
Whereas
after
using
DPP
the
recall
of
SVM
enhances
to
94.8%
in
this
w
ork.
Furthermore,
from
T
able
3
we
can
see
that,
DPP
is
increasing
the
recall
of
the
classiers
to
detect
anti-patterns.
T
able
3.
Comparison
of
proposed
w
ord
with
e
xisting
models
in
terms
of
recall
Authors
Model
Recall
(%)
Maig
a
et
al.
[5]
SMURF
84.09
Maig
a
et
al.
[6]
SVMDetect
84.09
Proposed
method
SVM
with
DPP
94.8
4.1.2.
RQ2:
Ho
w
do
the
classiers
perf
orm
in
terms
of
time?
In
this
w
ork,
sixteen
dif
ferent
classiers
are
tak
en
into
consideration.
Most
of
the
classiers
perf
ormed
well
in
terms
of
accurac
y
and
time.
NB,
DT
,
SVM,
LibSVM,
logistic,
and
LB
ha
v
e
tak
en
less
time
to
b
uild
a
model.
The
y
ha
v
e
tak
en
0.01
seconds
to
create
the
model.
NBT
ree
has
tak
en
the
most
time
to
b
uild
the
model
which
is
0.68
seconds.
The
v
oting
classier
also
tak
es
more
time
when
NBT
ree
is
added
with
other
classiers.
From
the
performance
analysis
of
dif
ferent
classiers
on
dif
ferent
projects,
it
can
be
said
that
NBT
ree
performs
poorly
in
terms
of
time
than
other
classiers.
Figure
3
sho
ws
the
detailed
result
of
the
time
that
is
tak
en
by
the
classiers
to
b
uild
the
model.
A
systematic
analysis
on
mac
hine
learning
classier
s
with
data
pr
e-pr
ocessing
to
...
(Nazneen
Akhter)
Evaluation Warning : The document was created with Spire.PDF for Python.
382
❒
ISSN:
2252-8938
Figure
3.
T
ime
comple
xity
to
b
uild
the
model
for
Ar
goUML
project
5.
CONCLUSION
Anti-pattern
is
a
signicant
nuisance
in
source
code
which
violates
the
standard
of
design
pattern.
Softw
are
quality
mak
es
the
softw
are
understandable,
reusable,
and
cost-ef
fecti
v
e.
Hence,
anti-pattern
detection
has
become
a
major
concern.
Ho
we
v
er
,
there
are
man
y
dif
culties
in
choosing
appropriate
classiers
to
identify
the
design
pattern.
In
this
paper
,
a
detailed
discussion
about
16
dif
ferent
well-kno
wn
classiers
is
done
to
detect
anti-patterns.
This
analysis
will
help
practitioners
to
choose
suitable
classiers.
In
addition,
to
impro
v
e
the
performance
of
the
class
iers,
SMO
TE
DPP
is
used.
Our
classiers
performed
better
with
DPP
than
the
e
xisting
models
without
DPP
.
Dagging,
KLR,
and
LibSVM
perform
better
in
terms
of
accurac
y
for
respecti
v
e
projects.
The
accurac
y
of
them
is
98.4%,
97%,
and
95.83%
respecti
v
ely
.
Ho
we
v
er
,
this
w
ork
presents
better
accurac
y
than
e
xisting
models.
Ho
we
v
er
,
the
data
size
used
for
this
w
ork
is
small.
In
the
future,
the
dataset
size
can
be
increased
by
adding
more
samples.
The
data
set
can
be
applied
to
a
deep
learning
model
adopting
an
attention
mechanism.
A
CKNO
WLEDGEMENTS
This
w
ork
w
as
supported
by
the
W
oosong
Uni
v
ersity
Academic
Research
Fund,
2024,
South
K
orea.
REFERENCES
[1]
G.
Rasool
and
Z.
Arshad,
“
A
re
vie
w
of
code
smell
mining
techniques,
”
J
ournal
of
Softwar
e:
Evolution
and
Pr
ocess
,
v
ol.
27,
no.
11,
pp.
867–895,
2015,
doi:
10.1002/smr
.1737.
[2]
F
.
P
alomba,
G.
Ba
v
ota,
M.
D.
Penta,
R.
Oli
v
eto,
A.
D.
Lucia,
and
D.
Posh
yv
an
yk,
“Detecting
bad
smells
in
source
code
using
change
history
information,
”
in
2013
28th
IEEE/A
CM
International
Confer
ence
on
A
utomated
Softwar
e
Engineering
(ASE)
,
2013,
pp.
268–278,
doi:
10.1109/ASE.2013.6693086.
[3]
M
.
F
o
wler
,
K.
Beck,
J.
Brant,
W
.
Opdyk
e,
D.
Roberts,
and
E.
Gamma,
Refactoring:
impr
o
ving
the
design
of
e
xisting
code
.
Ne
w
Y
ork,
United
States
of
America:
Pearson
Education,
1999.
[4]
N.
Akhter
,
S.
Rahman,
and
K.
A.
T
aher
,
“
An
anti-pattern
detection
technique
using
machine
learning
to
impro
v
e
code
quality
,
”
in
2021
International
Confer
ence
on
Information
and
Communication
T
ec
hnolo
gy
for
Sustainable
De
velopment
(ICICT4SD)
,
2021,
pp.
356–360,
doi:
10.1109/ICICT4SD50815.2021.9396937.
[5]
A.
Maig
a,
N.
Ali,
N.
Bhattacharya,
A.
Sabane,
Y
.
-G.
Gueheneuc,
and
E.
Aimeur
,
“SMURF:
A
SVM-based
incremental
anti-pattern
detection
approach,
”
in
2012
19th
W
orking
Confer
ence
on
Re
ver
se
Engineering
,
2012,
pp.
466–475,
doi:
10.1109/WCRE.2012.56.
[6]
A
.
Maig
a
et
al.
,
“Support
v
ector
machines
for
anti-pattern
detection,
”
in
2012
27th
IEEE/A
CM
International
Conf
er
ence
on
A
uto-
mated
Softwar
e
Engineering
,
2012,
pp.
278–281,
doi:
10.1145/2351676.2351723.
[7]
F
.
Pecorelli,
F
.
P
alomba,
D.
D.
Nucci,
and
A.
D.
Lucia,
“Comparing
heuristic
and
machine
learning
approaches
for
metric-based
code
smell
detection,
”
in
IEEE
International
Confer
ence
on
Pr
o
gr
am
Compr
ehension
,
2019,
pp.
93–104,
doi:
10.1109/ICPC.2019.00023.
[8]
F
.
A.
F
ontana,
M.
V
.
M
¨
antyl
¨
a,
M.
Zanoni,
and
A.
Marino,
“Comparing
and
e
xperimenting
machine
learning
techniques
for
code
smell
detection,
”
Empirical
Softwar
e
Engineering
,
v
ol.
21,
no.
3,
pp.
1143–1191,
2016,
doi:
10.1007/s10664-015-9378-4.
[9]
M.
K
essentini,
H.
Sahraoui,
M.
Boukadoum,
and
M.
W
immer
,
“Search-based
design
defects
detection
by
e
xample,
”
in
Fundamental
Appr
oac
hes
to
Softwar
e
Engineering
,
Berlin,
Heidelber
g,
2011,
pp.
401–415,
doi:
10.1007/978-3-642-19811-3
28.
[10]
I.
P
ol
´
a
ˇ
sek,
S.
Snopk
o,
and
I.
Kapustik,
“
Automatic
identication
of
the
anti-patterns
using
the
rule-based
approach,
”
in
2012
IEEE
10th
J
ubilee
International
Symposium
on
Intellig
ent
Systems
and
Informatics
,
2012,
pp.
283–286,
doi:
10.1109/SISY
.2012.6339530.
[11]
M
.
T
.
Aras
and
Y
.
E.
Selcuk,
“Metric
and
rule
based
aut
omated
detection
of
antipatterns
in
object-oriented
softw
are
sys-
tems,
”
in
2016
7t
h
International
Confer
ence
on
Computer
Science
and
Information
T
ec
hnolo
gy
(CSIT)
,
2016,
pp.
1–6,
doi:
10.1109/CSIT
.2016.7549470.
[12]
S
.
Uchiyama,
H.
W
ashizaki,
Y
.
Fukaza
w
a,
and
A.
K
ubo,
“Design
pattern
detection
using
softw
are
metrics
and
machine
learning,
”
F
ir
st
international
workshop
on
model-driven
softwar
e
migr
ation
(MDSM
2011)
,
v
ol.
708,
pp.
38–47,
2011.
[13]
A.
Barbez,
F
.
Khomh,
and
Y
.-G.
Gu
´
eh
´
eneuc,
“
A
machine-learning
based
ensemble
method
for
anti-patterns
detection,
”
J
ournal
of
Systems
and
Softwar
e
,
v
ol.
161,
2020,
doi:
10.1016/j.jss.2019.110486.
[14]
Y
.
Y
in,
Q.
Su,
and
L.
Liu,
“Softw
are
smell
detection
based
on
machine
learning
and
its
empirical
study
,
”
in
Second
T
ar
g
et
Reco
gni-
tion
and
Articial
Intellig
ence
Summit
F
orum
,
2020,
doi:
10.1117/12.2550500.
Int
J
Artif
Intell,
V
ol.
14,
No.
1,
February
2025:
376–384
Evaluation Warning : The document was created with Spire.PDF for Python.
Int
J
Artif
Intell
ISSN:
2252-8938
❒
383
[15]
U.
Azadi,
F
.
A.
F
ontana,
and
M.
Zanoni,
“Machine
learning
based
code
smell
detection
through
W
ekaNose,
”
in
Pr
oceed-
ings
of
the
40th
International
Confer
ence
on
Softwar
e
Engineering:
Companion
Pr
oceeedings
,
2018,
pp.
288–289,
doi
:
10.1145/3183440.3194974.
[16]
D.
Cruz,
A.
Santana,
and
E.
Figueiredo,
“Detecting
bad
smells
with
machine
learning
algorithms,
”
in
Pr
oceedings
of
the
3r
d
International
Confer
ence
on
T
ec
hnical
Debt
,
2020,
pp.
31–40,
doi:
10.1145/3387906.3388618.
[17]
L.
K
umar
and
A.
Sureka,
“
An
empirical
anal
ysis
on
web
service
anti-pattern
detection
using
a
machine
learning
frame
w
ork,
”
in
2018
IEEE
42nd
Annual
Computer
Softwar
e
and
Applications
Confer
ence
(COMPSA
C)
,
2018,
pp.
2–11,
doi:
10.1109/COMP-
SA
C.2018.00010.
[18]
N.
Pritam
et
al.
,
“
Assessment
of
code
smell
for
predicting
class
change
proneness
using
machine
learning,
”
IEEE
Access
,
v
ol.
7,
pp.
37414–37425,
2019,
doi:
10.1109/A
CCESS.2019.2905133.
[19]
A.
Jesudoss,
S.
Maneesha,
and
T
.
L.
N.
Dur
g
a,
“Identication
of
code
smell
using
machine
learning,
”
in
2019
International
Confer
-
ence
on
Intellig
ent
Computing
and
Contr
ol
Systems
(ICCS)
,
2019,
pp.
54–58,
doi:
10.1109/ICCS45141.2019.9065317.
[20]
M.
H.
-Kacem
and
N.
Bouassida,
“
A
h
ybrid
approach
to
detect
code
smells
using
deep
learning,
”
in
Pr
oceedings
of
the
13th
International
Confer
ence
on
Evaluation
of
No
vel
Appr
oac
hes
to
Softwar
e
Engi
neering
,
2018,
pp.
137–146,
doi:
10.5220/0006709801370146.
[21]
A.
K.
Das,
S.
Y
ada
v
,
and
S.
Dhal,
“Detecting
code
smells
using
deep
learning,
”
in
IEEE
Re
gion
10
Annual
International
Confer
ence
,
Pr
oceedings/TENCON
,
2019,
pp.
2081–2086,
doi:
10.1109/TENCON.2019.8929628.
[22]
A.
Barbez,
F
.
Khomh,
and
Y
.-G.
Gueheneuc,
“Deep
learning
anti-patterns
from
code
metrics
history
,
”
in
2019
IEEE
Internat
ional
Confer
ence
on
Softwar
e
Maintenance
and
Evolution
(ICSME)
,
2019,
pp.
114–124,
doi:
10.1109/ICSME.2019.00021.
[23]
R.
K.
Dash,
“Selection
of
the
best
classier
from
dif
ferent
datasets
using
WEKA,
”
International
J
ournal
of
Engineering
Resear
c
h
&
T
ec
hnolo
gy
(IJERT)
,
v
ol.
2,
no.
3,
pp.
1–7,
2013.
[24]
N.
V
.
Cha
wla,
K.
W
.
Bo
wyer
,
L.
O.
Hall,
and
W
.
P
.
K
e
gelme
yer
,
“SMO
TE:
Synthetic
minority
o
v
er
-s
ampling
technique,
”
J
ournal
of
Articial
Intellig
ence
Resear
c
h
,
v
ol.
16,
pp.
321–357,
2002,
doi:
10.1613/jair
.953.
[25]
V
.
Jakkul
a,
“T
utorial
on
support
v
ector
machine
(SVM),
”
Sc
hool
of
EECS,
W
ashington
State
Univer
sity
,
v
ol.
37,
pp.
1–13,
2011.
[26]
W
.
Chen,
X.
Y
an,
Z.
Zhao,
H.
Hong,
D.
T
.
Bui,
and
B.
Pradhan,
“Spatial
prediction
of
landslide
susceptibility
using
data
mining-
based
k
ernel
logistic
re
gression,
nai
v
e
Bayes
and
RBFNetw
ork
models
for
the
Long
County
area
(China),
”
Bulletin
of
Engineering
Geolo
gy
and
the
En
vir
onment
,
v
ol.
78,
no.
1,
pp.
247–266,
2019,
doi:
10.1007/s10064-018-1256-z.
[27]
C.
C.
Chang
and
C.
J.
Lin,
“LIBSVM:
A
library
for
support
v
ector
machines,
”
A
CM
T
r
ansactions
on
Intellig
ent
Systems
and
T
ec
hnolo
gy
,
v
ol.
2,
no.
3,
2011,
doi:
10.1145/1961189.1961199.
[28]
S.
L.
Cessie
and
J.
C.
V
.
Houwelingen,
“Ridge
estimators
in
logistic
re
gression,
”
Applied
statistics
,
v
ol.
41,
no.
1,
pp.
191-201,
1992.
[29]
I.
Rish,
“
An
empirical
study
of
the
nai
v
e
bayes
classier
,
”
IJCAI
2001
workshop
on
empirical
methods
in
articial
intellig
ence
,
v
ol.
3,
no.
22,
pp.
41–46,
2001.
[30]
R
.
K
oha
vi,
“Scaling
up
the
accurac
y
of
nai
v
e-Bayes
classiers:
A
decision-tree
h
ybrid,
”
in
2nd
International
Confer
ence
on
Knowl-
edg
e
Disco
very
and
Data
Mining
,
KDD
1996
,
1996,
pp.
202–207.
[31]
T
.
M.
Christian
and
M.
A
yub,
“Exploration
of
classication
using
NBT
ree
for
predicting
students’
performance,
”
in
2014
Interna-
tional
Confer
ence
on
Data
and
Softwar
e
Engineering
(ICODSE)
,
2014,
pp.
1–6,
doi:
10.1109/ICODSE.2014.7062654.
[32]
J.
R.
Quinlan,
C4.5:
Pr
o
gr
am
for
mac
hine
learning
.
San
Mateo,
California:
Mor
g
an
Kaufmann
Publishers,
1993.
[33]
C
.
Sammut
and
G.
I.
W
ebb,
“Decision
stump,
”
in
Encyclopedia
of
Mac
hine
Learning
,
Boston,
MA:
Springer
,
2011,
pp.
262–263,
doi:
10.1007/978-0-387-30164-8
202.
[34]
L.
Breiman,
“
Random
forests,
”
in
Mac
hine
Learning
,
Springer
,
2001,
pp.
5–32,
doi:
10.1023/A:1010933404324.
[35]
B.
Zi
´
ołk
o,
S.
Manandhar
,
R.
C.
W
ilson,
and
M.
Zi
´
ołk
o,
“Logitboost
WEKA
classi
er
speech
se
gmentation,
”
in
2008
IEEE
Interna-
tional
Confer
ence
on
Multimedia
and
Expo
,
2008,
pp.
1297–1300,
doi:
10.1109/ICME.2008.4607680.
[36]
J.
Friedman,
T
.
Hastie,
and
R.
T
ibshirani,
“
Additi
v
e
logisti
c
re
gression:
a
statistical
vie
w
of
boosting
(with
discussion
and
a
rejoinder
by
the
authors),
”
The
Annals
of
Statistics
,
v
ol.
28,
no.
2,
pp.
337–407,
2000,
doi:
10.1214/aos/1016218223.
[37]
L.
Breiman,
“
Bagging
predictors,
”
Mac
hine
Learning
,
v
ol.
24,
no.
2,
pp.
123–140,
1996,
doi:
10.1023/A:1018054314350.
[38]
P
.
Ro
y
,
R.
Chakrabortty
,
I.
Cho
wdhuri,
S.
Malik,
B.
Das,
and
S.
C.
P
al,
“De
v
elopment
of
dif
ferent
machine
learning
ensemble
classier
for
gully
erosion
s
usceptibility
in
Gandhesw
ari
W
atershed
of
W
est
Beng
al,
India
,
”
in
Mac
hine
Learning
for
Intellig
ent
Decision
Science
,
2020,
pp.
1–26,
doi:
10.1007/978-981-15-3689-2
1.
[39]
J.
Kittler
,
M.
Hatef,
R.
P
.
W
.
Duin,
and
J.
Matas,
“On
combining
classiers,
”
IEEE
T
r
ansactions
on
P
attern
Analysis
and
Mac
hine
Intellig
ence
,
v
ol.
20,
no.
3,
pp.
226–239,
1998,
doi:
10.1109/34.667881.
[40]
S.
Sri
v
asta
v
a,
“W
eka:
a
tool
for
data
preprocessing,
classication,
ensemble,
clustering,
and
association
rule
mining,
”
International
J
ournal
of
Computer
Applications
,
v
ol.
88,
no.
10,
pp.
26–29,
2014,
doi:
10.5120/15389-3809.
BIOGRAPHIES
OF
A
UTHORS
Nazneen
Akhter
completed
her
B.Sc.
in
Information
and
Communication
Engineeri
ng
and
M.Sc.
in
Information
and
Communication
Engineering
from
the
Department
of
Information
and
Communication
T
echnology
of
Bangladesh
Uni
v
ersity
of
Professionals,
besides
se
v
eral
professional
skills.
She
joined
the
Department
of
Information
and
Communication
T
echnology
as
a
lecturer
in
2021.
She
is
currently
serving
as
a
lecturer
in
the
Department
of
Computer
Science
and
Engineering
at
Bangladesh
Uni
v
ersity
of
Professionals,
Dhaka,
Bangladesh.
Her
research
areas
of
interest
include
softw
are
engineering,
articial
intelligence
and
applications,
and
natural
language
processing.
She
can
be
contacted
at
email:
nazneen.akhter@b
up.edu.bd.
A
systematic
analysis
on
mac
hine
learning
classier
s
with
data
pr
e-pr
ocessing
to
...
(Nazneen
Akhter)
Evaluation Warning : The document was created with Spire.PDF for Python.
384
❒
ISSN:
2252-8938
Afrina
Khatun
obtained
her
B.Sc.
in
Softw
are
Engineering
and
M.Sc.
in
Softw
are
Engineering
from
the
Institute
of
Information
T
echnology
,
Uni
v
ersity
of
Dhaka.
She
joined
the
Department
of
Information
and
Communication
T
echnology
as
a
lecturer
in
2021.
Currentl
y
,
she
is
w
orking
as
an
Assistant
Professor
at
the
Bangladesh
Uni
v
ersity
of
Professionals
in
the
Department
of
Computer
Science
and
Engineering.
Her
research
interest
is
the
application
of
machine
learning
and
deep
learning
in
softw
are
engineering.
She
can
be
contacted
at
email:
afrina.khatun@b
up.edu.bd.
Dr
.
Md.
Sazzadur
Rahman
is
w
orking
as
a
Professor
at
the
Institute
of
Information
T
echnology
(IIT),
Jahangirnag
ar
Uni
v
ersity
,
Sa
v
ar
,
Dhaka-1342,
Bangladesh.
He
recei
v
ed
hi
s
Ph.D.
in
Material
Science
(Surf
ace
Science)
from
K
yushu
Uni
v
ersity
,
Japan
as
a
MEXT
scholarship
grantee
in
2015.
Pre
viously
,
he
recei
v
ed
his
B.Sc.
(Hons.)
de
gree
and
M.Sc.
de
gree
in
Applied
Ph
ysics,
Elec-
tronics,
and
Communication
Engineeri
ng
from
the
Uni
v
ersity
of
Dhaka,
Bangladesh.
He
has
about
14
years
of
e
xperi
ence
in
research
and
teaching
in
electronics,
internet
of
things,
machine
learning,
communication
engineering,
and
material
science.
No
w
he
is
w
orking
as
a
co-in
v
estig
ator
on
the
DI-
VERSASIA
Erasmus+
project
(project
number
618615-EPP-1-2020-1-UKEPPKA2-CBHEJP).
He
can
be
contacted
at
email:
sazzad@juni
v
.edu.
A.
S.
M.
Sanwar
Hosen
recei
v
ed
the
M.S.
and
Ph.D.
de
grees
in
computer
science
and
engineering
from
Jeonb
uk
National
U
ni
v
ersity
,
Jeonju,
South
K
orea.
He
is
currently
an
Assistant
Professor
at
the
Department
of
Articial
Intelligence
and
Big
Data,
W
oosong
Uni
v
ersity
,
Daejeon,
South
K
orea.
He
has
published
se
v
eral
articles
in
journals
and
international
conferences.
His
current
research
interests
include
wireless
sensor
netw
orks,
the
internet
of
things,
fog-cloud
computing,
c
yber
security
,
articial
intelligence,
and
blockchain
technology
.
He
has
been
an
e
xpert
re
vie
wer
for
IEEE
T
ransactions,
Else
vier
,
Springer
,
and
MDPI
journals
and
mag
azi
nes.
He
has
also
been
in
vited
to
serv
e
as
the
T
echnic
al
Program
Committee
Member
f
or
se
v
eral
reputed
international
confer
ences,
such
as
IEEE
A
CM.
He
can
be
contacted
at
email:
sanw
ar@wsu.ac.kr
or
sanw
ar@jbnu.ac.kr
.
Mohammad
Shahidul
Islam
recei
v
ed
his
Ph.D.
in
Computer
Science
and
Information
Systems
from
the
Na
tional
Institute
of
De
v
elopment
Administration
(NID
A),
Bangk
ok,
Thailand,
B.T
ech.
in
Computer
Science
and
T
echnology
from
the
Indian
Institute
of
T
echnology-Roork
ee
(IITR),
Uttar
Pradesh,
India
in
2002,
M.Sc.
in
Mobile
Computing
and
Communication
from
Uni-
v
ersity
of
Greenwich,
London,
U.K
in
2008.
His
research
interests
include
articial
intelligence,
robotics,
machine
learning,
pattern
recognition,
and
rene
w
able
ener
gy
.
He
can
be
contacted
at
email:
suv
a.93@grads.nida.ac.th
or
suv
a93@gmail.com.
Int
J
Artif
Intell,
V
ol.
14,
No.
1,
February
2025:
376–384
Evaluation Warning : The document was created with Spire.PDF for Python.