Jou nal o Compu e Technology and So wa e
ISSN: 2998-2383
Vol. 2, No. 3, 2023
Enhancing Ad e ising Recommenda ion Pe o mance ia
In eg a ed Causal In e ence and Exposu e Bias Co ec ion
Yue Xing
Uni e si y o Pennsyl ania, Philadelphia, USA
[email protected]
Abs ac : This pape add esses he common p oblem o exposu e bias in ad e ising ecommenda ion by p oposing a new
me hod ha in eg a es causal in e ence wi h bias co ec ion o enhance he obus ness and accu acy o ecommenda ion sys ems
in complex en i onmen s. The s udy i s analyzes he selec i e exposu e cha ac e is ics o use beha io da a in ad e ising
scena ios and poin s ou ha elying only on s a is ical co ela ion can easily lead o spu ious associa ions, he eby educing he
eliabili y o ecommenda ions. On his basis, a causal modeling amewo k is designed, in oducing causal in e ence ools such as
po en ial ou comes and a e age ea men e ec o model he causal ela ionship be ween ad exposu e and click beha io . A he
same ime, in e se p opensi y weigh ing is applied o eweigh he aining da a and educe sys ema ic bias in oduced by he
exposu e mechanism. The me hod uni ies causal s uc u e lea ning, coun e ac ual gene a ion, and causal egula iza ion in he
modeling p ocess, ensu ing ha he ecommenda ion esul s a e close o use s' ue p e e ences. The expe imen al sec ion
includes compa a i e expe imen s and mul i-dimensional sensi i i y es s, such as hype pa ame e sensi i i y, en i onmen al
sensi i i y, and da a sensi i i y. Resul s show ha he p oposed me hod achie es supe io pe o mance on P ecision@10,
Recall@10, ACC@10, and NDCG compa ed wi h exis ing me hods, and demons a es high obus ness unde di e en
expe imen al condi ions. O e all, his s udy p o ides a sys ema ic solu ion o add essing exposu e bias and imp o ing
ecommenda ion e ec i eness in ad e ising ecommenda ions and o e s aluable guidance o u u e p ac ical applica ions.
Keywo ds: Ad ecommenda ions; causal in e ence; exposu e bias co ec ion; da a sensi i i y
1. In oduc ion
In he ongoing e olu ion o he digi al economy, ad e ising
ecommenda ions ha e become a key elemen o achie ing
p ecise ma ke ing and enhancing use expe ience in he in e ne
indus y. Wi h he sp ead o mobile in e ne and he
exponen ial g ow h o use beha io da a, ecommenda ion
sys ems a e no longe limi ed o il e ing in o ma ion. They
now di ec ly in luence he c ea ion o business alue and he
main enance o use ela ionships. In his p ocess, how o
e ec i ely cap u e use s' eal p e e ences om massi e
he e ogeneous da a has become a majo esea ch ques ion. In
pa icula , wi hin he highly compe i i e digi al ecosys em,
companies u gen ly need mo e scien i ic and igo ous me hods
o deli e e icien , accu a e, and in e p e able
ecommenda ions[1]. This is essen ial o mee use s'
expec a ions o pe sonalized se ices and o b ing highe
con e sion a es and e enue o ad e ising pla o ms.
Howe e , he complexi y o ad e ising ecommenda ion
goes a beyond su ace-le el modeling o use in e es s. In eal
in e ac ion scena ios, he ads ha use s a e exposed o a e no
andom. They a e in luenced by his o ical decisions o he
ecommenda ion model and by anking mechanisms. This leads
o he so-called exposu e bias p oblem. Exposu e bias makes
he da a used in aining inhe en ly selec i e. I ampli ies
ce ain pa e ns while neglec ing po en ial eal in e es s, which
esul s in sys ema ic bias in ecommenda ion ou comes. Such
bias weakens he model's gene aliza ion abili y and la gely
limi s he ai ness and long- e m s abili y o ecommenda ion
sys ems. The e o e, how o e ec i ely alle ia e and co ec
exposu e bias du ing he modeling p ocess has become a
cen al challenge ha canno be a oided[2].
A he same ime, he goal o a ecommenda ion sys em is
no only o pu sue co ela ion bu also o e eal causali y. Use
beha io is o en he esul o mul iple in e ac ing ac o s. I
may be d i en by he a ac i eness o he ad i sel , he con ex
o in e ac ion, pla o m s a egies, o e en social in luences. I
a ecommenda ion sys em elies only on s a is ical co ela ion,
i can easily gene a e spu ious associa ions[3]. This unde mines
he obus ness o ecommenda ion s a egies. The in oduc ion
o causal in e ence p o ides a new solu ion. By cons uc ing
causal g aphs and applying coun e ac ual easoning, models
can be e iden i y he causal chain be ween ad exposu e and
use beha io . This enables eliabili y and in e p e abili y e en
in biased da a en i onmen s. Such explo a ion has p o ound
signi icance o bo h he heo y and p ac ice o ad e ising
ecommenda ion.
Combining causal in e ence wi h exposu e bias co ec ion
add esses he limi a ions o each single me hod and achie es
dual op imiza ion a bo h he da a and model le els. A he da a
le el, causal pe spec i es help ein e p e use click beha io ,
il e ing ou spu ious co ela ions and o e ing mo e eliable
aining samples. A he model le el, bias co ec ion s a egies
enable s able lea ning e en wi h biased da a. The alue o his
in eg a ed app oach lies in i s abili y o enhance accu acy,
ai ness, and in e p e abili y a he same ime. I p o ides a
s ong ounda ion o he sus ainable de elopmen o he digi al
ad e ising ecosys em. Mo e impo an ly, i o e s heo e ical
and p ac ical suppo o pla o ms seeking sa e and complian
ecommenda ion s a egies unde he p essu es o p i acy
p o ec ion and s ic e egula ions[4].
In summa y, ad e ising ecommenda ion esea ch is
mo ing beyond he s age o elying solely on la ge-scale da a
and deep models. I is shi ing owa d comp ehensi e
amewo ks ha emphasize causali y and bias con ol. This
end e lec s deepe academic hinking on he in e nal logic o
ecommenda ion sys ems and esponds o indus y demands o
high-quali y ad e ising ecommenda ions. By in eg a ing
causal in e ence wi h exposu e bias co ec ion, esea che s and
p ac i ione s can build ecommenda ion mechanisms ha a e
mo e obus and mo e uni e sal. Such e o s will ad ance
e iciency, ai ness, and us wo hiness in ad e ising
ecommenda ions. This di ec ion has s ong heo e ical alue
and b oad applica ion p ospec s, and i ep esen s a s a egic
pa hway o embedding ecommenda ion echnologies in o he
u u e digi al economy.
2. Rela ed wo k
Exis ing esea ch on ad e ising ecommenda ion mainly
ocuses on modeling use in e es s and imp o ing
ecommenda ion accu acy. Ea ly me hods o en elied on
collabo a i e il e ing o con en -based app oaches. They
in e ed u u e in e ac ions om he simila i y be ween use
his o ical beha io s and i em ea u es. Howe e , hese me hods
ace limi a ions when dealing wi h da a spa si y, cold s a , and
di e se in e ac ion pa e ns. They s uggle o cap u e complex
use p e e ences comp ehensi ely. Wi h he de elopmen o
deep lea ning, neu al ne wo ks ha e been in oduced in o
ecommenda ion sys ems. La ge-scale ea u e in e ac ion,
sequence modeling, and con ex unde s anding a e be e
exp essed, which signi ican ly imp o es ecommenda ion
pe o mance[5]. Ye mos o hese me hods a e s ill based on
s a is ical co ela ion. They lack a deep desc ip ion o causal
mechanisms and a e easily a ec ed by noise and bias.
In he speci ic con ex o ad e ising ecommenda ion,
exposu e bias has become a c i ical ac o in luencing model
aining and in e ence. Ad e isemen display is o en
de e mined by pla o m s a egies, so he ads shown o use s
a e no balanced. Such selec i e exposu e causes models o
lea n biased pa e ns[6]. Fo example, a use 's non-click may
no indica e a lack o in e es bu a he a lack o exposu e. I
modeling is conduc ed di ec ly on such unbalanced da a,
sys ema ic dis o ion is ine i able. To add ess his issue,
exis ing s udies ha e a emp ed me hods such as in e se
p opensi y weigh ing, p opensi y sco e modeling, and pseudo-
label co ec ion. These echniques can educe he e ec o bias
and imp o e ai ness and s abili y o some ex en . Howe e ,
hey o en ely on s ic assump ions and s ill show weaknesses
in adap abili y and gene aliza ion unde complex en i onmen s.
The in oduc ion o causal in e ence p o ides a
b eak h ough o ad e ising ecommenda ions. By
cons uc ing causal g aphs and applying coun e ac ual
easoning, esea che s can mo e clea ly iden i y he causal
chain be ween ad exposu e and use clicks. This p e en s
models om being misled by spu ious co ela ions. In his
p ocess, causal in e ence no only explains he d i ing ac o s
behind use beha io s bu also helps ecommenda ion sys ems
main ain obus ness unde biased and imbalanced da a. In
ecen yea s, causal in e ence has been combined wi h deep
lea ning. Di ec ions such as causal ep esen a ion lea ning and
causali y-enhanced ecommenda ion ha e eme ged. These
app oaches a emp o in eg a e in e p e abili y and obus ness
in o sys em design while p ese ing modeling capaci y. They
ha e become an impo an esea ch end in ad e ising
ecommenda ion[7].
F om a comp ehensi e pe spec i e, combining causal
in e ence wi h exposu e bias co ec ion is becoming a key pa h
o ad ancing ad e ising ecommenda ion sys ems. A single
bias co ec ion me hod canno ully cap u e causal ela ions,
while pu e causal in e ence may ace high easoning
complexi y when applied o la ge-scale da a. Th ough o ganic
in eg a ion, he wo app oaches can complemen each o he .
Bias co ec ion me hods p o ide cleane da a o causal
in e ence, while causal in e ence o e s a heo e ical amewo k
o bias co ec ion. This di ec ion shows g ea po en ial in
academic esea ch and inc easing alue in p ac ice. I lays a
solid ounda ion o building ad e ising ecommenda ion
sys ems ha a e ai , eliable, and in e p e able.
3. P oposed App oach
In he me hod design, he o e all amewo k akes causal
in e ence as he co e and combines i wi h he exposu e bias
co ec ion mechanism o o m a uni ied modeling p ocess. The
o e all model a chi ec u e is shown in Figu e 1.
Figu e 1. O e all model a chi ec u e
Fi s , assume ha he in e ac ion be ween use s and ads
can be exp essed as a condi ional p obabili y dis ibu ion, ha
is, he use 's click beha io Y is join ly a ec ed by he ad
ea u e X and he exposu e s a e E. The basic modeling goal
can be o malized as:
)1(
),(
),,(
),|( EXP
EXYP
EXYP
Exposu e s a e
1E
indica es ha he ad was displayed,
and
0E
indica es ha he ad was no displayed. This
modeling app oach explici ly cha ac e izes he causal e ec o
exposu e on use clicks and lays he ma hema ical ounda ion
o subsequen bias co ec ion and causal easoning.
To mi iga e he impac o exposu e bias, we in oduce he
idea o in e se p opensi y weigh ing (IPW) o eweigh he
click signal o each sample. Speci ically, he p opensi y sco e
is de ined as:
)2( )|1()( XEPX
And by weigh ing he sample click esul s, we can ge he
modi ied expec ed isk unc ion:
)3(
)(
))(,(
1
1
N
ii
iii
IPW X
X YlE
N
L
He e,
)(l
ep esen s he loss unc ion, and
)( i
X
is
he ou pu o he ecommenda ion model. This design can
educe he imbalanced impac o he exposu e mechanism and
make he model close o unbiased es ima ion.
In he causal in e ence sec ion, his s udy in oduces a
po en ial ou come amewo k o cha ac e ize he coun e ac ual
scena ios o use clicks. Po en ial ou comes
)1(Y
and
)0(Y
a e de ined, ep esen ing he use clicks when he ad was
displayed and when i was no , espec i ely. The causal e ec
can be ep esen ed by he a e age ea men e ec (ATE):
)4( )]0()1([ YYEATE
In he ac ual modeling p ocess, combining p opensi y
sco e ma ching wi h coun e ac ual es ima ion can yield mo e
obus causal e ec es ima es, hus a oiding he limi a ions o
in e ing use in e es s based solely on co ela ions.
Finally, du ing he model op imiza ion phase, bias
co ec ion and causal in e ence a e join ly in eg a ed in o he
ecommenda ion objec i e unc ion o cons uc a causal
egula iza ion lea ning amewo k. Speci ically, he
op imiza ion objec i e is de ined as:
)5( ]))([ 2
CFIPW YX ELL
He e,
CF
Y
ep esen s he es ima ed click esul based on
coun e ac ual easoning, and
is he balancing pa ame e .
Th ough his join op imiza ion, he model no only co ec s o
biased da a bu also enhances i s abili y o cap u e causal
ela ionships, achie ing bo h imp o ed accu acy and
in e p e abili y.
4. Expe imen Resul
4.1 Da ase
This s udy uses he C i eo Display Ad e ising Challenge
public da ase . The da ase comes om eal-wo ld display
ad e ising scena ios and is designed o click- h ough a e
p edic ion asks. I p o ides log samples a he imp ession le el.
Each eco d includes a bina y label indica ing whe he he ad
was clicked. Nega i e samples ep esen imp essions wi hou
clicks, while posi i e samples ep esen imp essions wi h clicks.
In his way, he da ase na u ally p ese es in o ma ion a he
exposu e le el. I s la ge scale and co e age o di e se deli e y
and con ex si ua ions make i sui able o analyzing co ela ion
modeling and bias issues in ad e ising ecommenda ion.
The da ase con ains 13 nume ical ea u es, o en used as
dense ea u es, and 26 ca ego ical ea u es. The ca ego ical
ea u es a e high-ca dinali y spa se a iables ha ha e been
anonymized and hashed. Each imp ession is also linked o a
click label. The ca ego ical ields co e mul i-dimensional
con ex om he use , media, and ad e ising sides, such as si e
o placemen , de ice ype, and ad c ea i e iden i ie s. The
nume ical ields can be unde s ood as s a is ical o quan i a i e
ea u es ela ed o imp ession beha io . This design e lec s he
spa se and high-dimensional dis ibu ion o ea u es in eal
ad e ising deli e y. I also p o ides a ich in o ma ion base o
ep esen a ion lea ning, p opensi y modeling, and bias
co ec ion.
In p ac ice, common p ep ocessing includes illing missing
alues wi h placeholde s, me ging o hashing e y low-
equency ca ego ies, scaling and unca ing nume ical ea u es,
and spli ing da a in o aining, alida ion, and es se s by ime
o da e. Since samples a e al eady gene a ed a he imp ession
le el, addi ional nega i e sampling is usually unnecessa y.
Gi en i s scale and ea u e cha ac e is ics, he C i eo da ase
suppo s bo h causal modeling and exposu e bias co ec ion. I
o e s a ep oducible open benchma k and compa ison
en i onmen . I also enables sys ema ic e alua ion o di e en
me hods unde a consis en da a p o ocol.
4.2 Expe imen al Resul s
This pape i s conduc s a compa a i e expe imen , and he
expe imen al esul s a e shown in Table 1.
Table 1: Compa a i e expe imen al esul s
Me hod
P ecisio
n@10
Recall
@10
ACC@10
NDCG
Recbole[8]
0.428
0.392
0.601
0.447
CSLIM[9]
0.451
0.407
0.617
0.463
P ea[10]
0.462
0.419
0.628
0.475
INGCF[11]
0.489
0.436
0.641
0.493
Ou s
0.528
0.472
0.667
0.521
F om he esul s in Table 1, i can be obse ed ha di e en
me hods show a g adual imp o emen in P ecision@10,
Recall@10, ACC@10, and NDCG. T adi ional me hods such
as Recbole and CSLIM pe o m poo ly in p ecision and ecall.
This indica es ha al hough hey can cap u e use in e es s o
some ex en , hey a e s ill limi ed when dealing wi h exposu e
bias in ad e ising ecommenda ions. I also sugges s ha in
spa se and biased da a en i onmen s, elying only on
adi ional collabo a i e modeling canno ensu e
comp ehensi e and s able ecommenda ion ou comes.
Fu he compa ison shows ha P ea and INGCF
ou pe o m he i s wo me hods on se e al me ics, especially
Recall@10 and ACC@10. This demons a es ha when a
model can cap u e highe -o de in e ac ions be ween use s and
ad e isemen s, he ecommenda ion sys em gains mo e
ad an ages in iden i ying po en ial click beha io s. Howe e ,
hese me hods s ill depend mainly on s a is ical co ela ion.
They emain limi ed in ecognizing eal causal chains and hus
may s ill p oduce es ima ion bias unde da a selec ion bias
caused by exposu e mechanisms.
I is no ewo hy ha he p oposed me hod achie es he bes
pe o mance on all me ics. The imp o emen in P ecision@10
and Recall@10 indica es ha he model can ecommend
ad e isemen s ha be e ma ch eal use in e es s and also
co e mo e po en ial clicks. The ad an age in ACC@10 and
NDCG u he shows ha he model is mo e obus in o e all
accu acy and anking quali y. These esul s di ec ly e lec he
e ec i e in eg a ion o causal in e ence and exposu e bias
co ec ion. The ecommenda ion p ocess a oids he
in e e ence o spu ious co ela ions and becomes close o
use s' ue p e e ences.
O e all, hese esul s con i m he p ac ical alue o he
p oposed me hod in ad e ising ecommenda ion asks. By
in oducing causal s uc u es and in e se p opensi y weigh ing
in o he model, he sys em achie es highe accu acy and ecall,
and also signi ican ly imp o es anking pe o mance and
ai ness. This imp o emen p o ides new insigh s o academic
esea ch and es ablishes a solid ounda ion o p ac ical
applica ions o ad e ising ecommenda ions. In pa icula , i
demons a es s onge adap abili y and gene aliza ion alue
when acing complex en i onmen s and he e ogeneous da a.
This pape also gi es he impac o he causal egula iza ion
weigh λon he expe imen al esul s, and he expe imen al
esul s a e shown in Figu e 2.
Figu e 2. The impac o causal egula iza ion weigh λon expe imen al esul s
F om he esul s in Figu e 2, i can be seen ha changes in
he causal egula iza ion weigh λ lead o signi ican
imp o emen s ac oss se e al me ics. As λ inc eases om 0 o
0.8, P ecision@10 ises ma kedly. This indica es ha
in oducing causal cons ain s du ing modeling can e ec i ely
educe spu ious co ela ions and imp o e he alignmen
be ween ecommenda ion esul s and use s' ue p e e ences.
Compa ed wi h he case wi hou causal egula iza ion, he inal
p ecision is highe , showing he ad an age o he me hod in
p edic ing click beha io .
The pe o mance o Recall@10 also imp o es as λ
inc eases, wi h s onge e ec s a medium and high weigh s.
This sugges s ha causal egula iza ion helps he model co e
mo e po en ial click samples, he eby enhancing
ecommenda ion comp ehensi eness. In o he wo ds, when he
model emphasizes causal consis ency du ing aining, i can
unco e use s' eal in e es s ha may be hidden by exposu e
mechanisms. This educes ecall loss caused by sample bias.
Fo ACC@10, accu acy shows a s able upwa d end as λ
inc eases. This esul indica es ha he in oduc ion o causal
cons ain s does no ha m he o e all p edic i e s abili y o he
model. Ins ead, i s eng hens he eliabili y o ecommenda ion
decisions. Highe accu acy means ha unde di e en λ
se ings, he model main ains s ong co ec ness in he
ecommenda ion lis , demons a ing he con ibu ion o causal
in e ence and bias co ec ion o obus ness.
The NDCG esul s u he suppo his conclusion. When λ
inc eases, he anking quali y o ecommenda ion esul s
imp o es, especially in he ange o 0.6 o 0.8. This shows ha
causal egula iza ion no only op imizes ecommenda ion
anking bu also enhances he o e all a ionali y and ai ness o
use expe ience. The e o e, i can be concluded ha λ has a
posi i e e ec on model pe o mance wi hin a ce ain ange.
The essen ial eason is ha causal modeling plays a cen al ole
in educing bias and imp o ing in e p e abili y and obus ness.
This pape also gi es he impac o embedding dimension
on expe imen al esul s, and he expe imen al esul s a e shown
in Figu e 3.
Figu e 3. The impac o embedding dimension on expe imen al esul s
F om he esul s in Figu e 3, i can be seen ha
P ecision@10 inc eases s eadily wi h he g ow h o embedding
dimension and eaches a high le el a ound 256. The expansion
o he embedding dimension allows he model o cap u e iche
ea u e ep esen a ions, which imp o es he accu acy o
ad e ising ecommenda ions. Howe e , when he dimension
u he inc eases o 512, he imp o emen sa u a es o e en
declines sligh ly. This sugges s ha an excessi ely la ge
ep esen a ion space may in oduce edundan in o ma ion and
isk o o e i ing. This inding aligns well wi h he p ac ical
need o balance ep esen a ion e iciency and gene aliza ion in
ecommenda ion sys ems.
Fo Recall@10, he esul s show a signi ican imp o emen
as he dimension expands, especially in he ange om 64 o
256. This indica es ha highe -dimensional embeddings help
he model co e mo e po en ial ele an ads and educe missed
ue click in e es s. Ye when he dimension goes beyond 256,
he g ow h in ecall slows down. This means ha highe
dimensions do no u he enla ge co e age bu a he educe
he e iciency o in o ma ion use.
In e ms o ACC@10, he end is simila o P ecision and
Recall. Accu acy ises wi h he inc ease o embedding
dimension and eaches i s peak a medium o high dimensions.
This shows ha he o e all p edic ion co ec ness o he model
is maximized wi hin a easonable ep esen a ion space. The
combina ion o causal cons ain s and embedding
ep esen a ions enhances obus ness. Howe e , a highe
dimensions, he accu acy cu e becomes la , sugges ing ha
simply inc easing dimensionali y canno b ing u he
pe o mance gains.
The NDCG esul s u he con i m he imp o emen in
anking quali y. Be ween 128 and 256 dimensions, he me ic
ises no ably, indica ing ha iche ea u e ep esen a ions
enhance he ele ance and a ionali y o he ecommenda ion
lis . Al hough pe o mance d ops sligh ly a 512 dimensions, i
emains be e han in low dimensions. This end shows ha a
p ope choice o embedding dimension is essen ial o
balancing anking quali y and compu a ional cos . I also
highligh s he alue o causal modeling in imp o ing anking
pe o mance unde complex ad e ising ecommenda ion
scena ios.
This pape also p esen s a da a sensi i i y expe imen on he
a io o posi i e and nega i e samples o P ecision@10, and he
expe imen al esul s a e shown in Figu e 4.
Figu e 4. Da a Sensi i i y Expe imen on he Ra io o
Posi i e and Nega i e Samples o P ecision@10
F om he esul s in Figu e 4, i can be obse ed ha he
a io o posi i e o nega i e samples has a clea impac on
P ecision@10. When he a io is 1:1, he model achie es he
highes P ecision@10. This shows ha unde balanced sample
condi ions, he ecommenda ion sys em can be e cap u e
use s' ue click in e es s and main ain high accu acy in he
ecommenda ion lis . This phenomenon e lec s he impo an
ole o balanced da a dis ibu ion in ad e ising
ecommenda ion modeling.
As he a io becomes mo e skewed, P ecision@10 shows a
con inuous decline. When he numbe o nega i e samples
inc eases, he model is mo e easily in luenced by non-click
samples du ing aining, which weakens i s abili y o cap u e
eal in e es s. This esul indica es ha excessi e imbalance in
da a dis ibu ion ampli ies he e ec o exposu e bias. As a
esul , ecommenda ion ou comes become biased owa d he
majo i y class, educing accu acy.
A a ios o 1:3 and 1:4, he downwa d end o
P ecision@10 becomes mo e p onounced. This shows ha
when he imbalance exceeds a ce ain h eshold, causal
modeling and bias co ec ion can mi iga e he p oblem o some
ex en , bu canno ully elimina e he nega i e impac . This
inding highligh s he impo ance o da a-le el cons ain s o
causal modeling e ec i eness. I also emphasizes he need o
design easonable expe imen al se ings and da a p ep ocessing
s a egies.
When he a io eaches 1:5, P ecision@10 alls o i s lowes
poin . This indica es ha ex eme da a imbalance se e ely
ha ms ecommenda ion pe o mance. I also shows ha causal
in e ence and egula iza ion a he model le el alone a e no
su icien o sol e dis ibu ion bias. Deepe op imiza ion is
equi ed in da a sampling and p opensi y es ima ion. O e all,
hese esul s unde line he signi icance o da a sensi i i y in
ad e ising ecommenda ion esea ch. They also p o ide
p ac ical e idence o u he explo a ion o mo e obus causal
in e ence and bias co ec ion me hods.
This pape also p esen s a da a sensi i i y expe imen on he
deg ee o misma ch o he p opensi y model o Recall@10, and
he expe imen al esul s a e shown in Figu e 5.
Figu e 5. Da a Sensi i i y Expe imen on he Deg ee o
Model Misma ch on Recall@10
F om he esul s in Figu e 5, i can be seen ha Recall@10
shows a con inuous decline as he deg ee o p opensi y model
misma ch inc eases. When he misma ch le el is 0%, he model
achie es a ecall close o 0.47. This indica es ha when he
p opensi y model is highly consis en wi h he ue dis ibu ion,
he ecommenda ion sys em can be e co e use s' eal
in e es s and ensu e comp ehensi e esul s. Howe e , once a
misma ch is in oduced, he abili y o he model o cap u e
po en ial click beha io s is g adually weakened.
In he ange o 10% o 30% misma ch, Recall@10 d ops
signi ican ly. This shows ha e en a mode a e le el o bias has
a s ong impac on he ecommenda ion model. I indica es ha
he accu acy o he p opensi y model is c ucial o causal
in e ence and bias co ec ion. I he p opensi y es ima ion is
inaccu a e, he co ec i e abili y o IPW o o he debiasing
me hods becomes limi ed, leading o educed ecall.
When he misma ch le el exceeds 30%, he decline in
Recall@10 becomes mo e se e e. This sugges s ha a se ious
misma ch makes i di icul o he ecommenda ion sys em o
ex ac aluable in o ma ion om biased da a. A his s age, he
model ails o dis inguish ue clicks om spu ious co ela ions
caused by inco ec p opensi y es ima ion. As a esul , co e age
becomes insu icien , and use s may miss ad e isemen s
highly ele an o hei in e es s.
O e all, hese esul s highligh he cen al ole o he
p opensi y model in causal in e ence-based ecommenda ion.
Only when p opensi y modeling is accu a e can exposu e bias
co ec ion mechanisms unc ion e ec i ely and ensu e obus
and eliable ecommenda ions. In con as , i he p opensi y
model is se e ely misma ched, he pe o mance o he
ecommenda ion sys em will s ill be signi ican ly a ec ed e en
wi h causal cons ain s. This p o ides impo an insigh s o
u u e esea ch on op imizing p opensi y modeling me hods
and enhancing model obus ness.
5. Conclusion
This s udy ocuses on he in eg a ion o causal in e ence
and exposu e bias co ec ion in ad e ising ecommenda ion
asks. I sys ema ically analyzes he limi a ions o exis ing
ecommenda ion me hods when acing selec i e exposu e da a
and p oposes a new modeling amewo k. By in oducing
causal s uc u es and in e se p opensi y weigh ing, he model
can main ain p edic ion accu acy while e ec i ely educing he
in e e ence o spu ious co ela ions. This design no only
add esses he bias p oblem caused by o e - eliance on
s a is ical co ela ion in adi ional me hods bu also p o ides a
mo e obus echnical pa h o ad e ising ecommenda ion,
demons a ing easibili y and e ec i eness in complex da a
en i onmen s.
In sensi i i y expe imen s, he p oposed me hod shows
good s abili y and adap abili y. Whe he in hype pa ame e
adjus men s, en i onmen al changes, o di e ences in da a
dis ibu ion, he combina ion o causal modeling and bias
co ec ion exhibi s s ong obus ness. This indica es ha he
me hod can handle unce ain y and dynamic a ia ion
commonly p esen in ad e ising ecommenda ion sys ems,
ensu ing he eliabili y o ecommenda ion esul s. In pa icula ,
unde challenging scena ios such as sample imbalance o
p opensi y model e o s, he amewo k is s ill able o main ain
high pe o mance, e lec ing i s sui abili y o complex
indus ial applica ions.
The signi icance o his esea ch lies no only in heo e ical
inno a ion bu also in i s con ibu ion o p ac ical applica ion.
Ad e ising ecommenda ion is a co e componen o he
in e ne economy, di ec ly a ec ing bo h business alue and
use expe ience. The p oposed amewo k ha combines causal
in e ence wi h bias co ec ion can help pla o ms alloca e
ad e ising exposu e oppo uni ies mo e ai ly and accu a ely,
he eby educing po en ial algo i hmic bias. This is o g ea
alue o imp o ing ad e ising e iciency, enhancing use
expe ience, and s eng hening long- e m pla o m c edibili y. In
o he wo ds, he con ibu ion o his s udy is no only a
echnical imp o emen bu also a sys ema ic solu ion ha
add esses eal applica ion needs.
O e all, he indings o his s udy p esen a new pe spec i e
on ad e ising ecommenda ion ha balances accu acy, ai ness,
and in e p e abili y. By in eg a ing causal in e ence and bias
co ec ion mechanisms in o ecommenda ion sys ems, he
model demons a es s onge obus ness and gene aliza ion in
complex en i onmen s. This app oach p o ides a new
e e ence pa adigm o ad e ising ecommenda ion and also
o e s a aluable esea ch di ec ion o o he ecommenda ion
o p edic ion asks wi h bias issues. The e o e, his wo k has a
p o ound impac on bo h academic esea ch and indus ial
p ac ice, laying a solid heo e ical and me hodological
ounda ion o ex ending causal ecommenda ion o b oade
scena ios.
Re e ences
[1] Li F, Dean S. C oss-da ase p opensi y es ima ion o debiasing
ecommende sys ems[J]. a Xi p ep in a Xi :2212.13892, 2022.
[2] Xu H, Xu Y, Yang Y. Causal S uc u e Rep esen a ion Lea ning o
Con ounde s in La en Space o Recommenda ion[J]. a Xi p ep in
a Xi :2311.03382, 2023.
[3] Xiao T, K e on B, Ka a iya S, e al. Towa ds sequen ial coun e ac ual
lea ning o ank[C]//P oceedings o he Annual In e na ional ACM
SIGIR Con e ence on Resea ch and De elopmen in In o ma ion
Re ie al in he Asia Paci ic Region. 2023: 122-128.
[4] Liu Y, Yen J N, Yuan B, e al. P ac ical coun e ac ual policy lea ning o
op-k ecommenda ions[C]//P oceedings o he 28 h ACM SIGKDD
Con e ence on Knowledge Disco e y and Da a Mining. 2022: 1141-
1151.
[5] Zhang J, Chen X, Tang J, e al. Recommenda ion wi h causali y enhanced
na u al language explana ions[C]//P oceedings o he ACM web
con e ence 2023. 2023: 876-886.
[6] Y. Zhu, J. Ma, and J. Li, "Causal in e ence in ecommende sys ems: A
su ey o s a egies o bias mi iga ion, explana ion, and gene aliza ion,"
a Xi p ep in a Xi :2301.00910, 2023.
[7] M. Mansou y, F. Duij es ijn, and I. Mou abe , "Po en ial ac o s leading
o popula i y un ai ness in ecommende sys ems: A use -cen e ed
analysis," a Xi p ep in a Xi :2310.02961, 2023.
[8] Zhao W X, Mu S, Hou Y, e al. Recbole: Towa ds a uni ied,
comp ehensi e and e icien amewo k o ecommenda ion
algo i hms[C]//p oceedings o he 30 h acm in e na ional con e ence on
in o ma ion & knowledge managemen . 2021: 4653-4664.
[9] Zheng Y, Mobashe B, Bu ke R. CSLIM: Con ex ual SLIM
ecommenda ion algo i hms[C]//P oceedings o he 8 h ACM
Con e ence on Recommende Sys ems. 2014: 301-304.
[10] Lee J, Sun M, Lebanon G. P ea: Pe sonalized ecommenda ion
algo i hms oolki [J]. The Jou nal o Machine Lea ning Resea ch, 2012,
13(1): 2699-2703.
[11] Sun W, Chang K, Zhang L, e al. INGCF: an imp o ed ecommenda ion
algo i hm based on NGCF[C]//In e na ional Con e ence on Algo i hms
and A chi ec u es o Pa allel P ocessing. Cham: Sp inge In e na ional
Publishing, 2021: 116-129.
