A S udy on Con ex ual In luences on
Au oma ic Playlis Con inua ion
Anna GATZIOURA a,1, Miquel SÀNCHEZ-MARRÈaand Alípio Má io JORGE b,c
aKnowledge Enginee ing & Machine Lea ning G oup (KEMLG)
In elligen Da a Science and A i icial In elligence Cen e (IDEAI)
Uni e si a Poli ècnica de Ca alunya (UPC), Ba celona, Spain
bLabo a o y o A i icial In elligence and Decision Suppo (LIAAD)
INESC TEC, Po o, Po ugal
cFaculdade de Ciências da Uni e sidade do Po o (FCUP), Po o, Po ugal
Abs ac . Recommende sys ems s ill mainly base hei easoning on pai wise in-
e ac ions o in o ma ion on indi idual en i ies, like i em a ibu es o a ings, wi h-
ou p ope ly e alua ing he mul iple dimensions o he ecommenda ion p oblem.
Howe e , in many cases, like in music, i ems a e a ely consumed in isola ion, hus
use s a he need a se o i ems, selec ed o wo k well oge he , se ing a speci ic
pu pose, while ha ing some cogni i e p ope ies as a whole, ela ed o hei pe -
cep ion o quali y and sa is ac ion, unde gi en ci cums ances.
In his pape , we in oduce he e m o playlis concep in o de o cap u e he
implici cha ac e is ics o join music i em selec ions, ela ed o hei con ex , scope
and gene al pe cep ion by he use s. Al hough playlis consump ions may be asso-
cia ed wi h con ex ual a ibu es, hese may be o a ious ypes, di e en ly in lu-
encing use s’ p e e ences, based on hei cha ac e and emo ional s a e, he e o e
di e en ly e lec ed on hei inal selec ions. We highligh on he use o his e m in
HybA, ou hyb id ecommende sys em, o iden i y clus e s o simila playlis s able
o cap u e inhe i cha ac e is ics and seman ic p ope ies, no explici ly desc ibed
in hem. The expe imen al esul s p esen ed, show ha his concep ual clus e ing
esul s in playlis con inua ions o imp o ed quali y, compa ed o using explici
con ex ual pa ame e s, o he commonly used collabo a i e il e ing echnique.
Keywo ds. hyb id ecommende sys ems, au oma ic playlis con inua ion, con ex ual
dimensions, case-based easoning, la en opic models
1. In oduc ion
Music i ems a e a ely consumed in isola ion bu a he as sequences aiming o c ea e a
pa icula a mosphe e [15]. Mo e speci ic, playlis s a e se s o music i ems designed o be
consumed as a sequence, wi h speci ic p ope ies as a whole, simila o adi ional adio
b oadcas s [3].
The e o e, in playlis ecommenda ions, and simila domains, mo e han ecom-
mending isola ed i ems, o p esen ing an o de ed lis o he mos p omising al e na i es,
1Co esponding Au ho : Anna Ga ziou a, Knowledge Enginee ing & Machine Lea ning G oup (KEMLG),
In elligen Da a Science and A i icial In elligence Cen e (IDEAI), Uni e si a Poli ècnica de Ca alunya
(UPC), C/Jo di Gi ona 1-3, 08034, Ba celona, Spain; E-mail: [email p o ec ed].
like he majo i y o ecommende sys ems (RSs) do, he unde lying s uc u e o join i em
selec ions should be e alua ed, as i em in e ac ions wi hin a se may hea ily in luence
he esul [19]. The p esence o an i em wi hin a conc e e concep should be cap u ed,
in o de o ecommend se s o i ems, add essing quali y ela ed a ibu es, like cohe ence
and di e si y, while being ele an o he playlis ’s pu pose and c ea ion momen [18].
Popula ecommenda ion echniques, like collabo a i e il e ing (CF) and con en -
based (CB), mainly use he use -i em ma ix o explode pas in e ac ions and p edic
he sui abili y o an i em o a use , and no he sui abili y o an i em o a pa icula
concep . Mo e speci ic, CF echniques a e domain independen , ocusing only on use
a ings on i ems, while CB app oaches, when applied o music base hei analysis mainly
on sound ela ed a ibu es [4]. They nei he explo e he join i em selec ions no he
ci cums ances unde which hose we e pe o med. Ne e heless, use s may cons uc
di e en playlis s unde di e en ci cums ances, no always made o hei a ou i e o
he mos popula i ems. Thus, simply p edic ing i a song o an a is would be liked
by a use , wi hou e alua ing he whole concep , is no enough and usually esul s in
lowe pe o mance, especially in domains whe e use s pe o m a lo o ansac ions.
Fu he mo e, he in luence e en o e y simila con ex ual si ua ions, on use s’ music
pe cep ion and p e e ences, has been ound o hea ily depend on he use ’s cha ac e and
emo ional s a e [12] ha need o be cap u ed app op ia ely.
1.1. Mo i a ion
The mo i a ion o his wo k a ises om he lack o RSs o e icien ly ecommend se s
o i ems ha would i wi hin a s a ed concep , while add essing addi ional seman ic
cha ac e is ics, like cohe ence and di e si y, simila o au oma ic playlis con inua ion
(APC).
Mo e speci ic, we e alua e he in luence o explici and implici con ex ual dimen-
sions on playlis con inua ion ecommenda ions. HybA, a hyb id RS o APC, has been
designed wi h aim o gene a e ecommenda ions o “se s” o music i ems, no ela ed o a
speci ic use , bu o a speci ic concep . This RS uses Case-Based Reasoning (CBR) wi h
en i e playlis s modelled as cases, o iden i y hei s uc u es, combined wi h a La en
Di ichle Alloca ion (LDA) opic model, o cap u e he i ems’ s yles appea ing wi hin
di e en concep s. In his pape , we highligh on he way ha implici con ex ual ac o s
a e used in his RS o iden i y addi ional playlis cha ac e is ics. We emphasize mainly
on he e ie al s ep o he CBR cycle and complemen ou p e ious wo k [8]. The basic
con ibu ions o his wo k, can be summa ized as:
•The e m playlis concep , ex ending he explici con ex , is in oduced.
•We p esen an app oach o cap u ing implici playlis con ex h ough i s concep .
•We e alua e he abili y o explici and implici con ex ual ac o s o iden i y clus-
e s o simila playlis s, and hei sui abili y o he designed hyb id RS.
The es o he pape is s uc u ed as ollows: in he nex sec ion an o e iew o he
backg ound on music consump ion and he ela ed con ex ual dimensions can be ound.
Following, HybA is p esen ed, wi h emphasis on he e m o playlis concep and i s use
in he sys em. Finally, he e alua ion esul s, showing ha he p oposed concep ual il e -
ing is able o be e cap u e playlis simila i ies and p o ide imp o ed ecommenda ions,
compa ed o he use o explici con ex and use based CF echniques, can be ound.
2. Backg ound
2.1. Music Consump ion
When e e ing o music i ems hose can be songs,gen es,a is s,albums and adio
s a ions. The e o e, music ecommenda ions can be add essed a di e en le els o ab-
s ac ion [18]. As music i ems esul om a complica ed syn hesis p ocess, hei analy-
sis in e ms o con en cha ac e is ics, equi es deepe domain knowledge. In addi ion,
songs a e a ely lis ened o in isola ion [14]. Use s a he c ea e playlis s/sessions, being
sequences o songs, placing mo e impo ance on he songs and hei ela i e o de [19].
Playlis s con ain he no ion o i em sequences and se cha ac e is ics, while being
highly a ec ed by he in en and he con ex in which hey we e gene a ed and consumed
[11]. In addi ion, music is well known o e oke emo ions while a he same ime use s’
music needs a e in luenced by hei ac ual emo ional si ua ion [7]. Howe e , he e is
s ill a lack o solid me hods combining use s’ cogni i e pe cep ion o music wi h sound
cha ac e is ics, he e o e i becomes e en mo e di icul o cap u e hei pe cep ion o
a playlis and speci y he cha ac e is ics ha a “good” playlis should ha e. This no-
ion can be highly subjec i e, depending on pa ame e s like he use ’s music knowledge,
p e e ences, pe sonali y, emo ional s a e, con ex and in en [19].
As cu en ly mo e and mo e online si es ei he inco po a e some music ep oduc ion
in o hei en i onmen , o ocus pu ely on p esen ing music se s, au oma ic playlis gen-
e a ion (APG) and ecommenda ion has eme ged as among he in e es ing issues in he
music ecommenda ion domain. APG e e s o he au oma ic c ea ion o sequences o
music i ems based on some a ge cha ac e is ics. On he o he hand, au oma ic playlis
con inua ion (APC) which is a a ia ion, o sub-case, o APG, consis s in adding a se
o music i ems o a playlis in a way ha i would ma ch i s ini ial a ge cha ac e is ics.
The e o e, APC consis s in he selec ion o he mos app op ia e music i ems, and he
cons uc ion o a sequence o imp o ed quali y acco ding o cha ac e is ics in e ed om
a s a ed playlis [3,19].
Mo e p ecisely, gi en a s a ed lis , he aim is ecommending se s o songs able o
comple e i , while p o iding a mo e exci ing use expe ience. These ecommenda ions
a e gene a ed based on he cha ac e is ics o he s a ed lis , in e ms o music s yles
and a is a ie y, independen ly o he use who made i . The e o e, mo e han simply
p edic ing whe he a music i em would be highly a ed, he unde lying s uc u e o join
selec ions should be e alua ed, in o de o ecommend se s o songs sa is ying a he same
ime ele ance, and o he beyond accu acy dimensions, like cohe ence and di e si y [18].
2.2. Con ex ual Fac o s
Inco po a ing in o he ecommenda ion p oblem con ex ual in o ma ion ela ed o he
ecommenda ion momen , inc eases he inpu da a dimensions o h ee, namely use s,
i ems and con ex . Howe e , when co ec ly cap u ed, music con ex , may lead o a sig-
ni ican inc ease in ecommenda ion accu acy [13].
When e e ing o music, con ex , ini ially de ined as any “in o ma ion desc ibing
whe e you a e, whom you a e wi h, and wha esou ces a e nea by” [20], is mapped o
he use ’s si ua ion when consuming he music i ems, in e ms o ime, mood, ac i i y
and o he people’s p esence. I can be ca ego ized acco ding o se e al c i e ia, as:
•Fully obse able, pa ially obse able and unobse able
•P ima y and seconda y
•En i onmen al and use ela ed
Whe e p ima y and en i onmen al dimensions (loca ion, ime and wea he ) can be used
o de i e he seconda y and use ela ed ones (ac i i y, mood, social and cul u al) [6,11].
A con ex -awa e ecommenda ion p ocess e e s o he es ima ion o use con ex ual
p e e ences, and based on hose, on he gene a ion o he mos ele an ecommenda-
ions. Depending on he pa o he easoning p ocess ha he con ex ual in o ma ion is
aken in o accoun , i can be desc ibed as [1]:
•Con ex ual p e- il e ing: selec ion o he ele an da a, ha will be u he used
o he ecommenda ions’ gene a ion, based on a speci ic con ex .
•Con ex ual pos - il e ing: a ecommenda ion se is gene a ed om he en i e da a
se and is hen adjus ed based on he con ex ual in o ma ion o he ac i e use .
•Con ex ual modelling: con ex ual pa ame e s a e inse ed in o he ecommenda-
ion model o a e used o ans o m he i ems in o a di e en dimension.
Gillho e and Schedl [10] analyze he ela ionships be ween a ious con ex ual di-
mensions o iden i y whe he hose pe mi he accu a e p edic ion o use lis ening p e -
e ences. They e alua e hei in luence on he di e en ca ego ies o music i ems, namely
songs, gen es, a is s and mood. De ice, ask, wea he and ime ha e been ound as he
mos impo an a ibu es, being able o cap u e almos he same in o ma ion as all con-
ex ual ca ego ies when combined. Howe e , due o he da a spa si y in music ecom-
menda ion p oblems, s ill when ecommending songs he pe o mance is e y low, while
when i comes o gen e o a is p edic ions, he addi ional con ex ual in o ma ion indeed
imp o es p edic ion accu acy. In ou p e ious wo k [9], we ha e also ound ha e al-
ua ing he playlis c ea ion ime (hou o he day) leads o imp o ed ecommenda ion
accu acy o a is s and gen es.
As explici ly de ined con ex ela ed o playlis s may be o di e en ypes, o di -
e en ly e lec ed on use s’ selec ions, Pichl e al. [16] p opose he use o playlis names
o implici ly ex ac con ex ual in o ma ion. They c ea e con ex ual clus e s based on
playlis names, ha a e hen inco po a ed in o he ecommenda ion p ocess. Howe e ,
he e iciency o his app oach hea ily depends on he da a quali y, as highly subjec i e
names like “my a ou i e”, “bes ”, e c., lead o spa se clus e s ha a e no aluable o
he ecommenda ion p ocess.
3. Recommenda ion App oach
The designed RS aims o gene a e ecommenda ions o se s o i ems able o comple e
he ac i e use ’s expe ience, based on he ypes o i ems appea ing in simila concep s.
HybA aims o add ess simila seman ic concep s, a he han simila use s and elies on
he basic CBR idea, ha “Simila p oblems ha e simila solu ions” [17]. An o e iew
o i s easoning p ocess, ha ollows he gene al CBR Cycle, is shown in igu e 1.
Gi en a new playlis , HybA i s compa es he al eady selec ed i ems wi h hose in
he playlis s s o ed in he case base and e ie es he kmos simila o hem. Based on he
i ems ound in hose, weigh ed by he simila i y deg ee o he playlis (s) in which hey
appea wi h he new one, he con inua ion is cons uc ed and ecommended.
Figu e 1. HybA Reasoning and Recommenda ion Model
The p oblem en i ies can be desc ibed as:
•a se o songs I={i1,...,iz}
•a se o p e iously ep oduced playlis s L={l1,...,lk}whe e each can be w i en
as he se o songs ha i consis s o , being lj={ij1,...,ijn},ij ∈I, =1,...,n
•a se o use s U={u1,...,u } ha ha e o med hose playlis s
Each o hese en i ies may be associa ed wi h addi ional cha ac e is ics, like me a-
da a o edi o ial in o ma ion, empo al o con ex ual da a, p e e ences and demog aphic
cha ac e is ics, espec i ely, used o compu e hei simila i y deg ees, like in [8].
3.1. Playlis Gene al Concep
Aplaylis is a collec ion o music i ems, ep oduced as a meaning ul sequence ha
should ha e some special cha ac e is ics as a whole [3]. We p opose he ca ego iza ion
o playlis s’ cha ac e is ics in o:
•In e nal (o con en ela ed): he exac music i ems ha a playlis consis s o . The
s yles o hose i ems (which may be songs, a is s o gen es), de ine o a g and
ex end he cha ac e is ics o a playlis , like i s cohe ence o di e si y le el.
•Ex e nal (o en i onmen al ela ed): he pa ame e s ela ed o he playlis bu no
di ec ly desc ibed in i . Fo ins ance, he c ea ion pu pose, con ex , e c., and he
in luence o hose ac o s on he use ’s emo ional s a e and music pe cep ion.
Al hough no always explici ly e lec ed in e ms o o mula ion, he con ex , and
o he ex e nal pa ame e s ela ed o a playlis , ha e been ound o in luence i s s yle and
he i ems inally placed in i [10]. Howe e , as music pe cep ion is highly subjec i e, and
may be hea ily a ec ed by ci cums ances unde which is consumed, he pe sonali y and

he emo ional s a e o he use , i is ha d o es ablish a di ec connec ion be ween use s
and hei p e e ed playlis s unde a gi en con ex . Ne e heless, i has been ound [7]
ha depending on hei cha ac e , use s pe o m di e en music selec ions e en unde
simila ci cums ances. Fo ins ance, mo e ex o e people end o ely on “happy” music
unde sad o s ess ul si ua ions, in o de o “chee ” hemsel es up, while in o e s end
o lis en o “sad” o “dep essi e” music simila o hei emo ional s a e.
The e m playlis concep is in oduced as a wide e m o desc ibe he gene al cha -
ac e is ics o a playlis , being he esul o he combina ion o bo h in e nal and ex e nal
pa ame e s, beyond explicil y de ined con ex . Fu he mo e, as playlis s a e o mula ed
o se e a speci ic seman ic concep , hey a e ea ed as dis ibu ions o e music s yles.
The e o e each playlis can be w i en as lj={sj1,...,sjn}whe e each sj is he music
s yle o he song ij . In o de o cap u e he gene al concep o playlis s, wi hou ocusing
on hei speci ic con en , o when no addi ional in o ma ion, like hei names o explici ly
de ined con ex , is a ailable, he playlis s’ la en opic dis ibu ion, based on he music
s yles in hem, is p oposed. Thus, be o e analysing he playlis -song dis ibu ions, i s
he playlis -music s yles dis ibu ions a e iden i ied, wi h aim o cap u e he endencies
and pa e ns p esen in he playlis s, a he han inding he exac songs.
Figu e 2. Playlis s’ la en dis ibu ions o e music s yles
Ha ing playlis s desc ibed as dis ibu ions o e music s yles, a La en Di ichle Al-
loca ion (LDA) opic model [2] is buil o ex ac he unde lying la en opics, as in igu e
2. Playlis s a e hen cha ac e ized by hei dominan opic(s) as a ep esen a ion o hei
gene al concep , closely ela ed o hei cen al idea and cogni i e pe cep ion. We ha e
ound his a ibu e o be capable o cap u ing he gene al endencies in playlis s, and
hei simila i ies, and o imp o e ecommenda ion accu acy.
3.2. Candida es Re ie al
An impo an p ocess o a CBR sys em ha highly a ec s i s pe o mance, is i s abili y
o iden i y he cha ac e is ics o a new case, and e ie e, wi hin an accep able ime, he
pas cases ha seem as mos app op ia e o i s solu ion. The e o e, in o de o iden i y
globally accep able candida es o a new playlis , while educing he compu a ional ime,
i s a p ope clus e ing and p e- il e ing o he case base is pe o med.
Especially when applied o la ge scale da ase s he selec ion o he pa ame e s ha
mos ly cha ac e ize he playlis s and enable hei p ope clus e ing, is o high impo ance.
To his di ec ion, a con ex ual p e- il e ing based on he playlis gene a ion momen (hou
o he day) was ini ially es ed, and was ound o p o ide imp o ed esul s [9]. Howe e ,
as many imes con ex ual in o ma ion is no explici ly desc ibed, o e en i so i may
a ec di e en ly use s’ selec ions, a mo e gene al clus e ing would be mo e app op ia e.
The e o e, we ha e ex ended he con ex ual o a concep ual p e- il e ing, e alua ing i s
playlis s’ concep ual simila i ies.
Mo e speci ic, based on he music s yles’ dis ibu ion in a new playlis being lN=
{sN1,...,sNn}, i s opic dis ibu ion, speci ying i s concep , is iden i ied. Then he se
o ins ances wi h he same dominan opic,LC⊆L, ha add ess he same concep , a e
ini ially e ie ed and hen used o u he compu a ions, inally leading o he e ie al
o he kmos simila pas playlis s.
Gi en a e ie ed playlis lR={sR1,...,sRm},lR∈LC, i s simila i y Sim(lN,lR)wi h
lN, e e ed o as global simila i y, is compu ed as a unc ion o he local simila i y le els
o he i ems in hem. HybA inally e ie es he kmos simila playlis (s), hose whose
global simila i y ul ils equa ion (1), and cons uc s he playlis con inua ion using he
songs ha ini ially appea ed in hem, as in [8].
l0∈LC:∀lR∈LC,l0=a gmax{Sim(lN,lR)}(1)
4. E alua ion
In his sec ion we p esen g aphically he compa a i e esul s o HybA, using he p o-
posed concep ual clus e ing and p e- il e ing, wi h he use o h ee di e en con ex ual
clus e ings and a CF app oach, on h ee eal music da ase s.
Mo e speci ic, o a new playlis , cn xT, uses he clus e LHo playlis s gene a ed
in close hou s, cn xM, he clus e LMo playlis s ep oduced in close mon hs and cn xC
e alua es bo h he day hou and he mon h when he playlis was made. Fu he mo e, a
use -based CF2using Euclidean simila i y and he 10 nea es neighbou s o each use
has been es ed.
The da abases used o e alua ion come om he Po uguese music social ne wo k
Palco P incipal3and con ain in o ma ion on use s adding acks o hei lis ening sessions
o hei playlis s, on a gi en momen (e en ). Mo e de ails can be ound in able 1.
Da ase Palco Lis en1 Palco Lis en2 Palco Playlis s
E en s 1171849 295044 111942
Songs 29786 22986 26117
Use s 21815 5543 10392
Playlis s 86174 22108 22132
Table 1. Palco P incipal da ase s
To e alua e bo h ecommenda ion accu acy and quali y we ha e used he ecom-
menda ions’ a e age p ecision and cohe ence. Cohe en playlis s a e gene ally consid-
e ed as o “be e quali y”, while he di e si y and no el y deg ee ha a use enjoys may
be subjec o his/he ac ual music p e e ences [5,19].
•P ecision is he a io o co ec ecommenda ions: he missing i ems ha we e
success ully iden i ied, o e he o al numbe o ecommenda ions, calcula ed as:
P ecision =#Rele an RecommendedI ems/#RecommendedI ems
2h ps://mahou .apache.o g/use s/ ecommende / ecommende -documen a ion.h ml
3h p://palcop incipal.com/
•Cohe ence e alua es he homogeni y o a playlis R, and can be calcula ed as he
a e age simila i y (sim) among he pai s o consequen i ems in he lis , being:
Cohe ence =1
|R|−1∑
i∈R
sim(i,i+1)
Figu e 3. Recommenda ions’ p ecision and cohe ence on Palco Lis en1 da ase
Figu e 4. Recommenda ions’ p ecision and cohe ence on Palco Lis en2 da ase
Figu e 5. Recommenda ions’ p ecision and cohe ence on Palco Playlis s da ase
As i can be seen om igu es 3–5, CF shows a low pe o mance due o i s sligh ly
di e en scope, as i does no ake in o accoun any addi ional in o ma ion ela ed o
join i em selec ions. In addi ion, in he majo i y o cases including in o ma ion on i ems’
join selec ions and basic ime con ex , imp o es he esul s o e he CF app oach, while
going beyond explici con ex leads o highe imp o emen s. Mo e speci ic, he p oposed
concep ual clus e ing and p e il e ing p o ides an impo an imp o emen o bo h ec-
ommenda ion p ecision and cohe ence.
5. Conclusions
Al hough ecen ly con ex -awa e me hods ha e gained space among ecommenda ion
echniques, s ill hey ocus mainly on explici ly de ined con ex ual pa ame e s. Howe e ,
con ex may be o di e en ypes, di e en ly e lec ed in use s’ beha iou , especially in
domains ela ed o emo ional esponses, like in music consump ion.
In his pape we ha e p esen ed he e m o playlis concep , o implici ly cap u ing
he con ex o playlis gene a ions. This e m is used o cha ac e ize playlis s, cons uc ed
unde gi en ci cums ances, based on he music s yle combina ions in hem. Fu he mo e,
i has been ound o be e iden i y simila playlis s, compa ed o explici con ex ual
dimensions, and o p o ide imp o ed APC ecommenda ions compa ed wi h he widely
used CF app oach. As pa o ou u u e wo k, he exploi a ion o he la en concep s wi h
aim o pe o m a p ope “ ansla ion”, enabling he explana ion o he esul ing playlis
clus e s, will ake place.
Re e ences
[1] Gediminas Adoma icius and Alexande Tuzhilin. Con ex -awa e ecommende sys ems. In Recom-
mende sys ems handbook, pages 217–253. Sp inge , 2011.