Cha ac e izing p o ide ai ness in con en -based
e-se ice in elligen ecommenda ion
Raciel Ye a1[0000−0001−9759−261X], Bapi Du a1[0000−0002−2796−8914], F ancisco
J. Ma ´ınez1,2[0000−0001−7969−1737], and Luis Ma ´ınez1[0000−0003−4245−8813]
1Compu e Science Depa men , Uni e si y o Ja´e]n, 23007, Ja´en, Spain
[email p o ec ed]
2Ayesa Se icios Digi ales A anzados, 41005, Se illa, Spain
Abs ac . Fai ness is cu en ly ega ded as a ele an dimension o-
wa ds he goal o eaching us wo hy a i icial in elligence-based sys-
ems. In ecommende sys ems, ai ness is ocused on add essing biases
ha may disp opo iona ely bene i o ha m ce ain classes o use s and
i ems. In his con ibu ion we a e in e es ed on p o ide ai ness, which
aims a gua an eeing ha he p o ide s o he i ems would ha e he same
chance o he exposu e o hei i ems in he inal ecommenda ion lis s.
Pa icula ly, we will be ocused on cha ac e izing he p o ide ai ness as-
socia ed o in elligen con en -based ecommenda ion used o sugges ing
e-se ices in a ma ke place en i onmen in he egion o Ex emadu a,
Spain. He ein, he gene alized c oss-en opy has been used as me ic
o cha ac e izing ai ness associa ed o bo h basic and la en di ichle
alloca ion (LDA)-based con en -based ecommenda ion. As main ind-
ings, ou s udy has indica ed ha he ecommenda ion based on la en
di ichle alloca ion migh lead o be e ai ness alues o hose p o ide s
wi h la ge numbe o e-se ices. Fo he manage ial iewpoin , i sug-
ges s ha a highe p esence in online pla o ms o he p oduc s, migh
gua an ee be e associa ed ai ness alues. As a as we know, his con-
ibu ion p esen s one o he i s e o s on e alua ing p o ide ai ness
ecommenda ion in a conc e e e-se ice scena io.
Keywo ds: T us wo hy ecommenda ions ·p o ide ai ness ·e-se ices ·con en -
based ecommende sys ems ·la en di ichle alloca ion
1 In oduc ion
Recommende sys ems a e ele an in elligen -based sys ems ocused on p o id-
ing use s wi h hose in o ma ion i ems ha bes i hei p e e ences and needs
in sea ch spaces o e loaded wi h possible op ions. Cu en ly, ecommende sys-
ems ha e been ac i ely used in a di e si y o domains, such as e-comme ce,
e-lea ning, e- ou ism, e-heal h, and e-se ices [7]. Fo p o iding accu a e ecom-
menda ions in such con ex s, wo main ecommenda ion pa adigms ha e been
de eloped: con en -based ecommenda ion, and collabo a i e il e ing-based ec-
ommenda ion [7].
This p ep in has no unde gone pee e iew (when applicable) o any pos -submission imp o emen s o
co ec ions. The Ve sion o Reco d o his con ibu ion is published in A i icial In elligence in Human-Cen ic,
Resilien and Sus ainable Indus ies, P oceedings o he INFUS 2025 Con e ence, Volume 3, and is a ailable
online a h ps://doi.o g/10.1007/978-3-031-98565-2_62
2 R. Ye a e al.
Beyond ecommenda ion accu acy, he goal o de eloping us wo hy ec-
ommende sys ems has gained inc easing a en ion in ecen yea s [9], aligning
wi h he concep o us wo hy a i icial in elligen highligh ed by he AI Ac
3. Ac oss his objec i e, ai ness-awa e ecommenda ion has become one o he
key dimensions o ensu ing us wo hiness. Add essing his issue equi es iden-
i ying and mi iga ing challenges such as popula i y bias, he ma ginaliza ion o
mino i y p e e ences, and algo i hmic disc imina ion [4]. Implemen ing ai ness-
awa e models allows ecommende sys ems o deli e balanced and inclusi e
sugges ions while s ill main aining pe sonaliza ion.
The implemen a ion o ai ness c i e ia in ecommende sys ems has been
based in wo main dimensions. A i s , consume ai ness has been ocused on
a oiding ha he deli e ed ecommenda ion would be sensi i e o some demo-
g aphic class associa ed o he use s, such as age, na ionali y, o gende [10].
On he o he side, p o ide ai ness has been ocused on a oiding he monopoly
domina ion by ensu ing a simila chance o exposu e in he ecommenda ion lis s
o all he a ailable i ems dis ega ding hei associa ed p o ide [9].
The objec i e o he cu en con ibu ion is he explo a ion o p o ide ai -
ness in an e-se ice scena io. As no ed by [7], e-se ices ha e been iden i ied as
one key applica ion o ecommende sys ems. In ecen yea s, nume ous e-se ice
pla o ms ha e eme ged as ICT-d i en solu ions aimed no only a bene i ing co -
po a ions and key s akeholde s bu also a s imula ing economic g ow h, gen-
e a ing employmen , suppo ing local en e p ises, and enhancing o e all quali y
o li e. A no able example is D-Ru al4, a collabo a i e ini ia i e in ol ing mul-
iple ins i u ions dedica ed o c ea ing a se ice ma ke place o u al egions
ac oss Eu ope. The esea ch p esen ed in his wo k will be conduc ed wi hin
his amewo k, al hough i s indings and me hodologies can be ex ended o
o he con ex s. Conce ning he na u e o da a and he lack o use p e e ences,
ou me hodology will be based on he con en -based ecommenda ion app oach.
As a as we know, ou esea ch is one o he i s e o s ocused on e alua ing
p o ide ai ness in eal business scena ios.
The con ibu ion has he ollowing s uc u e. Fi s , Sec ion 2 illus a es he
ela ed wo ks associa ed o ou p oposal. Sec ion 3 in oduces he me hodology
o cha ac e izing ai ness in e-se ice ecommenda ion. Sec ion 4 pe o ms he
expe imen s, including da ase , p o ocol, and esul s. Sec ion 5 concludes he
con ibu ion.
2 Rela ed wo ks
The lack o ai ness in con empo a y ecommenda ion sys ems ac oss di e en
s akeholde s has been widely ecognized by esea che s [4, 8]. Fo example, job-
ma ching pla o ms o en ail o p esen use s wi h an unbiased selec ion o em-
ploymen oppo uni ies; musicians may s uggle o achie e ai exposu e in music
disco e y se ices; in o ma ion pla o ms end o disp opo iona ely highligh
3h ps://a i icialin elligenceac .eu/
4h ps://co dis.eu opa.eu/p ojec /id/101017304
Cha ac e izing p o ide ai ness in e-se ice ecommenda ion 3
ce ain opics on a ec ed communi ies; and c edi sco ing sys ems equen ly
a o pa icula socio-demog aphic g oups, among o he cases.
One o he ea lies e o s o add ess ai ness was p oposed by Fel e nig e al.
[5], who in oduced a ai ness-awa e app oach aimed a educing he dominance
o speci ic ca ego ies and ensu ing a mo e balanced ep esen a ion. I in ol es
i s anking i ems wi hin hei espec i e ca ego ies, be o e me ging he indi id-
ual lis s h ough a u n-based s a egy ha dis ibu es exposu e mo e equi ably.
In ecen yea s, esea ch has inc easingly ocused on p o ide ai ness (P-
ai ness). Bo a o e al. [2] analyzed dispa i ies among i em p o ide s in e ms
o ele ance, isibili y, and exposu e by simula ing di e en le els o mino i y
ep esen a ion in i em ca alogs and hei in e ac ions. Thei p oposed app oach
in eg a es obse a ion upsampling wi h loss egula iza ion o e ine i em a ing
p edic ions, ul ima ely enhancing ai ness by imp o ing he isibili y, exposu e,
and ep esen a ion o unde ep esen ed g oups.
Gomez e al. [6] add ess he issue o unde - ep esen ed p o ide s in online
pla o ms. By en iching i em me ada a wi h he con inen o o igin, hei p o-
posed me hod enhances equi y h ough a e- anking s a egy ha balances bo h
he p opo ion o ecommenda ions alloca ed o i ems om each con inen and
hei placemen wi hin he ecommenda ion lis .
The conduc ed analysis has highligh ed a ious esea ch ini ia i es aimed a
imp o ing p o ide ai ness (P- ai ness) in ecommende sys ems, mos o which
a e ailo ed o speci ic domains. In he con ex o e-se ices, whe e a medium-
sized i em da ase is expec ed, he e is a need o a anspa en cha ac e iza ion
o he associa ed ai ness. Howe e , he e iewed app oaches do no ully add ess
hese speci ic equi emen s.
3 The amewo k o cha ac e izing ai ness in e-se ice
in elligen ecommenda ions
This sec ion will p esen he amewo k o be used o cha ac e izing p- ai ness
in he e-se ices associa ed o he D-Ru al p ojec . I will be composed o he
ollowing s ages (Figu e 1): 1) In o ma ion modeling, 2) Con en -based ecom-
menda ion, and 3) Fai ness e alua ion.
Fig. 1. The amewo k o cha ac e izing ai ness in e-se ice in elligen ecommenda-
ions.
4 R. Ye a e al.
3.1 In o ma ion modeling
This s age will be ocused on iden i ying he in o ma ion ha will be used o
modeling e-se ices. S a ing o he in o ma ion a ailable in he D-Ru al pla -
o m, we will model e-se ices h ough he ollowing a ibu es:
–IdSe ice: E-se ice iden i ie .
–Se iceTi le: E-se ice name.
–Desc ip ion: E-se ice desc ip ion, men ioning he exac acili y ha is p o-
ided, and hei main ea u es and s eng hs.
–De ails: Fu he desc ip ion, ha some imes is illed and some imes missing.
–Ca ego y: Key opic whe e he se ice is associa ed, such as heal h, leisu e,
echnological, and so on.
–P o ide : Co po a ion/agency ha manage he se ice.
–Add ess: Place whe e he se ice is loca ed.
Nex s ages will use hese a ibu es.
3.2 Con en -based ecommenda ion
Taking in o accoun ha he e-se ice modeling is based on ex and i em a -
ibu es, we will use wo con en -based ecommenda ion app oaches o gene -
a ing ecommenda ions in his scena io (Figu e 2), ins ead o he collabo a i e
il e ing-based ha depends on use a ings [7].
The i s one ep esen s a basic app oach ha builds TF-IDF ec o s using
he e ms associa ed o he e-se ice desc ip ions [7], and is composed o he
ollowing s ages, which a e no de eloped in de ail due o space limi a ions.
Fig. 2. Schemes o he used con en -based ecommenda ion app oaches.
1. Use he con en associa ed o he ields Desc ip ion and De ails o he e-
se ice, o eaching a TF-IDF ma ix associa ed wi h he associa ed e ms.
Cha ac e izing p o ide ai ness in e-se ice ecommenda ion 5
2. Use he se ices consumed by he ac i e use , o building i s p o ile also
based on he ex ac ed e ms. Build he p o ile o each cu en use , based
on he e-se ice TF-IDF ec o s ha he has p e e ed/consumed.
3. Fo he ac i e use , ob ain he sco e used o gene a ing ecommenda ions
using he linea ke nel simila i y be ween he use and e-se ice p o iles.
4. Re ie e he ecommenda ion lis wi h he op sco ed se ices.
The second one employs la en di ichle alloca ion (LDA) [1] o ex ac ing
he opics associa ed o he desc ip ions, using hem o composing he i em
p o iles. In con as o he p e ious app oach, s ages 1 and 2 would ha e he
ollowing shape:
1. Use he con en associa ed o ields Desc ip ion and De ails o he e-se ices,
o eaching ele an seman ic opics using la en di ichle alloca ion.
2. Use he se ices consumed by he ac i e use , o building i s p o ile also
based on he ex ac ed e ms. Build he p o ile o each cu en use , based
on he e-se ice opic ec o s ha he has p e e ed/consumed.
3.3 Fai ness e alua ion
Deldjoo e al. [3] in oduce he use o gene alized c oss en opy o e alua ing
ai ness in indi idual ecommende sys ems, which will be he measu e used in
his wo k o cha ac e izing ai ness o he e-se ice ecommenda ion. Fo he
disc e e a ibu e case, au ho s de ine he gene alized c oss en opy (GCE) o
some pa ame e α= 0,1 as:
I(M, a) = 1
α(1 −α)X
aj
pα
(aj)p(1−α)(aj)−1 (1)
He e, a∈Ais he a ibu e o which he ai ness is e alua ed. pand p a e
he ac ual p obabili y dis ibu ion o he a ibu e aac oss he ecommenda ion
ou pu ; and he desi ed, ai , p obabili y dis ibu ion.
p is usually o malized as uni o m dis ibu ion, e.g. ega ding ha ai ness
means equali y. Deldjoo e al. [3] de ines p, o he a ibu e ajas:
p(aj) = Pi∈aj gi
Z(2)
whe e Z=Pi gi. Fu he mo e, gi, which is he ecommenda ion gain o
i, is o malized as:
gi=X
u∈U
ϕ(i, RecK
u)g(u, i, ) (3)
whe e ϕ(i, RecK
u) = 1 i i em iis in he e ie ed ecommenda ion lis .
g(u, i, ) is he gain o ecommending o he use u he i em iwi h he ank-
ing , and is calcula ed h ough he discoun ed cumula i e gain (DCG), keeping
el(u, i) = 1 : g(u, i, ) = 2 el(u,i)−1
log2( +1) . The ob ained DCG alues a e inally no -
malized in o he NDCG alues, which a e used o calcula ing gi
6 R. Ye a e al.
4 Expe imen s
This sec ion will be ocused on execu ing he p esen ed amewo k o cha ac e -
izing he p o ide ai ness associa ed o an e-se ice ecommenda ion da ase .
Sec ion 4.1 p esen s he da ase and he expe imen al p o ocol ha will be used.
Sec ion 4.2 p esen s and discusses he ob ained esul s.
4.1 Da ase and expe imen al p o ocol
As s a e abo e, we will use a da ase associa ed o a ma ke place o e-se ices
deployed a he Ex emadu a egion, in Spain. F om he a ibu es iden i ying
he da ase poin ed ou in Sec ion 3, we will be ocused on he company/ i m
ha p o ides he se ice (i.e. he se ice p o ide ).
O e all, ou da ase con ains i ems om 31 di e en p o ide s. Howe e , we
ha e also de ec ed p o ide s wi h a e y educed numbe o i ems, whe e i has
no sense he ai ness cha ac e iza ion. The e o e, we will limi ou analysis o
p o ide s ha ing a leas h ee associa ed se ices, which will be Ac i a-men e
Ex emadu a (7 se ices), BidInn (3 se ices), Gc Genomics (5 se ices), akin o
(3 se ices), BWell Labs (8 se ices), and ADI&SALU (5 se ices).
Ou expe imen al p o ocol is based on h ee s eps:
1. Assume ha each use has consumed one speci ic e-se ice. Then, we will
assume he p esence o 62 use s, each one e alua ing each o he 62 se ices.
2. Fo each use , we will apply he p ocedu e poin ed ou a Sec ion 3, o
gene a ing hei associa ed ecommenda ion and i s ai ness alue acco ding
o each o he speci ied p o ide s.
3. Fo each p o ide i is ob ained he a e age ai ness alues ac oss all he
ecommenda ion lis s, which a e he alues epo ed in his wo k.
Pa icula ly i will be conside ed h ee ecommenda ion app oaches: 1) he
simple con en -based app oach, 2) he LDA-based con en -based app oach wi h
5 ac o s, and 3) he LDA-based con en -based app oach wi h 15 ac o s. These
numbe s o ac o s ha e been se up based on he a e age leng h o he desc ip-
ions. Fu he mo e, i will be conside ed op 5 and op 10, as sizes o he ecom-
menda ion lis s. Ou expe imen al analysis will be mainly ocused on answe ing
he ollowing ques ions: 1) Is he e some ela ionships be ween he numbe o
e-se ices linked o he p o ide , and he associa ed ai ness?, and 2) Wha is he
e ec o using a mo e sophis ica ed ecommenda ion app oach, o e he ai ness
alues?
4.2 Resul s
Tables 1 and 2 illus a e he esul s o he h ee conside ed ecommenda ion
app oaches, o op 5 and op 10 ecommenda ion lis s.
The i s impo an inding o no ice, is ha o he basic app oach i was no
iden i ied a clea co ela ion be ween he numbe o se ices o each p o ide , and
Cha ac e izing p o ide ai ness in e-se ice ecommenda ion 7
Table 1. P o ide ai ness alues associa ed o he discussed con en -based ecommen-
da ion app oaches. Top 5 ecommenda ions. La ge alues sugges be e ai ness.
P o ide Basic con en -based LDA-based LDA-based
app oach app oach (5 ac o s) app oach (15 ac o s)
Ac i a-men e Ex emadu a (7 se ) -0.3586 -0.3549 -0.3214
BidInn (3 se ) -0.3296 -0.0577 -0.5021
Gc Genomics (5 se ) -0.2129 -0.5965 -0.3247
akin o (3 se ) -0.2125 ∞-0.2247
BWell Labs (8 se ) -0.3036 -0.4133 -0.2813
ADI & SALU (5 se ) -0.3393 -0.4303 -0.3392
he co esponding ai ness alues. As example, while o op 5 ecommenda ions
p o ide s wi h a smalle numbe o e-se ices (Gc Genomics and akin o) ob ain
he bes ai ness alues (-0.2129 and -0.2125), o op 10 i is in e es ing ha he
bes ai ness alues we e ob ained o wo p o ide s wi h espec i ely he la ges
and he smalles numbe o e-se ices (BWell Labs and akin o).
Table 2. P o ide ai ness alues associa ed o he discussed con en -based ecommen-
da ion app oaches. Top 10 ecommenda ions. La ge alues sugges be e ai ness.
P o ide Basic con en -based LDA-based LDA-based
app oach app oach (5 ac o s) app oach (15 ac o s)
Ac i a-men e Ex emadu a (7 se ) -0.8074 -0.1776 -0.5561
BidInn (3 se ) -0.9608 -0.2616 -1.1692
Gc Genomics (5 se ) -0.9153 -0.8457 -0.7374
akin o (3 se ) -0.6989 -1.3796 -0.7378
BWell Labs (8 se ) -0.5812 -0.3871 -0.5417
ADI & SALU (5 se ) -0.9666 -1.0565 -1.0215
Fu he mo e, i can be iden i ied ha o he scena ios wi h he la ge numbe
o e-se ices, such as BWell Labs and Ac i a-men e Ex emadu a, he use o a
mo e complex ecommenda ion app oach like he LDA-based, ends o lead o
an imp o emen o he ai ness alues. As example, in op 5 o Ac i a-men e
Ex emadu a he ai ness is imp o ed om -0.3586 o -0.3214 o LDA-based
(15 ac o s), and o BWell Labs i is imp o ed om -0.3036 o -0.2813 also
o LDA-based (15 ac o s). In he case o op 10 ecommenda ions, a Ac i a-
men e Ex emadu a i is imp o ed om -0.8074 o -0.1776 (5 ac o s) and a
BWell Labs om -0.5812 o -0.3871 (5 ac o s).
On he o he side, o hose p o ide s wi h a smalle numbe o e-se ices,
he use o a mo e complex ecommenda ion app oach was no necessa ily as-
socia ed o an imp o ed ai ness. Pa icula ly, in scena ios such as BidInn and
akin o, i ob ained a wo se ai ness alues. We hink ha his beha io should
be connec ed o he ac ha he educed numbe o se ices as well as hei
limi ed desc ip ions, do no allow o connec hem ( h ough he ecommenda-
ion app oach) o o he se ices belonging o he same p o ide . He ein, i is
in e es ing he case o akin o o op 5 ecommenda ions and 5 ac o s, whe e
LDA-based does no e ie e any se ice belonging o he same p o ide .
8 R. Ye a e al.
5 Conclusions
The cu en con ibu ion has been ocused on in oducing a amewo k o cha -
ac e izing p o ide ai ness in e-se ice in elligen ecommenda ion. I is com-
posed o h ee s ages ocused on in o ma ion modeling, con en -based ecom-
menda ion, and ai ness e alua ion. The gene alized c oss en opy measu e was
used o cha ac e izing ai ness. As ele an inding, i has been p o ed ha he
ecommenda ion based on la en di ichle alloca ion migh lead o be e ai -
ness alues o hose p o ide s wi h la ge numbe o e-se ices. The nex u u e
wo ks will be cen e on e alua ing ai ness in g oup ecommenda ions, as well as
p oposing me hods o imp o ing he ob ained ai ness in he cu en con ex .
Acknowledgemen
This p ojec is suppo ed by he Eu opean Union’s Ho izon Eu ope esea ch and
inno a ion p og am unde he Ma ie Sklodowska-Cu ie g an ag eemen numbe
101106164. Views and opinions exp essed a e howe e hose o he au ho (s)
only and do no necessa ily e lec hose o he Eu opean Union. Nei he he
Eu opean Union no he g an ing au ho i y can be held esponsible o hem.
We hank he suppo o Manuel Gimenez om Ayesa.
Re e ences
1. Blei, D.M., Ng, A.Y., Jo dan, M.I.: La en di ichle alloca ion. Jou nal o Machine
Lea ning Resea ch 3(Jan), 993–1022 (2003)
2. Bo a o, L., Fenu, G., Ma as, M.: In e play be ween upsampling and egula -
iza ion o p o ide ai ness in ecommende sys ems. Use Modeling and Use -
Adap ed In e ac ion 31(3), 421–455 (2021)
3. Deldjoo, Y., Anelli, V.W., Zamani, H., Bellogin, A., Di Noia, T.: A lexible ame-
wo k o e alua ing use and i em ai ness in ecommende sys ems. Use Modeling
and Use -Adap ed In e ac ion pp. 1–55 (2021)
4. Deldjoo, Y., Jannach, D., Bellogin, A., Di onzo, A., Zanzonelli, D.: Fai ness in
ecommende sys ems: esea ch landscape and u u e di ec ions. Use Modeling
and Use -Adap ed In e ac ion 34(1), 59–108 (2024)
5. Fel e nig, A., Tin a e , N., T an, T.N.T., S e inge , M.: Explana ions o g oups.
In: G oup Recommende Sys ems: An In oduc ion, pp. 105–126. Sp inge (2018)
6. G´omez, E., Bo a o, L., Salam´o, M.: P o ide ai ness ac oss con inen s in col-
labo a i e ecommende sys ems. In o ma ion P ocessing & Managemen 59(1),
102719 (2022)
7. Lu, J., Wu, D., Mao, M., Wang, W., Zhang, G.: Recommende sys em applica ion
de elopmen s: a su ey. Decision suppo sys ems 74, 12–32 (2015)
8. Sonboli, N., Bu ke, R., Eks and, M., Meh o a, R.: The mul isided complexi y o
ai ness in ecommende sys ems. AI magazine 43(2), 164–176 (2022)
9. Wang, Y., Ma, W., Zhang, M., Liu, Y., Ma, S.: A su ey on he ai ness o ecom-
mende sys ems. ACM T ansac ions on In o ma ion Sys ems 41(3), 1–43 (2023)
10. Ye a, R., Ba anco, M., Ma ´ınez, L.: A no el app oach o measu ing demog aphic
pa i y ai ness in g oup ecommenda ion. In: FLINS-ISKE 2024. pp. 195–202.
Wo ld Scien i ic (2024)
