Annals o Da a Science (2024) 11(5):1705–1739
h ps://doi.o g/10.1007/s40745-023-00474-4
A i icial In elligence Algo i hms o Collabo a i e Book
Recommende Sys ems
Clemens Tege meie 1·A ne Johannssen1·Na aliya Chukh o a2
Recei ed: 24 No embe 2022 / Re ised: 10 May 2023 / Accep ed: 13 May 2023 /
Published online: 8 June 2023
© The Au ho (s) 2023
Abs ac
Book ecommende sys ems p o ide pe sonalized ecommenda ions o books o use s
based on hei p e ious sea ches o pu chases. As online ading o books has become
inc easinglyimpo an in ecen yea s,a i icialin elligence(AI)algo i hmsa eneeded
o ecommend sui able books o use s and encou age hem o make pu chasing deci-
sions in he sho and he long un. In his pape , we conside AI algo i hms o
so called collabo a i e book ecommende sys ems, especially he ma ix ac o -
iza ion algo i hm using he s ochas ic g adien descen me hod and he book-based
k-nea es -neighbo algo i hm. We pe o m a comp ehensi e case s udy based on he
Book-C ossing benchma k da a se , and implemen a ious a ian s o bo h AI algo-
i hms o p edic unknown book a ings and o ecommend books o indi idual use s
based on he highes p edic ed a ings. This s udy aims o e alua e he quali y o he
implemen ed me hods in ecommending books by using selec ed e alua ion me ics
o AI algo i hms.
Keywo ds A i icial in elligence ·Book ecommende sys ems ·knn algo i hm ·
Machine lea ning ·Ma ix ac o iza ion algo i hm ·S ochas ic g adien descen
me hod
BA ne Johannssen
a ne.johannssen@uni-hambu g.de
Clemens Tege meie
[email p o ec ed]
Na aliya Chukh o a
na aliya.chukh o a@hcu-hambu g.de
1Uni e si y o Hambu g, Hambu g, Ge many
2Ha enCi y Uni e si y, Hambu g, Ge many
123
1706 Annals o Da a Science (2024) 11(5):1705–1739
1 In oduc ion
Book ecommende sys ems a e o en used by companies o p esen in e es ing and
pe sonalized book ecommenda ions o hei cus ome s. The ecommenda ions a e
supposed ocon ince hecus ome obuybooksin hesho unand ouse hebook ec-
ommende sys em o u he pu chases in he long un. As online ading o books has
conside ably inc eased in ecen yea s [1], book ecommende sys ems ha e become
mo e impo an . Online bookselle s such as Amazon, Ba nes & Noble, Wa e s ones,
and Thalia ha e played an impo an ole in his de elopmen . Fo ins ance, cu en
challenges o ecommende sys ems a e aking in o accoun he use ’s con ex , e.g.,
ime o mood [2–4], ensu ing di e si y [5] and including implici a ings o a g ea e
ex en [6] when gene a ing book ecommenda ions. Gene ally, he esea ch ocus has
u ned o A i icial In elligence (AI) algo i hms, e.g., he numbe o pape s conside ing
deep lea ning echniques has inc eased signi ican ly in ecen yea s and is equen ly
applied o ecommende sys ems [7–9]. No e ha AI, like machine lea ning, deep
lea ning, da a mining, and Big Da a analy ics, is based on Da a Science echniques,
so hese a eas a e closely ela ed [10–12]. While AI e e s o he de elopmen o
in elligen echniques ha can pe o m asks ha ypically equi e human in elligence
[13], Da a Science is an in e disciplina y ield ha in ol es he ex ac ion, p ocessing,
analysis, and in e p e a ion o la ge and complex da a se s. In pa icula , deep lea ning
and a ious a ian s o neu al ne wo ks o e a new way o add ess cu en challenges
o ecommende sys ems [14] and beyond [15–19]. Howe e , when using hese black
box algo i hms, he p oblem o missing explainabili y o how he ecommenda ions
a e gene a ed needs o be conside ed [20–22].
Inbook ecommende sys ems, AI algo i hms ha e he ask o sugges ingbooks ha
buye s a e po en ially in e es ed in and ha ha e no been ead by hem. Depending
on how he AI algo i hms a e supposed o ecommend books, a dis inc ion is made
be ween collabo a i e,con en -based, and hyb id book ecommende sys ems. In a
collabo a i e book ecommende sys em, AI algo i hms access all book a ings ha
ha ebeensubmi edbyuse so hebook ecommende sys em.Basedon hesubmi ed
book a ings, he AI algo i hms p edic o each use he a ings o he books hey ha e
no ye a ed. Then, he books wi h he highes p edic ed a ings can be ecommended
o each use [23]. Use s mos ly ha e a ed a e y small p opo ion o he books ha
exis in he da a. Thus, AI algo i hms ha e o p edic mo e o he book a ings han
a e known. In his pape , we ocus on AI algo i hms in collabo a i e ecommende
sys ems.
Two popula AI algo i hms in collabo a i e ecommende sys ems a e he ma ix
ac o iza ion algo i hm using he s ochas ic g adien descen me hod and he book-
based k-nea es -neighbo (knn) algo i hm [24]. In his pape , bo h hese algo i hms
a e conside ed in he amewo k o he modi ied Book-C ossing da a se . This da a
se om Cai–Nicolas Ziegle [25] is a kind o benchma k da a basis o esea ch on
AI algo i hms in collabo a i e ecommende sys ems [26]. We in es iga e a subse
consis ing o 42,137 explici a ings o he Book-C ossing da a se ha o ms he da a
basis. The ask o bo h AI algo i hms is o p edic he unknown book a ings o he
modi ied Book-C ossing da a se , and hen o ecommend he books wi h he highes
p edic ed a ings o each use . By using di e en a ian s o bo h AI algo i hms, his
123
Annals o Da a Science (2024) 11(5):1705–1739 1707
pape aims o e alua e bo h hese algo i hms in ecommending books based on he
modi ied Book-C ossing da a se . Fo his aim, he quali y o bo h AI algo i hms is
measu ed by selec ed e alua ion me ics o AI algo i hms.
This pape is o ganized as ollows. In Sec .2, a e in oducing he basics o collab-
o a i e book ecommende sys ems, a sho o e iew o AI algo i hms and common
e alua ion me ics is gi en. Sec ion3p esen s he book-based knn-algo i hm and he
ma ix ac o iza ion algo i hm using he s ochas ic g adien descen me hod. In Sec .4,
we p o ide a comp ehensi e case s udy based on he Book-C ossing da a se . In pa -
icula , we es ablish modi ica ions o he da a basis, p esen ou me hodology and
p oposed p ocedu e, gi e he esul s o he s udy, and discuss hem in de ail. In he
amewo ko hecases udyweshowhow hequali yo bo hAIalgo i hmsismeasu ed
using selec ed e alua ion me ics. Fo his pu pose, he s a is ical so wa e Rand he
co esponding package Recommende lab, which was de eloped o collabo a i e
ecommende sys ems [27], is used. Finally, Sec .5concludes he pape .
2 AI Algo i hms in Collabo a i e Book Recommende Sys ems
2.1 Essen ials
AI algo i hms in collabo a i e book ecommende sys ems a e equi alen o AI algo-
i hms ha a e gene ally used in en e ainmen ecommende sys ems (e.g., mo ie
ecommende sys ems), whe e he en e ainmen p oduc s a e e e ed o as i ems.
The esea ch on book ecommende sys ems depends e y much on he esea ch low
on en e ainmen ecommende sys ems, and mos o he esul s a e ans e able o
book ecommende sys ems. AI algo i hms need da a abou he book a ings by use s,
which can be explici o implici book a ings. A book a ing is called an explici
book a ing i a use ac i ely assigns a a ing on a speci ic scale (e.g., a scale om
1 o 10, whe e 10 ep esen s he mos posi i e expe ience and 1 is he mos nega i e
expe ience) o a book, see Table 1.
In con as , an implici book a ing is no di ec ly gi en by a use . Ins ead, he book
a ings a e p edic ed based on he use ’s beha io [28]. Fo example, a a ing o 1 is
assigned o a book i a use eads he comple e book whe eas he book ge s a a ing
o 0 i a use only spends a sho ime wi h he book. In he ollowing, howe e , we
ocus on explici book a ings.
Table 1 Example o explici
book a ings by use s (on a scale
om 1 o 10)
Book 1 Book 2 Book 3 Book 4 Book 5
Use 19––1–
Use 2–––64
Use 31010–1–
Use 4–109–9
Use 5–109––
123
1708 Annals o Da a Science (2024) 11(5):1705–1739
A collabo a i e book ecommende sys em includes da a abou pbook a ings.
The e a e nuse s u1,...,unwho ha e a ed books, and mbooks i1,...im ha ha e
ecei ed a a ing by a use . The pbook a ings a e ep esen ed in an n×m use -book
ma ix B,see(2.1).
B=
i1i2··· im
u1
u2
.
.
.
un
⎛
⎜
⎜
⎜
⎝
11 12 ··· 1m
21 22 ··· 2m
.
.
.....
.
.
n1··· ··· nm
⎞
⎟
⎟
⎟
⎠
(2.1)
A use can a e a book only once. E e y ow shows he a ings by one use and
e e y column ep esen s he a ings o one book. Thus, e e y en y in he use -book
ma ix is a a ing by one use o one book. Fo mally his means ha he en y ui is
he a ing by use u o book i. Ra ing all mbooks is he maximum amoun o a ings
a use can gi e. A book can ecei e a maximum o n a ings meaning ha e e y use
has a ed his book.
In his pape , we conside a ing p edic ions o use s who ha e al eady a ed a
leas one book and o books ha ha e al eady ecei ed a leas one a ing by a use .
The e o e, he cold s a p oblem [29] ha deals wi h he ques ion o how o p edic a
a ingo ause o o a book wi hou anyknowledgeabou pas a ings is no add essed.
Mos use s ha e a ed only a small pe cen age o he mbooks, which implies ha B
is a spa se ma ix. The densi y
DB=p
nm ·100
o he use -book ma ix measu es he pe cen age o he known pbook a ings in
ela ion o he heo e ically possible book a ings (i.e., nm a ings). I is impo an o
he AI algo i hms o be able o p edic he la ge numbe o unknown a ings by a small
numbe o known a ings.
I should be no ed ha each use has a di e en iew on he ques ion o which
a ing co esponds o a ce ain book quali y. One use may a gue ha he a ing 6 is
a good a ing on a scale om 1 o 10, whe eas ano he use conside s only a ings
g ea e o equal o 9 as good a ings. The mean u ep esen s he a e age book a ing
by a use . Based on he s anda d de ia ion σ( u)o a use ’s book a ings, conclusions
can be d awn whe he a use has gi en simila book a ings (low s anda d de ia ion)
o a ying book a ings (high s anda d de ia ion). The mean o he a ings o a book
iindica es how well use s ha e a ed he book on a e age. Addi ionally, he s anda d
de ia ion σ( i)desc ibes he size o he a ing ange o a speci ic book. The mean o
all p a ings o he use -book ma ix is gi en by μ.
Based on he gi en book a ings, AI algo i hms p edic he unknown a ings o
he use -book ma ix B. This enables he AI algo i hms o c ea e an o de ed lis o
each use wi h he Nbooks ha ecei ed he highes p edic ion. As a consequence,
he books on he lis a e ecommended o each use . How AI algo i hms deal wi h
di e en pe cep ions o he a ing scale by use s is explained in Sec s.3.1 and 3.2.
123
Annals o Da a Science (2024) 11(5):1705–1739 1709
2.2 Memo y-Based Ve sus Model-Based AI Algo i hms
The AI algo i hms in book ecommende sys ems a e dis inguished be ween memo y-
based and model-based AI algo i hms. Memo y-based AI algo i hms access he en i e
use -book ma ix o ecommend books o use s [23,30]. In con as , model-based
AI algo i hms c ea e a model om he use -book ma ix. Based on his model, use s
ge book ecommenda ions. Wi hin he memo y-based AI algo i hms, a dis inc ion is
made be ween use -based and book-based AI algo i hms. They employ wo di e en
app oaches o o ecas he unknown a ings o he use -book ma ix. Use -based AI
algo i hms p edic he missing a ings o e e y use based on simila use a ings. In
con as , book-based AI algo i hms o ecas he unknown a ings o e e y book by
conside ing he a ings o simila a ed books. Based on he o ecas s o he unknown
book a ingse e yuse ge s hebooks ecommended ha ecei ed hehighes p edic ed
a ings. The knn-algo i hm is a popula memo y-based AI algo i hm in collabo a i e
book ecommende sys ems [23,31].
Model-based AI algo i hms can be classi ied in o he ields o eg ession, clus e -
ing, neu al ne wo ks, deep lea ning and dimensionali y educ ion [31]. Model-based
AI algo i hms a e mos ly dimension- educing algo i hms, and ma ix ac o iza ion
algo i hms a e o en applied in his con ex . The ma ix ac o iza ion algo i hm using
he s ochas ic g adien descen me hod and he ma ix ac o iza ion algo i hm using
he al e na ing leas squa es me hod a e wo popula ma ix ac o iza ion algo i hms
[23,24]. Addi ionally, esea ch on neu al ne wo ks and deep lea ning in collabo a-
i e ecommende sys ems has inc eased signi ican ly in ecen yea s and can also be
applied o collabo a i e book ecommende sys ems [7–9].
2.3 E alua ion Me ics o AI Algo i hms in Book Recommende Sys ems
AI algo i hms in book ecommende sys ems p edic he unknown book a ings o he
use -book ma ix B. Based on he p edic ions, he AI algo i hms sugges Nbooks
o e e y use as an o de ed lis . The quali y o he AI algo i hms depends on he
g ade o sa is ac ion o he use s in ela ion o he p oposed books. Howe e , he
sa is ac ion is ha dly measu able in eali y. Thus, he quali y o he algo i hms can
only be app oxima ed by online o o line es s [28,32]. In he ollowing, we ocus
on o line es s and he co esponding e alua ion measu es. The e alua ion me ics
can be di ided in o he ields o p edic ion accu acy,classi ica ion accu acy and
di e si y [23]. The alues o he e alua ion me ics depend on he cha ac e is ics o
he conside ed da a se (e.g., he ange o he a ing scale). The e o e, i is impo an
o compa e he quali y o di e en AI algo i hms using he same da a se [28].
Spli ,boo s apping and c oss- alida ion a e me hods ha can be used o e alua e
he AI algo i hms [27]. In pa icula , in a c oss- alida ion, use s a e di ided in o a
p ede e mined numbe o g oups o equal size. The numbe o g oups is equi alen o
he numbe o i e a ions pe o med. In each i e a ion, one g oup is he es g oup and
all o he g oups a e conside ed o be he aining g oups. The es g oup is changed in
e e yi e a ion, so ha a e all i e a ions each use wasin he es g oup once. The use s
o he aining g oups a e e e ed o as aining use s and he use s o he es g oup
123

1710 Annals o Da a Science (2024) 11(5):1705–1739
Table 2 O e iew o he classi ica ion accu acy (con usion ma ix)
Recommended No ecommended All
Rele an books m e (TP) m n (FN) m
I ele an books mie (FP) min (TN) mi
All memnm
a e e e ed o as es use s. In each i e a ion, a model is de eloped based on he gi en
a ings o he aining use s. This model is hen es ed on he es use s. In his p ocess,
some o he known a ings a e employed o es he model and some o he known
a ings o he es use s a e wi hheld o alida e he model. Mos o line e alua ion
me ics a e measu ed by he p edic ed alues o he wi hheld a ings. This is mos ly
done by aking he mean o he alues ega ding he o line assessmen me ics in all
i e a ions [27].
P edic ion accu acy me ics measu e how p ecisely an AI algo i hm es ima es he
a ings. The la ge he de ia ion o he p edic ed alue om he ue alue is, he la ge
is he alue o he me ics, and he wo se he p edic ion accu acy o he AI algo i hm.
In he ollowing, we conside h ee common p edic ion accu acy me ics, i.e., he Roo
Mean Squa e E o (RMSE), he Mean Squa e E o (MSE), and he Mean Absolu e
E o (MAE):
RMSE = ui ∈ es (ˆ ui − ui)2
p es
MSE = ui ∈ es (ˆ ui − ui)2
p es
MAE = ui ∈ es |ˆ ui − ui|
p es
He e, es deno es he a ings o he es da a se ha need o be alida ed and p es
deno es hei numbe . The eal a ings a e e e ed o as ui, whe eas ˆ ui ep esen s he
p edic ed a ings o use u o book i.
Classi ica ion accu acy me ics de e mine whe he a use ecei es book ecom-
menda ions ha a e ele an o he use [23]. The la ge he alue o a classi ica ion
accu acy me ic, he la ge he classi ica ion accu acy. P ecision and Recall a e he
wo mos popula classi ica ion accu acy me ics and can be compu ed based on he
con usion ma ix gi en in Table 2:
P ecision =m e
me
Recall =m e
m
No e ha i holds 0 ≤P ecision,Recall ≤1.
123
Annals o Da a Science (2024) 11(5):1705–1739 1711
In Table 2, he ollowing ac onyms a e used: T ue Posi i es (TP), False Nega i es
(FN), False Posi i es (FP), T ue Nega i es (TN). Following Table 2, he ecommended
bu i ele an numbe o books mie co esponds o he ype I e o . Addi ionally, he
numbe o books m n ha is no ecommended bu ele an can be in e p e ed as
ype II e o . P ecision measu es how many o he ecommended books a e ele an ,
whe easRecall speci ies howmanyo he ele an booksa e ecommended. The e o e,
P ecision minimizes he ype I e o , while Recall minimizes he ype II e o . No e ha
minimizing one e o ype inc eases he o he e o ype in many cases. This leads o a
ade-o be ween he maximiza ion o Recall on he one hand and he maximiza ion
o P ecision on he o he hand [23].
Ano he impo an e alua ion me ic is Di e si y. The main idea is ha use s do
no app ecia e o ha e he same books sugges ed o e and o e again. Di e si y can
be measu ed in di e en ways [5,33]. One app oach is o de e mine how many o he
mbooks o he da a se a e ecommended o he use s. This abili y is called Co e age.
I he e a e many books ha a e no ecommended o any use , his could mean a lack
o di e si ica ion. Bobadilla e al. [34] de ined a use ’s co e age as he p opo ion o
books no a ed by he use ha ha e been a ed by one o he use ’s nea es neighbo s
(a use ’s nea es neighbo s a e use s who ha e a ed books simila ly o he conside ed
use ). Yang e al. [23] p esen ed a di e en app oach o measu e Di e si y ha akes
in o accoun he simila i y o books ecommended o one use and he simila i y o
books ecommended o wo di e en use s.
3 AI Algo i hms
3.1 Book-Based knn Algo i hm
The knn algo i hm is a nonpa ame ic algo i hm [35]. In collabo a i e book ecom-
mende sys ems, i is used as a eg ession algo i hm o es ima e he missing alues o
he use -book ma ix B. A dis inc ion is made be ween he use -based and he book-
based knn-algo i hm. We will mainly ocus on he book-based knn-algo i hm in he
ollowing.
Fi s , o e e y book, he simila i y o all o he books is measu ed by a simila i y
measu e. Two books a e conside ed o be simila , i use s ha e gi en hem a simila
a ing. The knea es neighbo s o a book a e he kbooks ha a e mos simila o
he book. E e y book has use s who ha e no a ed he book. Based on one use ’s
a ings o he knea es neighbo s o he book he use ’s a ing o he book can be
p edic ed. The e o e, he use ’s a ings o he knea es neighbo s a e weigh ed wi h
he alue o he co esponding simila i y measu e. Fo example, in Table 1, books 2
and 3 ha e ecei ed simila a ings by use s 4 and 5. Book 2 has been highly a ed by
use 3. The e o e, he p edic ed a ing o use 3 o book 3 could also be high. Wi h his
app oach, he knn algo i hm ies o p edic all unknown book a ings. A e ob aining
he p edic ions, he Nbooks wi h he highes p edic ed a ings a e sugges ed o each
use [23].
123
1712 Annals o Da a Science (2024) 11(5):1705–1739
The simila i y o wo books is measu ed by a simila i y measu e. A common sim-
ila i y measu e is he B a ais-Pea son co ela ion coe icien [36]
wi
gh =u∈U ug − g·( uh − h)
u∈U ug − g2·u∈U( uh − h)2(3.1)
wi h−1≤wi
gh ≤1,whe ewi
gh indica es hesimila i yo hebooksigandihmeasu ed
by he B a ais-Pea son co ela ion coe icien . No e ha only he use s who ha e a ed
bo h books (i.e., u∈U) a e used o he calcula ion. Some books may gene ally ha e
been a ed highe han o he books. This is aken in o accoun by sub ac ing he
espec i e mean. The mean o book igis gi en by gand he mean o book ihis gi en
by h. The e o e, a use ’s a ing coun s as a posi i e a ing only i i exceeds he mean
o he book’s a ings. The denomina o con ains he s anda d de ia ion o he a ings
om he mean. A la ge s anda d de ia ion indica es ha a book has ecei ed di e en
a ings. In con as , a low s anda d de ia ion means ha a book has ecei ed mos ly
he same a ings. In his way he di e ences o he a ings om he mean conside ed
in he nume a o a e scaled. Thus, he B a ais-Pea son co ela ion coe icien akes
in o accoun gene al a ing di e ences be ween books. A alue o “1” implies a high
simila i y be ween wo books, whe eas a alue o “−1” means ha wo books ha e
ecei ed opposi e a ings by use s and a e he e o e no simila . The adjus ed cosine
simila i y and he Euclidean dis ance a e also popula simila i y measu es bu a e no
conside ed in his pape [37].
To imp o e he quali y o he simila i y measu e, he numbe o use s who ha e
a ed bo h books could be conside ed. This ensu es ha books a e only coun ed as
simila i hey ha e been a ed simila ly by mul iple use s [23,37]. This app oach can
be ep esen ed o he calcula ion o he simila i y o wo books as ollows [23]:
wi
gh =2·|Ug∩Uh|
|Ug|+|Uh|·wi
gh (3.2)
In (3.2), |Ug| ep esen s he numbe o use s who ha e a ed he book ig,|Uh|indica es
how many use s ha e a ed he book ihand |Ug∩Uh|co esponds o he numbe o
use s who ha e a ed bo h books. The ac ion ge s smalle i ewe use s ha e a ed
bo hbooks. Mul iplying he ac ion by wi
gh ensu es ha e y ewcommonuse a ings
esul in a lowe simila i y o wo books [23].
A e calcula ing he simila i y measu es o all books, he knea es neighbo s
a e de e mined o each book. The knea es neighbo s o a book a e he books ha
ha e he highes simila i y alue [23]. The unknown a ings o use s o a book a e
de e mined by weigh ing he a ings o hese use s a he knea es neighbo s o he
book wi h he simila i y measu e [36,37]:
P(ug)= g+σ( g)·ga∈N(g) uga− ga
σ( ga)·wi
gga
ga∈N(g)wi
gga
(3.3)
123
Annals o Da a Science (2024) 11(5):1705–1739 1713
He e, P(ug)is he p edic ion o he a ing o a use o book ig.Thekmos simila
books o book iga ein hese N(g). These books a e deno ed by ga,...,gk.Only
he books ha ha e been a ed by a gi en use a e conside ed as nea es neighbo s
in he p edic ion. To accoun o di e ences in a ings be ween books, he a ings o
he books a e no malized. Addi ionally, he no malized a ings a e weigh ed by he
simila i y measu e wi
gga.
The use ’s weigh ed no malized a ings o he books a e ans o med in o he
book’s a ing scale by mul iplying he s anda d de ia ion σ( g)o he book’s a ings.
The esul ing alue is added o he mean go book ig. I book ighas a la ge s anda d
de ia ion, he mul iplica ion by he s anda d de ia ion ensu es ha a posi i e alue
should cause a g ea e de ia ion o he p edic ed alue om he mean go book ig.
Thisapp oachisknownas z-sco e.Ano he commonapp oachisgi enby hede ia ion
om he mean. This app oach does no ake in o accoun he s anda d de ia ion o he
conside ed book and he book’s nea es neighbo s [36,37]. In his way, he unknown
a ings o use s a e es ima ed o each book. As explained la e in his sec ion, i may
no be possible o p edic all a ings. Fo each use , he books no ye a ed by he use
a e so ed in a descending o de acco ding o he heigh o he p edic ed a ing.
The book-based knn algo i hm can be classi ied as ei he a memo y-based [23,31]
o a model-based AI algo i hm [38]. I depends on whe he each ime a lis o book
ecommenda ions is c ea ed o a use , he simila i y measu es a e ecalcula ed using
he use -book ma ix. I his is ue, he book-based knn algo i hm is a memo y-based
AI algo i hm. In con as , he book-based knn algo i hm can be conside ed as a model-
based algo i hm i he simila i y measu es a e ecompu ed only a egula in e als.
The model is he simila i y ma ix ha con ains he simila i y be ween he books.
The classi ica ion o he book-based knn algo i hm as a model-based AI algo i hm
is suppo ed by esea ch esul s showing ha he simila i ies be ween he books a e
s able o e ime [38].
Thequali yo he knnalgo i hm dependson he choiceo hesimila i ymeasu e and
he possible conside a ion o he numbe o common use s o wo books. Addi ionally,
he choice o he numbe o knea es neighbo s plays an impo an ole: choosing a
small numbe o nea es neighbo s could esul in an o e i ing o he a ings o he
nea es neighbo s [28]. Mo eo e , he e is a isk ha he nea es neighbo s ha e no
ecei edany a ings by he use . This implies ha i is no possible o p edic a a ing o
he book. Co e age (see Sec .2.3) is a measu e o de e mine he ex en o he p oblem
[34,36]. In con as , choosing a la ge numbe o nea es neighbo s could lead o he
p oblem o unde i ing [35]. Use ’s a ings o books ha a e no simila enough o
he book migh in luence he p edic ion oo much. In ex eme cases his could lead o
unsa is ac o y book ecommenda ions. The e o e, i is o en sugges ed o ake a alue
in he ange be ween 20 and 50 o he numbe o knea es neighbo s o sol e he
ade-o be ween o e i ing and unde i ing [36,37].
3.2 Ma ix Fac o iza ion Algo i hm Using he S ochas ic G adien Descen Me hod
The ma ix ac o iza ion algo i hm using he s ochas ic g adien descen me hod is a
model-based AI algo i hm. The main assump ion behind he ma ix ac o iza ion is
123
1720 Annals o Da a Science (2024) 11(5):1705–1739
examine he impac o di e si y in book ecommenda ions on use ’s sa is ac ion wi h
he ecommended books. Adamopoulos/Tuzhilin [42] conside ed di e en subse s o
he da a se and p oposed unexpec edness as an addi ional oppo uni y o imp o e
he ecommenda ion quali y. Pa k/Tuzhilin [43] p esen ed an app oach o sol ing he
long- ail p oblem (e.g., how o deal wi h books wi h ew a ings) o ecommende
sys ems. Deldjoo e al. [44] used he da a se as pa o hei s udy which examined
he in luence o da a cha ac e is ics on he accu acy and ai ness (e.g., measu ing o
wha ex en he quali y o he ecommenda ion depends on being in a speci ic g oup
as age) o ecommende sys ems.
4.2 P ocedu e and Me hodology
In his sec ion, he me hodology and he p ocedu e o analyze he quali y o he
book-based knn-algo i hm and he ma ix ac o iza ion algo i hm using he s ochas ic
g adien descen me hod o ecommend books om he modi ied Book-C ossing da a
se is p esen ed.
4.2.1 P ocedu e
On he one hand, he quali y o 31 a ian s o book-based knn-algo i hms, in which he
numbe o knea es neighbo s is a ied om 20 o 50, is measu ed. On he o he hand,
he quali y o 11 a ian s o he ma ix ac o iza ion algo i hm using he s ochas ic
g adien descen me hod, in which he numbe o la en ac o s is a ied om 5 o 15,
is measu ed. In o de o measu e he quali y o he a ian s o bo h algo i hms, he
alues o he p edic ion accu acy me ics RMSE, MSE, and MAE and o he classi-
ica ion me ics P ecision and Recall (see Sec .2.3) a e conside ed. As an addi ional
check on he quali y o he a ian s, hey a e compa ed wi h he alues o he e alua-
ion me ics in a “ andom” algo i hm ( ecommends books andomly) and a “popula ”
algo i hm ( ecommends equen ly a ed books). The alues o he e alua ion me -
ics a e de e mined using he R-package Recommende lab ha was de eloped by
Michael Hahsle o es and e alua e collabo a i e ecommende sys ems [27].
To compa e he a ian s o bo h AI algo i hms, hey need o ha e he same ain-
ing and es da a se . Addi ionally, mo e han one aining da a se and one es
da a se should be used o he e alua ion. This may educe he isk ha he di i-
sion in o a aining and a es da a se would a ec he quali y o he algo i hms.
To sa is y hese impo an equi emen s o quali y compa abili y, he op ion o he
Recommende lab package o de e mine an e alua ion scheme is employed. Using
he command se .seed ensu es ha he e alua ion scheme is he same o all es ed
a ian s.
We apply c oss- alida ion wi h 10 pa i ions and 10 i e a ions o each a ian o
bo h AI algo i hms. The e o e, he 1842 use s o he modi ied Book-C ossing da a se
we e di ided in o 10 pa i ions consis ing o abou 184 use s. In each i e a ion, use s
o 9 pa i ions o m he aining da a se and de elop a model. This model is es ed
using he es da a se , which consis s o he use s o one pa i ion. Thus, each use is
nine imes in he aining da a se and once in he es da a se (see Table 6).
123

Annals o Da a Science (2024) 11(5):1705–1739 1721
Table 6 O e iew o he se ings a he a ing scheme
Me hod Pe cen age o aining use s I e a ions Gi en GoodRa ing
C oss- alida ion 90% 10 9 9
4.2.2 E alua ion Me ics
To be able o measu e he e alua ion me ics, he a ing scheme uses he op ion Gi en
o speci y how many o a es use ’s known a ings should be used o es ing and how
many should be used o alida ion. The alue “9” is se o Gi en. Thus, om
each es use , 9 o he known a ings a e u ilized o es he model de eloped by he
aining da a se . Based on he es ima ion o he emaining known a ings o he es
use s, he algo i hms a e alida ed. Since each use has submi ed a leas 10 a ings
in he modi ied Book-C ossing da a se , a leas one a ing is used o alida ion o
e e y es use . The p edic ion accu acy me ics RMSE, MSE, and MAE a e measu ed
by he known a ings used o alida ion. A e 10 i e a ions, he p edic ion accu acy
me ics a e de e mined as he mean o hei alues om hese i e a ions.
To measu e he classi ica ion accu acy me ics P ecision and Recall, he alue
o GoodRa ing was decisi e o he a ing scheme. The alue o GoodRa ing
indica es he a ing om which on a book belonging o a alida ing a ing is so ele an
o a es use ha i should be ecommended o he es use . This is a hypo he ical
assump ion since, in eali y, he use has al eady a ed he book. The book a ings o
he modi ied Book-C ossing da a se a e on a scale o 1 o 10, whe e a alue o “9”
is chosen o GoodRa ing. This choice is based on he assump ion ha he use
would like o ecei e a ecommenda ion o a book ha he use has a ed 9 o 10.
Addi ionally, his assumes ha he use would no know he book ye . The numbe
o a ings o alida e wi h a a ing o 9 o 10 is de e mined o each es use . Then,
o each es use , he a ings o he 2056 emaining books a e p edic ed, since 9 o
he known a ings a e used o es he model. O he 2056 books, each es use was
ecommended once he 10 and once he 20 books wi h he highes p edic ed a ings.
The main assump ion o choosing he wo lis sizes is ha a use would mos ly only
look a he ecommenda ions placed a he op o he lis .
Fo each es use , P ecision is measu ed as he p opo ion o books in he lis om
ecommended books ha we e p e iously de e mined o be ele an books o alida e.
Recall is measu ed as he p opo ion o he ele an books o be alida ed. A e 10
i e a ions, he alues o Recall and P ecision a e aken as he mean o he esul s om
he 10 i e a ions.
4.2.3 Book-Based knn Algo i hm
P e ious esea ch on collabo a i e ecommende sys ems (as discussed in Sec .3.1)
conside sanumbe o 20–50knea es neighbo sasop imum.Basedon hissugges ion,
31 a ian s wi h alues o k om k=20 o k=50 o he book-based knn algo i hm
a e es ed.
123
1722 Annals o Da a Science (2024) 11(5):1705–1739
Table 7 O e iew o he a ian s o he book-based knn-algo i hm
na.as.ze o kNo malize Me hod no malize_sim_ma ix alpha
FALSE 20–50 z-sco e Pea son FALSE 0.5
Table 8 O e iew o he a ian s o he ma ix ac o iza ion algo i hm using he s ochas ic g adien descen
me hod
kγλmin_epochs max_epochs min_imp o emen No malize
5–15 0.001 0.015 50 200 0.000001 z-sco e
The B a ais-Pea son co ela ion coe icien (3.1) is chosen as simila i y measu e.
Fo no maliza ion, he z-sco e app oach (3.3) is used. The unknown book a ings
a e no se o 0 (op ion na.as.ze o) because he simila i y o wo books in he
B a ais-Pea son co ela ion coe icien is only based on he use s who a ed bo h
books. The meaning o alpha is no de ined in he Recommende lab package
and ela ed ins uc ions, so he alue was le a he de aul alue o 0.5 (no e ha
p e- es s showed no change in he sco ing me ics a di e en alues o alpha). The
op ion o no malize he simila i y ma ix o he books is no se , as gene al di e ences
in he a ings a e al eady aken in o accoun when calcula ing he simila i ies o he
books (see Sec .3.1). Fo an o e iew o he se ings see Table 7.
4.2.4 Ma ix Fac o iza ion Algo i hm Using he S ochas ic G adien Descen Me hod
Fo he a ian s o he ma ix ac o iza ion algo i hm using he s ochas ic g adien
descen me hod, he numbe o la en ac o s is a ied om 5 o 15 la en ac o s. Funk
[40], as he ounde o he me hod, s a ed in his blog en y 25 and 40 as alues o a
easonable numbe o la en ac o s o he Ne lix da a se , whe e he use -i em ma ix
has a size o 8.5 billion en ies. Ko en e al. [24] men ioned a numbe o 20 o 100
la en ac o s o he same da a se . Since he modi ied Book-C ossing da a se has
app oxima ely 3.8 million en ies, alues be ween 5 and 15 a e chosen o he numbe
o la en ac o s.
Fo no maliza ion, he z-sco e is used, as o he a ian s o he book-based knn-
algo i hm. The z-sco e is chosen, since he common app oach o conside ing he bias
bu, he bias bi, and he mean o all known a ings μin he p edic ion o he a ings
could no be selec ed, see Sec .3.2. The o he pa ame e s (see Table 8) a e le a he
de aul alues [27,40].
4.3 Resul s
In his sec ion, he esul s om all a ian s o bo h AI algo i hms a e p esen ed o
he p edic ion accu acy me ics RMSE, MSE, and MAE as well as he classi ica ion
accu acy me ics P ecision and Recall. In o de o addi ionally check he quali y o
heseAIalgo i hmsc i ically, heya e also compa ed wi h he esul s o he wocon ol
123
Annals o Da a Science (2024) 11(5):1705–1739 1723
Table 9 Compa ison o he
alues o he bes a ian s o
bo h AI algo i hms o RMSE,
MSE, and MAE wi h he alues
o he con ol algo i hms
RMSE MSE MAE
Random_1 2.092 4.383 1.563
Popula _1 1.562 2.443 1.178
SVDF_15 1.555 2.420 1.148
kNN_24 2.040 4.180 1.396
kNN_50 2.036 4.157 1.407
algo i hms“popula ” and “ andom”. The esul s a e ounded o he hi ddecimalplace.
The alues o he e alua ion me ics o all a ian s o bo h AI algo i hms (see Tables
12–16) and he wo con ol algo i hms “popula ” and “ andom” (see Tables 17–18)
can be ound in Appendix B.
Fo he ma ix ac o iza ion algo i hm, he a ian wi h 15 la en ac o s ecei ed
he lowes alues (RMSE =1.555, MSE =2.420, MAE =1.148, see Table 9) and
he a ian wi h 5 la en ac o s he highes alues (RMSE =1.560, MSE =2.437,
MAE =1.159, see Table 12) o all h ee p edic i e accu acy me ics. In each case, he
inc ease in a la en ac o sligh ly imp o ed he p edic i e accu acy me ics. The e o e,
he di e ence be ween he wo s and he bes a ian is abou 0.005 o RMSE, abou
0.017 o MSE, and abou 0.011 o MAE. Fo he book-based knn algo i hm, he
a ian wi h 24 nea es neighbo s ecei ed he bes alue o he p edic ion accu acy
me ic MAE wi h a alue o 1.396. The a ian wi h 50 nea es neighbo s pe o med
he bes o he p edic i e accu acy me ics RMSE (2.036) and MSE (4.157) (see Table
9). Table 9and Fig. 3show he a ian s o bo h AI algo i hms wi h he bes alues o
RMSE, MSE and MAE and he alues o he con ol algo i hms.
The ma ix ac o iza ion algo i hm using he s ochas ic g adien descen me hod
achie ed o all a ian s lowe alues in he p edic ion accu acy me ics compa ed
o all a ian s o he book-based knn algo i hm and han he con ol algo i hms. All
a ian s o he book-based knn algo i hm had lowe sco es on he h ee p edic i e
accu acy me ics han he con ol algo i hm “ andom”. Compa ed o he con ol algo-
i hm “popula ”, all a ian s o he book-based knn algo i hm had highe alues. The
minimum di e ence be ween he p edic ion accu acy me ics o bo h AI algo i hms
is 0.476 o RMSE, 1.72 o MSE, and 0.248 o MAE. The maximum di e ence is
0.506 o RMSE, 1.854 o MSE, and 0.262 o MAE (see Tables 12 and 14).
In he ollowing, he esul s o P ecision and Recall a e discussed o bo h a lis o
10 and 20 ecommended books. As o he book-based knn algo i hm, o he lis wi h
10 ecommended books, he a ian wi h 29 nea es neighbo s achie ed he bes alues
o P ecision (0.006) and Recall (0.017). In con as , o he lis o 20 ecommended
books, he a ian wi h 22 nea es neighbo s ecei ed he highes alue o P ecision
(0.006), and he a ian wi h 26 nea es neighbo s ecei ed he highes alue o Recall
(0.024) (see Tables 10–11). Fo he ma ix ac o iza ion algo i hm using he s ochas ic
g adien descen me hod, he a ian wi h 5 la en ac o s ob ained he highes alues
o P ecision (0.016) and Recall (0.035) o he lis o 10 ecommended books. Fo he
lis o 20 ecommended books, he a ian wi h 5 la en ac o s a P ecision (0.013) and
he a ian wi h 9 la en ac o s a Recall (0.060) pe o med bes (see Tables 10–11).
123
1724 Annals o Da a Science (2024) 11(5):1705–1739
Fig. 3 G aphical compa ison o he alues o he bes a ian s o bo h AI algo i hms o RMSE, MSE, and
MAE wi h he alues o he con ol algo i hms
Table 10 Compa ison o he alues o he bes a ian s o bo h algo i hms o P ecision and Recall a he
op 10 lis wi h he alues o he con ol algo i hms
TP FP FN TN NP ecision Recall
Random_1 0.020 9.980 5.740 2040.260 2056 0.002 0.003
Popula _1 0.249 9.751 5.511 2040.489 2056 0.025 0.059
SVDF_5 0.158 9.842 5.603 2040.397 2056 0.016 0.035
kNN_29 0.061 9.751 5.699 2040.489 2056 0.006 0.017
Table 11 Compa ison o he alues o P ecision and Recall o he bes a ian s o bo h AI algo i hms a
he op 20 lis wi h he alues o he con ol algo i hms
TP FP FN TN NP ecision Recall
Random_1 0.054 19.946 5.706 2030.294 2056 0.003 0.010
Popula _1 0.360 19.640 5.400 2030.600 2056 0.018 0.082
SVDF_5 0.264 19.736 5.496 2030.504 2056 0.013 0.057
SVDF_9 0.264 19.736 5.496 2030.504 2056 0.013 0.060
kNN_22 0.111 19.511 5.649 2030.729 2056 0.006 0.023
kNN_26 0.104 19.519 5.656 2030.720 2056 0.005 0.024
All a ian s o he ma ix ac o iza ion algo i hm using he s ochas ic g adien
descen me hod had highe alues o P ecision and Recall han all a ian s o he
book-based knn algo i hm. Addi ionally, all a ian s o bo h AI algo i hms had highe
alues o P ecision and Recall han he con ol algo i hm andom (see Tables 10–11).
Fo he lis o 10 ecommended books, he minimum di e ence be ween bo h AI
algo i hms o P ecision is 0.007, while he minimum di e ence o Recall is 0.012.
The maximum di e ence is 0.011 o P ecision and 0.021 o Recall. Fo he lis o 20
ecommended books, he minimum di e ence be ween he AI algo i hms is 0.006 o
P ecision and 0.029 o Recall. The maximum di e ence is 0.008 o P ecision and
0.038 is Recall (see Tables 13,15,16).
123
Annals o Da a Science (2024) 11(5):1705–1739 1725
Fig. 4 G aphical compa ison o he alues o he bes a ian s o bo h AI algo i hms o P ecision and
Recall
Fo bo h AI algo i hms, a highe alue o Recall is obse ed o all a ian s o
he lis o 20 ecommended books. The alue o P ecision is highe o all a ian s
in he ma ix ac o iza ion algo i hm using he s ochas ic g adien descen me hod o
he lis wi h 10 ecommended books. Fo he book-based knn algo i hm, his is ue
o mos a ian s, al hough he di e ence is much smalle he e. Figu e 4shows his
endency by looking a he a ian s o he AI algo i hms ha sco ed he highes o
Recall o P ecision.
4.4 Discussion
Fo he chosen se ings o bo h AI algo i hms, all a ian s o he ma ix ac o iza ion
algo i hm using he s ochas ic g adien descen me hod show supe io pe o mance
compa ed o all conside ed a ian s o he book-based knn algo i hm.
The ma ix ac o iza ion algo i hm using he s ochas ic g adien descen me hod
led o be e esul s o he p edic ion accu acy me ics compa ed o bo h con ol algo-
i hms. As o P ecision and Recall, he ma ix ac o iza ion algo i hm showed a be e
pe o mance han he “ andom” algo i hm and a wo se pe o mance han he “popu-
la ” algo i hm. Thus, he quali y o he ma ix ac o iza ion algo i hm applied o he
modi ied Book-C ossing da a se can be conside ed as good. The a ian s o he knn
algo i hm led o be e sco es han he con ol algo i hm “ andom” and wo se sco es
han he con ol algo i hm “popula ” on he p edic ion accu acy and classi ica ion
accu acy me ics. One eason o he poo pe o mance o he book-based knn algo-
i hm could be a possibly low co e age o he books o he modi ied Book-C ossing
da a se [34]: he co e age o a book is he p opo ion o use s who ha e no a ed a
book and a he same ime ha e a ed one o he knea es neighbo s o he book (see
123

1726 Annals o Da a Science (2024) 11(5):1705–1739
Sec .2.3). Low co e age means a high p obabili y ha i a use has no a ed a book,
he use has a ed only a e y small ac ion o he knea es neighbo s, o in ex eme
cases, none o he knea es neighbo s o he book. In he i s case, he p oblem o
o e i ing he p edic ed a ing o he use ’s a ing a he ew knea es neighbo s may
occu [28]. In he second case, no p edic ion can be made o he a ing. The good
esul s o he con ol algo i hm “popula ” on he classi ica ion accu acy me ics ecall
and p ecision and on he p edic ion accu acy me ics migh be ela ed o he ac ha
use s migh like books ha ha e been a ed by many use s. Based on Fig.4, whe e he
bes alues o bo h AI algo i hms o he lis s o 10 and 20 ecommended books a e
plo ed, he ade-o be ween a high alue o Recall and a high alue o P ecision
desc ibed in Sec .2.3 can be seen.
Mo eo e , he compu a ion o P ecision and Recall in he Recommende lab
package can be conside ed o be c i ical: when c ea ing he lis o ecommended
books o a es use , he 10 o 20 books wi h he highes p edic ed a ings we e
ecommended. This in ol es p edic ing a ings o books wi h a a ing o be alida ed
and a ings o books whe e he ue a ing is unknown. He e, an AI algo i hm lead
o a high alue o Recall i a la ge p opo ion o he books a e ecommended wi h a
ele an a ing o alida e and a high alue o P ecision i a la ge p opo ion o he
ecommended books a e books wi h a ele an a ing o alida e (see Sec .4.2). He e,
i is no possible o s a e wi h ce ain y, whe he he use migh ind he ecommended
books wi h he p edic ed a ing mo e in e es ing han he a ings o be alida ed o
he books o which he use has gi en a a ing o 9 o 10. The alues o Recall and
P ecision we e he e o e calcula ed bu should only be in e p e ed wi h cau ion.
5 Conclusions
In his pape , we in es iga ed he pe o mance o wo popula AI algo i hms o col-
labo a i e book ecommende sys ems using he Book-C ossing benchma k da a se .
We implemen ed di e en a ian s o he book-based knn algo i hm and he ma ix
ac o iza ion algo i hm using he s ochas ic g adien descen me hod based on selec ed
p edic ion and classi ica ion accu acy me ics as well as using wo con ol algo i hms.
These a ian s a e cha ac e ized by a ia ions in he numbe o knea es neighbo s
in he book-based knn algo i hm and in he numbe o jla en ac o s in he ma ix
ac o iza ion algo i hm using he s ochas ic g adien descen me hod. We pe o med a
comp ehensi ecase s udy o analyze he quali yo bo hAIalgo i hms o collabo a i e
book ecommende sys ems o ecommend books om he modi ied Book-C ossing
da a se .
Fo he in es iga ed a ian s o bo h AI algo i hms, he a ian s o he ma ix ac-
o iza ion algo i hm using he s ochas ic g adien descen me hod showed supe io
pe o mance. In con as , he book-based a ian s pe o med wo se han he a ian s
o he ma ix ac o iza ion algo i hm using he s ochas ic g adien descen me hod
and han he con ol algo i hm “popula ”. I seems ha he poo pe o mance o he
book-based knn algo i hm migh be ela ed o he p oblem o poo co e age o he
book-based knn algo i hm.
123
Annals o Da a Science (2024) 11(5):1705–1739 1727
This pape conside ed use s who ha e al eady a ed books and books ha ha e
al eady ecei ed a ings. Fo AI algo i hms, he e is also he ques ion o how o deal
wi h new use s who ha e no ye submi ed a ings and new books ha ha e no ye
ecei ed a ings. This p oblem is known as he cold s a p oblem. I deals wi h he
ques ion o which books a e sugges ed o a new use and o which use s a new book is
sugges ed. Ano he in e es ing ques ion is how AI algo i hms deal wi h he g ey sheep
p oblem. This p oblem deals wi h use s whose a ing beha io is di icul o explain
by any pa e ns, which makes i e y di icul o AI algo i hms o ecommend sui able
books o hem.
In addi ion o hese aspec s, u u e esea ch could ocus on he pe o mance o bo h
algo i hms when es ed on o he ecen book da a se s such as he Goodbooks-10k
da a se [45] and he Good eads da a se [46,47], which a e also equen ly used in
esea ch abou book ecommende sys ems [6,21].1
Acknowledgemen s The au ho s hank bo h anonymous e iewe s o hei aluable eedback and sugges-
ions, which we e impo an and help ul o imp o e he pape .
Au ho con ibu ions ClemensTege meie :Concep ualiza ion, Me hodology, So wa e, Valida ion, Fo mal
analysis, In es iga ion, W i ing - O iginal D a , Visualiza ion A ne Johannssen: Valida ion, Fo mal analy-
sis, In es iga ion, W i ing - O iginal D a , W i ing - Re iew & Edi ing, Supe ision, P ojec adminis a ion
Na aliya Chukh o a: Valida ion, Fo mal analysis, In es iga ion, W i ing - Re iew & Edi ing, Supe ision.
Funding Open Access unding enabled and o ganized by P ojek DEAL.
Da a a ailibili y The da a ha suppo he indings o his s udy a e a ailable om he espec i e e e ences
as men ioned in he main ex .
Code A ailabili y The code is a ailable om he au ho s upon eques .
Decla a ions
Compliance wi h E hical S anda ds This a icle does no con ain any s udies wi h human pa icipan s o
animals pe o med by he au ho s.
Con lic o in e es The au ho s decla e ha hey ha e no known compe ing inancial in e es s o pe sonal
ela ionships ha could ha e appea ed o in luence he wo k epo ed in his pape .
Disclosu e o po en ial con lic s o in e es (i) This manusc ip is he au ho s’ o iginal wo k, which has no
been published no submi ed simul aneously elsewhe e; (ii) all au ho s ha e checked he manusc ip and
ag eed o he submission, and (iii) he e is no con lic o in e es .
Open Access Thisa icleislicensedunde aC ea i eCommonsA ibu ion4.0In e na ionalLicense,which
pe mi s use, sha ing, adap a ion, dis ibu ion and ep oduc ion in any medium o o ma , as long as you gi e
app op ia e c edi o he o iginal au ho (s) and he sou ce, p o ide a link o he C ea i e Commons licence,
and indica e i changes we e made. The images o o he hi d pa y ma e ial in his a icle a e included
in he a icle’s C ea i e Commons licence, unless indica ed o he wise in a c edi line o he ma e ial. I
ma e ial is no included in he a icle’s C ea i e Commons licence and you in ended use is no pe mi ed
by s a u o y egula ion o exceeds he pe mi ed use, you will need o ob ain pe mission di ec ly om he
copy igh holde . To iew a copy o his licence, isi h p://c ea i ecommons.o g/licenses/by/4.0/.
1The da a se s a e a ailable a h p:// as ml.com/goodbooks-10k-a-new-da ase - o -book-
ecommenda ions/ and h ps://si es.google.com/eng.ucsd.edu/ucsdbookg aph/home?pli=1.
123
1728 Annals o Da a Science (2024) 11(5):1705–1739
Appendix A
See Figs. 5,6,7,8,9,10,11 and 12.
Fig. 5 F equencies o he numbe o a ings pe use in he modi ied Book-C ossing da a se ( he use wi h
964 submi ed a ings is no included in his igu e)
Fig. 6 F equencies o he numbe o a ings pe book in he modi ied Book-C ossing da a se
123
Annals o Da a Science (2024) 11(5):1705–1739 1729
Fig. 7 F equencies o he medians o one use ’s a ings in he modi ied Book-C ossing da a se
Fig. 8 F equencies o he medians o one book’s a ings in he modi ied Book-C ossing da a se
123
1736 Annals o Da a Science (2024) 11(5):1705–1739
Table 16 Values o P ecision and Recall o he book-based knn-algo i hm o 36–50 neighbo s
TP FP FN TN NP ecision Recall n
kNN_36 0.054 9.758 5.706 2040.482 2056 0.005 0.014 10
kNN_36 0.099 19.524 5.661 2030.716 2056 0.005 0.022 20
kNN_37 0.054 9.758 5.706 2040.482 2056 0.006 0.014 10
kNN_37 0.101 19.523 5.660 2030.717 2056 0.005 0.022 20
kNN_38 0.052 9.760 5.708 2040.480 2056 0.005 0.014 10
kNN_38 0.098 19.525 5.662 2030.715 2056 0.005 0.022 20
kNN_39 0.051 9.761 5.709 2040.479 2056 0.005 0.014 10
kNN_39 0.099 19.525 5.661 2030.715 2056 0.005 0.022 20
kNN_40 0.052 9.760 5.708 2040.480 2056 0.005 0.014 10
kNN_40 0.097 19.527 5.663 2030.713 2056 0.005 0.022 20
kNN_41 0.053 9.759 5.707 2040.481 2056 0.005 0.015 10
kNN_41 0.098 19.526 5.662 2030.714 2056 0.005 0.023 20
kNN_42 0.056 9.755 5.704 2040.484 2056 0.006 0.016 10
kNN_42 0.099 19.524 5.661 2030.716 2056 0.005 0.023 20
kNN_43 0.058 9.754 5.703 2040.485 2056 0.006 0.016 10
kNN_43 0.101 19.523 5.660 2030.717 2056 0.005 0.024 20
kNN_44 0.058 9.754 5.702 2040.486 2056 0.006 0.016 10
kNN_44 0.099 19.524 5.661 2030.716 2056 0.005 0.023 20
kNN_45 0.059 9.753 5.702 2040.487 2056 0.006 0.016 10
kNN_45 0.101 19.523 5.660 2030.717 2056 0.005 0.023 20
kNN_46 0.060 9.752 5.700 2040.488 2056 0.006 0.016 10
kNN_46 0.102 19.522 5.658 2030.718 2056 0.005 0.024 20
kNN_47 0.058 9.754 5.703 2040.485 2056 0.006 0.016 10
kNN_47 0.103 19.521 5.658 2030.719 2056 0.005 0.024 20
kNN_48 0.059 9.753 5.702 2040.487 2056 0.006 0.015 10
kNN_48 0.104 19.520 5.656 2030.720 2056 0.005 0.024 20
kNN_49 0.059 9.753 5.702 2040.487 2056 0.006 0.015 10
kNN_49 0.103 19.521 5.658 2030.719 2056 0.005 0.023 20
kNN_50 0.057 9.755 5.703 2040.485 2056 0.006 0.015 10
kNN_50 0.101 19.523 5.659 2030.717 2056 0.005 0.023 20
123

Annals o Da a Science (2024) 11(5):1705–1739 1737
Table 17 Values o RMSE,
MSE, and MAE o he con ol
algo i hms “ andom” and
“popula ”
RMSE MSE MAE
Random_1 2.092 4.383 1.563
Popula _1 1.562 2.443 1.178
Table 18 Values o P ecision and Recall o he con ol algo i hms “ andom” and “popula ”
TP FP FN TN NP ecision Recall n
Random_2 0.020 9.980 5.740 2040.260 2056 0.002 0.003 10
Random_2 0.054 19.946 5.706 2030.294 2056 0.003 0.010 20
Popula _2 0.249 9.751 5.511 2040.489 2056 0.025 0.059 10
Popula _2 0.360 19.640 5.400 2030.600 2056 0.018 0.082 20
Re e ences
1. Bö senbla (2021) Online-Buchhandel mi hohen Wachs ums a en. h ps://www.boe senbla .ne /
news/buchhandel-news/online-buchhandel-mi -hohen-wachs ums a en-185281
2. Adoma icius G, Tuzhilin A (2011) Con ex -awa e ecommende sys ems. In: Ricci F, Rokach L,
Shapi a B, Kan o PB (eds) Recommende sys ems handbook. Sp inge , Bos on, pp 217–253
3. Villegas NM, Sánchez C, DÍaz-Cely J, Tamu a G (2018) Cha ac e izing con ex -awa e ecommende
sys ems: a sys ema ic li e a u e e iew. Knowl Based Sys 140:173–200. h ps://doi.o g/10.1016/j.
knosys.2017.11.003
4. Raza S, Ding C (2019) P og ess in con ex -awa e ecommende sys ems—an o e iew. Compu Sci
Re 31:84–97. h ps://doi.o g/10.1016/j.cos e .2019.01.001
5. Kuna e M, PoŽ l T (2017) Di e si y in ecommende sys ems—a su ey. Knowl Based Sys 123:154–
162. h ps://doi.o g/10.1016/j.knosys.2017.02.009
6. Ramak ishnan G, Saicha an V, Chand aseka an K, Ra hnamma MV, Ramana VV (2020) Collabo a i e
il e ing o book ecommenda ion sys em. In: Das KN, Bansal JC, Deep K, Naga AK, Ponnambalam
P, Naidu RC (eds) So compu ing o p oblem sol ing. Sp inge , Singapo e, pp 325–338
7. Da’u A, Salim N (2019) Recommenda ion sys em based on deep lea ning me hods: a sys ema ic e iew
and new di ec ions. A i In ell Re 53(4):2709–2748. h ps://doi.o g/10.1007/s10462-019-09744-1
8. Zhang S, Yao L, Sun A, Tay Y (2019) Deep lea ning based ecommende sys em: a su ey and new
pe spec i es. ACM Compu Su 52:1. h ps://doi.o g/10.1145/3285029
9. Wu L, He X, Wang X, Zhang K, Wang M (2021) A su ey on neu al ecommenda ion: om collabo-
a i e il e ing o con en and con ex en iched ecommenda ion. a Xi p ep in a Xi : 2104.13030
10. Olson DL, Shi Y (2007) In oduc ion o business da a mining. McG aw-Hill/I win, New Yo k
11. Shi Y, Tian Y, Kou G, Peng Y, Li J (2011) Op imiza ion based da a mining: heo y and applica ions.
Sp inge , Be lin
12. Shi Y (2022) Ad ances in big da a analy ics. h ps://doi.o g/10.1007/978-981-16-3607-3
13. Tien JM (2017) In e ne o hings, eal- ime decision making, and a i icial in elligence. Ann Da a Sci
4:149–178. h ps://doi.o g/10.1007/s40745-017-0112-5
14. Ba maz Z, Yu ekli A, Bilge A, Kaleli C (2018) A e iew on deep lea ning o ecommende sys ems:
challenges and emedies. A i In ell Re 52(1):1–37. h ps://doi.o g/10.1007/s10462-018-9654-y
15. Foe chB enesR, JohannssenA, Chukh o aN(2022) Anin elligen bank up cy p edic ionmodel using
a mul ilaye pe cep on. In ell Sys Appl 16:200136. h ps://doi.o g/10.1016/j.iswa.2022.200136
16. Yeganeh A, Shadman A, Abbasi SA, Pou panah F, Johannssen A, Chukh o a N (2022) An ensemble
neu al ne wo k amewo k o imp o ing he de ec ion abili y o a base con ol cha in non-pa ame ic
p o ile moni o ing. Expe Sys Appl 204:117572. h ps://doi.o g/10.1016/j.eswa.2022.117572
17. Ku ani A, Doshi P, Vakha ia A, Shah M (2023) A comp ehensi e compa a i e s udy o a i icial neu al
ne wo k (ANN) and suppo ec o machines (SVM) on s ock o ecas ing. Ann Da a Sci 10(1):183–
208. h ps://doi.o g/10.1007/s40745-021-00344-x
123
1738 Annals o Da a Science (2024) 11(5):1705–1739
18. Yeganeh A, Chukh o a N, Johannssen A, Fo uhi H (2023) A ne wo k su eillance app oach using
machine lea ning based con ol cha s. Expe Sys Appl 219:119660. h ps://doi.o g/10.1016/j.eswa.
2023.119660
19. Yeganeh A, Johannssen A, Chukh o a N, Abbasi SA, Pou panah F (2023) Employing machine lea n-
ing echniques in moni o ing au oco ela ed p o iles. Neu al Compu Appl. h ps://doi.o g/10.1007/
s00521-023-08483-3
20. Zhang Y, Chen X (2020) Explainable ecommenda ion: a su ey and new pe spec i es. Found T ends
In Re ie al 14(1):1–101. h ps://doi.o g/10.1561/1500000066
21. Ghazima in A, Balalau O, Saha Roy R, Weikum G (2020) PRINCE: p o ide -side in e p e abili y
wi h coun e ac ual explana ions in ecommende sys ems. In: P oceedings o he 13 h in e na ional
con e ence on web sea ch and da a mining. Associa ion o Compu ing Machine y, pp 196–204
22. Li X, Xiong H, Li X, Wu X, Zhang X, Liu J, Bian J, Dou D (2022) In e p e able deep lea ning:
in e p e a ion, in e p e abili y, us wo hiness, and beyond. Knowl In Sys 64(12):3197–3234. h ps://
doi.o g/10.1007/s10115-022-01756-8
23. Yang Z, Wu B, Zheng K, Wang X, Lei L (2016) A su ey o collabo a i e il e ing- based ecommende
sys ems o mobile in e ne applica ions. IEEE Access 4:3273–3287. h ps://doi.o g/10.1109/access.
2016.2573314
24. Ko en Y, Bell R, Volinsky C (2009) Ma ix ac o iza ion echniques o ecommende sys ems. Com-
pu e 42(8):30–37. h ps://doi.o g/10.1109/mc.2009.263
25. Ziegle C-N, McNee S, Kons an J, Lausen G (2005) Imp o ing ecommenda ion lis s h ough opic
di e si ica ion. In: P oceedings o he 14 h in e na ional Wo ld Wide Web con e ence. ACM P ess,
London
26. Alha hi H, Inkpen D, Szpakowicz S (2017) A su ey o book ecommende sys ems. J In ell In Sys
51(1):139–160. h ps://doi.o g/10.1007/s10844-017-0489-9
27. Hahsle M(2021) ecommende lab: lab o de elopingand es ing ecommende algo i hms.Rpackage
e sion 0.2-7. h ps://gi hub.com/mhahsle / ecommende lab
28. He locke JL,Kons anJA, Te eenLG,RiedlJT(2004)E alua ingcollabo a i e il e ing ecommende
sys ems. ACM T ans In Sys 22(1):5–53. h ps://doi.o g/10.1145/963770.963772
29. Lika B, Kolom a sos K, Hadjie hymiades S (2014) Facing he cold s a p oblem in ecommende
sys ems. Expe Sys Appl 41(4 Pa 2):2065–2073. h ps://doi.o g/10.1016/j.eswa.2013.09.005
30. B eese JS, Hecke man D, Kadie C (1998) Empi ical analysis o p edic i e algo i hms o collabo a i e
il e ing. In: P oceedings o he ou een h con e ence on unce ain y in a i icial in elligence. Mo gan
Kau mann, pp 43–52
31. Wenga C, Fansi M, Chab ie S, Ma i J-M, Gabillon A (2021) A comp ehensi e e iew on non-neu al
ne wo kscollabo a i e il e ing ecommenda ionsys ems.In:a Xi p ep in a Xi :2106.10679.h ps://
doi.o g/10.48550/a Xi .2106.10679
32. Beel J, Lange S (2015) A compa ison o o line e alua ions, online e alua ions, and use s udies
in he con ex o esea ch-pape ecommende sys ems. In: Kapidakis S, Mazu ek C, We la M (eds)
P oceedings o he 19 h in e na ional con e ence on heo y and p ac ice o digi al lib a ies. Lec u e
No es in Compu e Science, ol 9316, pp 153–168
33. B adley K (2001) Imp o ing ecommenda ion di e si y. In: P oceedings o he AICS ’01
34. Bobadilla J, O ega F, He nando A, Gu ié ez A (2013) Recommende sys ems su ey. Knowl Based
Sys 46:109–132. h ps://doi.o g/10.1016/j.knosys.2013.03.012
35. James G, Wi en D, Has ie T, Tibshi ani R (2013) An in oduc ion o s a is ical lea ning. Sp inge , New
Yo k
36. Des osie s C, Ka ypis G (2010) A comp ehensi e su ey o neighbo hood-based ecommenda ion
me hods. In: Recommende sys ems handbook. Sp inge , Be lin, pp 107–144
37. He locke J, Kons an JA, Riedl J (2002) An empi ical analysis o design choices in neighbo hood-
based collabo a i e il e ing algo i hms. In Re ie al 5(4):287–310. h ps://doi.o g/10.1023/A:
1020443909834
38. Sa wa B, Ka ypis G, Kons an J, Reidl J (2001) I em-based collabo a i e il e ing ecommenda ion
algo i hms. In: P oceedings o he en h in e na ional con e ence on Wo ld Wide Web—WWW ’01.
ACM P ess, London
39. Ng A, Soo K (2017) Numsense! Da a science o he Layman: no ma h added. Ligh ning Sou ce Inc
40. Funk S (2006) Ne lix upda e: y his a home. h ps://si e .o g/~simon/jou nal/20061211.h ml
41. BookC ossing, Abou BookC ossing. (2021). h ps://www.bookc ossing.com/abou
123
Annals o Da a Science (2024) 11(5):1705–1739 1739
42. Adamopoulos P, Tuzhilin A (2014) On unexpec edness in ecommende sys ems: o how o be e
expec he unexpec ed. ACM T ans In ell Sys Technol 5/4:1–32. h ps://doi.o g/10.1145/2559952
43. Pa k Y-J, Tuzhilin A (2008) The long ail o ecommende sys ems and how o le e age i . In: P oceed-
ings o he 2008 ACM con e ence on ecommende sys ems. Associa ion o Compu ing Machine y,
pp 11–18
44. Deldjoo Y, Bellogin A, Di Noia T (2021) Explaining ecommende sys ems ai ness and accu acy
h ough he lens o da a cha ac e is ics. In P ocess Manag 58:5. h ps://doi.o g/10.1016/j.ipm.2021.
102662
45. Zajac Z (2017) Goodbooks-10k: a new da ase o book ecommenda ions. In: Fas ML
46. Wan M, McAuley JJ (2018) I em ecommenda ion on mono onic beha io chains. In: Pe a S, Eks and
MD, Ama iain X, O’Dono an J (eds) P oceedings o he 12 h ACM con e ence on ecommende
sys ems. Associa ion o Compu ing Machine y, pp 86–94
47. Wan M, Mis a R, Nakashole N, McAuley JJ (2019) Fine-g ained spoile de ec ion om la ge-scale
e iew co po a. In: Ko honen A, T aum DR, Mà quez L (eds) P oceedings o he 57 h con e ence o he
associa ion o compu a ional linguis ics. Associa ion o Compu a ional Linguis ics, pp 2605–2610
Publishe ’s No e Sp inge Na u e emains neu al wi h ega d o ju isdic ional claims in published maps
and ins i u ional a ilia ions.
123