Recei ed 27 Ap il 2023, accep ed 6 June 2023, da e o publica ion 9 June 2023, da e o cu en e sion 19 June 2023.
Digi al Objec Iden i ie 10.1109/ACCESS.2023.3284975
BukaGini: A S abili y-Awa e Gini Index Fea u e
Selec ion Algo i hm o Robus Model
Pe o mance
MOHAMED ALY BOUKE 1, (Membe , IEEE), AZIZOL ABDULLAH 1,
JAROSLAV FRNDA 2,3, (Senio Membe , IEEE), KORHAN CENGIZ 4,5, (Senio Membe , IEEE),
AND BASHIR SALAH 6
1Depa men o Communica ion Technology and Ne wo k, Facul y o Compu e Science and In o ma ion Technology, Uni e si i Pu a Malaysia, Se dang 43400,
Malaysia
2Depa men o Quan i a i e Me hods and Economic In o ma ics, Facul y o Ope a ion and Economics o T anspo and Communica ions, Uni e si y o Žilina,
01026 Žilina, Slo akia
3Depa men o Telecommunica ions, Facul y o Elec ical Enginee ing and Compu e Science, VSB—Technical Uni e si y o Os a a, 70800 Os a a, Czech
Republic
4Depa men o Compu e Enginee ing, Is inye Uni e si y, 34010 Is anbul, Tu key
5Depa men o In o ma ion Technologies, Facul y o In o ma ics and Managemen , Uni e si y o H adec K álo é, 500 03 H adec K álo é, Czech Republic
6Depa men o Indus ial Enginee ing, College o Enginee ing King Saud Uni e si y, Riyadh 11421, Saudi A abia
Co esponding au ho : Mohamed Aly Bouke ([email p o ec ed]g)
This s udy ecei ed unding om King Saud Uni e si y, Saudi A abia h ough esea che s suppo ing p ojec numbe (RSP2023R145).
And was suppo ed by he Minis y o Educa ion, You h and Spo s o he Czech Republic unde he g an SP2023/007 conduc ed by VSB -
Technical Uni e si y o Os a a, Czechia, and pa ially suppo ed by Ins i u ional esea ch o he Facul y o Ope a ion and Economics o
T anspo and Communica ions–Uni e si y o Zilina, no. 2/KKMHI/2022. Acknowledgemen s: The au ho s would like o hank King Saud
Uni e si y, Riyadh, Saudi A abia, wi h esea che s suppo ing p ojec numbe RSP2023R145.
ABSTRACT Fea u e in e ac ion is a i al aspec o Machine Lea ning (ML) algo i hms, and gaining a deep
unde s anding o hese in e ac ions can signi ican ly enhance model pe o mance. This pape in oduces
he BukaGini algo i hm, an inno a i e and obus app oach o ea u e in e ac ion analysis ha capi alizes
on he Gini impu i y index. By exploi ing he unique p ope ies o he BukaGini index, ou p oposed
algo i hm e ec i ely cap u es bo h linea and nonlinea ea u e in e ac ions, p o iding a iche and mo e
comp ehensi e ep esen a iono heunde lyingda a.We ho oughly e alua e he BukaGini algo i hm agains
adi ional Gini index-based me hods on a ious eal-wo ld da ase s. These da ase s include he High School
S uden s’ Pe o mance (HSSP) da ase , which examines ac o s a ec ing s uden pe o mance; Cance Da a,
which ocuses on iden i ying cance ypes based on gene exp ession; Spambase, which a ge s spam email
classi ica ion; and he UNSW-NB15 da ase , which add esses ne wo k in usion de ec ion. Ou expe imen al
esul s demons a e ha he BukaGini algo i hm consis en ly ou pe o ms adi ional Gini index-based
me hods in e ms o accu acy. Ac oss he es ed da ase s, he BukaGini algo i hm achie es imp o emen s
anging om 0.32% o 2.50%, unde sco ing i s e ec i eness in handling di e se da a ypes and p oblem
domains. This pe o mance gain highligh s he po en ial o he BukaGini algo i hm as a aluable ool o
ea u e in e ac ion analysis in a ious ML applica ions.
INDEX TERMS BukaGini algo i hm, Gini index, ensemble lea ning, ea u e in e ac ion analysis, da a
mining.
I. INTRODUCTION
The apid g ow h o da a in a ious domains, including
inance, heal hca e, social media, and IoT, has led o an
The associa e edi o coo dina ing he e iew o his manusc ip and
app o ing i o publica ion was Yongming Li .
inc eased demand o e icien and e ec i e me hods o
p ocess la ge-scale da ase s. ML echniques a e essen ial in
his con ex , p o iding powe ul ools o ex ac aluable
insigh s om complex, high-dimensional da a [1]. Howe e ,
high dimensionali y poses nume ous challenges, such as
inc eased compu a ional complexi y, o e i ing, and educed
59386
This wo k is licensed unde a C ea i e Commons A ibu ion-NonComme cial-NoDe i a i es 4.0 License.
Fo mo e in o ma ion, see h ps://c ea i ecommons.o g/licenses/by-nc-nd/4.0/ VOLUME 11, 2023
M. A. Bouke e al.: BukaGini: A S abili y-Awa e Gini Index Fea u e Selec ion Algo i hm
in e p e abili y o he esul ing models. To add ess hese
issues, ea u e selec ion echniques ha e been de eloped o
iden i y he mos ele an a iables, educe dimensionali y,
and enhance model pe o mance [2].
Fea u e selec ion can be iewed as a sea ch p oblem o ind
he op imal subse o ea u es ha maximizes an ML model’s
pe o mance [3]. The sea ch space is ypically de ined by
he powe se o all a ailable ea u es, esul ing in a combi-
na o ial p oblem ha g ows exponen ially wi h he numbe
o ea u es. As exhaus i e sea ch becomes compu a ionally
in easible, a ious sea ch s a egies ha e been p oposed o
na iga e he sea ch space mo e e icien ly. Fea u e selec ion
me hods can be b oadly ca ego ized in o il e , w appe , and
embedded app oaches [4].
Ensemble lea ning, a popula ML echnique, combines
mul iple models o imp o e o e all pe o mance and obus -
ness. I has been success ully employed in nume ous appli-
ca ions, such as classi ica ion, eg ession, and clus e ing.
In ensemble lea ning, he combina ion o base models o
lea ne s can exploi he complemen a y s eng hs o each
model o p oduce an ensemble model wi h be e p edic i e
pe o mance han i s cons i uen s [5].
Fea u e in e ac ion analysis is essen ial o unde s anding
complex ela ionships be ween he ea u e in da ase s. I helps
e eal how ea u es in e ac and in luence he a ge a iable,
aiding in iden i ying signi ican ea u e in e ac ions ha can
imp o e model pe o mance. Fea u e in e ac ion analysis is
p ecious in high-dimensional da ase s, whe e indi idual ea-
u e impo ance may be challenging o disce n. Techniques
such as F-ANOVA and Hie a chical G oup-Lasso ha e been
de eloped o iden i y ea u e in e ac ions [6], [7], [8]
The Gini index, a widely-used impu i y measu e in deci-
sion ee algo i hms, has shown p omise in ea u e selec ion
due o i s abili y o quan i y inequali y o impu i y in a
da ase . The Gini index, ini ially de eloped by Co ado Gini
in 1912 as a measu e o inequali y, has been widely used
in a ious domains, such as economics, ecology, and ML
[9]. In ML, he Gini index is employed as an impu i y mea-
su e o assess he quali y o a spli in decision ee lea ning
algo i hms, such as CART (Classi ica ion and Reg ession
T ees [10], [11], [12].
Se e al Gini-based ea u e selec ion app oaches ha e been
p oposed, exhibi ing good accu acy and compu a ional e i-
ciency esul s [10]. Howe e , hese me hods a e no wi hou
limi a ions. One majo d awback is hei sensi i i y o noise,
which can lead o he selec ion o i ele an o edundan
ea u es. Addi ionally, exis ing Gini-based app oaches o en
employ subop imal sea ch s a egies, hinde ing hei abili y o
iden i y he bes ea u e subse e ec i ely. Fu he mo e, mos
Gini-based ea u e selec ion me hods a e ailo ed o speci ic
lea ning scena ios, limi ing hei adap abili y o di e se appli-
ca ions and p oblem domains.
To add ess he abo emen ioned issues, his pape in o-
duces BukaGini, a no el enhanced ea u e selec ion algo-
i hm based on he Gini index ha add esses hese limi a ions.
By inco po a ing ad anced op imiza ion echniques and a
e sa ile amewo k, BukaGini o e comes he sho comings
o exis ing Gini-based ea u e selec ion me hods, o e ing
imp o ed sea ch e iciency, con e gence, and adap abili y o
a ious lea ning scena ios.
The main con ibu ions o his pape a e:
•A no el algo i hm, Bukagini,enhances Gini index-based
ea u e selec ion by add essing i s limi a ions and le e -
aging he bene i s o ensemble lea ning.
•The in oduc ion o ea u e in e ac ion analysis in he
ea u e selec ion p ocess o iden i y and e ain essen ial
in e ac ions be ween ea u es.
•An e alua ion o he algo i hm’s s abili y using esam-
pling echniques like c oss- alida ion.
•The e ec i eness o he Bukagini algo i hm was ho -
oughly e alua ed using nume ous benchma k da ase s,
demons a ing i s supe io i y in pe o mance, s abili y,
and in e p e abili y o e adi ional Gini index-based
ea u e selec ion me hods. This e alua ion in ol ed he
use o c oss- alida ion o model gene alizabili y, Ran-
dom Fo es as an ensemble me hod o enhanced pe o -
mance and accu a e ea u e impo ance measu emen ,
ea u e in e ac ion analysis o unde s anding a iable
in e dependencies, and s abili y analysis o es he algo-
i hm’s esilience o changes in he inpu da a.
The emainde o he pape is s uc u ed as ollows. Sec-
ion II comp ehensi ely e iews ela ed wo k on Gini-based
ea u e selec ion me hods, highligh ing hei s eng hs and
limi a ions. Sec ion III discusses he Gini index-based ea u e
selec ion and i s challenges, which led o he de elopmen
o he no el BukaGini algo i hm. Sec ion IV in oduces he
BukaGini algo i hm, p o iding an o e iew and de ailing i s
componen s, such as ensemble lea ning, ea u e in e ac ion,
and s abili y analysis. Sec ion Vdesc ibes he expe imen al
se up, including da ase selec ion, p ep ocessing s eps, imple-
men a ion, and e alua ion me ics, o compa e he BukaGini
algo i hm wi h he adi ional Gini index ea u e selec ion.
Sec ion VI p esen s he esul s and discussion, ocusing on
he compa ison wi h con en ional Gini index-based me hods,
and highligh s he imp o emen s achie ed by he BukaGini
algo i hm. Finally, Sec ion VII concludes he pape and ou -
lines po en ial u u e esea ch di ec ions.
II. RELATED WORK
Fea u e selec ion is an essen ial p ep ocessing s ep in ML,
aiming o iden i y he mos ele an a iables, educe dimen-
sionali y, and enhance model pe o mance. Nume ous ea u e
selec ion echniques ha e been p oposed in he li e a u e,
including il e , w appe , and embedded me hods. This li e -
a u e e iew o e iews ecen ad ances in ea u e selec ion
echniques, ocusing on hei me hodologies, applica ions,
s eng hs, and limi a ions. The pape s e iewed we e selec ed
based on hei ele ance o he p oposed BukaGini algo i hm
and hei con ibu ions o he ield o ea u e selec ion.
Macedo e al. [13] p opose he Decomposed Mu ual In o -
ma ion Maximiza ion (DMIM) me hod, a no el sequen ial
VOLUME 11, 2023 59387
M. A. Bouke e al.: BukaGini: A S abili y-Awa e Gini Index Fea u e Selec ion Algo i hm
o wa d ea u e selec ion echnique based on Mu ual
In o ma ion (MI). DMIM add esses he limi a ions o exis -
ing MI-based me hods by o e coming he complemen a -
i y penaliza ion issue. Ex ensi e e alua ions demons a e
DMIM’s supe io pe o mance o o he MI-based ea u e
selec ion me hods, making i a p e e ed choice in his
domain.
Shaheen e al. [14] in oduce a new ea u e selec ion ech-
nique called he ‘‘Rele ance-Di e si y Algo i hm,’’ which
selec s impo an ea u es based on ele ance and di e si y
measu es o op imize he numbe o ea u es and educe
sea ch ime. This app oach o e comes some limi a ions o
exis ing ea u e selec ion echniques ha p ima ily ocus on
he in o ma ion con ained wi hin a ea u e.
Kou e al. [15] explo e he e alua ion o ea u e selec ion
me hods o ex classi ica ion wi h small sample da ase s as
a mul iple c i e ia decision-making (MCDM) p oblem. The
au ho s p opose using MCDM-based me hods and compa e
i e MCDM me hods wi h en ea u e selec ion me hods
and h ee classi ie s on en small da ase s. The esul s high-
ligh he e ec i eness o he MCDM-based app oach, wi h
Documen equency (DF) as he p e e ed ea u e selec ion
me hod.
Liu e al. [16] p esen a no el ensemble ea u e selec-
ion me hod wi h c oss-class sample g anula ion, ocusing
on local ea u e signi icance. The echnique consis s o wo
phases: (1) c oss-class sample g anula ion, whe e da a is
sepa a ed in o mul iple g anules based on sample loca ions
in hei espec i e classes, and (2) ensemble ea u e selec-
ion, whe e localized ea u e signi icance e alua ions a e
in eg a ed. Expe imen s on 20 UCI da ase s demons a e he
me hod’s supe io i y in e ms o accu acy and ime e iciency
compa ed o exis ing ea u e selec ion schemes.
H. Zhang e al. [17] add ess he p oblem o pa ially
labeled he e ogeneous ea u e selec ion in la ge-scale eal-
wo ld da ase s. I in oduces h ee mono onic unce ain y
measu es based on equi alence classes and neighbo hood
classes o explo e nonlinea co ela ions in he da a. Consis-
en en opy and mono onic neighbo hood en opy a e p o-
posed, along wi h a maximal neighbo hood en opy s a egy.
Two ea u e selec ion algo i hms a e p esen ed using hese
measu es. Expe imen al esul s demons a e he e ec i e-
ness and supe io i y o he p oposed ea u e selec ion mea-
su es in e ms o classi ica ion accu acy and compu a ional
complexi y.
P. Liu e al. [18] p esen a simple ye e ec i e ea u e
selec ion s a egy called Loss Reweigh in Scale Dimension
(LRSD) o aining single-s age ancho - ee objec de ec o s.
Using a eweigh unc ion, LRSD dynamically eweigh s he
aining loss o posi i e samples om selec ed op-k ea u e
le els.
Zhang e al. [19] add ess he gap in in o ma ion- heo e ic-
based mul i-label ea u e selec ion me hods by in oducing
wo assump ions, Label Independence Assump ion (LIA)
and Pai ed-label Independence Assump ion (PIA). The p o-
posed me hod, MFSJMI, uses join mu ual in o ma ion
and an in e ac ion weigh o conside mul iple-label
co ela ions.
Qu e al. [20] p oposed an algo i hm ha combines In o -
ma ion Gain and decision in o ma ion o ea u e selec ion o
imp o e classi ica ion accu acy and educe ime complexi y.
I in oduces neighbo hood in o ma ion en opy measu es
based on join in o ma ion g anules and p oposes a nonmono-
onic algo i hm ha u ilizes decision in o ma ion. To han-
dle high-dimensional da ase s, In o ma ion Gain is used o
p elimina y dimensionali y educ ion. Expe imen s on wel e
public da ase s show he algo i hm’s low ime cos and high
classi ica ion accu acy.
Zhu e al. [21] p opose a hyb id ea u e selec ion
me hod (HFSIA) based on a i icial immune algo i hms
o high-dimensional da a, combining he il e and
me aheu is ic-based sea ch s a egies. The p ocess in oduces
a le hal mu a ion mechanism, adap i e adjus men ac o s,
and a Cauchy mu a ion ope a o o imp o e sea ch pe o -
mance and di e si y. Expe imen s on 22 high-dimensional
da ase s compa e HFSIA o 23 o he ea u e selec ion me h-
ods, e ealing i s compe i i e compu a ional cos and be e
a e age classi ica ion accu acy.
Shi e al. [22] p oposed a hie a chical ea u e selec ion
me hod ha balances in e -class independence and in a-class
edundancy by conside ing class hie a chy and ea u e co e-
la ions. I u ilizes s uc u al ela ion egula iza ion o max-
imize independence be ween un ela ed classes and ea u e
ela ion egula iza ion o minimize edundancy wi hin each
class.
Ba e al. [23] p esen Glee, a no el G anula Compu ing
(G C) based amewo k o e icien and e ec i e ea u e
selec ion. Glee calcula es he g anula i y alue o each ea-
u e, eo de s hem acco dingly, and adds ea u es o he
selec ion pool one by one un il a e mina ion condi ion is me .
This app oach elimina es i e a i e calcula ions o in o ma ion
g anula ion, p o ides a sequence o ea u es insensi i e o
da a pe u ba ion, and is compa ible wi h a ious exis ing e -
mina ion condi ions. Expe imen s on 20 UCI da ase s show
Glee’s supe io i y in educing ime consump ion, imp o ing
ea u e s abili y, and main aining compe i i e classi ica ion
pe o mance.
Zheng e al. [24] p opose a no el s eaming ea u e selec-
ion me hod o unlabeled da a, in oducing a dynamic sim-
ila i y g aph o e alua e i ele an ea u es adap i ely. The
echnique consis s o wo s ages: minimum edundancy and
maximum ele ance and le e ages simila i y g aph di usion
o elimina e un eliable simila i ies. Expe imen s show he
p oposed S eaming Fea u e Selec ion ia G aph Di usion
(SFS-GD) me hod ou pe o ms exis ing unsupe ised ea u e
selec ion me hods.
Se e al ea u e selec ion echniques ha e been p oposed
in he li e a u e, which can be b oadly ca ego ized in o il e ,
w appe , and embedded me hods. Fil e me hods, such as he
Chi-squa e es , in o ma ion gain, co ela ion coe icien , and
Relie F, assess he ele ance o ea u es independen ly o any
p edic i e model. In con as , w appe me hods, including
59388 VOLUME 11, 2023
M. A. Bouke e al.: BukaGini: A S abili y-Awa e Gini Index Fea u e Selec ion Algo i hm
sequen ial o wa d selec ion and ecu si e ea u e elimina-
ion, ely on he pe o mance o a speci ic p edic i e model
o e alua e ea u e subse s. Embedded me hods, like LASSO
and decision ees, in eg a e ea u e selec ion di ec ly in o he
lea ning p ocess.
The Gini index, widely adop ed in decision ee algo i hms
like CART and andom o es s, is a popula ea u e selec ion
echnique. Howe e , i has limi a ions, such as bias owa ds
ea u es wi h many ca ego ies. Resea che s ha e p oposed
al e na i es, like he Gain Ra io and Symme ical Unce ain y
me hods, o add ess his issue.
In conclusion, his li e a u e e iew highligh s he di e se
ange o ea u e selec ion echniques and hei applica ions
ac oss a ious domains. Each me hod has s eng hs and lim-
i a ions, making i c ucial o iden i y and de elop imp o ed
s a egies ha o e come hese challenges. The p oposed Buk-
aGini algo i hm aims o add ess he limi a ions o exis -
ing Gini-based ea u e selec ion me hods by inco po a ing
ensemble lea ning, ea u e in e ac ion analysis, and s abili y
analysis, p o iding a e sa ile and obus amewo k.
III. GINI INDEX-BASED FEATURE SELECTION
The Gini index has been p o en e ec i e in iden i ying ele-
an ea u es in many applica ions. Howe e , i su e s om
some limi a ions, including sensi i i y o ea u e in e ac ions,
suscep ibili y o o e i ing, and ins abili y when aced wi h
mino changes in he da ase [25]. Mo eo e , he Gini index
may no ully exploi he bene i so ensemble lea ning, which
can imp o e gene aliza ion and obus ness. This has mo i-
a ed he de elopmen o a no el algo i hm named Bukagini,
which seeks o enhance Gini index-based ea u e selec ion by
inco po a ing ensemble lea ning, ea u e in e ac ion analysis,
and s abili y analysis.
Gini Index De ini ion and Calcula ion: The Gini index is a
widely used impu i y measu e in decision ee algo i hms o
ea u e selec ion. I quan i ies a da ase ’s deg ee o impu i y
o diso de , wi h lowe alues indica ing pu e subse s. The
Gini index is de ined as [26]:
Gini(P)=1−X(Pi)2(1)
whe e Pi ep esen s he p opo ion o ins ances belonging o
he class iin he da ase . When applied o ea u e selec ion,
he Gini index can ank ea u es based on hei abili y o
disc imina e be ween di e en classes. A highe Gini index
alue o a ea u e indica es g ea e disc imina o y powe .
A decision ee is cons uc ed by i e a i ely spli ing he
da a, and he ea u e impo ance can be de i ed om he
Gini index educ ion caused by each ea u e. The Gini index
de e mines he bes spli ing poin o each ea u e in decision
ee algo i hms. The ea u e and he spli poin ha esul in
he lowes Gini index is selec ed, leading o he pu es child
nodes.
IV. THE BUKAGINI ALGORITHM
The Bukagini algo i hm is a no el enhanced Gini index-based
ea u e selec ion me hod combining ensemble lea ning,
FIGURE 1. BukaGini low cha .
ea u e in e ac ion, and s abili y analysis. By add essing he
limi a ions o adi ional Gini index-based me hods, he Buk-
agini algo i hm aims o imp o e model pe o mance, gen-
e aliza ion, and in e p e abili y. The lowcha in Figu e 1
p esen s an o e iew o he BukaGini algo i hm’s wo k low.
The p ocess begins wi h da a p ep ocessing, including clean-
ing, no maliza ion, and o he necessa y da a ans o ma ions.
Nex , he algo i hm calcula es he Gini index and anks he
ea u es based on hei impo ance. Following his s ep, he
algo i hm applies an ensemble-based app oach o cons uc
mul iple models ha exploi he complemen a y s eng hs o
each model. A he nex s age, he BukaGini algo i hm analy-
ses ea u e in e ac ion o iden i y and quan i y he in e ac ions
be ween di e en ea u es. This in o ma ion aids in iden i y-
ing signi ican ea u e in e ac ions ha can imp o e model
pe o mance. Subsequen ly, s abili y analysis is conduc ed o
assess he obus ness and consis ency o he selec ed ea u es
ac oss di e en da a samples. Finally, he algo i hm p oceeds
o model aining and e alua ion, whe e i e alua es he pe -
o mance o he selec ed ea u es using a ious pe o mance
me ics such as accu acy, p ecision, ecall, and F1 sco e. This
i e a i e p ocess allows he BukaGini algo i hm o add ess he
limi a ions o adi ional Gini index-based ea u e selec ion
me hods and imp o e model pe o mance, gene aliza ion,
and in e p e abili y.
A. MATHEMATICAL REPRESENTATION
Le D be a da ase con aining nsamples and m ea u es, whe e
each sample iis ep esen ed as a ec o (x1i,x2i,...,xmi) and
belongs o one o c a ge classes. The Gini index o ea u e
VOLUME 11, 2023 59389
M. A. Bouke e al.: BukaGini: A S abili y-Awa e Gini Index Fea u e Selec ion Algo i hm
jis de ined as:
Gini(j)=1−X(P(ck|xj))2
whe e P(ck|xj) is he p obabili y o class ckGi en he ea u e
j, which is calcula ed by di iding he numbe o samples by
class ckBy he o al numbe o samples o each spli . The
Gini index is calcula ed o each ea u e and anks hem based
on impo ance.
In he ensemble-based app oach, le Ebe an ensemble o
Tbase lea ne s (e.g., decision ees). The inal p edic ion o
sample i is de e mined by agg ega ing he p edic ions o each
base lea ne :
yi=F((x1i,x2i,...,xmi)
whe e Fis an agg ega ion unc ion (e.g., a majo i y o e o
classi ica ion o an a e age o eg ession).
Fo ea u e in e ac ion analysis, le xj1and xj2be wo
selec ed ea u es. The in e ac ion e m Ij1j2is de ined as:
Ij1j2=xj1×xj2
This in e ac ion e m is added o he selec ed ea u es, and
he pe o mance o he ensemble model is e alua ed wi h he
in e ac ion e m included.
The s abili y analysis is pe o med using esampling ech-
niques like c oss- alida ion. Le CV be he c oss- alida ion
sco e o he ensemble model o each esampled da ase . The
a e age s abili y sco e S is calcula ed as ollows:
S=1
L×XCV i
whe e L is he numbe o esampled da ase s, and CV iis he
c oss- alida ion sco e o he i h esampled da ase .
The Bukagini algo i hm combines he Gini index-based
ea u e selec ion wi h ensemble lea ning, ea u e in e ac ion
analysis, and s abili y analysis o c ea e a mo e obus and
in e p e able model.
B. PREPROCESSING
The p ep ocessing s ep in ol es handling missing alues,
con e ing ca ego ical a iables, no malizing o scaling da a,
and o he necessa y p ep ocessing asks ailo ed o he spe-
ci ic da ase .
C. GINI INDEX CALCULATION AND FEATURE RANKING
The Gini index is calcula ed o each ea u e in he da ase ,
and he ea u es a e anked based on hei Gini index al-
ues. The op-k ea u es a e hen selec ed o u he analy-
sis, whe e k is a use -de ined pa ame e de e mined h ough
c oss- alida ion.
D. ENSEMBLE-BASED APPROACH
An ensemble o base lea ne s (e.g., decision ees) is
employed o imp o e he gene aliza ion and obus ness o
he model. Ensemble me hods, such as Random Fo es , Bag-
ging, o Boos ing, can be used in his s ep o agg ega e
he p edic ions o mul iple base lea ne s and achie e be e
pe o mance. In his s udy, we u ilized he Random Fo es
ensemble me hod.
E. FEATURE INTERACTION ANALYSIS
Fea u e in e ac ion analysis gene a es in e ac ion e ms o
he op-k selec ed ea u es. The in e ac ion e m be ween wo
ea u es is he p oduc o hei alues. The pe o mance o
he ensemble model is e alua ed wi h he in e ac ion e ms
included, and he mos impo an in e ac ions a e e ained in
he inal ea u e se .
F. STABILITY ANALYSIS
S abili y analysis is conduc ed using esampling echniques,
such as c oss- alida ion. This s ep aims o e alua e he s abil-
i y o he Bukagini algo i hm by measu ing i s pe o mance
on di e en esampled da ase s. The a e age s abili y sco e is
calcula ed o assess he algo i hm’s s abili y o e all.
G. MODEL TRAINING AND EVALUATION
The inal selec ed ea u es, including he essen ial ea u e
in e ac ions, a e used o ain he ensemble model. The
model’s pe o mance is e alua ed on a es da ase using
a ious me ics, such as accu acy, p ecision, ecall, and F1
sco e.
V. EXPERIMENTAL SETUP
This sec ion will p o ide a de ailed o e iew o compa -
ing he p oposed BukaGini algo i hm wi h he adi ional
Gini index. This compa ison will enable us o assess he
e ec i eness o he BukaGini algo i hm in add essing he
limi a ions o he con en ional Gini index and enhancing
o e all model pe o mance. The se up includes he selec ion
o an app op ia e da ase , p ep ocessing s eps, implemen ing
bo h he BukaGini algo i hm and adi ional Gini index-based
ea u e selec ion, and e alua ing he esul s using a ious
pe o mance me ics. Th oughou his sec ion, we will ou line
he s eps o be ollowed, om da a acquisi ion o model
e alua ion, ensu ing a comp ehensi e unde s anding o he
expe imen al p ocedu e.
Mo eo e , Figu e 2p o ides a isual ep esen a ion o he
expe imen al se up o compa ing he pe o mance o he
BukaGini algo i hm and adi ional Gini index-based me h-
ods. The igu e illus a es he wo k low o he expe imen al
p ocess, including da a p ep ocessing, ea u e selec ion using
bo h he BukaGini algo i hm and he adi ional Gini index,
model aining, and e alua ion o he pe o mance me ics.
A. LAB ENVIRONMENT
The expe imen s in his s udy we e conduc ed using a pe -
sonal compu e wi h he ollowing speci ica ions:
•Ope a ing Sys em: Windows 10 64-bi
•P ocesso : In el Co e i7
•Memo y: 32 GB RAM
The lab en i onmen was se up o acili a e he implemen-
a ion and es ing o he BukaGini algo i hm and i s compa -
59390 VOLUME 11, 2023

M. A. Bouke e al.: BukaGini: A S abili y-Awa e Gini Index Fea u e Selec ion Algo i hm
FIGURE 2. Expe imen se up o BukaGini algo i hm and adi ional Gini
index compa ison.
ison wi h he adi ional Gini Index. To ensu e consis ency
and ep oducibili y o he esul s, we u ilized he ollowing
so wa e ools and lib a ies:
•Py hon 3: The p ima y p og amming language o
implemen ing he algo i hms, conduc ing da a p ep o-
cessing, and unning expe imen s.
•NumPy: A popula Py hon nume ical compu ing
lib a y o e icien a ay ope a ions and ma hema ical
calcula ions.
•Pandas: A da a manipula ion and analysis lib a y used
o loading, cleaning, and p ocessing he spambase
da ase .
•Sciki -lea n: A machine lea ning lib a y ha p o ides a
ange o ools o da a mining and da a analysis, includ-
ing classi ica ion, eg ession, and clus e ing algo i hms.
In his s udy, we used sciki -lea n o implemen he
decision ee classi ie s and e alua e hei pe o mance.
•Ma plo lib: A plo ing lib a y o c ea ing s a ic, in e -
ac i e, and anima ed isualiza ions in Py hon. We used
Ma plo lib o gene a e he igu es o isualizing he
s abili y sco es and compa ing he pe o mance o he
models.
The lab en i onmen was con igu ed o ensu e ha all
equi ed dependencies we e ins alled and ha he necessa y
da a and code iles we e o ganized in a s uc u ed manne .
This acili a ed a smoo h execu ion o he expe imen s and
allowed o easy analysis and compa ison. The codebase
was e sion-con olled using Gi , ensu ing ha all changes
and upda es we e acked and could be easily e e ed o
modi ied.
B. DATASETS DESCRIPTION
To e alua e he pe o mance o he Bukagini algo i hm,
we conduc ed expe imen s on se e al benchma k da ase s
om di e en domains. These da ase s ha e a ying numbe s
o ea u es, samples, and a ge classes, which helps assess
he e sa ili y and e ec i eness o he algo i hm. The da ase s
used in he expe imen s a e as ollows:
1. HSSP Da ase : This da ase con ains in o ma ion on
he pe o mance o high school s uden s in ma hema ics,
including hei g ades and demog aphic in o ma ion. The
da a was collec ed om h ee high schools in he Uni ed
S a es.
Sou ce: h ps://www.kaggle.com/da ase s/ kia isak/
s uden -pe o mance-in-ma hema ics
2. Cance Da a: This da ase con ains in o ma ion on
570 cance cells and 30 ea u es o de e mine whe he
he cance cells a e benign o malignan . The cance da a
includes wo ypes o cance s: 1. benign cance (B) and 2.
malignan cance (M).
Sou ce: h ps://www.kaggle.com/da ase s/e dem aha/
cance -da a
3. Spambase: The Spambase da ase con ains a ound 4,600
emails labeled as spam o ham. The da ase was c ea ed by
collec ing spam emails om pos mas e s and indi iduals,
while non-spam emails came om iled wo k and pe sonal
emails.
Sou ce: h ps://a chi e.ics.uci.edu/ml/da ase s/spambase
4. UNSW-NB15: The UNSW-NB15 da ase con ains aw
ne wo k packe s gene a ed by he IXIA Pe ec S o m ool
in he Cybe Range Lab o he Aus alian Cen e o Cybe
Secu i y (ACCS) o c ea e a hyb id o eal mode n no mal
ac i i ies and syn he ic a ack beha io s.
Sou ce:h ps:// esea ch.unsw.edu.au/p ojec s/unsw-nb15-
da ase
C. EVALUATION METRICS
To assess he pe o mance o he Bukagini algo i hm, we used
a ious e alua ion me ics. These me ics comp ehensi ely
unde s and he algo i hm’s pe o mance, including accu acy,
gene aliza ion, and s abili y. The e alua ion me ics used in
he expe imen s a e:
•Accu acy: The p opo ion o co ec ly classi ied sam-
ples o he o al numbe o samples.
•P ecision: The p opo ion o ue posi i e p edic ions o
he sum o ue posi i e and alse posi i e p edic ions.
•Recall: The p opo ion o ue posi i e p edic ions o he
sum o ue posi i e and alse nega i e p edic ions.
•F1 Sco e: The ha monic mean o p ecision and ecall,
p o iding a balanced e alua ion o bo h me ics.
•S abili y sco e: The a e age c oss- alida ion sco e
ob ained du ing s abili y analysis, indica ing he algo-
i hm’s s abili y.
D. PARAMETER SETTINGS
Fo he expe imen s, we se he ollowing pa ame e s o he
BukaGini algo i hm:
•Top-k ea u es: We selec ed he bes en ea u es a e
anking based on he Gini index o bo h he BukaGini
algo i hm and he adi ional Gini index-based me h-
ods. The use can se his pa ame e o de e mine using
c oss- alida ion o o he ea u e selec ion e alua ion
VOLUME 11, 2023 59391
M. A. Bouke e al.: BukaGini: A S abili y-Awa e Gini Index Fea u e Selec ion Algo i hm
echniques, such as ecu si e ea u e elimina ion o g id
sea ch.
•Ensemble me hod: In ou expe imen s, we used Ran-
dom Fo es as he ensemble lea ning me hod due o
i s obus ness, abili y o handle high-dimensional da a,
and excellen pe o mance in a ious applica ions. The
pa ame e s o he base lea ne s and he ensemble
me hod should be uned o achie e op imal pe o mance.
I is wo h no ing ha he BukaGini algo i hm can be
ex ended o wo k wi h o he ensemble me hods, such
as Bagging o Boos ing, o explo e hei impac on
pe o mance.
•Numbe o esampled da ase s o s abili y analysis:
We used 5- old c oss- alida ion o s abili y analysis,
which esul ed in 5 esampled da ase s. The numbe o
olds o esampling echniques can be adjus ed acco d-
ing o he da ase size and compu a ional esou ces a ail-
able o balance s abili y assessmen and compu a ional
e iciency.
•Tes se a io: Fo all da ase s, we used a es se a io o
20% o e alua e he pe o mance o he selec ed ea u es
on unseen da a. This allowed us o assess he gene al-
iza ion capabili ies o he BukaGini algo i hm and he
adi ional Gini index-based me hods.
•Da a p ep ocessing: We used LabelEncode o encode
all nominal ea u es in he da ase s and applied S an-
da dScale o scale he ea u es. This ensu ed all ea u es
we e on a simila scale and p e en ed biases due o he
di e en measu emen uni s o anges.
The expe imen s speci ically compa ed he BukaGini algo-
i hm wi h adi ional Gini index-based ea u e selec ion
me hods o demons a e i s supe io pe o mance. Fu he -
mo e, he expe imen s we e conduc ed on a ious da ase s
omdi e en domains, highligh ing he e sa ili yand adap -
abili y o he BukaGini algo i hm.
VI. RESULTS AND DISCUSSION
The expe imen s we e conduc ed o compa e he pe o -
mance o he BukaGini algo i hm wi h he adi ional Gini
index-based ea u e selec ion me hod. Resul s showed ha
he BukaGini algo i hm consis en ly ou pe o med adi ional
Gini index-based me hods ega ding Accu acy, P ecision,
Recall, and F1 sco e ac oss all da ase s. This imp o emen
can be a ibu ed o inco po a ion o ensemble lea ning,
ea u e in e ac ion analysis, and s abili y analysis in he
BukaGini algo i hm, which add esses he limi a ions o a-
di ional me hods and enhances hei pe o mance.
A. HSSP DATASET
In he HSSP da ase , we applied he BukaGini algo i hm and
compa ed i s esul s o hose ob ained using he adi ional
Gini Index. He e is a de ailed discussion o he esul s:
The s abili y sco es o he HSSP da ase using he Buk-
aGini algo i hm a e shown in Figu e 3. These sco es ep e-
sen he model’s pe o mance on di e en esampled da ase s
FIGURE 3. S abili y sco es o 5 olds HSSP.
TABLE 1. Resul s o adi ional Gini index on HSSP da ase .
( olds) using c oss- alida ion. The a e age s abili y sco e is
85%, which indica es ha he model pe o ms consis en ly
well ac oss di e en olds, wi h only sligh a ia ions in pe -
o mance. The s abili y analysis in he BukaGini algo i hm
helps assess he consis ency o ea u e impo ance ac oss di -
e en esampled da ase s o olds. In o he wo ds, i measu es
how s able he signi icance o a ea u e is when he model is
ained on sligh ly di e en subse s o he da a. A highe s a-
bili y sco e indica es ha a ea u e’s impo ance is consis en
ac oss di e en da a samples. This is c ucial because, in eal-
wo ld si ua ions, da a dis ibu ions can change o may con ain
noise. A s able ea u e is mo e likely o gene alize well on
unseen da a and is less p one o o e i ing. The BukaGini
algo i hm selec s ea u es con ibu ing o a mo e obus and
accu a e model by ocusing on ea u es wi h high s abili y
sco es.
The model’s pe o mance using he adi ional Gini index
on he HSSP da ase is shown in Table 1. The o e all accu acy
o he model is 88%. The p ecision, ecall, and F1-sco e o
class 0 (non-passing s uden s) a e 87.36%, while o class 1
(passing s uden s), hese me ics a e 88.57%. These me ics’
mac o a e age and weigh ed a e age a e also qui e simila ,
a ound 87.96% and 88%, espec i ely.
Table 2p esen s he model’s pe o mance using he Buk-
aGini algo i hm on he HSSP da ase . The o e all accu acy
o his model is 90.5%, which is 2.5% highe han he model
using he adi ional Gini index. The p ecision, ecall, and
F1-sco e o class 0 a e 88.77%, 91.57%, and 90.15%, espec-
i ely, while o class 1, hese me ics a e 92.15%, 89.52%,
and 90.82%, espec i ely. These me ics’ mac o a e age and
weigh ed a e age a e close, a ound 90.46% and 90.55%,
espec i ely.
59392 VOLUME 11, 2023
M. A. Bouke e al.: BukaGini: A S abili y-Awa e Gini Index Fea u e Selec ion Algo i hm
FIGURE 4. S abili y sco es o 5 olds cance da ase .
TABLE 2. Resul s o BukaGini on HSSP da ase .
In conclusion, he BukaGini algo i hm ou pe o ms he
adi ional Gini index ega ding accu acy and o he pe o -
mance me ics on he HSSP da ase . The BukaGini model
has a be e p ecision, ecall, and F1-sco e o bo h classes,
and he s abili y sco es indica e ha he model pe o ms
consis en ly well ac oss di e en olds. This demons a es
ha he addi ional s abili y and ea u e in e ac ion analysis
in he BukaGini algo i hm con ibu es o imp o ed model
pe o mance and obus ness.
B. CANCER DATASET
We applied he BukaGini algo i hm o he cance da ase and
compa ed i s esul s o hose ob ained using he adi ional
Gini Index. He e is a de ailed discussion o he esul s:
The BukaGini algo i hm compu es s abili y sco es o
assess he eliabili y and obus ness o he selec ed ea u es.
The s abili y sco es o he cance da ase a e shown in
Figu e 4.
The a e age s abili y sco e o 94% indica es ha he
selec ed ea u es consis en ly con ibu e o he model’s pe -
o mance ac oss di e en da a samples. This p o es ha he
BukaGini algo i hm can iden i y obus and eliable ea u es
o he gi en da ase .
We compa ed he pe o mance o he models buil using he
adi ional Gini Index and BukaGini algo i hm. The esul s
a e summa ized in Table 3and Table 4. The BukaGini model
achie es an accu acy o 96.49%, which is highe han he
adi ional Gini Index model’s accu acy o 95.61%. This
imp o emen can be a ibu ed o he addi ional s abili y and
ea u e in e ac ion analyses inco po a ed in he BukaGini
TABLE 3. Resul s o adi ional Gini index on cance da ase .
TABLE 4. Resul s o BukaGini on cance da ase .
algo i hm. These addi ional analyses con ibu e o a mo e
obus and accu a e model.
In summa y, he esul s demons a e ha he BukaGini
algo i hm ou pe o ms he adi ional Gini Index ega ding
s abili y and accu acy. The BukaGini algo i hm’s addi ional
s abili y and ea u e in e ac ion analysis con ibu e o i s
imp o ed pe o mance and obus ness, making i a be e
choice o his da ase .
C. SPAMBASE
In he spambase da ase , we applied he BukaGini algo i hm
and compa ed i s esul s o hose ob ained using he adi-
ional Gini Index. The s abili y sco es compu ed by he Buk-
aGini algo i hm, which assesses he eliabili y and obus ness
o he selec ed ea u es, yielded an a e age s abili y sco e o
93% (Figu e 5), indica ing ha he selec ed ea u es consis-
en ly con ibu e o he model’s pe o mance ac oss di e en
da a samples. This demons a es ha he BukaGini algo i hm
can iden i y obus and eliable ea u es o he gi en da ase .
We compa ed he pe o mance o he models buil using
he adi ional Gini Index and he BukaGini algo i hm. The
esul s a e summa ized in Figu e 5and Table 6. In e es ingly,
while he BukaGini model’s s abili y sco es we e highe ,
he o e all pe o mance me ics showed some di e ences
be ween he adi ional Gini Index and BukaGini models.
The adi ional Gini Index model achie ed an accu acy o
91.86%, while he BukaGini model achie ed a sligh ly highe
accu acy o 92.94%. One possible explana ion o he a ying
pe o mance o he adi ional Gini Index and he BukaGini
algo i hm is he speci ic cha ac e is ics o he spambase
da ase . The na u e o he da a, including he dis ibu ion o
ea u es and he class imbalance, can signi ican ly impac
he e ec i eness o ea u e selec ion me hods. The cha ac-
e is ics o he spambase da ase may be be e sui ed o
VOLUME 11, 2023 59393
M. A. Bouke e al.: BukaGini: A S abili y-Awa e Gini Index Fea u e Selec ion Algo i hm
TABLE 5. Resul s o adi ional Gini index on Spambase da ase .
TABLE 6. Resul s o BukaGini on spambase da ase .
FIGURE 5. S abili y sco es o 5 olds spambase da ase .
he BukaGini algo i hm, which places g ea e emphasis on
s abili y and ea u e in e ac ion analysis.
Ano he explana ion is ha he bene i s o inco po a ing
s abili y and ea u e in e ac ion analysis in he BukaGini
algo i hm may depend on he da ase used. While he spam-
base da ase demons a ed he e ec i eness o he BukaGini
algo i hm in iden i ying eliable ea u es, i is possible ha
o he da ase s may no bene i om hese addi ional analyses.
The e o e, i is impo an o conside he speci ic p ope ies
o he da ase when selec ing he mos app op ia e ea u e
selec ion me hod.
D. UNSWNB15 DATASET
In he UNSW-NB15 da ase , he BukaGini algo i hm com-
pu es s abili y sco es o assess he eliabili y and obus ness
o he selec ed ea u es.
The s abili y sco es o he UNSW-NB15 da ase a e
depic ed in Figu e 6. The a e age s abili y sco e o 93%
indica es ha he chosen ea u es consis en ly con ibu e o
he model’s pe o mance ac oss di e en da a samples. This
FIGURE 6. S abili y sco es o 5 olds UNSWNB15 da ase .
TABLE 7. T adi ional Gini index on UNSWNB15 da ase .
TABLE 8. Resul s o BukaGini on UNSWNB15 da ase .
p o es ha he BukaGini algo i hm can iden i y obus and
eliable ea u es o he gi en da ase .
We compa ed he pe o mance o he models buil using he
adi ional Gini Index and BukaGini algo i hm. The esul s
a e summa ized in Table 7and Table 8. The BukaGini model
ou pe o ms he con en ional Gini Index model in e ms
o accu acy (93.09% s 92.77%). This imp o emen can
be a ibu ed o he addi ional s abili y analysis and ea u e
in e ac ion analysis inco po a ed in he BukaGini algo i hm,
which helps o iden i y mo e obus and eliable ea u es. The
esul s demons a e ha he BukaGini algo i hm leads o a
mo e accu a e and obus model o he UNSW-NB15 da ase ,
as e idenced by be e pe o mance me ics and highe s abil-
i y sco es.
E. PERFORMANCE ANALYSIS OF THE BUKAGINI
ALGORITHM
The pe o mance o he BukaGini algo i hm was ho -
oughly analyzed by e alua ing i s e ec i eness ac oss a i-
ous da ase s, such as he HSSP da ase , he Cance da ase ,
he Spambase da ase , and he UNSW-NB15 da ase . The
59394 VOLUME 11, 2023