Indonesian
J
our
nal
of
Electrical
Engineering
and
Computer
Science
V
ol.
36,
No.
2,
No
v
ember
2024,
pp.
1052
1069
ISSN:
2502-4752,
DOI:
10.11591/ijeecs.v36.i2.pp1052-1069
r
1052
Bee-inspir
ed
kno
wledge
transfer:
synthesizing
data
f
or
enhanced
deep
lear
ning
explainability
Kritanat
Chungnoy
1
,
T
anator
n
T
anantong
1,2
,
P
okpong
Songmuang
1,2
1
Department
of
Computer
Science,
F
aculty
of
Science
and
T
echnology
,
Thammasat
Uni
v
ersity
,
P
athum
Thani,
Thailand
2
Thammasat
Uni
v
ersity
Research
Unit
in
Data
Inno
v
ation
and
Artificial
Intelligence,
Thammasat
Uni
v
ersity
,
P
athum
Thani,
Thailand
Article
Inf
o
Article
history:
Recei
v
ed
Mar
13,
2024
Re
vised
Jul
3,
2024
Accepted
Jul
14,
2024
K
eyw
ords:
Bee
algorithm
Deep
learning
Explainability
Explainable
AI
Synthesize
data
ABSTRA
CT
This
paper
presents
the
generation
method
for
an
e
xplainable
model
based
on
the
gi
v
en
information
of
a
black
box
model
using
a
concept
of
kno
wledge
trans-
fer
to
synthesize
a
dataset.
The
proposed
method
applies
with
GAN
and
Bee
algorithm
(B
A)
for
data
synthesis
technique
to
synthesize
a
dataset
by
consid-
ering
loss
v
a
lue
in
a
kno
wledge
transferring
process
to
inherit
the
significance
of
features.
The
synthesized
dataset
is
used
to
train
for
a
proxy
model
as
an
e
x-
plainable
model.
The
result
of
the
e
xperiment
indicates
that
kno
wledge
transfer
from
Bee
algo
better
than
generati
v
e
adv
ersarial
netw
ork
(GAN)
in
terms
of
the
coef
ficient
of
determination
R2.
In
addition,
e
xplainable
models
from
the
syn-
thesized
data
of
the
Bee-based
method
obtai
ns
F1
score
superior
to
those
from
the
GAN-based
method
in
all
datasets
and
settings.
The
dataset
synthesized
from
the
Bee-based
method
produces
the
e
xplainable
prediction
model
that
has
similar
top-10
features
according
to
similarity
score
of
0.6718
using
shaple
y
ad-
diti
v
e
e
xplanations
(SHAP)
feature
importance
which
is
higher
than
those
from
GAN-based
method
for
0.4218
in
a
v
erage.
Additionally
,
e
xperimental
result
to
e
v
aluate
accurac
y
sho
ws
that
F1
score
from
e
xplainable
models
from
the
Bee-
based
method
are
closed
to
F1
score
from
a
model
generated
from
the
original
dataset.
This
is
an
open
access
article
under
the
CC
BY
-SA
license
.
Corresponding
A
uthor:
T
anatorn
T
anantong
Thammasat
Uni
v
ersity
Research
Unit
in
Data
Inno
v
ation
and
Artificial
Intelligence,
Thammasat
Uni
v
ersity
P
athum
Thani,
Thailand
Email:
tanatorn@sci.tu.ac.th
1.
INTR
ODUCTION
Artificial
intelligence
(AI)
has
been
de
v
eloped
and
used
in
man
y
tasks
and
domains
[1].
Among
man
y
approaches,
deep
learning
approach
has
g
ained
significant
popularity
and
widespread
adoption
due
to
scalability
,
v
ersatility
,
and
state-of-the-art
performance
[2]–[12].
Ho
we
v
er
,
models
from
deep
learning
are
a
black
box
model
with
deep,
computationally
e
xpensi
v
e
layers
and
ha
v
e
been
recently
found
vulnerable
to
spoofing
with
well-designed
input
samples
in
man
y
safety-critical
applicati
ons
[13].
The
black
box
model
refers
to
the
lack
of
transparenc
y
and
interpretability
from
containing
a
lar
ge
number
of
comple
x
parameters,
making
it
challenging
to
underst
and
wh
y
the
model
chooses
particular
predictions.
Unfortunately
,
transparenc
y
and
accountability
are
ones
of
crucial
f
actors
to
w
ards
ethics
guidelines
for
trust
w
orth
y
AI
[14]
by
European
Commission
standard.
Addressing
the
black
box
nature
of
deep
learning
models
is
an
ongoing
research
area,
and
o
ne
of
the
solutions
is
e
xplanatory
artificial
intelligence
(XAI)
which
is
a
technique
to
mak
e
these
models
more
interpretable
without
sharply
sacrificing
their
performance.
J
ournal
homepage:
http://ijeecs.iaescor
e
.com
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
r
1053
XAI
refers
to
techniques
and
approaches
that
aim
to
mak
e
AI
systems,
particularly
deep
le
arning
models,
more
transparent
and
understandable
[15].
The
goal
of
XAI
is
to
pro
vide
insights
into
wh
y
AI
systems
mak
e
specific
decisions
or
predictions,
enhancing
their
interpretability
and
enabli
ng
users
to
trust
and
v
erify
their
outputs.
XAI
t
hus
plays
a
role
in
bridging
the
g
ap
between
the
technical
comple
xity
of
AI
models
and
the
need
for
human
understanding,
accountability
,
and
trust.
As
the
use
of
deep
learning
models
continues
to
inte
grate
into
critical
applications,
the
demand
for
XAI
techniques
gro
ws
accordingly
and
has
become
one
of
focused
task
in
AI
de
v
elopment
[16].
There
are
se
v
eral
XAI
methods
such
as
rule-based
e
xplanations,
attention
mechanism,
and
feature
importance.
Ho
we
v
er
,
most
of
t
he
e
xisting
method
requires
an
original
dataset
to
iden-
tify
which
input
features
or
v
ariables
ha
v
e
the
most
significant
impact
on
a
model’
s
predictions.
T
echniques
such
as
feature
importance
scores,
permutation
importance,
and
SHAP
(shaple
y
addi
ti
v
e
e
xplanations)
v
alues
can
help
to
quantify
the
contrib
ution
of
each
feature
to
the
model’
s
output
for
e
xplanation.
Some
e
xisting
re-
search
focus
on
e
xplaining
indi
vidual
predictions
using
local
interpretable
model-agnostic
e
xplanations
(LIME)
to
approximate
the
model’
s
beha
vior
for
a
specific
input
by
generating
a
simpler
and
interpretable
model
(such
as
linear
re
gression
and
decision
tree
(DT))
around
that
input
to
e
xplain
its
prediction.
Unfortunately
,
most
of
the
practical
AI
models
may
not
share
their
training
data
due
to
cop
y
right
or
trade
secret.
Thus,
it
becomes
dif
ficult
to
generate
an
XAI
model
without
the
training
dataset.
In
this
paper
,
we
propose
a
method
to
generate
XAI
model
based
on
the
concept
of
kno
wledge
t
ransfer
inspired
by
kno
wledge
dis
tillation
(KD)
architecture.
The
method
does
not
require
the
training
dataset
of
the
black
box
model,
b
ut
the
comple
x
model
itself
as
a
teacher
model.
The
method
in
v
olv
es
transferring
kno
wledge
directly
from
a
teacher
model
(lar
ge
machine
learning
model)
to
generate
a
synthesized
dataset
that
ensembles
the
characteristics
of
a
teacher
model
for
traini
ng
for
a
simpler
and
interpretable
prediction
model
as
an
e
xplainable
model
to
w
ards
the
original
comple
x
model.
The
obtained
e
xplainable
model
hence
can
be
used
to
e
xplain
the
prediction
and
trace
for
bias
in
decision-making
of
deep
learning-based
AI.
The
core
h
ypothesis
of
this
method
is
that
kno
wledge
within
a
comple
x
model
contains
can
be
transferred
to
its
synthesized
data,
and
the
kno
wledgeable
synthesi
zed
data
can
be
used
to
generate
a
prediction
model
that
resembles
to
the
original
comple
x
model
in
terms
of
similarity
of
feature
importance
and
capability
to
predict
similar
to
the
original
model.
The
chosen
data
synthesis
method
in
this
w
ork
includes
the
frequently
used
generati
v
e
adv
ersarial
netw
ork
(GAN)
and
Bee
algorithm
(B
A).
The
scope
of
data
types
in
this
w
ork
limits
to
tab
ular
and
numerical
data.
The
contrib
utions
of
this
paper
includes:
W
e
propose
a
no
v
el
m
ethod
to
generate
an
XAI
model
from
a
comple
x
black
box
model
bases
on
KD-based
data
synthesis
call
“kno
wledge-transferred
data
synthesis
for
e
xplainable
AI”
(KT
-XAI).
W
e
demonstrate
on
ho
w
to
apply
GAN
and
B
A
for
kno
wledge-transferred
data
synthesis,
called
KT
-GAN
and
KT
-Bee,
respecti
v
ely
.
W
e
demonstrate
that
the
proposed
KT
-Bee
contrib
utes
to
synthesize
a
dataset
that
inherits
kno
wledge
from
a
teacher
model
superior
to
KT
-GAN.
W
e
demonstrate
that
synthesized
datasets
from
KT
-Bee
are
systematically
similar
to
the
model
generated
from
the
original
dataset
by
sharing
top-k
important
features
and
can
be
used
as
an
e
xplainable
model
to
the
black
box
model.
W
e
observ
e
the
kno
wledge
transfer
parameters
including
data
distrib
ution,
transfer
accurac
y
and
sharing
of
top-k
important
features
and
found
that
the
size
of
synthesized
dataset
does
not
af
fect
the
performance
of
kno
wledge-transferred
data
synthesis.
2.
B
A
CKGR
OUND
2.1.
Explainable
artificial
intelligent
XAI
is
an
approach
to
AI
that
aims
to
mak
e
machine
learning
models
and
their
decisions
more
under
-
standable
and
interpretable
by
humans
since
man
y
AI
models
,
particularly
deep
learning
models,
are
dif
ficult
to
understand
ho
w
the
y
arri
v
e
at
their
conclusions.
XAI
thus
helps
to
identify
and
mitig
ate
bias
in
AI
models
by
re
v
ealing
which
features
or
data
points
are
i
nfluencing
decisions
to
impro
v
e
transparenc
y
and
allo
ws
for
the
identification
and
rectification
of
unf
air
or
discriminatory
outcomes
[17].
DTs
is
another
approach
often
used
as
a
proxy
model.
There
are
se
v
eral
w
orks
mentioned
on
decomposing
neural
netw
ork
models
into
DT
models.
Deep
rule
e
xtraction
from
DTs
(DeepRED)
[18]
demonstrates
the
e
xtending
of
the
comprehensible
Bee-inspir
ed
knowledg
e
tr
ansfer:
synthesizing
data
for
enhanced
...
(Kritanat
Chungnoy)
Evaluation Warning : The document was created with Spire.PDF for Python.
1054
r
ISSN:
2502-4752
rule
e
xtraction
from
DTs
(CRED)
[19]
algorithm
which
is
designed
for
shallo
w
netw
orks
to
arbitrarily
man
y
hidden
layers.
DeepRED
aims
to
simplify
using
RxREN
[20]
to
reduce
unnecessary
input
and
applies
algo-
rithm
C4.5
[21]
to
create
a
DT
.
DeepRED
is
able
to
construct
complete
trees
closely
f
aithful
to
the
original
neural
netw
ork
model,
b
ut
the
generated
trees
are
lar
ge
and
requires
time
and
memory;
thus,
it
is
dif
ficult
in
terms
of
scalability
.
Artificial
neural
netw
ork-decision
tree
(ANN-DT)
[22]
is
another
DT
-based
proxy
model.
The
ANN-
DT
e
xtracts
binary
DT
from
a
trained
neural
netw
ork
and
generat
e
outputs
for
samples
interpolated
from
the
training
dataset.
The
criterion
of
an
attrib
ute
selec
tion
is
based
on
a
significance
analys
is
of
the
v
ariables
on
the
neural-netw
ork
output.
The
ANN-DT
is
able
to
e
xtract
rules
from
feedforw
ard
neural
netw
orks
with
continuous
outputs.
These
e
xtracted
rules
are
from
the
neural
netw
ork
wit
ho
ut
making
assumptions
about
the
internal
structure
of
the
neural
netw
ork
or
the
features
of
the
data.
2.2.
Kno
wledge
distillation
KD
[23]
is
a
technique
in
machine
learning
to
compress
a
lar
ge
machine
learning
model
as
teacher
model
to
a
smaller
and
more
understandable
model
as
student
model
b
ut
able
to
maintain
the
beha
vior
of
the
original
model.
KD
is
specifically
used
in
scenarios
where
processing
po
wer
,
memory
,
and
storage
are
limited.
The
goal
of
KD
is
to
transfer
the
kno
wledge
and
insights
learned
by
the
teacher
model
to
the
student
model,
resulting
in
a
more
compact
and
ef
ficient
model.
The
teacher
model
is
typically
a
lar
ge
comple
x
model
that
has
been
trained
on
a
lar
ge
dataset
using
a
deep
learning
and
achie
v
es
acceptable
accurac
y
b
ut
is
computationally
e
xpensi
v
e
and
memory-intensi
v
e.
The
student
model
is
a
smaller
and
simpler
model
that
is
designed
to
mimic
the
beha
vior
of
the
teacher
model
with
fe
wer
paramete
rs
to
be
more
lightweight
and
suitable
for
deplo
yment
on
resource-restricted
de
vices.
During
the
training,
the
loss
function
i
s
used
for
training
the
student
model
aiming
to
compare
the
soft
tar
gets
produced
by
the
teacher
model
with
the
student
model’
s
predictions.
The
distillation
loss
encourages
the
student
model
to
align
its
predictions
with
those
of
the
teacher
model.
Through
the
training
process,
the
student
model
learns
from
the
teacher
model’
s
decision
boundaries,
pattern
recognition,
and
generalization
capabilities.
This
allo
ws
the
student
model
to
capture
the
underlying
kno
wledge
of
the
teacher
model.
W
ith
the
ability
to
transfer
kno
wledge
from
a
comple
x
model
to
a
more
compact
model
of
KD,
we
are
interested
to
apply
the
method
to
synthes
ize
a
dataset
from
a
comple
x
model
as
a
teacher
model.
The
synthesized
data
with
kno
wledge
from
the
teacher
model
thus
can
be
a
representati
v
e
as
an
e
xplainable
model
of
the
black
box
model.
2.3.
Generati
v
e
adv
ersarial
netw
ork
A
GAN
is
a
machine
learning
technique
well-kno
wn
for
the
generation
of
ne
w
data
instances
that
resemble
a
gi
v
en
dataset
as
synthesized
data
[24].
GAN
consi
sts
of
tw
o
neural
netw
orks
as
generator
and
discriminator
.
The
former
is
responsible
for
generating
ne
w
data
instances
that
mimic
the
data
by
considering
random
noise
and
include
it
into
data
samples
.
The
latter
then
observ
es
both
original
data
samples
and
gen-
erated
data
samples
and
attempts
to
distinguish
between
them
whether
which
samples
a
re
real,
and
which
are
generated.
Upon
checking
the
discriminating
results,
loss
v
alue
is
calculated
to
impro
v
e
generator
criteria
until
the
generator
becomes
proficient
at
creating
data
that
is
dif
ficult
for
the
discriminator
to
distinguish.
In
prac-
tical,
Addepalli
et
al.
[25]
applied
GAN
to
generate
data
enriching
dataset
and
solv
e
imbalanced
data
issues
for
training
data
for
deep
learning.
F
or
its
ability
to
reliably
generate
resembled
data,
conditional
tab
ular
-GAN
(CT
-GAN)
algorithm
[26]
is
thus
selected
to
apply
for
kno
wledge
transferred
data
synthesis
for
this
w
ork.
2.4.
Bee
algorithm
B
A
is
a
population-based
optimization
technique
inspir
ed
by
the
foraging
beha
vior
of
bees
to
find
optimal
solutions
for
optimization
problems
[27],
[28].
It
is
f
amous
for
an
ability
to
e
xplore
the
search
space
ef
ficiently
and
handle
comple
x
problems
successfully
.
As
the
concept
of
food
foraging
beha
vior
,
the
parameters
in
the
algorithm
include
the
number
of
scout
bees
(n),
elite
bees
(e),
number
of
food
areas
for
foraging
(m),
number
of
tar
get
food
areas,
and
number
of
bees
for
foraging
[27].
The
steps
of
B
A
are
as
follo
ws.
The
algorithm
starts
by
generating
random
bee
population
and
en
vironment
based
on
setting
parame-
ters.
Then,
scout
bees
are
sent
out
to
scout
for
solutions,
and
the
solutions
are
e
v
aluated
using
fitness
function.
Elite
bees
select
the
solution
based
on
fitness
scores
where
the
higher
the
score,
the
more
probability
the
solu-
tion
is
selected.
The
neighborhood
of
the
selected
solution
is
searched
for
e
xpanding
the
possible
solutions
for
scout
bees
to
forage,
and
the
y
are
also
e
v
aluated
using
fitness
function.
Then
the
solution
among
all
solutions
is
selected
as
representati
v
e
solution.
Indonesian
J
Elec
Eng
&
Comp
Sci,
V
ol.
36,
No.
2,
No
v
ember
2024:
1052–1069
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
r
1055
In
2019,
Chungno
y
et
al.
[29]
adapted
B
A
to
solv
e
missing
data
by
imputing
the
data
systematically
.
The
B
A-based
data
imputation
sho
wed
the
ability
to
generate
missing
data
ef
fecti
v
ely
.
The
imputed
dataset
w
as
e
v
aluated
to
yield
the
highest
accurac
y
and
superior
to
other
techniques
for
23%
in
a
v
erage.
It
signifies
that
B
A
has
capability
to
generate
a
data
v
alue
with
high
accurac
y
and
impact
to
a
machine
learning
model.
Hence,
we
select
B
A
to
be
used
as
one
of
the
techniques
for
kno
wledge
transferred
data
synthesis
for
this
w
ork.
3.
MA
TERIALS
AND
METHODS
This
paper
presents
a
method
for
de
v
eloping
XAI
for
a
black
box
model
based
on
KD
concepts
and
data
synthesis.
The
synthesized
data
that
are
transferred
from
a
black
box
model
are
used
to
train
for
AI
can
become
a
model
that
can
e
xplain
the
reason
wh
y
the
AI
chooses
the
output.
Thus,
the
method
consists
of
tw
o
main
parts
as
kno
wledge
transfer
and
e
xplainable
model
generation
as
sho
wn
in
Figure
1.
Figure
1.
An
o
v
ervie
w
of
the
process
3.1.
Kno
wledge
transfer
F
or
the
kno
wledge
transfer
part,
we
generate
synthesis
data
from
the
teacher
model
which
is
a
bl
ack
box
model
de
v
eloped
from
a
deep
learning
cl
assification.
The
synthesized
data
as
a
student
model
are
used
for
prediction
along
with
a
prediction
from
the
teacher
model.
Loss
function
calculates
for
loss
v
alue
and
uses
it
to
adjust
the
student
model
for
better
predict
ion
performance.
The
method
is
repeated
until
the
assigned
epoch
iteration
is
met.
The
result
of
this
part
is
the
final
synthesized
dataset
to
be
used
for
de
v
eloping
an
XAI
model.
In
this
paper
,
tw
o
techniques
of
data
synthesis
with
kno
wledge
transfer
from
a
teach
model
are
pre-
sented.
First,
we
present
kno
wledge
transfer
based
on
GAN
(KT
-GAN),
which
applies
model
generator
from
CT
-GAN
[26].
This
technique
synthesizes
tab
ular
data.
Second,
we
propose
a
kno
wledge
transfer
based
on
Bee
algorithm
(KT
-Bee)
in
data
synthesis.
3.1.1.
Kno
wledge
transfer
based
on
generati
v
e
adv
ersarial
netw
ork
KT
-GAN
applies
the
method
of
a
model
generator
from
CT
-GAN
[26].
W
ith
the
method,
the
tab
ular
data
are
synthesized
as
a
student
model.
Ins
tances
in
synthesized
dataset
are
used
in
prediction
and
the
y
are
Bee-inspir
ed
knowledg
e
tr
ansfer:
synthesizing
data
for
enhanced
...
(Kritanat
Chungnoy)
Evaluation Warning : The document was created with Spire.PDF for Python.
1056
r
ISSN:
2502-4752
compared
to
a
prediction
from
the
teacher
model
to
find
dif
ference.
Loss
then
is
calculated
using
binary
cross
entrop
y
loss
(1).
Loss
=
1
N
N
X
i
=1
y
i
l
og
(
p
(
y
i
))
+
(1
y
i
)
l
og
(1
p
(
y
i
))
(1)
Where
y
is
the
label
(whereas
1
re
fers
to
matching
label
between
predicti
ons
of
a
teacher
model
and
a
student
model
and
0
for
not-matching
label)
and
p(y)
is
a
probability
of
label
being
1.
In
step-wise,
KT
-GAN
processes
are
gi
v
en
in
Figure
2.
The
pseudo
code
of
all
steps
is
as
follo
ws.
i)
Initial
step
Assign
classes
Assign
i
number
of
instances
and
f
number
of
features
for
synthesis
for
all
assigned
classes
Create
class
label
for
all
instances
(in
this
w
ork,
we
create
class
labels
as
balance
class
type)
Generate
data
for
each
feature
at
random
within
kno
wn
boundary
(min-max)
for
all
i
instances
and
f
features
ii)
Use
the
generated
dataset
to
test
for
predi
ction
result
using
the
teacher
model
and
collect
class
labels
for
all
instances
iii)
Compare
the
result
of
predictions
from
teacher
model
and
class
label
from
the
generated
data
whether
the
y
are
matched
or
not
i
v)
Assign
matching
label
if
the
y
are
matched
as
1
and
unmatched
as
0
v)
Apply
matching
label
to
calculation
with
loss
function
in
(1)
and
use
loss
res
ult
to
calculate
for
adjust
weight
for
generator
model.
vi)
Use
generator
model
to
synthesize
generated
dataset
vii)
Repeat
2
-
6
for
assigned
epoch
viii)
END
Figure
2.
Processes
of
KT
-GAN
3.1.2.
Kno
wledge
transfer
based
on
Bee
algorithm
KT
-Bee
e
xploits
BE
to
synthesize
data
according
to
gi
v
en
information
in
the
teacher
model.
The
core
process
is
to
let
bees
to
find
the
best
solution
of
ha
ving
the
lo
west
loss
v
alue.
There
are
tw
o
types
of
bees
including
scout
bee
and
follo
wer
bee.
The
scout
bees
are
task
ed
to
randomly
generate
a
data
v
alue
based
on
a
specified
data
type
as
a
solution.
The
follo
wer
bees
are
to
choose
a
solution
by
considering
fitness
function
of
a
scout
bee
based
on
probability
.
The
higher
score
from
fitness
function
represents
the
higher
quality
of
a
Indonesian
J
Elec
Eng
&
Comp
Sci,
V
ol.
36,
No.
2,
No
v
ember
2024:
1052–1069
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
r
1057
solution
and
has
more
probability
to
be
chosen,
and
vice
v
ersa.
Once
a
follo
wer
bee
chooses
a
solution,
the
data
v
alue
as
the
chosen
solution
is
adjusted.
After
all
solutions
are
adjusted,
fitness
function
will
be
updated
to
reflect
the
change.
A
ne
w
set
of
scout
bees
then
be
gin
another
round
of
foraging
to
generate
a
solution
from
the
updated
dataset
until
the
assigned
iteration
is
met.
The
objecti
v
e
function
of
this
process
is
to
minimize
the
loss
v
alue
from
binary
cross
entrop
y
loss
using
(2).
Loss
bee
(
j
)
=
1
N
N
X
i
=1
y
i
l
og
(
m
i
)
+
(1
y
i
)
l
og
(1
m
i
)
(2)
Where
y
is
the
label
(the
y
v
alue
is
al
w
ays
1,
where
1
means
matching),
and
m
is
a
label
is
assigned
to
if
predictions
of
a
teacher
model
and
a
student
model
are
matched
(1)
and
for
not-matching
(0)
label.
In
step-wise
(Figure
3),
KT
-Bee
processes
are
gi
v
en
in
Figure
3(a).
It
can
be
split
into
2
phases
as
process
of
scout
bees
in
Figure
3(b)
and
process
of
follo
wer
bees
in
Figure
3(c).
The
pseudo
code
of
all
steps
is
as
follo
ws.
1)
Initial
step
(a)
Assign
classes
(b)
Assign
i
number
of
Instances
and
f
number
of
features
for
synthesis
for
all
assigned
classes
(c)
Create
class
label
for
all
instances
(in
this
w
ork,
we
create
class
labels
as
balance
class
type)
2)
Scout
bee
step
(Figure
3(b))
(a)
Generate
data
for
each
feature
at
random
within
kno
wn
boundary
(Min-Max)
for
all
i
instances
and
f
features
of
b
number
of
scout
bees
(b)
Use
the
generated
dataset
to
test
for
prediction
result
using
the
teacher
model
and
collect
class
labels
for
all
instances
(c)
Compare
the
result
of
predicti
ons
from
teacher
model
and
class
label
from
the
generated
data
whether
the
y
are
matched
or
not
(d)
Assign
matching
label
if
the
y
are
matched
as
1
and
unmatched
as
0
(e)
Apply
matching
label
to
calculation
with
loss
function
in
(2)
and
use
Loss
result
to
calculate
for
fitness
function
of
each
scout
bee
using
(3).
The
higher
fitness
v
alue
of
a
solution
from
a
scout
bee,
the
more
probability
it
is
chosen
by
follo
wer
bees.
3)
F
ollo
wing
bee
step
(Figure
3(c))
(a)
m
number
of
follo
wer
bee
chooses
a
solution
re
g
arding
fitness
v
alue
based
on
probability
calculated
from
(4)
where
p
(
bee
j
)
represents
probability
of
choosing
by
each
follo
wer
bee.
(b)
Each
follo
wer
bee
randomly
select
s
a
feature
to
change
based
on
probability
calculated
gi
v
en
in
(7)
where
p
(
f
eatur
e
k
)
is
a
probability
of
feature
k,
and
k
refers
to
k
th
feature
while
m
represents
a
number
of
all
features.
In
the
first
itera
tion,
weight
of
all
features
is
set
as
1,
and
the
chosen
solution
for
data
change
is
limited
to
the
instances
that
are
under
the
matching
label
of
0
(not
matching).
(c)
Find
top
k
similarity
instance
between
unmatching
group
and
matching
group
with
results
from
(8)
where
p
is
a
considered
v
ector
,
and
q
is
a
tar
get
v
ector
to
find
similarity
.
r
represents
r
th
feature,
and
s
is
a
number
of
all
features.
(d)
Use
the
top
K
instances
(from
3(c))
to
train
for
re
gression
model
to
find
prediction
v
alues
to
change
data
for
the
selected
feature
(from
3(b))
(e)
Repeat
step
3(c)
and
3(d)
for
all
instances
with
matching
label
of
0
(f)
Use
the
changed
solution
to
predict
class
label
using
the
teacher
model
and
assign
the
ne
w
matching
label.
Then
re-calculate
for
loss
and
fitness
function
and
update
the
weight
of
each
feature
using
(6).
If
the
ne
w
weight
is
less
than
0,
the
weight
is
set
as
0.
4)
Repeat
2
and
3
according
to
the
assigned
iteration
and
if
iteration
is
greater
than
n
,
the
follo
wing
bee
has
a
probability
to
randomly
cop
y
solution
from
the
pre
vious
iteration.
5)
Repeat
2,
3,
and
4
until
all
designated
classes
are
processed
6)
END
Bee-inspir
ed
knowledg
e
tr
ansfer:
synthesizing
data
for
enhanced
...
(Kritanat
Chungnoy)
Evaluation Warning : The document was created with Spire.PDF for Python.
1058
r
ISSN:
2502-4752
(a)
(b)
(c)
Figure
3.
Processes
of
KT
-Bee:
(a)
o
v
ervie
w
processes
of
KT
-Bee,
(b)
processes
of
scout
bees,
and
(c)
processes
of
follo
wing
bees
f
itness
(
bee
j
)
=
1
Loss
(
bee
j
)
(3)
p
(
bee
j
)
=
f
itness
(
bee
j
)
P
n
j
=1
f
itness
(
bee
j
)
(4)
f
itness
dif
f
=
f
itness
(
bee
j
)
new
f
itness
(
bee
j
)
ol
d
(5)
w
eig
ht
new
=
w
eig
ht
cur
r
ent
+
f
itness
dif
f
(6)
p
(
f
eatur
e
k
)
=
w
eig
ht
(
f
eatur
e
k
)
P
m
l
=1
w
eig
ht
(
f
eatur
e
l
)
(7)
d
pq
=
dist
(
x
p
;
x
q
)
=
v
u
u
t
s
X
r
=1
(
x
pr
x
q
r
)
2
(8)
Indonesian
J
Elec
Eng
&
Comp
Sci,
V
ol.
36,
No.
2,
No
v
ember
2024:
1052–1069
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
r
1059
3.2.
T
raining
f
or
explainable
model
T
o
obtain
an
e
xplainable
model
that
represents
the
comple
x
AI
model,
the
synthesized
dataset
i
s
trained
for
classification
to
resembl
e
the
original
model.
T
o
k
eep
an
e
xplainable
model
simple
and
inherently
interpretable,
DT
technique
is
chosen
as
the
model
generated
by
DT
representing
decisions
and
decision-
making
processes
in
a
human-readable
tree-lik
e
structure,
with
each
node
representing
a
decision
based
on
a
feature,
and
each
branch
representing
the
outcome
of
that
decision.
Furthermore,
the
transparenc
y
of
DT
model
mak
es
it
easy
to
follo
w
the
logic
behind
a
model’
s
predictions.
Thus,
it
will
help
user
to
visually
trace
the
path
from
the
root
node
to
a
leaf
node
to
understand
ho
w
the
model
arri
v
ed
at
a
particular
decision.
Ho
we
v
er
,
we
are
a
w
are
that
DT
might
be
too
simplistic
to
capture
i
ntricate
patterns
in
the
data
and
may
not
of
fer
the
same
predicti
v
e
accurac
y
as
more
comple
x
models
lik
e
deep
neural
netw
orks.
4.
EXPERIMENT
SETTING
In
this
w
ork,
an
e
xperiment
is
separated
into
5
parts
as
sho
wn
in
Figure
4
including:
i)
dat
a
prepa-
ration,
ii)
training
teacher
model,
ii)
kno
wledge
transfer
to
synthesis
dataset
usi
ng
KT
-GAN
or
KT
-Bee,
i
v)
de
v
elop
an
e
xplainable
model
from
synthesized
data,
v)
e
v
aluation
of
the
e
xplainable
model,
and
vi)
e
v
aluation
comparison.
Figure
4.
Experiment
o
v
ervie
w
4.1.
Data
pr
eparation
In
this
e
xperiment,
we
selected
6
datasets
including
Higgs
dataset
(HI)
[20],
Jannis
dataset
(J
A)
[21],
HLS4ML
LHC
Jet
data
set
(HLS)
[30],
[31],
MagicT
elescope
(MT)
[32],
MiniBooNE
(MB)
[33]
and
numerai
(NU)
[34].
The
data
were
prepared
by
discarding
a
class
with
less
than
15%
of
the
total
data.
W
e
also
remo
v
ed
instances
with
a
missing
v
alue(s)
and
processed
the
dataset
to
become
a
balance
dataset
as
a
training
set
for
a
teacher
model.
Statisti
cs
of
prepared
datasets
are
sho
wn
in
T
able
1.
F
or
the
testing
set,
20%
of
datasets
are
separated
for
model
e
v
aluation.
T
able
1.
Dataset
details
Dataset
Number
instances
Number
Features
Number
class
Higgs
dataset
92,446
28
2
Jannis
dataset
44,202
54
3
HLS4ML
LHC
Jet
dataset
805,330
16
5
MiniBooNE
72,998
50
2
Numerai
95,324
21
2
MagicT
elescope
13,376
10
2
Bee-inspir
ed
knowledg
e
tr
ansfer:
synthesizing
data
for
enhanced
...
(Kritanat
Chungnoy)
Evaluation Warning : The document was created with Spire.PDF for Python.
1060
r
ISSN:
2502-4752
4.2.
T
raining
teacher
models
T
o
obtain
teacher
models,
we
train
each
dataset
with
general
deep
learning
for
clas
sification.
The
obtained
models
are
considered
as
a
teacher
model
that
represents
a
high
quality
classification
m
odel
in
terms
of
accurac
y
in
a
form
of
comple
x
and
lar
ge
black
box
model.
P
arameter
setting
for
training
the
teacher
models
is
as
follo
ws:
i)
layer:
5
layer
(32,16,8,4,2);
ii)
acti
v
ation
function:
rectified
linear
unit
(ReLU);
iii)
weight
op-
timization:
Adam;
i
v)
learning
rate:
constant;
and
v)
epoch:
10,000
epoch.
The
teacher
models
were
e
v
aluated
for
performance
using
10
fold-cross-v
alidation
and
yielded
the
accurac
y
result
as
sho
wn
in
T
able
2.
T
able
2.
Accurac
y
performance
of
teacher
models
trained
with
10-fold
cross-v
alidation
Dataset
Higgs
Jannis
HLS4ML
LHC
Jet
MiniBooNE
Numerai
MagicT
elescope
General
deep
0.7054
0.6231
0.6748
0.8695
0.5090
0.7397
4.3.
Kno
wledge
transfer
T
o
obtain
the
synthesized
data
with
kno
wledge-transferred
from
the
teacher
model,
we
applied
either
KT
-GAN
or
KT
-Bee
separately
to
compare
their
performances.
The
parameter
setting
for
data
synthesis
and
kno
wledge
transferring
is
as
follo
ws.
i)
KT
-GAN
Using
“ctg
an”
library
in
p
ython
v
ersion
0.7.3
[26]
as
data
generator
with
all
def
ault
parameter
settings
and
epoch
is
set
to
100
ii)
KT
-Bee
Number
of
scout
bee:
20
Number
of
follo
wer
bee:
100
T
op
k
similarity:
10
Iteration:
100
The
kno
wledge-transferred
synthesized
data
are
generated
into
3
sizes
re
g
arding
number
of
instances
as
i)
same
instance
number
to
t
he
original
dataset
(original
size);
ii)
50%
less
instances
than
the
original
dataset
(50%
less
size);
and
iii)
50%
more
instances
than
the
original
dataset
(50%
more
size).
The
sizes
are
to
study
whether
the
size
of
synthesized
data
ha
v
e
the
ef
fect
on
classification
performance
or
not.
4.4.
T
raining
explainable
model
The
synthesized
datasets
are
generated
to
train
for
an
e
xplainable
model
to
enhance
interpretability
of
prediction
decision
of
a
black
box
model.
In
this
e
xperiment,
we
chose
the
DT
classification
approach
based
on
scikit-learn
v
ersion
1.3.0
[35]
to
represent
the
classification
model
as
the
model
from
DT
is
easy
to
interpret
and
follo
w
.
The
DT
classification
has
the
parameter
setting
sho
wn
in
T
able
3.
W
e
then
e
v
aluated
performance
of
the
generated
e
xplainable
models
using
10
fold-cross-v
alidation
as
sho
wed
in
T
able
4.
T
able
3.
P
arameter
setting
for
e
xplainable
model
P
arameter
V
alues
Criterion
Gini
Splitter
Best
Maximum
depth
None
Minimum
samples
split
2
Minimum
samples
leaf
1
Minimum
weight
fraction
leaf
0.0
Maximum
features
None
Random
state
None
Maximum
leaf
nodes
None
Minimum
impurity
decrease
0.0
Class
weight
None
Cost-comple
xity
pruning
alpha
0.0
Indonesian
J
Elec
Eng
&
Comp
Sci,
V
ol.
36,
No.
2,
No
v
ember
2024:
1052–1069
Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian
J
Elec
Eng
&
Comp
Sci
ISSN:
2502-4752
r
1061
T
able
4.
Performance
of
e
xplainable
models
with
kno
wledge-transferred
synthesized
data
Dataset
KT
-Bee
KT
-GAN
HI
50%
les
s
size
0.8184
1.000
Original
size
0.8268
1.000
50%
more
size
0.8364
1.000
J
A
50%
l
ess
size
0.7363
0.9932
Original
size
0.7651
0.9966
50%
more
size
0.7733
0.9979
HLS
50%
less
size
0.8975
0.9997
Original
size
0.8897
0.9999
50%
more
size
0.9999
0.9999
MT
50%
less
size
0.8686
0.9983
Original
size
0.8766
0.9991
50%
more
size
0.8728
0.9994
MB
50%
less
size
0.8466
1.000
Original
size
0.8677
1.000
50%
more
size
0.8536
1.000
NU
50%
less
size
0.7689
1.000
Original
size
0.7742
1.000
50%
more
size
0.7652
1.000
4.5.
Ev
aluation
and
comparison
T
o
e
v
aluate
and
compare
the
e
xplainable
models
from
the
tw
o
presented
methods,
tw
o
aspects
are
con-
sidered
including
completeness
of
kno
wledge
transferring
and
completeness
the
models
as
sho
wn
in
Figure
5.
The
kno
wledge
transfer
method
presented
in
this
w
ork
dra
ws
inspiration
from
KD,
a
technique
for
transferring
kno
wledge
from
a
lar
ge
deep
learning
model
(teacher
model)
to
a
smaller
deep
learning
model
(student
model).
The
goal
is
to
enable
the
student
model
to
achie
v
e
performance
close
to
that
of
the
teacher
model,
allo
wing
it
to
mak
e
predictions
similar
to
the
teacher
model
and
thus
be
suitable
for
deplo
yment
on
mobile
de
vices.
T
o
e
v
aluate
the
ef
fecti
v
eness
of
the
proposed
kno
wledge
transfer
method,
we
de
v
eloped
performance
metrics
tailored
to
assess
the
successful
transfer
of
kno
wledge
from
a
black-box
model
(teacher
model)
to
synthetic
dataset
belie
v
ed
to
represent
the
dataset
used
to
train
the
black-box
model.
Figure
5.
Ev
aluation
o
v
ervie
w
The
aspect
of
completeness
of
kno
wledge
transferring
considers
distrib
ution
of
the
data
of
each
feat
ure
and
performance
of
the
kno
wledge
transferring
(KTP).
F
or
distrib
ution
of
data,
we
assume
that
the
distrib
ution
of
data
of
transferred
data
should
be
similar
or
close
to
the
original
data,
and
we
e
v
aluated
by
comparing
the
distrib
uti
on
of
the
original
dataset
and
the
synthesized
datas
et.
KTP:
this
metric
in
v
olv
es
using
the
dataset
obtained
through
the
kno
wledge
transfer
process
to
train
a
prediction
model
and
then
e
v
aluat
ing
its
performance
on
testset.
The
prediction
results
should
ideally
match
those
of
the
teacher
model.
A
higher
KTP
v
alue
indicates
that
the
synthetic
dataset
captures
the
kno
wledge
of
the
teac
her
model
more
accurately
.
This
aligns
with
the
h
ypothesis
that
if
kno
wledge
transfer
is
successful,
the
model
trained
using
synthetic
dataset
should
produce
prediction
results
similar
to
those
of
the
teacher
model.
F
or
performance
of
the
kno
wledge
transferring,
KTP
can
be
calculated
using
(9).
Bee-inspir
ed
knowledg
e
tr
ansfer:
synthesizing
data
for
enhanced
...
(Kritanat
Chungnoy)
Evaluation Warning : The document was created with Spire.PDF for Python.