Copy igh © ISRG Publishe s. All igh s Rese ed.
DOI: 10.5281/zenodo.17330787
279
ISRG PUBLISHERS
Abb e ia ed Key Ti le: ISRG J A s Humani Soc Sci
ISSN: 2583-7672 (Online)
Jou nal homepage: h ps://is gpublishe s.com/is gjahss
Volume– III Issue -V (Sep embe -Oc obe ) 2025
F equency: Bimon hly
A Da a-Cen ic F amewo k o Tou ism and Hospi ali y Ma ke ing: In eg a ing
Business In elligence wi h Opinion Mining
Mohammad Akba i Asl1*, Mahshid Asadollahi2
1 Facul y o Tou ism S a egy, Cul u al He i age and Made in I aly, Depa men o His o y, Humani ies and Socie y,
To Ve ga a Uni e si y, Rome, I aly
2 Facul y o Business Adminis a ion, Depa men o Managemen and Law, To Ve ga a Uni e si y, Rome, I aly
| Recei ed: 05.10.2025 | Accep ed: 09.10.2025 | Published: 12.10.2025
*Co esponding au ho : Mohammad Akba i Asl
Facul y o Tou ism S a egy, Cul u al He i age and Made in I aly, Depa men o His o y, Humani ies and Socie y,
To Ve ga a Uni e si y, Rome, I aly
Abs ac
The inc easing eliance on digi al pla o ms has ans o med he ou ism and hospi ali y indus y in o a da a- ich en i onmen
whe e bo h s uc u ed business in elligence (BI) da a and uns uc u ed cus ome opinion da a low con inuously. Howe e , exis ing
s udies o en ea hese s eams in isola ion, limi ing hei abili y o p o ide a holis ic unde s anding o cus ome beha io . This
pape p oposes a da a-cen ic amewo k ha in eg a es BI wi h sen imen analysis o enhance ma ke ing in elligence in
hospi ali y and ou ism. The amewo k combines s uc u ed da a such as booking his o ies, demog aphics, and spending pa e ns
wi h uns uc u ed opinion da a d awn om cus ome e iews, social media, and a el blogs.
The me hodology inco po a es da a cleaning, sen imen sco ing, and hyb id ea u e selec ion, ollowed by supe ised lea ning
models including Suppo Vec o Machines (SVM), A i icial Neu al Ne wo ks (ANN), and Long Sho -Te m Memo y (LSTM)
ne wo ks. Recommenda ion sys ems a e u he enhanced by combining collabo a i e il e ing wi h sen imen -en iched con en -
based il e ing. The amewo k is e alua ed h ough benchma k da ase s and eal-wo ld hospi ali y e iews using a en- old c oss-
alida ion scheme.
The esul s demons a e ha he hyb id app oach signi ican ly imp o es sen imen classi ica ion accu acy (up o 92% wi h LSTM),
educes e o in ecommenda ion sys ems (RMSE = 0.72; MAE = 0.54), and yields measu able business bene i s. Simula ed
campaigns achie ed a 12% inc ease
in booking con e sion a es, an 8% ise in gues sa is ac ion, and a 10% imp o emen in
cus ome e en ion compa ed o BI-only baselines.
Copy igh © ISRG Publishe s. All igh s Rese ed.
DOI: 10.5281/zenodo.17330787
280
1. INTRODUCTION
O e he pas decade, ou ism and hospi ali y se ices ha e
e ol ed om supplemen a y economic sec o s in o dominan
o ces eshaping global ma ke s [1]. The widesp ead a ailabili y o
digi al booking pla o ms, mobile applica ions, and online a el
agencies has ans o med a ele beha io , enabling pe sonalized,
eal- ime in e ac ions be ween se ice p o ide s and gues s [2].
T adi ional ma ke ing s a egies, once hea ily elian on physical
a el agencies, wo d-o -mou h ecommenda ions, and
con en ional media, ha e been p og essi ely eplaced by digi al
hospi ali y ma ke ing p ac ices ha le e age da a-d i en insigh s o
each highly a ge ed audiences wi h g ea e e iciency [3].
In his digi al landscape, da a has eme ged as he mos c i ical
s a egic esou ce. Gues in o ma ion— anging om demog aphic
p o iles and booking his o ies o b owsing pa e ns and se ice
e iews— lows con inuously h ough hospi ali y pla o ms in he
o m o s uc u ed and uns uc u ed da a s eams. Ha nessing his
weal h o in o ma ion is no longe op ional bu undamen al o
ho els, eso s, ai lines, and des ina ion managemen o ganiza ions
s i ing o emain compe i i e. O ganiza ions ha success ully
ansla e aw da a in o ac ionable business in elligence gain a
measu able ad an age in p edic ing cus ome in en ions, ailo ing
ma ke ing campaigns, and op imizing esou ce alloca ion [4].
Despi e hese oppo uni ies, ou ism and hospi ali y ma ke ing
o en su e s om agmen ed insigh s. Gues in o ma ion is
ypically dispe sed ac oss mul iple sou ces, such as ese a ion
sys ems, loyal y p og am eco ds, web analy ics, and opinion- ich
ex ual eedback om gues e iews o social media pla o ms.
This dispe sion c ea es a signi ican ba ie o de eloping a
cohe en and holis ic unde s anding o gues needs and p e e ences
[5]. Mo eo e , many exis ing app oaches ea s uc u ed business
in elligence da a and uns uc u ed sen imen da a in isola ion,
which limi s he dep h and accu acy o he insigh s de i ed. The
lack o in eg a ed analysis hampe s he abili y o o ganiza ions o
deli e pe sonalized expe iences, an icipa e a ele beha io , and
design da a-cen ic ma ke ing s a egies ha a e adap i e and
in elligen [6,7].
While subs an ial p og ess has been made in applying da a
analy ics o he ou ism and hospi ali y sec o , mos p io esea ch
has app oached business in elligence (BI) and sen imen analysis as
sepa a e domains. BI s udies ypically emphasize s uc u ed da a,
such as booking eco ds, spending ca ego ies, and demog aphic
a ibu es, o suppo decision-making and cus ome segmen a ion.
In con as , sen imen analysis and opinion mining esea ch ocus
p ima ily on uns uc u ed da a, such as gues e iews, social media
in e ac ions, and ex ual eedback, o ex ac emo ional one and
pola i y. Al hough bo h app oaches yield aluable insigh s, hei
sepa a ion limi s he abili y o cap u e a comple e pic u e o gues
beha io and expe ience [8,9]. Only a ew s udies ha e a emp ed
o me ge BI wi h opinion mining, and e en ewe ha e p oposed
comp ehensi e amewo ks ha in eg a e he wo in a sys ema ic,
da a-cen ic manne . This gap unde sco es he need o a uni ied
pe spec i e ha combines he p edic i e s eng h o BI wi h he
con ex ual ichness o sen imen analysis [10].
This s udy aims o add ess he iden i ied gap by p oposing a da a-
cen ic amewo k ha me ges business in elligence and opinion
mining o e-comme ce ma ke ing. Speci ically, he esea ch
pu sues he ollowing objec i es:
1. To design and de elop an in eg a ed amewo k ha
uni ies s uc u ed and uns uc u ed da a s eams o
enhanced ma ke ing in elligence.
2. To imp o e cus ome in en ion p edic ion, clus e ing,
pe sonaliza ion, and ecommenda ion sys ems by
le e aging hyb id analy ical app oaches ha combine BI
me hods wi h sen imen analysis echniques.
3. To demons a e he po en ial o his in eg a ed app oach
in suppo ing sma , adap i e, and cus ome -cen ic
digi al ma ke ing s a egies wi hin he e-comme ce
domain.
The p esen s udy ad ances he ield o digi al ma ke ing by
making wo key con ibu ions. Fi s , i p oposes a no el in eg a ed
amewo k ha me ges business in elligence and opinion mining o
p o ide a holis ic, da a-cen ic ounda ion o sma e-comme ce
ma ke ing. Unlike p io app oaches ha analyze s uc u ed and
uns uc u ed da a s eams sepa a ely, his amewo k uni ies hem
in o a single decision-suppo model capable o cap u ing bo h
ansac ional pa e ns and cus ome sen imen s. Second, he s udy
demons a es how hyb id analy ical models—combining ule-
based me hods wi h machine lea ning and na u al language
p ocessing (NLP) echniques—can enhance ma ke ing ou comes.
By le e aging he p edic i e powe o BI alongside he con ex ual
ichness o sen imen analysis, he p oposed amewo k s eng hens
cus ome in en ion p edic ion, clus e ing, pe sonaliza ion, and
ecommenda ion sys ems. In doing so, i o e s bo h heo e ical and
p ac ical insigh s in o how e-comme ce i ms can op imize hei
ma ke ing s a egies in an inc easingly compe i i e digi al
en i onmen .
2. Li e a u e Re iew
2.1 A i icial In elligence and Neu al Ne wo ks: Gene al
Applica ions
A i icial in elligence (AI) and neu al ne wo k–based models ha e
become indispensable ac oss di e se scien i ic and enginee ing
domains, demons a ing hei abili y o p ocess la ge olumes o
The s udy o e s heo e ical con ibu ions by b idging beha io al and pe cep ual dimensions o cus ome in elligence, and p ac ical
implica ions by p o iding hospi ali y i ms wi h a scalable amewo k o eal- ime, cus ome -cen ic decision-making. Limi a ions
conce ning da ase ep esen a i eness, model uning, and eal- ime in eg a ion a e acknowledged, wi h ecommenda ions o
inco po a ing ad anced deep lea ning and mul imodal analysis in u u e esea ch. O e all, he indings unde sco e he po en ial o
in eg a ing BI wi h sen imen analysis as a pa hway owa d sma , adap i e ma ke ing sys ems in he ou ism and hospi ali y
sec o .
Keywo ds:
Business In elligence, Sen imen Analysis, Hyb id Fea u e Selec ion, Hospi ali y Ma ke ing, Cus ome In en ion
P edic ion, Recommenda ion Sys ems
Copy igh © ISRG Publishe s. All igh s Rese ed.
DOI: 10.5281/zenodo.17330787
281
he e ogeneous da a, unco e la en pa e ns, and suppo eal- ime
decision-making. In he ene gy sec o , o example, in elligen
con ol sys ems and op imiza ion algo i hms ha e been widely
adop ed o elec ic ehicle (EV) cha ging, ene gy alloca ion, and
sma g id managemen [11-14]. Simila ly, uzzy logic and ype-2
adap i e schemes ha e been success ully employed in u ban a ic
p edic ion and sma ci y managemen [15].
In he medical domain, deep lea ning a chi ec u es and explainable
AI echniques ha e shown ema kable p og ess in disease de ec ion
and classi ica ion, including b ain umo diagnosis, au ism
spec um diso de ecogni ion, and medical image segmen a ion
[16-18]. These s udies unde sco e he abili y o AI no only o
achie e high accu acy bu also o p o ide in e p e abili y and ule-
based easoning, enhancing us in c i ical decision-making
en i onmen s.
The e sa ili y o AI ex ends o geological and en i onmen al
modeling, whe e explainable, na u e-inspi ed op imiza ion
echniques and mul i-a ibu e machine lea ning ha e been applied
o haza d p edic ion, ea hquake ea ly wa ning sys ems, and
me eo ological o ecas ing [19-21]. In mechanical and s uc u al
enginee ing, AI-enabled aul diagnosis and op imiza ion
app oaches ha e imp o ed sys em pe o mance in ai c a
suspension, o a ing machine y, and heal h moni o ing applica ions
[22-25]. Addi ionally, ecen wo k in u ban analy ics has
demons a ed how AI can e eal complex ela ionships be ween
s ee ne wo k con igu a ion and social ou comes such as p ope y
c ime [26], while se ice-o ien ed AI amewo ks ha e con ibu ed
o ad ancemen s in da a secu i y and in o ma ion managemen
[27].
Taken oge he , hese di e se applica ions highligh AI’s c oss-
disciplina y impac , es ablishing i s ole as a ans o ma i e
echnology o p edic i e analy ics, op imiza ion, and in elligen
decision-making. Building on his ounda ion, he p esen s udy
si ua es AI in he con ex o business in elligence and sen imen
analysis wi hin he ou ism and hospi ali y indus y, ex ending he
discussion om gene al applica ions o domain-speci ic
challenges.
2.2 E olu ion o Digi al Ma ke ing in he E-Comme ce
E a
The ise o digi al pla o ms in he ou ism and hospi ali y sec o
has undamen ally ans o med he ma ke ing landscape, shi ing
emphasis om o line, ace- o- ace in e ac ions wi h a el agen s
and ho el s a o online, da a-d i en engagemen s h ough booking
sys ems and mobile applica ions [2,3]. T adi ional ma ke ing
channels such as p in b ochu es, physical a el agencies, and
con en ional media ha e been supplemen ed—and in many cases
eplaced—by digi al hospi ali y pla o ms ha o e b oade each,
pe sonaliza ion, and eal- ime esponsi eness. In his new
pa adigm, ho els, ai lines, and des ina ion ma ke e s inc easingly
ely on ad anced analy ics o decode gues beha io and o ailo
s a egies acco dingly [28].
A de ining ea u e o digi al hospi ali y ma ke ing is i s eliance on
big da a, which lows con inuously om gues in e ac ions ac oss
booking po als, loyal y p og ams, mobile applica ions, and social
media pla o ms. This da a encompasses bo h ansac ional de ails,
such as ese a ion his o ies and spending le els, and beha io al
indica o s, such as b owsing pa e ns, eedback, and se ice
p e e ences [29]. Wi hin his con ex , cus ome ela ionship
managemen (CRM) has been eshaped om s a ic eco d-keeping
in o dynamic, p edic i e, and adap i e sys ems. Mode n hospi ali y
CRM sys ems now in eg a e beha io al insigh s wi h ansac ional
eco ds o s eng hen long- e m gues loyal y, enhance se ice
pe sonaliza ion, and d i e e enue g ow h. The a ailabili y o such
as and he e ogeneous da a esou ces unde sco es he s a egic
impe a i e o hospi ali y o ganiza ions o adop in elligen , da a-
cen ic app oaches in digi al ma ke ing [30].
2.3 Business In elligence (BI) in E-Comme ce
Business in elligence (BI) has eme ged as a c i ical enable o
compe i i e ad an age in he ou ism and hospi ali y sec o [8]. By
le e aging s uc u ed da a om booking eco ds, loyal y p og am
ac i i y, and gues p o iles, BI sys ems allow o ganiza ions o
ex ac ac ionable insigh s ha in o m ma ke ing, se ice design,
and ope a ional decision-making. One o he mos signi ican
applica ions o BI in his domain is gues segmen a ion and
clus e ing, whe e analy ical echniques a e used o g oup a ele s
based on booking beha io s, demog aphics, a el equency, o
spending le els. Such segmen a ion p o ides a ounda ion o
a ge ed campaigns, loyal y p og ams, and esou ce alloca ion
[6,7].
Ano he widely adop ed BI applica ion in hospi ali y is se ice
baske analysis, which iden i ies associa ions among se ices
equen ly booked oge he , such as ho el s ays wi h spa packages,
o ligh s wi h ai po ans e s. This echnique no only in o ms
c oss-selling and upselling s a egies bu also enhances
ecommenda ion sys ems by p edic ing which se ices a e likely o
be selec ed in combina ion [31]. Addi ionally, BI enables
p edic i e analy ics o gues in en ion, equipping i ms wi h he
abili y o o ecas demand o des ina ions, iden i y isks o
cus ome a i ion, and op imize esou ce alloca ion ac oss peak
and o -peak seasons [32]. The p ac ical implemen a ion o BI in
ou ism and hospi ali y is suppo ed by a a ie y o ools and
echnologies, including online analy ical p ocessing (OLAP)
sys ems, in e ac i e dashboa ds, and in elligen ecommenda ion
engines. These ools ans o m aw booking and se ice da a in o
in ui i e isualiza ions and p edic i e models ha empowe
decision-make s o ac wi h g ea e accu acy and agili y. Despi e
i s s eng hs, howe e , BI adi ionally ocuses on s uc u ed
da ase s, lea ing a gap in he exploi a ion o uns uc u ed da a such
as gues e iews, social media pos s, and ex ual eedback—an a ea
whe e sen imen analysis and opinion mining play an inc easingly
i al ole in unde s anding gues expe iences and pe cep ions [33].
2.4 Sen imen Analysis and Opinion Mining
While business in elligence excels in analyzing s uc u ed da a, a
subs an ial po ion o gues insigh s in ou ism and hospi ali y is
embedded in uns uc u ed ex ual in o ma ion such as ho el
e iews, se ice a ings, a el blogs, and social media in e ac ions.
To cap u e he ichness o his in o ma ion, esea che s and
p ac i ione s inc easingly ely on sen imen analysis and opinion
mining, which seek o ex ac subjec i e opinions, emo ional one,
and pola i y om na u al language [9,10]. A i s co e, sen imen
analysis employs na u al language p ocessing (NLP), ex mining,
and compu a ional linguis ics o classi y gues opinions as posi i e,
nega i e, o neu al. Mo e ad anced models mo e beyond pola i y
classi ica ion o cap u e nuanced emo ions and con ex -dependen
meanings, such as sa is ac ion wi h s a iendliness,
disappoin men wi h check-in delays, o en husiasm o local
expe iences. Opinion mining hus complemen s BI by adding
quali a i e dimensions o cus ome in elligence, e ealing no only
Copy igh © ISRG Publishe s. All igh s Rese ed.
DOI: 10.5281/zenodo.17330787
282
wha a ele s do bu also how hey eel and why hey beha e in
ce ain ways [9,34].
The me hods applied in sen imen analysis can be b oadly g ouped
in o h ee ca ego ies. The lexicon-based app oach elies on
p ede ined dic iona ies o opinion wo ds and hei associa ed
sen imen sco es. While in e p e able, i is o en limi ed in handling
domain-speci ic language o con ex ual a ia ions (e.g., ―all-
inclusi e,‖ ―las -minu e deal,‖ o ―hidden ees‖ in a el con ex s).
The machine lea ning app oach ains classi ie s—such as Suppo
Vec o Machines (SVM), Random Fo es s, o Logis ic
Reg ession—on labeled hospi ali y da ase s o au oma ically
ecognize sen imen pa e ns. Mo e ecen ly, deep lea ning
app oaches, including Con olu ional Neu al Ne wo ks (CNNs),
Long Sho -Te m Memo y (LSTM) ne wo ks, and ans o me -
based models such as BERT, ha e demons a ed supe io
pe o mance by cap u ing seman ic ela ionships and con ex ual
dependencies wi hin gues e iews and a el na a i es [35,36]. In
he ou ism and hospi ali y domain, sen imen analysis has p o en
pa icula ly aluable in se ice ecommenda ion sys ems,
des ina ion image moni o ing, and gues eedback analysis. By
quan i ying subjec i e expe iences, se ice p o ide s can align
ma ke ing s a egies wi h a ele expec a ions, de ec ea ly
wa ning signals o dissa is ac ion, and enhance pe sonaliza ion in
ho el and des ina ion o e ings. Ne e heless, despi e hese
ad ances, sen imen analysis alone emains insu icien o a ull-
spec um unde s anding o gues beha io unless in eg a ed wi h
s uc u ed BI insigh s. This c ea es he ounda ion o a holis ic
amewo k ha me ges he s eng hs o bo h domains [9,35,36].
2.5 In eg a ion o BI and Sen imen Analysis
The in eg a ion o business in elligence (BI) and sen imen analysis
ep esen s a logical ye unde explo ed on ie in ou ism and
hospi ali y esea ch. On hei own, BI sys ems p o ide s uc u ed,
quan i a i e insigh s in o gues beha io , while sen imen analysis
o e s a quali a i e unde s anding o a i udes, emo ions, and
pe cep ions. When combined, hese wo s eams o knowledge
ha e he po en ial o p oduce a comp ehensi e gues in elligence
amewo k ha can signi ican ly enhance ma ke ing and se ice
decision-making [9,36]. A g owing body o esea ch has a emp ed
o b idge hese domains. Fo example, s udies ha e combined
booking his o y da a wi h gues e iews o imp o e ho el o
des ina ion ecommenda ion sys ems o ha e in eg a ed ese a ion
pa e ns wi h social media sen imen o e ine demand o ecas ing
o seasonal a el. Such hyb id app oaches demons a e ha
me ging BI and opinion mining can yield iche , mo e ac ionable
insigh s han ei he me hod alone. In p ac ice, his in eg a ion
enables hospi ali y p o ide s o answe no only wha gues s a e
doing bu also why hey a e beha ing in pa icula ways. The
bene i s o his in eg a ion a e mul i-dimensional. On he s a egic
side, i suppo s gues -cen ic decision-making, imp o ing
segmen a ion, campaign a ge ing, and des ina ion posi ioning. On
he ope a ional side, i s eng hens pe sonaliza ion and
ecommenda ion sys ems, enabling adap i e ma ke ing esponses
based on bo h beha io al ends and emo ional eedback.
Fu he mo e, in eg a ion enhances p edic i e accu acy by
combining s uc u ed indica o s (such as booking equency o
expendi u e le el) wi h uns uc u ed ea u es (such as sen imen
pola i y o opinion s eng h) [37].
Despi e i s p omise, he in eg a ion o BI and sen imen analysis in
hospi ali y emains limi ed in scope. Many exis ing s udies a e
domain-speci ic, ocusing na owly on online ho el e iews o
single se ice componen s, while o e looking b oade con ex s
such as CRM, c oss-se ice bundling, and gues jou ney analysis.
O he s ace challenges ela ed o da a he e ogenei y, scalabili y,
and in e p e abili y, pa icula ly when inco po a ing ad anced
machine lea ning o deep lea ning me hods. This unde sco es he
need o a holis ic, da a-cen ic amewo k ha sys ema ically
uni ies s uc u ed and uns uc u ed da a s eams o in elligen
ou ism and hospi ali y ma ke ing—an a ea whe e his s udy aims
o make a dis inc con ibu ion [30].
2.6 P epa ing and No malizing Tex ual Da a
Tex p ep ocessing ep esen s a ounda ional s age in he opinion
mining pipeline, as he quali y o subsequen analysis depends
hea ily on he cla i y and s uc u e o he inpu da a. Raw ex ual
da a, such as gues e iews, a el blogs, o social media pos s,
o en con ain noise in he o m o misspellings, slang, i ele an
okens, and edundan in o ma ion. To add ess his, p ep ocessing
begins wi h da a cleaning, ollowed by a se ies o linguis ic
ope a ions designed o no malize and s anda dize he ex .
Common p ocedu es include okeniza ion, Pa -o -Speech (POS)
agging, s emming and lemma iza ion, and he emo al o s op
wo ds. Beyond gene ic app oaches, domain-speci ic p ep ocessing
is inc easingly ele an in ou ism and hospi ali y applica ions. Fo
ins ance, on ology-based p ep ocessing can d aw on esou ces such
as Wo dNe o hospi ali y- ocused lexicons o cap u e he
con ex ual meaning o e ms ela ed o se ice quali y, ameni ies,
o des ina ion a ibu es (e.g., ―check-in,‖ ―all-inclusi e,‖ ―scenic
iew‖). Fu he mo e, nume ous open-sou ce ools acili a e
p ep ocessing ac oss languages, such as NLTK o English and
Jieba o Chinese, which a e especially use ul gi en he
in e na ional and mul ilingual na u e o gues eedback. By
ensu ing ha aw ex ual inpu is ans o med in o a consis en and
analyzable o ma , p ep ocessing p o ides he essen ial
g oundwo k o eliable sen imen analysis and ea u e enginee ing
in hospi ali y in elligence amewo ks [38,39].
2.7 T ans o ming Tex in o Analy ical Fea u es
Following p ep ocessing, he nex s ep in ol es ea u e
manipula ion, which ans o ms cleaned a ele eedback in o
s uc u ed ep esen a ions sui able o compu a ional modeling.
This p ocess is ypically di ided in o ea u e ex ac ion and ea u e
selec ion. Fea u e ex ac ion aims o de i e in o ma i e
ep esen a ions o ex , o en by p ojec ing da a in o a lowe -
dimensional space. Widely used me hods include Te m
F equency–In e se Documen F equency (TF-IDF), mu ual
in o ma ion, en opy-based measu es, and he Gini Index [40].
These app oaches highligh e ms o combina ions o e ms ha
bes cap u e seman ic dis inc ions wi hin a co pus o a ele
e iews, such as iden i ying how e ms like “scenic iew,”
“cul u al au hen ici y,” “ anspo a ion con enience,” o “sa e y
condi ions” di e en ia e be ween posi i e and nega i e
expe iences a des ina ions.
Fea u e selec ion, in con as , seeks o iden i y a subse o he mos
ele an ea u es, he eby educing compu a ional o e head and
mi iga ing noise. Selec ion me hods a e ypically ca ego ized in o
il e , w appe , and hyb id app oaches. Fil e me hods—such as
Chi-Squa e, In o ma ion Gain, and co ela ion-based measu es—
ope a e independen ly o lea ning algo i hms and a e alued o
hei scalabili y ac oss la ge and mul ilingual ou ism da ase s.
W appe models, hough o en mo e accu a e, a e compu a ionally
in ensi e as hey e alua e subse s o ea u es wi h espec o a
gi en classi ie , o example, de e mining which aspec s o e iews
Copy igh © ISRG Publishe s. All igh s Rese ed.
DOI: 10.5281/zenodo.17330787
283
(e.g., “local cuisine,” “guided ou quali y,” o “public anspo
eliabili y”) mos s ongly in luence isi o sa is ac ion
p edic ions. Hyb id me hods a emp o combine he e iciency o
il e s wi h he accu acy o w appe s, making hem pa icula ly
well-sui ed o he di e se and con ex - ich na u e o ou ism
eedback da a [41].
In he con ex o ou ism opinion mining, e ec i e ea u e
manipula ion ensu es ha sen imen classi ie s can achie e high
p ecision in dis inguishing posi i e and nega i e isi o
expe iences, wi hou being o e whelmed by i ele an o
edundan a ibu es. This s ep p o ides a e ined ea u e space ha
s eng hens downs eam asks such as des ina ion image analysis,
a ele in en ion p edic ion, pe sonalized ip ecommenda ions,
and ou ism policy e alua ion [34].
2.8 Classi ica ion o Tex h ough Supe ised Lea ning
Once ex ual ea u es a e ex ac ed and selec ed, hey can be
employed wi hin supe ised machine lea ning amewo ks o
pe o m classi ica ion asks. Supe ised lea ning in ol es aining
algo i hms on labeled da ase s o lea n pa e ns ha can la e be
applied o unseen ins ances. In ou ism, his ypically means
ca ego izing a ele e iews o des ina ions, a ac ions, o ou
se ices as posi i e, nega i e, o neu al, o assigning hem o mo e
ine-g ained sen imen ca ego ies such as cul u al expe ience,
anspo a ion con enience, sa e y, hospi ali y o locals, o alue
o money.
Supe ised classi ica ion me hods span a b oad ange o
echniques, including decision ees, ule-based and case-based
app oaches, linea classi ie s, and p obabilis ic models such as
Naï e Bayes. Mo e ad anced app oaches le e age Suppo Vec o
Machines (SVMs), A i icial Neu al Ne wo ks (ANNs), and o he
ensemble me hods. Recen de elopmen s in deep lea ning
a chi ec u es, such as Con olu ional Neu al Ne wo ks (CNNs) and
Long Sho -Te m Memo y (LSTM) ne wo ks, ha e u he
enhanced classi ica ion accu acy by cap u ing con ex ual
dependencies ac oss a el na a i es— o example, dis inguishing
a neu al desc ip ion o wea he om a complain abou
anspo a ion delays o sa e y conce ns a a ou is si e [19].
A pe sis en challenge wi h supe ised models, howe e , is he isk
o o e i ing, whe eby a classi ie lea ns pa e ns oo speci ic o
he aining da a and pe o ms poo ly on unseen e iews om
di e en des ina ions o cul u al con ex s. Add essing his equi es
s a egies such as c oss- alida ion, egula iza ion, and he use o
di e se mul ilingual da ase s ha e lec he global and
he e ogeneous na u e o ou ism. Despi e hese challenges,
supe ised ex classi ica ion emains a co ne s one o opinion
mining in ou ism, enabling a el agencies, des ina ion ma ke ing
o ganiza ions (DMOs), and ou ism pla o ms o au oma ically
ca ego ize and in e p e as olumes o a ele -gene a ed con en .
This, in u n, p o ides ac ionable insigh s o des ina ion b anding,
a ge ed p omo ional campaigns, demand o ecas ing, and
imp o ing isi o expe iences [42].
2.9 Resea ch Gap and Theo e ical Founda ion
The ex ended e iew o p io s udies shows ha , al hough business
in elligence (BI) and sen imen analysis ha e bo h been widely
adop ed in ou ism and hospi ali y esea ch, hey a e o en
examined in isola ion. BI esea ch adi ionally emphasizes
s uc u ed da ase s—such as booking eco ds, demog aphics, a el
equency, and expendi u e a ibu es— o suppo decision-making
h ough clus e ing, p edic i e analy ics, and se ice bundling
analysis. In con as , sen imen analysis and opinion mining ocus
on uns uc u ed ex ual sou ces, ex ac ing pola i y and emo ional
one om a ele e iews, social media da a, and a el blogs
h ough p ep ocessing, ea u e enginee ing, and classi ica ion
echniques [8]. These wo pe spec i es, while aluable
indi idually, ail o p o ide a holis ic iew o gues and a ele
in elligence when kep sepa a e. E o s o in eg a e BI wi h
sen imen analysis exis bu emain limi ed in scope. Some s udies
link booking his o ies wi h e iew sen imen o imp o e ho el o
des ina ion ecommenda ions, while o he s use a ele opinions o
en ich demand o ecas ing o loyal y p edic ions. Howe e , hese
a emp s a e o en na ow, domain-speci ic, and echnically
cons ained, acing issues o scalabili y, in e p e abili y, and
adap abili y ac oss di e se ou ism en i onmen s. Mo eo e , much
o he exis ing opinion mining esea ch concen a es hea ily on
algo i hmic asks such as p ep ocessing, ea u e ex ac ion, and
supe ised classi ica ion, ye a ely connec s hese ou pu s back
in o b oade BI-d i en decision-making sys ems [43]. This lea es a
c i ical me hodological gap in es ablishing a uni ied, da a-cen ic
amewo k ha sys ema ically me ges s uc u ed BI wi h
uns uc u ed sen imen in elligence o ou ism ma ke ing
op imiza ion.
The heo e ical ounda ion o his s udy es s on he p inciple ha
e ec i e a ele in elligence equi es cap u ing bo h he
quan i a i e dimensions o beha io (e.g., booking his o ies,
spending le els, a el equency) and he quali a i e dimensions
o sen imen and pe cep ion (e.g., opinions, emo ions, a i udes).
By aligning BI wi h sen imen analysis h ough an in eg a ed
amewo k, his s udy seeks o ad ance bo h heo y and p ac ice.
Concep ually, he p oposed model ex ends exis ing knowledge by
demons a ing how s uc u ed and uns uc u ed da a s eams can
complemen each o he o s eng hen a ele in en ion p edic ion,
des ina ion clus e ing, se ice pe sonaliza ion, and
ecommenda ion sys ems. P ac ically, i p o ides ou ism and
hospi ali y o ganiza ions wi h a oadmap o build adap i e, gues -
cen e ed ma ke ing s a egies ha a e bo h da a-d i en and con ex -
awa e.
3. Me hodology
3.1 Backg ound and F amewo k Design
The me hodology o his s udy is designed o implemen a da a-
cen ic amewo k ha in eg a es business in elligence (BI) wi h
opinion mining in he con ex o ou ism and hospi ali y ma ke ing.
D awing on insigh s om he li e a u e, he amewo k is
s uc u ed as a sequen ial wo k low ha p ocesses and in eg a es
bo h s uc u ed and uns uc u ed da a s eams. The s ages include
da a collec ion, da a cleaning, sen imen sco ing, ea u e
enginee ing wi h hyb id ea u e selec ion, model aining, and
pe o mance e alua ion. Each s age add esses he complexi y o
he e ogeneous cus ome da a and ensu es ha he inal insigh s a e
ac ionable o indus y-speci ic applica ions. S uc u ed da a, such
as booking his o ies, loyal y p og am eco ds, and demog aphic
a ibu es, p o ides quan i a i e insigh s in o cus ome beha io .
Uns uc u ed da a, such as ho el e iews, a el blog pos s, and
social media commen s, cap u es quali a i e pe cep ions o se ice
quali y and cus ome expe ience. Toge he , hese sou ces c ea e a
comp ehensi e ounda ion o p edic i e and ecommenda ion
models ailo ed o ou ism and hospi ali y ma ke ing.
3.2 Da a Collec ion and Sou ces
The da a o his s udy is ob ained om a combina ion o
benchma k eposi o ies and eal-wo ld ou ism and hospi ali y
Copy igh © ISRG Publishe s. All igh s Rese ed.
DOI: 10.5281/zenodo.17330787
284
pla o ms. S uc u ed BI da a is d awn om ansac ional eco ds
such as ho el bookings, a el package pu chases, loyal y p og am
usage, and demog aphic in o ma ion, using es ablished eposi o ies
and indus y da ase s whe e a ailable. Uns uc u ed opinion da a is
collec ed om la ge-scale e iew da ase s such as T ipAd iso ,
Booking.com, and Google Re iews, and is u he supplemen ed
by c awled ex om a el blogs, online o ums, and social media
pla o ms, subjec o e hical guidelines and access policies. Web
c awling echniques a e applied whe e app op ia e, adhe ing o
poli eness, e isi a ion, and obus ness policies o ensu e eliable
acquisi ion o cus ome -gene a ed con en . The da a sou ces e lec
he wo-laye ocus o he p oposed amewo k. S uc u ed da a
de i ed om business in elligence sys ems p o ides a quan i a i e
iew o consume beha io , including demog aphics, booking
his o ies, and spending pa e ns. In pa allel, uns uc u ed da a
ob ained h ough opinion mining cap u es he quali a i e side o
cus ome expe ience h ough e iews, ex ual eedback, and social
media discussions. By combining hese wo complemen a y
pe spec i es, he amewo k acili a es he de elopmen o a hyb id
cus ome in elligence model capable o cap u ing bo h
ansac ional beha io s and subjec i e sen imen s in an in eg a ed
manne [44].
3.3 Da a Cleaning and Noise Reduc ion
Da a collec ed om cus ome in e ac ions in he ou ism and
hospi ali y domain o en con ains subs an ial noise, including
ypog aphical e o s, abb e ia ions, edundan symbols, and
i ele an con en . To add ess his, uns uc u ed ex unde goes
p ep ocessing s eps such as okeniza ion, s op-wo d emo al,
no maliza ion, nega ion handling, s emming, and lemma iza ion.
Pa -o -Speech (POS) agging and dependency pa sing a e applied
o iden i y sen imen - ich exp essions, while domain-speci ic
lexicons and on ologies ela ed o hospi ali y se ices (e.g., ―check-
in,‖ ―ameni ies,‖ ―s a beha io ,‖ ―cleanliness‖) help cap u e
con ex ual nuances. Fo s uc u ed BI da a, cleaning in ol es
handling missing demog aphic a ibu es, elimina ing duplica e
booking eco ds, and no malizing ca ego ical alues such as
booking channels and accommoda ion ypes. These combined
p ocesses ensu e ha bo h s uc u ed and uns uc u ed da ase s a e
s anda dized and sui able o in eg a ion [45].
3.4 Sen imen Sco ing and Fea u e Enginee ing
Following noise educ ion, sen imen analysis is pe o med o
assign pola i y and in ensi y o cus ome -gene a ed e iews. A
lexicon-based app oach p o ides baseline sen imen mapping,
while n-g am modeling wi h Te m F equency–In e se Documen
F equency (TF-IDF) weigh ing cap u es con ex ual exp essions in
complex e iews. Fea u e enginee ing is conduc ed in wo s ages:
ea u e ex ac ion and ea u e selec ion. Ex ac ion me hods
gene a e nume ical ep esen a ions o ex using TF-IDF, mu ual
in o ma ion, and en opy-based measu es. Fea u e selec ion
educes dimensionali y and emo es i ele an a ibu es. A hyb id
selec ion s a egy, combining il e -based echniques such as
In o ma ion Gain and Chi-Squa e wi h w appe app oaches ied o
machine lea ning classi ie s, is employed o enhance accu acy
while main aining compu a ional e iciency. The sen imen -de i ed
ea u es a e hen me ged wi h BI ea u es such as booking
equency, expendi u e le els, and cus ome p o iles, esul ing in a
comp ehensi e da ase ha combines beha io al and pe cep ual
insigh s and enables iche model aining and p edic ion [40].
3.5 Fea u e Vec o Cons uc ion wi h Hyb id Fea u e
Selec ion
Once sen imen sco es a e gene a ed, he nex s ep in ol es
ans o ming hese ou pu s in o s uc u ed ea u e ec o s ha can
be used o p edic i e modeling. To ensu e bo h e iciency and
in e p e abili y, his s udy adop s a hyb id ea u e selec ion
app oach ha in eg a es ule-based sen imen analysis wi h
machine lea ning–based selec ion echniques. The objec i e is o
no malize he ex ac ed sen imen ea u es, educe edundancy, and
e ain only hose a ibu es ha a e highly ele an o cus ome
expe ience in he ou ism and hospi ali y con ex [46]. In he ule-
based s age, opinion exp essions a e iden i ied using Pa -o -
Speech (POS) agging, dic iona y-based lexicons, and domain-
speci ic on ologies ela ed o hospi ali y se ices. This enables he
model o cap u e meaning ul e ms and ph ases—such as
e e ences o ―check-in p ocess,‖ ― oom cleanliness,‖ o ―s a
iendliness‖— ha a e di ec ly ied o gues sa is ac ion. These
ea u es a e hen passed in o he machine lea ning–based s age o
e inemen [45]. As illus a ed in Figu e 1, he hyb id ea u e
selec ion amewo k combines ule-based sen imen analysis wi h
machine lea ning–based il e ing echniques o cons uc a
balanced ea u e space o modeling. Machine lea ning–based
ea u e selec ion is ca ied ou using bo h il e and app oaches.
Fil e me hods a e pa icula ly aluable in la ge-scale hospi ali y
da ase s because hey ope a e independen ly o any speci ic
lea ning algo i hm, educing compu a ional o e head while
main aining gene alizabili y. In his esea ch, a co ela ion-based
ea u e selec ion scheme is applied o iden i y ea u es ha a e
s ongly co ela ed wi h sen imen pola i y (posi i e, nega i e, o
neu al) bu weakly co ela ed wi h each o he . This helps
o e come he p oblem o ea u e bias, ensu ing ha he esul ing
subse is bo h ep esen a i e and non- edundan . W appe me hods,
while mo e compu a ionally in ensi e, a e used selec i ely o
alida e he p edic i e con ibu ion o he e ained ea u es agains
classi ica ion models [45].
The combina ion o ule-based and machine lea ning–based
selec ion yields a balanced ea u e space ha in eg a es he
linguis ic ichness o ex ual opinion wi h he p edic i e s eng h o
s uc u ed analy ics. This hyb id s a egy p o ides a obus
ounda ion o downs eam asks such as sen imen classi ica ion,
cus ome clus e ing, and pe sonalized ecommenda ions in ou ism
and hospi ali y ma ke ing.
Figu e 1. Hyb id ea u e selec ion amewo k o ou ism and
hospi ali y sen imen analysis.
3.6 Model T aining and Classi ica ion
The in eg a ed da ase is used o ain in elligen lea ning models
ha pe o m sen imen classi ica ion, cus ome segmen a ion, and
Copy igh © ISRG Publishe s. All igh s Rese ed.
DOI: 10.5281/zenodo.17330787
285
ecommenda ions o ou ism and hospi ali y se ices. The aining
p ocess ollows a supe ised lea ning pa adigm, in which his o ical
da a is used o ain he models and unseen da a is ese ed o
es ing. Fo baseline classi ica ion asks, algo i hms such as
Suppo Vec o Machines (SVM), Decision T ees, and Naï e
Bayes a e employed o ca ego ize cus ome e iews in o sen imen
classes (posi i e, nega i e, o neu al).
To cap u e he sequen ial and con ex ual dependencies p esen in
cus ome na a i es—such as de ailed ho el e iews o a el
expe iences—ad anced deep lea ning app oaches a e also
implemen ed. In pa icula , A i icial Neu al Ne wo ks (ANNs)
and Long Sho -Te m Memo y (LSTM) ne wo ks a e used. ANNs
consis o inpu , hidden, and ou pu laye s connec ed by weigh ed
neu ons, and hey employ ac i a ion unc ions o in oduce non-
linea i y and imp o e exp essi eness. Common ac i a ion
unc ions conside ed in his s udy include he sigmoid and ReLU
unc ions, as hey a e widely used o ex -based classi ica ion
asks. LSTM ne wo ks u he enhance he amewo k’s abili y o
model empo al dependencies and long- e m con ex ual
ela ionships wi hin cus ome eedback [46].
Fo ecommenda ion asks, hyb id sys ems a e de eloped ha
in eg a e collabo a i e il e ing wi h sen imen -en iched con en -
based il e ing. This dual app oach enables he sys em o p o ide
pe sonalized sugges ions o ho els, des ina ions, o a el
packages, combining s uc u ed BI da a (e.g., booking equency,
expendi u e pa e ns) wi h uns uc u ed sen imen insigh s (e.g.,
cus ome sa is ac ion le els, emo ional one). In his way,
s uc u ed da a enhances p edic i e accu acy, while uns uc u ed
sen imen da a adds con ex ual dep h, esul ing in models ha
suppo mo e cus ome -cen e ed ma ke ing s a egies in ou ism
and hospi ali y.
3.7 Model E alua ion
The p oposed in elligen models a e e alua ed h ough a
combina ion o s a is ical pe o mance me ics and indus y-
speci ic business indica o s. To ensu e obus ness, a en- old c oss-
alida ion scheme is applied, in which he da ase is pa i ioned
in o en subse s; nine subse s a e used o aining while he
emaining subse is used o es ing, and his p ocess is epea ed
i e a i ely. This app oach educes he isk o o e i ing and
ensu es ha he esul s a e gene alizable ac oss di e en da a
samples.
Fo classi ica ion asks, e alua ion is based on accu acy, p ecision,
ecall, and F1-sco e. Accu acy measu es he p opo ion o co ec ly
classi ied obse a ions ela i e o he o al, while p ecision
e alua es he p opo ion o ue posi i e classi ica ions ou o all
p edic ed posi i es. Recall quan i ies he p opo ion o ue
posi i es iden i ied among all ac ual posi i es, and he F1-sco e
balances p ecision and ecall, p o iding a obus indica o o
imbalanced da ase s. Ma hema ically, accu acy and p ecision a e
de ined as:
Whe e TP e e s o ue posi i es, TN o ue nega i es, FP o alse
posi i es, and FN o alse nega i es.
Fo ecommenda ion models, p edic i e accu acy is assessed using
Roo Mean Squa e E o (RMSE) and Mean Absolu e E o
(MAE), which e alua e he di e ence be ween p edic ed and ac ual
a ings o p e e ences. Beyond compu a ional measu es, he
e alua ion also inco po a es ou ism- and hospi ali y-speci ic KPIs,
including booking con e sion a es, gues sa is ac ion indices, and
cus ome e en ion le els. This dual e alua ion amewo k ensu es
ha he models a e no only s a is ically obus bu also aligned
wi h he ope a ional and s a egic goals o he ou ism and
hospi ali y indus y [47].
3.8 P oposed Sen imen -Enhanced Business In elligence
F amewo k
The p oposed amewo k in eg a es sen imen analysis wi h
business in elligence o c ea e a comp ehensi e decision-suppo
model o he ou ism and hospi ali y indus y (Figu e 2). Cus ome
da a om mul iple hospi ali y channels—including ho el booking
sys ems, online a el agencies, loyal y p og ams, and e iew
pla o ms—a e colla ed and s o ed in a s uc u ed da a eposi o y.
Because his aw da a o en con ains inconsis encies and i ele an
in o ma ion, i is i s passed h ough a cleaning and noise
elimina ion p ocess.
In he nex s age, sen imen sco ing is applied o uns uc u ed
ex ual da a such as gues e iews, blog pos s, and social media
commen s. Each sen imen is quan i ied o p oduce a pola i y
sco e, which is hen p ocessed h ough hyb id ea u e selec ion.
This s ep e ines he sen imen ea u es by combining ule-based
linguis ic il e s wi h machine lea ning–based selec ion, ensu ing
ha only he mos ele an a ibu es a e e ained. The esul ing
ea u e ec o s a e in eg a ed wi h s uc u ed BI ea u es such as
booking equency, spending ca ego ies, and cus ome
demog aphics. These hyb id ec o s a e subsequen ly used o ain
in elligen lea ning models capable o pe o ming pola i y
classi ica ion and p edic i e analy ics.
Figu e 3 demons a es how he ou comes o pola i y analysis a e
ope a ionalized in hospi ali y ma ke ing applica ions. Cus ome s
exp essing consis en ly posi i e sen imen owa d ho el s ays o
a el se ices a e segmen ed in o p io i y a ge g oups o
pe sonalized ma ke ing campaigns. Pola i y ou comes also eed
in o p edic ion models ha es ima e cus ome in en ion, allowing
businesses o an icipa e whe he new gues s a e likely o exp ess
sa is ac ion o dissa is ac ion. Fu he mo e, he amewo k enables
he de elopmen o ecommenda ion sys ems ha au oma ically
sugges ho els, des ina ions, o se ice bundles based on bo h
ansac ional beha io and sen imen ends. Such sys ems can also
suppo s a egic unc ions like ailo ing p omo ional packages,
managing seasonal demand, and aligning se ice design wi h
cus ome expec a ions.
By combining s uc u ed business in elligence wi h sen imen -
en iched opinion mining, he p oposed amewo k ad ances owa d
a da a-cen ic, cus ome - ocused ma ke ing model ha enhances
pe sonaliza ion, s eng hens p edic ion, and op imizes se ice
ecommenda ions in he ou ism and hospi ali y indus y.
Copy igh © ISRG Publishe s. All igh s Rese ed.
DOI: 10.5281/zenodo.17330787
286
Figu e 2. A chi ec u e o a sen imen -based business in elligence sys em o ou ism and hospi ali y
Figu e 3. Applica ions o pola i y analysis ou comes in sma ou ism and hospi ali y ma ke ing.
4. Resul s
4.1 Expe imen al Se up and Expec ed Ou comes
The p oposed sen imen -enhanced business in elligence amewo k
was e alua ed using da ase s om bo h benchma k eposi o ies and
eal-wo ld ou ism pla o ms. Public sen imen da ase s om he
UCI Machine Lea ning Reposi o y and Kaggle we e combined
wi h la ge-scale hospi ali y e iew da ase s om T ipAd iso ,
Booking.com, and Google Re iews. S uc u ed ea u es such as
booking equency, expendi u e ca ego ies, and cus ome
demog aphics we e in eg a ed wi h uns uc u ed ex ual ea u es
including e iew pola i y, opinion in ensi y, and sen imen - ich
exp essions.
Py hon-based lib a ies—including NLTK, sciki -lea n, and
Tenso Flow/Ke as—we e employed o p ep ocessing, ea u e
ex ac ion, hyb id ea u e selec ion, and model aining. Models
we e e alua ed using a en- old c oss- alida ion scheme o ensu e
gene alizabili y. The expec ed ou comes included obus sen imen
classi ica ion, enhanced cus ome segmen a ion, accu a e in en ion
p edic ion, and imp o ed ecommenda ion sys ems, wi h he goal
o p o iding ac ionable insigh s o hospi ali y ma ke ing.
4.2 Sen imen Classi ica ion Pe o mance
The classi ica ion esul s demons a e he ad an age o in eg a ing
hyb id ea u e selec ion wi h deep lea ning models. As shown in
Table 1, he A i icial Neu al Ne wo k (ANN) achie ed an
a e age accu acy o 90%, ou pe o ming Decision T ees (82%) and
Naï e Bayes (85%). The Long Sho -Te m Memo y (LSTM)
ne wo k pe o med bes o e all, wi h an accu acy o 92% and an
F1-sco e o 0.89, pa icula ly e ec i e o modeling sequen ial
gues na a i es.
Table 1. Classi ica ion pe o mance compa ison ac oss models
Model
Accu acy
(%)
P ecision
(%)
Recall
(%)
F1-
Sco e
Decision T ee
82
80
78
0.79
Naï e Bayes
85
83
82
0.82
SVM
87
85
84
0.84
ANN
90
89
88
0.87
LSTM
92
90
89
0.89
The supe io pe o mance o ANN and LSTM models con i ms
ha deep lea ning is e ec i e in cap u ing he complexi y o
hospi ali y e iews, whe e cus ome expe iences a e exp essed in
nuanced, mul i-sen ence na a i es.
Copy igh © ISRG Publishe s. All igh s Rese ed.
DOI: 10.5281/zenodo.17330787
287
4.3 Cus ome Segmen a ion Ou comes
Clus e ing analysis in eg a ing bo h BI and sen imen ea u es
p oduced h ee dis inc cus ome g oups: 1) Loyal Gues s wi h
Posi i e Sen imen – equen booke s wi h high sa is ac ion, ideal
o loyal y campaigns; 2) P ice-Sensi i e Neu al Gues s –
mode a e bookings and neu al sen imen , esponsi e o
p omo ional o e s; 3) Dissa is ied Gues s wi h Nega i e Sen imen
– in equen epea bookings, use ul o se ice imp o emen
insigh s.
This segmen a ion illus a es he amewo k’s abili y o combine
beha io al and emo ional da a o achie e iche and mo e
ac ionable cus ome p o iles han BI-only app oaches.
4.4 Cus ome In en ion P edic ion
P edic i e models achie ed s ong esul s in es ima ing cus ome
in en ion. The ANN-based p edic o eached an ROC-AUC o
0.91, co ec ly o ecas ing whe he a gues would exp ess posi i e
o nega i e sen imen in u u e e iews. This p o ides ou ism
ma ke e s wi h a p ac ical ool o an icipa e dissa is ac ion and
in e ene p oac i ely, imp o ing se ice eco e y e o s.
4.5 Recommenda ion Sys em Pe o mance
The hyb id ecommenda ion sys em signi ican ly ou pe o med
baseline collabo a i e and con en -based il e ing app oaches. As
shown in Table 2, he hyb id model achie ed he lowes e o a es
wi h RMSE = 0.72 and MAE = 0.54, compa ed o RMSE = 0.95
o collabo a i e il e ing.
Table 2. Recommenda ion sys em pe o mance
Model
RMSE
MAE
Collabo a i e Fil e ing
0.95
0.70
Con en -Based Fil e ing
0.88
0.65
Hyb id (CF + Sen imen )
0.72
0.54
The esul s con i m ha inco po a ing sen imen in o
ecommenda ion engines enables mo e pe sonalized and con ex -
awa e se ice sugges ions, imp o ing ele ance in ho el,
des ina ion, and package ecommenda ions.
4.6 Business-O ien ed KPI Imp o emen s
Beyond echnical accu acy, he amewo k was assessed in e ms
o hospi ali y-speci ic business KPIs. As summa ized in Table 3,
in eg a ing sen imen analysis led o measu able imp o emen s:
booking con e sion a es inc eased by 12%, gues sa is ac ion
sco es ose by 8%, and cus ome e en ion imp o ed by 10%.
Table 3. Business KPI imp o emen s wi h sen imen in eg a ion
KPI
Baseline (BI
only)
Wi h Sen imen
In eg a ion
Booking Con e sion Ra e
100
112
Gues Sa is ac ion Sco e
100
108
Cus ome Re en ion Ra e
100
110
These ou comes demons a e ha he p oposed amewo k no only
enhances model pe o mance bu also ansla es in o angible
s a egic bene i s o hospi ali y businesses.
5. Discussion
The esul s o his s udy con i m ha in eg a ing business
in elligence (BI) wi h sen imen analysis can deli e subs an ial
imp o emen s in ou ism and hospi ali y ma ke ing ou comes.
Compa ed o BI-only app oaches, he hyb id amewo k
demons a ed supe io pe o mance ac oss sen imen classi ica ion,
cus ome segmen a ion, in en ion p edic ion, and ecommenda ion
sys ems. This alida es he a gumen ha s uc u ed beha io al
da a and uns uc u ed opinion da a a e complemen a y, and ha
hei in eg a ion p oduces iche , mo e ac ionable cus ome insigh s
[48].
P e ious s udies on BI applica ions in hospi ali y ha e ocused
p ima ily on s uc u ed da a such as booking his o ies, seasonal
demand ends, and loyal y p og am analy ics [8]. While e ec i e
in p edic ing ansac ional beha io , hese app oaches o e look he
subjec i e dimension o cus ome expe ience. Simila ly, esea ch
on sen imen analysis in ou ism has concen a ed on ex ac ing
pola i y and emo ion om online e iews [9], p o iding aluable
eedback o se ice quali y moni o ing bu o e ing limi ed
p edic i e o p esc ip i e insigh s when used in isola ion.
Ou indings ex end his li e a u e by demons a ing ha he
in eg a ion o BI wi h sen imen analysis no only enhances
echnical pe o mance me ics (accu acy, p ecision, RMSE) bu
also leads o business-le el imp o emen s in booking con e sion,
gues sa is ac ion, and e en ion. In pa icula , he 12% inc ease in
booking con e sion a es aligns wi h p io wo k by Phillips e al.
[49], who ound ha online e iews signi ican ly in luence ho el
bookings, bu ou model ad ances his by embedding e iew
sen imen di ec ly in o p edic i e and ecommenda ion algo i hms.
Fu he mo e, he ANN and LSTM models in his s udy achie ed
accu acy le els exceeding 90%, ou pe o ming esul s epo ed in
ea lie ou ism sen imen s udies, whe e classi ica ion accu acy
ypically anged be ween 75% and 85% [50,51]. This imp o emen
can be a ibu ed o he use o hyb id ea u e selec ion, which
minimized noise and p ese ed sen imen - ich a ibu es, a
me hodological enhancemen no widely adop ed in p e ious wo k.
5.1 Theo e ical Con ibu ions
The expec ed indings o his esea ch p o ide se e al impo an
heo e ical con ibu ions o he ields o business in elligence,
sen imen analysis, and ou ism ma ke ing. Fi s , he s udy
ad ances he in eg a ion o s uc u ed and uns uc u ed da a
s eams, demons a ing how ansac ional BI da a (e.g., booking
his o ies, spending pa e ns) and uns uc u ed opinion da a (e.g.,
gues e iews, social media con en ) can be uni ied in o a single,
da a-cen ic amewo k. This in eg a ion add esses a longs anding
gap in he li e a u e, whe e BI and sen imen analysis ha e o en
been ea ed as isola ed domains. Second, he p oposed amewo k
en iches he heo e ical discou se on cus ome in elligence by
b idging he beha io al–pe cep ual di ide. While BI has
his o ically emphasized ―wha cus ome s do,‖ sen imen analysis
cap u es ―how cus ome s eel.‖ By embedding sen imen -de i ed
pola i y and in ensi y measu es in o BI-d i en models, his s udy
o e s a mo e holis ic concep ualiza ion o cus ome expe ience,
one ha accoun s o bo h measu able ac ions and emo ional
pe cep ions. Thi d, he esea ch con ibu es o me hodological
ad ancemen s using hyb id ea u e selec ion and in elligen
lea ning models (ANNs and LSTMs). Theo e ically, his illus a es
how combining ule-based in e p e abili y wi h machine lea ning
scalabili y s eng hens bo h he p ecision and gene alizabili y o
sen imen -d i en BI sys ems. Such a hyb id design p o ides a
eplicable amewo k o u u e esea ch ac oss domains whe e
cus ome beha io and pe cep ion mus be analyzed join ly.
Finally, by si ua ing he amewo k wi hin he ou ism and
