Applica ion Design o Cus ome Chu n P edic ion
Using Random Fo es and XGBoos Algo i hms o
Telecommunica ion Indus y in Indonesia
Imanuel Re elino Mu man o1
In o ma ion Sys em
Uni e si as Bina Nusan a a
Jaka a, Indonesia
imanuel. e [email protected]
Hani Se iawan2
In o ma ion Sys em
Uni e si as Bina Nusan a a
Jaka a, Indonesia
[email p o ec ed]
Je ba Runggu Hamonangan3
Doloksa ibu
In o ma ion Sys em
Uni e si as Bina Nusan a a
Jaka a, Indonesia
[email p o ec ed]
Nau al Ya i4
In o ma ion Sys em
Uni e si as Bina Nusan a a
Jaka a, Indonesia
[email p o ec ed]
Abs ac - Cus ome chu n is a p ima y challenge in he
elecommunica ions indus y, di ec ly impac ing e enue and
inc easing acquisi ion cos s. This esea ch designs a chu n
p edic ion sys em o Indonesian ISP se ices using Machine
Lea ning algo i hms, speci ically Random Fo es and
XGBoos . To add ess he p e alen issue o class imbalance in
chu n da a, we implemen a hyb id app oach combining
Clus e -Based Unde sampling wi h Cos -Sensi i e Lea ning.
Explainable AI (XAI) me hods we e applied o in e p e model
p edic ions, speci ically LIME and SHAP, o p o ide
anspa en in e p e a ions o model p edic ions a bo h global
and local le els. Cus ome segmen a ion using K-Means
clus e ing is in eg a ed o suppo pe sonalized e en ion
s a egies. The inal ou pu is an in e ac i e dashboa d buil
wi h S eamli , se ing as a decision-suppo ool o
managemen . Ou esul s show ha he XGBoos model
ou pe o ms o he s wi h a ROC-AUC o 0.900 and Recall o
0.907. The s udy includes a discussion on he business impac
and limi a ions, highligh ing he po en ial o signi ican cos
sa ings h ough p oac i e e en ion.
Keywo ds— Cus ome Chu n, Random Fo es , XGBoos ,
Explainable AI, Cos -Sensi i e Lea ning, Telecommunica ion.
I. INTRODUCTION
The In e ne Se ice P o ide (ISP) indus y in Indonesia
aces inc easing compe i i e p essu e and high cus ome
chu n a es, which signi ican ly educe p o i abili y and
ele a e cus ome acquisi ion cos s [1], [2]. Acco ding o
indus y epo s, chu n a es in Sou heas Asia’s elecom
sec o can exceed 20%, leading o subs an ial e enue losses
and declining cus ome li e ime alue (CLV) [3]. The
challenge is u he compounded by he limi ed adop ion o
p edic i e analy ics o p oac i e chu n managemen ,
esul ing in eac i e and cos ly e en ion campaigns [4].
Machine Lea ning (ML) has eme ged as a obus
app oach o ea ly chu n p edic ion, p o iding au oma ed
insigh s in o complex beha io al pa e ns wi hin la ge-scale
da ase s [5]. Recen s udies ha e shown ha ensemble
algo i hms such as Random Fo es (RF) and XGBoos
(XGB) ou pe o m adi ional classi ie s due o hei abili y
o handle non-linea in e ac ions, manage high-dimensional
da a, and main ain gene aliza ion ac oss imbalanced da ase s
[6], [7]. Manzoo e al. (2024) con i med he dominance o
ensemble me hods in elecom chu n p edic ion, epo ing
consis en AUC alues abo e 0.88 in compa a i e
e alua ions [8]. Simila ly, esea ch in Malaysia and Vie nam
has alida ed hese algo i hms’ egional e ec i eness,
achie ing high ROC-AUC sco es be ween 0.87 and 0.90 [9],
[10].
Despi e hese ad an ages, ensemble models a e o en
c i icized o hei “black-box” na u e, limi ing manage ial
us and in e p e abili y in business decision-making [11].
Explainable AI (XAI) has he e o e become essen ial o
b idging his gap, allowing p ac i ione s o unde s and he
easoning behind model p edic ions [12]. Among he mos
e ec i e XAI ools a e SHAP (SHapley Addi i e
exPlana ions), which quan i ies each ea u e’s con ibu ion
based on coope a i e game heo y, and LIME (Local
In e p e able Model-agnos ic Explana ions), which p o ides
local app oxima ion o indi idual ins ances [13], [14]. When
in eg a ed wi h p edic i e modeling, hese ools can
ans o m chu n p edic ion sys ems in o anspa en and
ac ionable decision-suppo solu ions. Ou esea ch makes
h ee key con ibu ions:
1. Building a obus p edic ion model using RF and
XGBoos wi h a hyb id app oach o handle imbalanced
da a (Clus e -Based Unde sampling and Cos -Sensi i e
Lea ning).
2. Applying Explainable AI (XAI) ia LIME ( o local
in e p e a ion) and SHAP ( o global in e p e a ion) o
elucida e chu n causes.
3. In eg a ing p edic ions, XAI, and cus ome segmen a ion
(K-Means) in o an in e ac i e S eamli dashboa d o
suppo da a-d i en decision-making and pe sonalized
e en ion s a egies.
II. METHODOLOGY
This esea ch ollows he CRISP-DM (C oss-Indus y
S anda d P ocess o Da a Mining) amewo k combined
wi h Sc um me hodology o enable i e a i e and adap i e
de elopmen [15]. The combina ion ensu es s uc u ed
expe imen a ion, s akeholde eedback, and inc emen al
e inemen h oughou he p ojec li ecycle.
A. Da a Collec ion and P epa a ion
The esea ch popula ion comp ises ac i e cus ome s o
he ICONNET se ice in he Cen al Ja a Regional. A
s a i ied andom sample o 100.000 cus ome s was
selec ed om his o ical da a co e ing Janua y 2020-
Sep embe 2025, collec ed om CRM (cus ome
p o iles, complain s), Billing (paymen his o y), and
Ne wo k Ope a ion (usage da a) sys ems. The da ase
was spli in o h ee subse s using s a i ied sampling:
70% T aining Se (70.000 samples), 15% Valida ion Se
(15.000 samples), and 15% Tes Se (15.000 samples).
B. Da a P ep ocessing
P ep ocessing was conduc ed o handle missing alues,
ou lie s, and ea u e inconsis encies. Missing nume ical
alues (e.g., Tenu e) we e impu ed using he median,
while ca ego ical a iables we e illed using he mode
[16]. Ou lie s in con inuous a iables such as
Paymen _Delay we e capped a he 99 h pe cen ile o
mi iga e skewness. Fea u e enginee ing in oduced
de i ed a ibu es such as A e age_Mon hly_Spend and
Complain _Ra io o s eng hen he model’s
disc imina i e powe [17].
Explo a o y Da a Analysis (EDA) e ealed a s ong
imbalance be ween chu ned (17.55%) and non-chu ned
(82.45%) cus ome s (Fig. 1). Co ela ion analysis (Fig.
1) indica ed ha a iables such as
Paymen _Delay_Days, Down ime_Minu es, and
Complain _Coun we e highly co ela ed wi h chu n
likelihood, con i ming hei ele ance as p edic i e
ea u es.
Fig. 1. Co ela ion hea map o nume ical ea u es showing
ela ionships be ween key a iables like Paymen _Delay and
Complain _Coun .
C. Imbalanced Da a Handling
Cus ome chu n da a ypically su e om class
imbalance, whe e chu ne s ep esen a small mino i y.
To add ess his, we employed a hyb id esampling and
cos -weigh ing s a egy [18]:
1. Clus e -Based Unde sampling: he majo i y class
(non-chu n) was segmen ed in o clus e s using K-
Means; ep esen a i e samples we e hen selec ed
om each clus e o educe edundancy wi hou
losing dis ibu ional in eg i y.
2. Cos -Sensi i e Lea ning: models we e ained wi h
adjus ed misclassi ica ion penal ies h ough
pa ame e s class_weigh ='balanced' (Random
Fo es ) and scale_pos_weigh (XGBoos ),
calib a ed o he chu n a io o 1:4.7 [19].
D. Machine Lea ning Modeling
Two ensemble lea ning algo i hms we e used:
1. Random Fo es (RF): a bagging-based model ha
cons uc s mul iple decision ees on boo s ap
samples and agg ega es p edic ions h ough
majo i y o ing [20].
2. XGBoos (Ex eme G adien Boos ing): a boos ing
algo i hm ha builds ees sequen ially, minimizing
esidual e o s h ough g adien op imiza ion [21].
A Logis ic Reg ession model se ed as a baseline o
pe o mance compa ison and was e alua ed be o e he
applica ion o he hyb id sampling echnique. This
abla ion s udy con i med he necessi y o using
ensemble me hods and he cus om sampling echnique.
E. Model Valida ion and Hype pa ame e Tuning
Model alida ion used a igo ous ain- alida ion- es
spli (70%:15%:15%). Hype pa ame e uning was
sys ema ically pe o med using G id Sea ch (o
Randomized Sea ch) wi h 5- old S a i ied C oss-
Valida ion on he aining se . This s a i ica ion ensu ed
ha he mino i y class dis ibu ion was main ained
ac oss all olds, add essing he imbalance isk du ing
uning.
The hype pa ame e space explo ed included a wide
ange o alues o ind he op imal con igu a ion:
• XGBoos : The sea ch space included n_es ima o s
( anging om 50 o 500, s ep 50), max_dep h (3 o
15), lea ning_ a e (0.01 o 0.3), subsample (0.6 o
1.0), and scale_pos_weigh ( uned be ween 4.0 and
6.0 based on he 1:4.7 imbalance a io). Ea ly
s opping (wi h a pa ience o 10 ounds) was
implemen ed using he alida ion se o p e en
o e i ing and op imize he numbe o boos ing
ounds. The op imal con igu a ion ound was
n_es ima o s=100, max_dep h=7, lea ning_ a e
=0.1, subsample=0.8, colsample_by ee=0.8, scale_
pos_weigh =4.7, and gamma=0.1.
• Random Fo es : The sea ch space ocused on
n_es ima o s (100 o 300), max_dep h (10 o 20),
and min_samples_spli (2 o 10). The op imal
con igu a ion selec ed was n_es ima o s=200,
max_dep h=15, min_samples_spli =4, min_
samples_lea =2, class_weigh ='balanced', and
max_ ea u es='sq '.
The op imal hype pa ame e s selec ed we e hose ha
maximized he ROC-AUC sco e on he c oss- alida ion
olds, p io i izing obus class sepa a ion o e aw
accu acy. Table I p o ides he inal chosen con igu a ion
o each model.
F. Model In e p e a ion (XAI) and Segmen a ion
To enhance model anspa ency, we employed wo XAI
echniques:
1. SHAP (SHapley Addi i e exPlana ions): p o ides
bo h global and local in e p e abili y by es ima ing
he ma ginal con ibu ion o each ea u e o he
p edic ion [22].
2. LIME (Local In e p e able Model-agnos ic
Explana ions): gene a es in e p e able local
app oxima ions, enabling case-by-case explana ion
o chu n p obabili y o indi idual cus ome s [23].
K-Means Clus e ing was used o cus ome
segmen a ion based on beha io pa e ns, enabling
a ge ed e en ion s a egies. All componen s we e
in eg a ed in o an in e ac i e S eamli dashboa d,
acili a ing isualiza ion and decision-making.
G. E alua ion Me ics
Gi en he imbalanced da a, we used P ecision, Recall,
F1-Sco e, and ROC-AUC. Recall (sensi i i y) is
pa icula ly c ucial in chu n p edic ion o minimize alse
nega i es (cus ome s who chu n bu a e no iden i ied).
III. RESULTS AND DISCUSSION
A. P edic ion Model Pe o mance
A e ex ensi e hype pa ame e op imiza ion and
alida ion, bo h Random Fo es and XGBoos models we e
e alua ed on he es da ase using ou me ics: P ecision,
Recall, F1-Sco e, and ROC-AUC. The esul s (Table I)
con i m ha XGBoos achie ed supe io o e all
pe o mance, pa icula ly in ecall and class-sepa a ion
capabili y.
Table I. Model Pe o mance Compa ison
Model
P ecision
Recall
F1
-
Sco e
ROC
-
AUC
Random
Fo es 0.776 0.879 0.825 0.898
XGBoos 0.747 0.907 0.819 0.900
The XGBoos model’s high Recall (0.907) demons a es
i s e ec i eness in iden i ying ac ual chu ne s, minimizing
alse nega i es ha could esul in inancial losses. The
ROC-AUC o 0.900 indica es s ong disc imina i e
pe o mance be ween chu n and non-chu n classes. These
esul s align wi h indings om p e ious esea ch in he
egion, such as Lee and Singh (2024) in Malaysia (AUC =
0.88) and T an and Nguyen (2023) in Vie nam (AUC = 0.89)
[9], [10].
The Con usion Ma ix (Fig. 2) u he illus a es he high
ue-posi i e a e achie ed by XGBoos ’s cos -sensi i e
lea ning con igu a ion. The emphasis on ecall op imiza ion
ensu es ha cus ome s wi h a high p obabili y o chu n a e
a ely o e looked, which is c ucial o p oac i e e en ion
p og ams in elecom ope a ions [24].
Fig. 2. Con usion Ma ix o he inal XGBoos model on he es se ,
showing he coun o T ue Nega i es, False Posi i es, False Nega i es, and
T ue Posi i es.
B. Model In e p e a ion (XAI)
To enhance anspa ency and manage ial in e p e abili y,
Explainable AI (XAI) echniques—SHAP and LIME—we e
applied o he ained XGBoos model.
1. Global In e p e a ion (SHAP): The SHAP summa y plo
(Fig. 3) iden i ies he op con ibu ing ea u es:
Paymen _Delay_Days, Down ime_Minu es, and
Complain _Coun . These ea u es collec i ely explain
mos o he a iance in chu n p edic ion. No ably, longe
paymen delays and equen se ice down ime a e
posi i ely co ela ed wi h highe chu n p obabili y [25].
The SHAP global ea u e impo ance con i ms he
ope a ional ele ance o billing and se ice quali y
ac o s, aligning wi h elecom business p io i ies [26].
Fig. 3. SHAP summa y plo showing he global ea u e impo ance
based on mean absolu e SHAP alues. paymen _delay_days is he
s onges p edic o o chu n.
2. Local In e p e a ion (LIME): LIME p o ides g anula
in e p e abili y by explaining indi idual p edic ions. Fo
ins ance, a speci ic cus ome (ID 131100728911) was
p edic ed o chu n wi h 60.3% p obabili y, p ima ily
in luenced by Paymen _Delay_Days = 4.7 and
Down ime_Minu es = 194.6 (Fig. 4). This local
explana ion enables pe sonalized e en ion ac ions—such
as p o iding a ge ed paymen lexibili y o p oac i e
main enance [27]
The combined use o SHAP and LIME b idges he
in e p e abili y gap, empowe ing manage s o jus i y
in e en ions based on quan i iable ea u e con ibu ions
a he han opaque s a is ical ou pu s. This aligns wi h
ecen ends emphasizing us wo hy AI in business
analy ics [28].
Fig. 4. LIME ea u e plo o speci ic cus ome .
The dashboa d p o ides a ea u e con ibu ion analysis
o indi idual cus ome s (Fig. 5), deli e ing clea and
ac ionable insigh s o suppo he in e en ion eam in
making da a-d i en e en ion decisions.
Fig. 5. Ac ionable Insigh o speci ic cus ome
C. Cus ome Segmen a ion Resul s (K-Means)
K-Means clus e ing iden i ied h ee dis inc cus ome
segmen s (Table II), enabling e icien esou ce alloca ion
o e en ion campaigns.
TABLE II. CUSTOMER SEGMENTATION PROFILES
P o ile
Segmen A
Loyal
Cos ume
Segmen B
A -Risk
Cos ume
Segmen C
New
Cos ume
A e age Tenu e 38 Mon hs 15 Mon hs 6 Mon hs
Paymen Delay Low High Medium
Chu n Ra e 2% 45% 18%
Recommenda ion
Loyal y
P og am &
Upselling
P oac i e
in e en ion &
Complain
Resolu ion
Se ice &
Onboa ding
Segmen B (A -Risk Cus ome s) cons i u es he highes
chu n isk g oup, equi ing p io i y in e en ion campaigns.
The segmen a ion ou comes alida e ha chu n isk is no
uni o m and mus be add essed h ough di e en ia ed
s a egies. This app oach aligns wi h mode n Cus ome
Expe ience Managemen (CXM) amewo ks emphasizing
beha io al segmen a ion o esou ce op imiza ion [29].
D. Business Impac and Cos -Bene i Implica ions
The dashboa d p o ides Cos -Bene i Analysis o
Ta ge ed Chu n P e en ion (Fig. 6). The model’s high ecall
(90.7%) is c i ical o ealizing signi ican cos sa ings and
Re u n on In es men (ROI). Ou analysis, based on he
Cus ome Li e ime Value (CLV) me ic de i ed om he
a e age cus ome li e ime o 31.3 mon hs, indica es ha a
success ul e en ion campaign a ge ing he 90.7% co ec ly
iden i ied high- isk chu ne s could sa e app oxima ely Rp
3,035,876,270 mon hly in p e en ed CLV loss. The Cos -
Bene i implica ion is calcula ed by con as ing he cos o
in e en ion (e.g., o e ing a 1-mon h se ice discoun o
p o iding p oac i e echnical suppo , es ima ed a Rp X pe
cus ome ) agains he po en ial e enue p ese a ion
(A e age CLV pe cus ome , es ima ed a Rp 3,035,876).
When in e en ion cos s - such as o e ing discoun s o
cus ome suppo - a e ac o ed in, he model’s e u n on
in es men (ROI) emains s ongly posi i e. By p io i izing
high- isk cus ome s iden i ied h ough p edic i e modeling,
companies can alloca e e en ion budge s mo e e icien ly,
educing was ed expendi u es on low- isk cus ome s [30].
This ou come is consis en wi h p e ious economic
e alua ions o AI-d i en e en ion sys ems in elecom
en e p ises [31].
Fig. 6. Cos -Bene i Analysis o Ta ge ed Chu n P e en ion
E. Limi a ions and Fu u e Wo k
This s udy, while deli e ing s ong p edic i e
pe o mance, has se e al acknowledged limi a ions ha
impac i s gene alizabili y. Fi s ly, he sample size o
100,000 cus ome s om a single egion (Cen al Ja a) is
ela i ely small o he highly di e se Indonesian
elecommunica ions ma ke . This single egion ocus means
he model may no be ep esen a i e o cus ome beha io in
o he egions (e.g., Suma a o Eas e n Indonesia), which
may ha e di e en ne wo k in as uc u es, compe i i e
landscapes, o consume p o iles. Fu u e esea ch will ocus
on:
1. Enhancing Gene alizabili y: Valida ing he model on
la ge , mul i- egional da ase s (e.g., 200,000+ cus ome s)
o ensu e ep esen a i e pe o mance ac oss Indonesia.
2. Ad anced Fea u e In eg a ion: Inco po a ing addi ional
da a sou ces, such as social media sen imen analysis and
eal- ime ne wo k pe o mance me ics, o imp o ed
p edic i e powe .
3. Abla ion S udies: Fu he expe imen s isola ing he
impac o indi idual imbalance-handling componen s
(Clus e -Based s. Cos -Sensi i e) could cla i y hei
espec i e con ibu ions [32].
4. E hical Conside a ions: Da a ai ness and p i acy emain
key conce ns o p edic i e analy ics, equi ing
anspa en go e nance amewo ks [33].
IV. CONCLUSION
This s udy success ully designed and alida ed an
Explainable AI–d i en cus ome chu n p edic ion sys em
ailo ed o he Indonesian ISP indus y. By le e aging
Random Fo es and XGBoos wi h hyb id da a-balancing and
XAI in e p e abili y mechanisms (SHAP and LIME), he
sys em achie ed compe i i e pe o mance (ROC-AUC =
0.900, Recall = 0.907).
The in eg a ion o p edic i e modeling, segmen a ion,
and isualiza ion in o a uni ied S eamli dashboa d p o ides
managemen wi h ac ionable, anspa en insigh s o guide
p oac i e cus ome e en ion. These indings con i m ha
combining echnical igo wi h in e p e abili y and business
alignmen leads o measu able economic bene i s.
Fu u e esea ch should expand da a di e si y, alida e
c oss- egional scalabili y, and explo e mul imodal AI
a chi ec u es o dynamic cus ome beha io p edic ion in
eal ime.
Imanuel Re elino Mu man o: Concep ualiza ion,
Me hodological O e sigh , and Manusc ip Supe ision.
Hani Se iawan: Da a P ep ocessing, Model E alua ion, and
Business Impac Analysis. Je ba Runggu Hamonangan
Doloksa ibu: Model De elopmen , In eg a ion o XAI
Componen s, and Dashboa d Implemen a ion. Nau al Ya i:
O iginal D a P epa a ion, Segmen a ion Analysis, and XAI
Visualiza ion.
ACKNOWLEDGMENT
The au ho s exp ess hei g a i ude o Bina Nusan a a
Uni e si y o esea ch suppo and acili ies, and o PT PLN
Icon Plus (Cen al Ja a Regional) o p o iding access o
ope a ional da a and business insigh s. Special app ecia ion
is ex ended o colleagues and amilies o hei con inuous
encou agemen and aluable eedback h oughou he
esea ch p ocess.
REFERENCES
[1] U. Manzoo e al., “A Re iew o Machine Lea ning Techniques o
Cus ome Chu n P edic ion,” Jou nal o Business Analy ics, ol. 1,
no. 1, pp. 10–25, 2024.
[2] P. Anggo o, F. R. Yunia o, and A. Wibowo, “P edic ion and
Analysis o Cus ome Chu n in Indonesian Telecommunica ion
Company Using Machine Lea ning and Fea u e Selec ion,” Jou nal o
Physics: Con e ence Se ies, ol. 1823, 2021.
[3] A. Das ile, T. Celik, and S. Po sane, “S a is ical and Machine
Lea ning Models in C edi Sco ing: A Re iew,” Applied So
Compu ing, ol. 91, 2020.
[4] G. Kau and N. Kuma , “Chu n P edic ion in Telecom Using Machine
Lea ning Techniques: A Re iew,” In e na ional Jou nal o
In o ma ion Technology, ol. 15, no. 2, pp. 775–784, 2023.
[5] L. B eiman, “Random Fo es s,” Machine Lea ning, ol. 45, no. 1, pp.
5–32, 2001.
[6] T. Chen and C. Gues in, “XGBoos : A Scalable T ee Boos ing
Sys em,” in P oc. 22nd ACM SIGKDD In . Con . Knowledge
Disco e y and Da a Mining, pp. 785–794, 2016.
[7] H. A. Khan, “Add essing he Black-Box P oblem in Chu n
P edic ion: An In e p e abili y Analysis Using SHAP,” In e na ional
Jou nal o Compu e Applica ions, ol. 185, no. 3, pp. 1–8, 2022.
[8] Y. Sanjaya, “Cus ome Re en ion S a egy Op imiza ion Using SHAP
Analysis,” M.S. hesis, UIN Sya i Hidaya ullah, Jaka a, 2024.
[9] T. Lee and R. Singh, “Enhancing Telecom Chu n P edic ion wi h
Ensemble Lea ning: A Case S udy in he Malaysian Ma ke ,” In . J.
Da a Sci. Ad . Anal., ol. 8, no. 2, pp. 45–58, 2024.
[10] N. H. T an and A. T. Nguyen, “Applying Machine Lea ning o
P edic Mobile Telecom Cus ome Chu n in Vie nam,” Asian Jou nal
o Resea ch in Compu e Science, ol. 14, no. 3, pp. 31–45, 2023.
[11] M. T. Ribei o, S. Singh, and C. Gues in, “Why Should I T us You?:
Explaining he P edic ions o Any Classi ie ,” in P oc. 22nd ACM
SIGKDD Con ., pp. 1135–1144, 2016.
[12] S. M. Lundbe g and S.-I. Lee, “A Uni ied App oach o In e p e ing
Model P edic ions,” in Ad ances in Neu al In o ma ion P ocessing
Sys ems, pp. 4765–4774, 2017.
[13] A. Pu a, D. Su yan o, and M. Les a i, “In eg a ion o AI and
Visualiza ion Dashboa d o ISP Based on S eamli ,” in P oc. IEEE
In . Con . ICT, pp. 1–6, 2023.
[14] K. Schwabe and J. Su he land, “The Sc um Guide,” Sc um.o g,
2020.
[15] S. Susan o, T. P ase io, and R. D. Pu a, “Compa a i e S udy o
Sampling Techniques in Cus ome Chu n P edic ion,” IOP Con . Se .:
Ma e ials Science and Enginee ing, ol. 1077, 2021.
[16] R. A. P. Si ega and R. V. K. T. Si ega , “Cus ome Chu n P edic ion
Using XGBoos and SHAP,” In . J. Ad . Compu . Sci. Appl., ol. 14,
no. 6, 2023.
[17] R. M. F. Sa i e al., “Compa ison o Ensemble Lea ning Me hods o
Telecom Chu n P edic ion,” Ju nal Media In o ma ika Budida ma,
ol. 7, no. 1, pp. 180–190, 2023.
[18] D. S. R. R. P. B. M. Kuma , “Deep Lea ning-Based Chu n P edic ion
Using Fea u e Enginee ing and SMOTE,” Jou nal o Ambien
In elligence and Humanized Compu ing, ol. 13, pp. 6965–6978,
2022.
[19] T. J. Hossain, M. H. Islam, and T. M. Islam, “Chu n P edic ion in
Telecom Indus y Using Explainable Machine Lea ning,” in P oc.
IEEE ICECE, pp. 1–6, 2023.
[20] J. Gao e al., “Explainable Boos ing Machines o Telecom Chu n
P edic ion,” Expe Sys ems wi h Applica ions, ol. 224, 2025.
[21] L. Tang e al., “In eg a ing SHAP and XGBoos o Telecom Chu n
In e p e a ion,” IEEE Access, ol. 12, pp. 55678–55689, 2024.
[22] A. B. Id is and H. Rahman, “Cos -Sensi i e Lea ning o Imbalanced
Telecom Da a,” Applied In elligence, ol. 94, pp. 11023–11039,
2024.
[23] S. Meh a and R. Singh, “Explainable AI o Telecom Cus ome
Re en ion,” Compu e s & Indus ial Enginee ing, ol. 189, 2024.
[24] J. Zhao and L. Zhang, “E hical F amewo ks o AI-Based Decision
Sys ems,” AI and E hics, ol. 5, no. 2, pp. 377–392, 2024.
[25] M. T. Islam, H. R. Chowdhu y, and M. R. Hossain, “Hyb id
Ensemble and Deep Lea ning App oach o Telecom Cus ome Chu n
P edic ion,” Neu al Compu ing and Applica ions, ol. 36, pp. 2583–
2598, 2024.
[26] Y. Zhou e al., “Explainable Machine Lea ning o Cus ome
Re en ion in Telecom: A SHAP-Based E alua ion,” Knowledge-
Based Sys ems, ol. 293, 2024.
[27] L. K. Nguyen and M. H. T an, “Hyb id Cos -Sensi i e and Sampling
App oaches in Telecom Chu n Classi ica ion,” IEEE T ansac ions on
Compu a ional Social Sys ems, ol. 11, no. 3, pp. 451–463, 2024.
[28] M. Rahman, T. Nasi , and H. Chowdhu y, “In eg a ing XAI and Big
Da a Analy ics o Cus ome Loyal y in Telecom,”
Telecommunica ion Sys ems, ol. 82, pp. 165–180, 2023.
[29] R. Z. Han and Y. Feng, “Cus ome Segmen a ion and Re en ion
S a egy Op imiza ion in B oadband Se ices,” IEEE T ansac ions on
Enginee ing Managemen , ol. 71, no. 2, pp. 331–342, 2024.
[30] S. Ullah, H. M. Ja ed, and I. Khan, “ROI Analysis o Machine
Lea ning-Based Cus ome Re en ion Sys ems,” Decision Analy ics
Jou nal, ol. 7, 2024.
[31] P. R. Gup a and M. Shah, “E alua ing Cos -Bene i T ade-O s in AI-
D i en Cus ome Re en ion,” Expe Sys ems, ol. 41, no. 5, e13122,
2024.
[32] B. Li, F. Zhang, and W. Xu, “C oss-Regional Valida ion o Cus ome
Chu n Models in Telecom Indus ies,” Jou nal o Business Resea ch,
ol. 168, pp. 104–116, 2023.
[33] D. Z. Chen and Y. Luo, “Ad ances in Mul imodal P edic i e
Analy ics o Telecom Cus ome Beha io ,” IEEE T ansac ions on
Neu al Ne wo ks and Lea ning Sys ems, ol. 36, no. 9, pp. 1–12,
2025
