Recei ed No embe 22, 2020, accep ed Decembe 9, 2020, da e o publica ion Decembe 14, 2020,
da e o cu en e sion Decembe 23, 2020.
Digi al Objec Iden i ie 10.1109/ACCESS.2020.3044670
Con ex -Awa e P ocess Pe o mance
Indica o P edic ion
ALFONSO E. MÁRQUEZ-CHAMORRO 1, KATE REVOREDO2,
MANUEL RESINAS 1, (Membe , IEEE), ADELA DEL-RÍO-ORTEGA1,
FLÁVIA M. SANTORO3, AND ANTONIO RUIZ-CORTÉS 1, (Membe , IEEE)
1Sma Compu e Sys ems Resea ch and Enginee ing Lab (SCORE), Resea ch Ins i u e o In o ma ics
Enginee ing (I3US), Uni e sidad de Se illa, 41012 Se ille, Spain
2Depa men o In o ma ion Sys ems and Ope a ions, Vienna Uni e si y o Economics and Business (WU), 1020 Vienna, Aus ia
3G adua e P og am o In o ma ics, Uni e si y o he S a e o Rio de Janei o, Rio de Janei o 21941-901, B azil
Co esponding au ho : Al onso E. Má quez-Chamo o ([email p o ec ed])
This wo k was suppo ed in pa by he Eu opean Union’s Ho izon 2020 Resea ch and Inno a ion P og amme h ough he Ma ie
Sklodowska-Cu ie unde G an 645751 RISE_BPM, and in pa by he Spanish Go e nmen h ough p ojec s HORATIO and
OPHELIA unde G an RTI2018-101204-B-C21/C22. The wo k o Ka e Re o edo was suppo ed
by he Os e eichische Akademie de Wissenscha en.
ABSTRACT I is well-known ha con ex impac s unning ins ances o a p ocess. Thus, de ining and using
con ex ual in o ma ion may help o imp o e he p edic i e moni o ing o business p ocesses, which is one
o he main challenges in p ocess mining. Howe e , iden i ying his con ex ual in o ma ion is no an easy
ask because i migh change depending on he a ge o he p edic ion. In his pape , we p opose a no el
me hodology named CAP3 (Con ex -awa e P ocess Pe o mance indica o P edic ion) which in ol es wo
phases. The i s phase guides p ocess analys s on iden i ying he con ex o he p edic i e moni o ing o
p ocess pe o mance indica o s (PPIs), which a e quan i iable me ics ocused on measu ing he p og ess
o s a egic objec i es aimed o imp o e he p ocess. The second phase in ol es a con ex -awa e p edic i e
moni o ing echnique ha inco po a es he ele an con ex in o ma ion as inpu o he p edic ion. Ou
me hodology le e ages con ex -o ien ed domain knowledge and expe s’ eedback o disco e he con ex ual
in o ma ion use ul o imp o e he quali y o PPI p edic ion wi h a dec ease o e o a es in mos cases,
by adding his in o ma ion as ea u es o he da ase s used as inpu o he p edic i e moni o ing p ocess.
We expe imen ally e alua ed ou app oach using wo- eal-li e o ganiza ions. P ocess expe s om bo h
o ganiza ions applied CAP3 me hodology and iden i ied he con ex ual in o ma ion o be used o p edic ion.
The model lea ned using his in o ma ion achie ed lowe e o a es in mos cases han he model lea ned
wi hou con ex ual in o ma ion con i ming he bene i s o CAP3.
INDEX TERMS Business p ocess managemen , p ocess mining, p edic i e moni o ing, con ex awa eness,
p ocess indica o p edic ion.
I. INTRODUCTION
P ocess mining [1] allows he ex ac ion o use ul in o ma ion
om e en logs and his o ical da a o business p ocesses.
This in o ma ion can be used o imp o e he pe o mance o
hese business p ocesses. One o he applica ions o p ocess
mining is he p edic i e moni o ing o business p ocesses
[2], which p edic s di e en aspec s o he execu ion o a
business p ocess, such as he nex ac i i y [3], [4], o he
alue o a p ocess pe o mance indica o [5]–[7]. P ocess
pe o mance indica o s a e quan i iable me ics ocused on
measu ing he p og ess owa ds a goal o s a egic objec i e
The associa e edi o coo dina ing he e iew o his manusc ip and
app o ing i o publica ion was Albe o Cano .
aimed a con olling and imp o ing he business p ocess [8].
Some examples a e he emaining execu ion ime o a p ocess
ins ance, he likelihood o a aul in he sys em o he abno -
mal e mina ion o a unning ins ance. These p edic ions
enable he applica ion o p oac i e and co ec i e ac ions o
imp o e p ocess pe o mance and mi iga e possible isks in
eal ime.
Recen ly, he e ha e been many esea ch e o s ocused
on imp o ing he quali y o hese p edic ions. One s eam
o wo k has success ully used da a om he con ex asso-
cia ed o a unning p ocess o imp o e he p edic i e pe -
o mance [9]–[14]. This con ex associa ed o a p ocess is
he knowledge po en ially ele an o guide i s execu ion
[15]. This knowledge can be associa ed wi h an ac i i y o
222050 This wo k is licensed unde a C ea i e Commons A ibu ion 4.0 License. Fo mo e in o ma ion, see h ps://c ea i ecommons.o g/licenses/by/4.0/ VOLUME 8, 2020
A. E. MÁRQUEZ-CHAMORRO e al.: CAP3
wi h he whole p ocess. Fo ins ance, he loca ion whe e a
p ocess occu s o he le el o p io i y o a speci ic ac i -
i y can be conside ed as knowledge associa ed wi h a p o-
cess o ac i i y, espec i ely. P ocess con ex in o ma ion
plays an impo an ole in p ocess mining as epo ed in he
li e a u e [16], [17].
The use o con ex in hese wo ks is limi ed o build-
ing a p edic i e model using all o he a ailable con ex ual
in o ma ion in he e en log wi hou conside ing whe he
his da a is he mos ele an o p ocess p edic ion. This
may be a p oblem because including oo much in o ma ion
migh no always be bene icial o he p edic i e quali y [18].
Mo eo e , p edic i e moni o ing algo i hms may conside
ce ain ea u es as ele an while hey a e no [19]. The e o e,
he iden i ica ion o adequa e con ex in o ma ion, which
can le e age he o ecas ing o p ocess indica o s, becomes
impe a i e. Howe e , his is no an easy ask because depend-
ing on he indica o , di e en con ex a ibu es can be el-
e an . Fo ins ance, i we wan o p edic he s a e o an
ac i i y, he in ol ed human esou ce can be he con ex o
be conside ed while o p edic ing he emaining ime o a
p ocess execu ion he p io i y a iable may be he con ex o
conside . In his pape , we add ess his issue by iden i ying he
con ex in o ma ion o a business p ocess ela ed o a ce ain
indica o so ha i can be used o imp o e i s p edic ion.
In [20], a me hodology named ORGANON is p oposed o
iden i ying business p ocess- ele an con ex ual in o ma ion
which could impac on he p ocess goals. Based on a se o
c i e ia and a ma ix o analyzing on ological ansac ions,
his me hodology disco e s he essen ial ac i i ies and hen
hei main a ibu es a e examined. I he a ia ion in he alue
o hese a ibu es impac s he goal o he p ocess, hey will
be iden i ied as elemen s o he immedia e/in e nal con ex .
This me hodology iden i ies exis ing con ex elemen s o a
p ocess, bu does no ocus on he con ex ha is ele an o
p ocess pe o mance indica o s.
In his pape , we p opose a me hodology named CAP3
(Con ex -awa e P ocess Pe o mance indica o P edic ion)
ha comp ises wo phases: (1) an ex ension o he
ORGANON me hodology o he de ini ion o he con ex
necessa y o p edic ing p ocess indica o s. Ou main goal
is o injec con ex -o ien ed domain knowledge and expe s’
eedback o imp o ing he e ec i eness o cu en p edic i e
solu ions. (2) A con ex -awa e p edic i e moni o ing ech-
nique ha uses he ele an con ex as inpu . Expe imen al
esul s on he applica ion o ou app oach in wo eal-li e
o ganiza ions con i m he bene i o he app oach.
This wo k has implica ions o he ope a ional manage-
men o o ganiza ions, by sugges ing a me hodology o de ine
he con ex in o ma ion which p o ides in o ma ional suppo
o decision make s abou when, whe e and why business
p ocesses need o be adap ed. In addi ion, an imp o emen in
he p edic ion e o o he pe o mance indica o s, in many
occasions, also means sa ings in human and economic
esou ces and p e en ion o impo an loss o u no e o he
companies [21].
The emainde o his pape is o ganized as ollows.
Sec ion II includes some de ini ions and wo ks ela ed o
he con ex in BPM and in oduces p edic i e moni o -
ing. Sec ion III summa izes he ela ed wo ks in his a ea.
Sec ion IV p esen s he con ibu ions o ou wo k. The expe -
imen and he discussion o he ob ained esul s a e p esen ed
in Sec ion V. Finally, Sec ion VI concludes he wo k and
p esen s possible u u e di ec ions.
II. BACKGROUND
This in oduc o y sec ion p o ides some backg ound on he
con ex iden i ica ion in BPM and he p edic i e moni o ing
o business p ocesses. Speci ically, Sec ion II.A includes
some de ini ions and wo ks ela ed o he concep and ole
o con ex in BPM. Then, Sec ion II.B in oduces some basic
concep s o he p edic i e moni o ing o business p ocesses
la e used in his pape .
A. IDENTIFYING CONTEXT IN BPM
Gene ally speaking, con ex can be de ined as he ci cum-
s ances in which an e en occu s. Acco ding o [22], con ex
is an open concep , since i is no limi ed o he imagina ion
o a pe son, while [23] explains con ex as "any in o ma ion
ha can be used o cha ac e ize he si ua ion o an en i y." Ye ,
[24] s a es ha con ex es ic s one s ep a a oubleshoo ing
wi hou in e ening in i explici ly. In o he wo ds, con ex
is use ul in o ma ion o he pe o mance o ac i i ies and
in e ac ions ha occu in a wo k p ocess [25].
Con ex o business p ocesses suppo s he unde s anding
o he a ia ions in each ins ance, i.e., each p ocess execu ion
could ha e a dis inc se o con ex in o ma ion associa ed.
Mo eo e , i helps o explain why decisions we e made.
In business p ocesses, con ex can be de ined as he minimum
se o a iables ha con ains all he impo an in o ma ion ha
impac s hei design, implemen a ion and execu ion [26].
In [27], au ho s p esen a me amodel s uc u ed in h ee
laye s, ha oge he , a e able o suppo he ep esen a ion
o p ocess con ex in a pa icula domain. The i s laye is
he Con ex Me amodel, whe e p ocess con ex is o mally
de ined. I e e s o he elemen s ela ed o he manipula ion
o con ex and hei ela ionship. Among hese elemen s a e
Con ex ual Elemen and Focus. A con ex is de ined as he
se o con ex ual elemen s and hose con ex ual elemen s a e
ela ed o a ocus, o ins ance o a pa icula ac i i y o
he p ocess. Acco ding o he au ho s, each ins ance o a
business p ocess is subjec o changes in con ex , and as well,
con ex ual knowledge can add ele an in o ma ion o suppo
he execu ion o ac i i ies.
De ining he co ec con ex is a challenge. In [20],
a me hodology o iden i ying business p ocess- ele an con-
ex ual in o ma ion called ORGANON is desc ibed. This
me hodology is based on a ques ionnai e, a se o c i e ia and
a ma ix o analyzing on ological ansac ions. ORGANON
is di ided in o wo s eps. Fi s , essen ial ac i i ies a e dis-
co e ed, i.e. he ones ha ha e a di ec in luence on he
p ocess goal [28]. These ac i i ies a e selec ed acco ding
o a semi-s uc u ed guide consis ing o a se o ques ions
VOLUME 8, 2020 222051
A. E. MÁRQUEZ-CHAMORRO e al.: CAP3
answe ed by expe s o he p ocess. Then, acco ding o
[29], an on ological ansac ion ma ix is buil , egis e ing
he ela ionship ha keeps he essen ial ac i i ies oge he .
Thus, i is necessa y o conside whe he hese ac i i ies
o m comple e cycles o an on ological ansac ion. A cycle
is desc ibed by ou phases ( eques , p omise, s a e and
accep ) which comp ises a se o ac i i ies pe o med by
an ini ia o (clien /applican ) and an execu o ha aims o
achie e a ce ain goal [20]. Ac i i ies which o m comple e
cycles o an on ological ansac ion a e conside ed essen ial
ac i i ies.
Once hese essen ial ac i i ies ha e been de ec ed, i is
necessa y o elici inne a ibu es om hem (i.e. all he
inpu s and ou pu s in he business p ocess ac i i ies modeled,
ex e nal da a, a i ac s o business ules as desc ibed in [27])
and analyze he impac o each a ibu e in he p ocess goal.
The impac analysis e i ies wha may occu wi h he goal
o p ocess (achie ed/no achie ed) i he alue o an a ibu e
a ies in an unp edic able way. I he a ia ion in he alue o
hese a ibu es impac s he goal o he p ocess, hey will be
iden i ied as con ex ual elemen s.
B. PREDICTIVE MONITORING OF PROCESS INDICATORS
P edic i e moni o ing o business p ocesses p o ides he
o ecas o p ocess pe o mance indica o s o a unning p o-
cess ins ance wi h a p edic i e model and can be used as sup-
po o decision making in an o ganiza ion [21]. Examples o
cases whe e p edic i e moni o ing can be used a e: an insu -
ance company wan s o p edic he emaining execu ion ime
o a p ocess ins ance (e.g. comple e ime o esol e a claim),
o an IT company wan s o p edic he numbe o inciden s
sol ed in one mon h o know i a ce ain se ice ag eemen
will be sa is ied.
P edic i e moni o ing elies on building a p edic i e model
om an e en log o he business p ocess. An e en log (L)
is composed o a se o aces (T). Each ace (Ti) e lec s an
execu ion o a p ocess ins ance. Fo mally, we can exp ess a
ace as an o de ed lis o e en s Ti=[Ei1,...,Eim] whe e
Ei1 ep esen s he i s e en and Eim he inal e en o ace
Ti. Simila ly, a log can be exp essed as he se o aces o
he ins ances ha ha e s a ed and inished in an in e al o
ime L=[T1,...,Tn] whe e T1 ep esen s he i s execu ed
ace and Tn he las in he ime in e al. Finally, an e en
ep esen s he execu ion o jus an ac i i y o he p ocess.
Each e en con ains a se o a ibu es (a), which ep esen s
all he in o ma ion o he de ini ion o such e en , e.g.
imes amp, he name o he ac i i y, he esou ce ha execu es
he ac i i y, o he alue o some da a used h oughou he
ins ance, Ej=[aj1,...,ajo] whe e ode e mines he o al
numbe o a ibu es o he e en . An example o a ypical
e en log is depic ed in Table 1. Each ace o his e en log
con ains an e en id, which is a unique iden i ie o each
e en , a imes amp, ha indica es he ime and da e o he
execu ion o an ac i i y, he name o his ac i i y, he esou ce
o pe son who execu es he ac i i y, and inally he cos o he
ac i i y.
TABLE 1. E en log example.
A p ocess pe o mance indica o (I) is a quan i iable me -
ic ocused on measu ing he p og ess owa d a goal o
s a egic objec i e. Indica o s can be classi ied in o wo ypes:
single-ins ance indica o s o agg ega ed indica o s. The o -
me is compu ed o each ace in he log using he alues
o he a ibu es o he e en s ha compose his ace. The e-
o e, i can be de ined as a unc ion o a ace T,i.e. I(T).
This unc ion can e u n a bina y alue, e.ga de e mined
condi ion ul illed by he ace, o a eal alue, e.g. he
du a ion o an ac i i y. Ins ead, an agg ega ed indica o is
compu ed o a se o aces by agg ega ing mul iple alues o
a single-ins ance indica o using some agg ega ion unc ion,
e.g.sum o a e age. An example o his ype o indica o
could be he pe cen age o inciden s sol ed in a ce ain pe iod
o ime. In his pape , we conside bo h single-ins ance and
agg ega ed indica o s.
One o he main issues add essed in p edic i e moni o ing
o business p ocesses is he p edic ion o he alue o an
indica o be o e a p ocess ins ance inishes by means o a p e-
dic i e model. The e o e, a p edic i e model o an indica o
Iis a unc ion PI([Ei1,...,Eil]), ha compu es a p edic ion
o I om he ace [Ei1,...,Eil], whe e Ei1is he i s e en
and Eilis he las e en ha ha e occu ed in ace Tia a
gi en momen .
III. RELATED WORK
Some me hods in he li e a u e ha e employed con ex-
ual in o ma ion o business p ocess p edic i e moni o ing.
In [30], a clus e ing o ien ed me hod ha p edic s p ocessing
imes and associa ed SLA (Se ice Le el Ag eemen ) io-
la ions is p esen ed. The unning ins ance is assigned o a
e e ence scena io (clus e ), which is used o he p edic ion.
The p edic i e model is based on decision ees, called P edic-
i e Clus e ing T ee (PCT). The de ini ion o hese clus e s,
gene a ed by P edic i e Clus e ing sub-module, can be ep-
esen ed as a se o logical decision ules and g oups aces
acco ding o simila a ge alues. The inpu s o he me hod
a e a log e en wi h da a a ibu es and en i onmen ea u es,
a a ge measu e and a h eshold o isk. P edic ion accu acy
is e alua ed using Roo Mean Squa ed E o (RMSE) and
Maximum Dwell Time (MDT).
In [31], a p ocess ace is con e ed in o a se o con ex
p ope ies and a ibu es o p ocess. A clus e ing me hod
is used o selec he mos signi ican s uc u al pa e ns o
make he o ecas . This clus e ing me hod conside s he con-
ex da a and a ge a iables de i ed om pe o mance al-
ues. Th ee di e en eg ession algo i hms (Linea eg ession,
222052 VOLUME 8, 2020
A. E. MÁRQUEZ-CHAMORRO e al.: CAP3
RepT ee and IB-k) a e used o he p edic ion. The inpu s o
he algo i hm a e he aces o a log e en , and a a ge pe o -
mance measu e (in his case, he emaining p ocessing ime).
The uples a e cons uc ed om he e en da a in o ma ion o
he aces. Some de i ed a ibu es and con ex in o ma ion
a e also included in he encoding.
In [6], s a is ical echniques o he p edic ion o e en s and
hei co ela ion wi h con ex ual elemen s o anspo a ion
p ocesses, such as wea he condi ions o oad a ic, a e
applied. An in eg a ion pla o m named FInes , ha inco po-
a es he p edic i e moni o ing module, was de eloped. The
me hod ecei es 3 di e en da a sou ces: sys em messages
om he p ocesses, agg ega ed da a wi h addi ional in o ma-
ion o he p ocess, such as es ima ed ime o a i al s. ac ual
a i al o he cause o delays, and quali y indica o s om he
CARGO 2000 sys em. The sys em e u ns a p edic ion o he
delay in he deli e ies.
Al hough he wo k in [9] does no p o ide p edic ion
pe o mance measu emen s, i conside s he p ocess con ex
o he analysis o key p ocess pe o mance indica o s. The
au ho s pe o med a s a is ical analysis o ex ac signi ican
di e ences in pe o mance measu es o di e en analyzed
con ex s. These pe o mance measu es a e calcula ed using
he p ocess en i ies labeled wi h di e en con ex a ibu es.
In [32], a me hod o ca ego ize possible en i onmen al
condi ions and case p ope ies in o con ex ca ego ies which
a e meaning ul o he p ocess execu ion was p oposed. I is
ela ed o ou p oposal in he sense ha i sea ches o knowl-
edge which in luences he execu ion o a business p ocess,
bu i di e s in he sense ha he main goal is o g oup his
knowledge.
In ecen yea s, he e has been an e e g owing in e es
in he a ea o con ex -awa e p ocess p edic i e moni o ing,
wi h a numbe o wo ks app oaching his challenge om
di e en angles. Yeshchenko e al. explo e in [10] he idea
o in eg a ing he ex e nal uns uc u ed con ex o business
p ocesses in o p edic ion me hods. In pa icula , hey p opose
a echnique o en ich e en logs wi h sen imen in o ma ion
ex ac ed om media con en by means o sen imen analysis
echniques. As e alua ion, XGBoos is applied o he p edic-
ion o he emaining ime o a p ocess in ou di e en e en
logs, compa ing he esul s be ween he pu e and he en iched
e en logs wi h posi i e esul s.
In [14], a echnique o a documen -awa e p edic i e busi-
ness p ocess moni o ing is p esen ed. In his case, he e en
log is en iched wi h s uc u ed con ex om documen s,
ex ac ed using a ex -based app oach o au oma ed in o -
ma ion ex ac ion. The au ho s plan o use long-sho e m
memo y neu al ne wo k (LSTM) o p edic nex ac i i y, bu
no ac ual e alua ion is epo ed in he pape .
The wo k in [11] examines he impac and e ec s o
inco po a ing disc e e and con inuous con ex da a a ibu es
on p edic ion quali y and accu acy. The au ho s e alua e
he applica ion o a LSTM ne wo k wi h di e en inpu
con igu a ions on a eal-li e e en log o p edic he nex
occu ing e en . They show ha p edic ion accu acy can be
signi ican ly imp o ed by inco po a ing addi ional e en da a
a ibu es in LSTM based p ocess p edic ion.
Sende o ich e al. [12] a gue he impo an ole o
in e -case dependencies in p edic i e p ocess moni o ing.
They p esen a me hod o ea u e encoding o p ocess
cases ha elies on a bi-dimensional s a e space ep esen-
a ion, including in a- and in e -case dependencies. Fo he
in e -case encoding hey p opose o pa i ion he eco ded
(and unning) cases in o case ypes, and use a de i a ion
unc ion o a oid ea u e space explosion. They e alua e hei
app oach in wo eal e en logs and show he imp o emen o
he p edic ion o he emaining p ocess ime in unning cases,
using linea eg ession, Lasso, andom o es s and g adien
ee boos ing. RMSE and MAE (Mean Absolu e E o ) a e
he p edic ion accu acy measu es selec ed.
Finally, Hinkka e al.’s main goal in [13] is o imp o e he
p edic ion accu acy o p edic ion models, o any case-le el
p edic ion ask, by exploi ing addi ional e en a ibu es
ha a e o en a ailable in he e en logs while also ak-
ing in o accoun he scalabili y. The au ho s in oduce a
‘‘me hod o exploi e en a ibu es in o Recu en Neu al
Ne wo k (RNN) p edic ion models by clus e ing e en s by
hei e en a ibu e alues and using he clus e labels in he
RNN inpu ec o s, ins ead o he aw e en da a’’. In ou
ou o he i e da ase s e alua ed, he p oposed app oach
ou pe o med ha ing he ac ual a ibu e alues in he inpu
ec o , also educing aining and p edic ion imes.
Al hough some o he a o emen ioned wo ks explo e he
idea o exploi ing he con ex in o ma ion o he p edic ion,
some a e ocused on a ce ain ype o con ex ual in o ma ion,
e.g. sen imen s, o da a sou ces, e.g. documen s, o p edic ion
ac i i y, e.g. emaining p ocess ime. O he s seek o iden i y-
ing he dependencies amongs cases o imp o ing he pe o -
mance o he p edic ion i sel . Gene ally, mos o hese wo ks
do no in o m how he con ex ual a ibu es we e chosen o
compose he log. We go a s ep u he and aim a guiding
he p ocess o iden i ying which he app op ia e con ex ual
in o ma ion o imp o e he p edic i e moni o ing is, since
his has p o en no o be a i ial ask [20]. In ha sense,
many o hem can be used o complemen ou p oposal he e,
whose main con ibu ion is p o iding an ex ended me hodol-
ogy based on ORGANON [20] o ex ac con ex a ibu es
om business p ocesses o he p edic i e moni o ing using
domain knowledge o he p ocess. This knowledge can be
ob ained om expe s o manage s o he p ocess.
IV. PROPOSAL
Ou p oposal CAP3 (Con ex -awa e P ocess Pe o mance
indica o P edic ion) has wo majo pa s: (i) a me hodology
o elici he ele an con ex ual elemen s o he p ocess
moni o ing p esen ed in Sec ion IV-A; (ii) a con ex -awa e
p edic i e moni o ing echnique ha uses he ele an con ex
as inpu , which is desc ibed in Sec ion IV-B.
A. CONTEXT IDENTIFICATION METHODOLOGY
ORGANON me hodology p esen ed in [20] disco e s he
con ex ual in o ma ion associa ed o a business p ocess which
VOLUME 8, 2020 222053
A. E. MÁRQUEZ-CHAMORRO e al.: CAP3
FIGURE 1. Ex ended ORGANON me hodology p ocedu al model.
a e aligned o he objec i es o he p ocess. Howe e , we need
o ex end his p oposal o ex ac he con ex in o ma ion
necessa y o imp o e he p edic i e pe o mance o PPIs. Bu
i is impo an o no e ha including oo much in o ma ion o
adding he inco ec ea u es is no bene icial o he p edic-
i e quali y [18]. The e o e, we ha e adap ed he ORGANON
me hodology o exclusi ely iden i y hose con ex a ibu es
wi h a di ec in luence on he PPIs o be p edic ed. Focused
on his pu pose, new asks ha e been added o ORGANON.
Figu e 1depic s he ex ension o ORGANON me hodol-
ogy ha we p opose in his esea ch. All he s eps o he
me hodology a e new wi hin he excep ion o S ep 3.
Fi s o all, we need o analyze he p ocess model and
he PPIs we wan o p edic (S ep 1). As inpu s, we ake all
iden i ied PPIs and he p ocess model. PPIs can be ob ained
om he documen a ion o he p ocess o can be di ec ly
eques ed om he p ocess analys . One o mo e PPIs o he
p ocess can be selec ed o p edic ion. Ou me hod wo ks o
bo h simple and agg ega ed indica o s which a e compu ed
using p e ious measu es de ined o e mul i-p ocess ins ances
[33]. We can ollow some speci ic c i e ia o he elici a ion
o PPIs such as he answe o he ques ion: wha is he PPI
ha is ela ed o a highe numbe o ac i i ies? Secondly,
an in e iew wi h he business p ocess analys is ca ied ou
(S ep 2). The ques ionnai e (de ailed in Table 2) used as inpu
o his ac i i y collec s some o he ques ions e lec ed in
ORGANON o de e mine some in o ma ion abou he exe-
cu ion o he p ocess and some new ones ela ed o ex e nal
con ex a ibu es o he p edic ion o PPIs. This ques ionnai e
is gene ic (i.e. independen o he p ocess). The ou pu o his
ac i i y is he ques ionnai e illed wi h he answe s p o ided
by he p ocess analys s. These answe s will be use ul du ing
he es o he p ocedu e. Following ou me hodology and
a e answe ing he ques ionnai e, we ocus on ques ions
3 and 7 o ob ain a p elimina y lis o Essen ial Business
En i y (EBE) candida es. These a e he essen ial elemen s o
he business p ocess, such as i ems o a i ac s, which should
be handled by he p ocess [34]. They a e ep esen ed wi h he
da a objec named EBE handled by he p ocess.
The nex s eps in ol e iden i ying he a ibu es ela ed
o he PPIs om he in o ma ion p o ided in he ques ion-
nai e. ORGANON me hodology jus iden i ies he in e nal
a ibu es, which a e all he inpu s and ou pu s in he business
p ocess ac i i ies modeled, ex e nal da a, a i ac s, business
ules, among o he s classi ied in [27]. This is pe o med in
S ep 3, which is a subp ocess ha g oups h ee ac i i ies
p e iously de ined in he ORGANON me hodology. The goal
o he i s wo ac i i ies (iden i y which ac i i ies consume
EBEs and iden i y on ological blocks) is o iden i y he essen-
ial ac i i ies ela ed wi h an EBE. The essen ial ac i i ies [28]
a e hose which ha e a di ec in luence on he goal o he
p ocess. The hi d ac i i y in ol es he elici a ion o a ibu es
om each essen ial ac i i y. The de ails on how o pe o m
hese ac i i ies a e p o ided a [27].
Howe e , o he p edic ion o PPIs, we also need o
iden i y ex e nal and p ocess a ibu es ha a e no de i ed
om he essen ial ac i i ies. Ex e nal a ibu es a e hose ha
e lec e en s un ela ed o he execu ion o he p ocess bu
can ha e a di ec in luence on he p ocess, e.g. he wea he .
P ocess a ibu es a e ela ed o inhe i ed cha ac e is ics o
he p ocess usually e lec ed in he e en logs o he o ga-
niza ions, such as he name o he ac i i y o imes amp.
The elici a ion o hese a ibu es is ca ied ou in S eps
4 and 5, espec i ely. These wo ac i i ies ecei e as inpu s
he ques ionnai e illed by p ocess expe s and he lis o
EBEs. Speci ically answe s o ques ions 10 o 14 p o ide us
222054 VOLUME 8, 2020
A. E. MÁRQUEZ-CHAMORRO e al.: CAP3
TABLE 2. Semi-s uc u ed guide o elici a ion o con ex a ibu es. Ques ions 1 o 9 ha e been upda ed om ORGANON p e ious e sion. Ques ion 10 o
14 ha e been included o his new e sion.
TABLE 3. PPI-a ibu e ma ix. Once in e nal/ex e nal con ex a ibu es ha e been ex ac ed, he ollowing able is illed.
FIGURE 2. Expe imen al p ocedu e o CAP3 me hodology.
in o ma ion abou he essen ial ex e nal and p ocess a ibu es
which could be conside ed as con ex a ibu es.
Once all he essen ial a ibu es (in e nal, p ocess and ex e -
nal) a e de ined, he p ocess business analys assesses he
impac o each a ibu e on he PPI (S ep 6). To do his,
we ecei e as inpu he lis o essen ial a ibu es, he p ocess
model and he lis o PPIs. The p ocess analys is esponsible
o linking he di e en a ibu es o he PPIs ha we wan
VOLUME 8, 2020 222055
A. E. MÁRQUEZ-CHAMORRO e al.: CAP3
o p edic , acco ding o hei ela ionship. Each a ibu e can
be ela ed wi h one o mo e PPIs. Table 3depic s he poin s
ha should be conside ed by he p ocess analys . Fi s column
s a es he selec ed PPI o be p edic ed. Second column ep e-
sen s he name o he a ibu es ob ained in he penul ima e
s ep o he me hodology depic ed in Figu e 1(i.e. in e -
nal, ex e nal and p ocess essen ial a ibu es). Thi d column
shows he possible alues o he a ibu es (nume ical o ca -
ego ical). Fou h column indica es he ype o a ibu e (in e -
nal, p ocess o ex e nal). Fi h column e lec s he answe o
he yes/no ques ion: ‘‘Do di e en alues o his a ibu e
(i.e. changes) ha e a di ec impac on he alue o he PPI?’’
and inally las column p o ides he eason o he answe o
he p e ious column. Fo ins ance, in an inciden esolu ion
p ocess, in which a echnician has o go o di e en places o
sol e he inciden , he physical loca ion in which he inciden
akes place may ha e an in luence on he indica o we a e
p edic ing, such as he esolu ion ime. The e o e, i will
be a con ex ual elemen use ul o he p edic i e moni o ing
p ocess. The e o e, i a a ia ion in he alue o he a ibu es
has a di ec impac on he p edic ion o he PPI, hey will
be iden i ied as con ex ual elemen s. A his poin , we also
analyze he g anula i y (i.e he le el o de ail conside ed o
each a ibu e) o con ex a ibu es. A ine g anula i y o
a ibu es usually leads o a mo e p ecise easoning while a
coa se g anula i y leads o a less p ecise one wi h he ben-
e i o being less compu a ionally demanding. Fo example,
a p ocess a ibu e can speci y he own we e an inciden has
occu ed. Maybe his le el o de ail is no aluable o he
p edic ion. We can g oup all he owns o he same egion.
In his way, we educe he numbe o possible alues o his
a ibu e and compu ing cos dec eases. The inal ou pu o
he me hodology would be a PPI-a ibu e ma ix wi h he
subse o con ex a ibu es de i ed om Table 3which ha e
a posi i e answe in he las column.
B. CAP3: A CONTEXT-AWARE PREDICTIVE MONITORING
TECHNIQUE
This sec ion desc ibes ou Con ex -awa e P ocess Pe o -
mance indica o P edic ion echnique (CAP3). This me hod,
depic ed in Figu e 2, includes a i s s ep which ep esen s
he iden i ica ion o con ex a ibu es o PPI p edic ion
p esen ed in he p e ious sec ion as an ex ended e sion o
ORGANON. The inpu s o he ac i i y a e he p ocess model
and he documen a ion o PPIs. In his ac i i y, we analyze
he impac o con ex a ibu es on he PPIs o be p edic ed.
As we ha e desc ibed, p ocess analys s ha know he de ails
on how he p ocess beha es, analyze his impac and hen
decide on he app op ia e con ex . The ou pu o his ac i i y
is he PPI-a ibu e ma ix, whe e we can ind he selec ed
PPI o be p edic ed and he con ex a ibu es necessa y o
he p edic ion. Then, a second ac i i y, named P ep ocess
he e en log, il e s he e en log L( o med by he se
o aces T) o emo e unnecessa y in o ma ion, en iches
he e en log wi h addi ional in o ma ion adding con ex
a ibu es, and ans o ms some a ibu es o he e en log.
The inpu s o his ac i i y a e he e en log o he p ocess,
he ex e nal a ibu es and he PPI-a ibu es ma ix whe e
con ex a ibu es a e ound. The ou pu o his ac i i y is he
da ase wi h all he a ibu es o he e en log. As de ined in
[35], S age 1 o Figu e 2 ep esen s he lea ning phase. In his
s age, he da ase is gene ally encoded in ea u e ec o s ha
can be in e p e ed by he p edic i e algo i hm. One o he
di e en echniques o he encoding o da a applied in he
li e a u e [7], [36] can be used. As a esul o his ac i i y we
ob ain a se o ea u e ec o s ha ep esen s he se o aces
T, whe e each ace is o med by a se o e en s Eand each
e en is composed by a se o a ibu es a.
Then, he p edic i e me hod is execu ed and gene a es a
p edic ion model Pas ou pu da a, based on he knowledge o
he aces To he e en log. This model is e alua ed o asses
i s alidi y, using he di e en aces o p ocess ins ances as
a es se , by means o quali y me ics. S age 2 o Figu e 2
ep esen s he p edic ion phase o a ypical p edic i e mon-
i o ing me hod. A un ime, he gene a ed model is applied
o ongoing ins ances in a gi en momen o he execu ion.
Then, he p edic i e model will de e mine he alue o he
p edic ed ou pu s o his p ocess ins ance, i.e. he esul o
he unc ion PI([Ei1,...,Eil]), ha compu es a p edic ion o
I om he ace [Ei1,...,Eil], whe e Eilis he las e en ha
ha e occu ed in ace Tia a gi en momen . .
V. EVALUATION
In o de o es he alidi y and applicabili y o ou app oach,
we applied he p oposed me hodology o he iden i ica ion
o con ex in o ma ion in wo eal-li e o ganiza ions. Once we
ha e ex ac ed con ex in o ma ion, we apply ou p edic i e
moni o ing echnique and also p o ide an expe imen al anal-
ysis o he ele ance o he de ini ion and inclusion o con ex
a ibu es o he p edic i e moni o ing o business p ocesses.
The es o he sec ion is o ganized as ollows: a desc ip ion
o he wo eal-li e o ganiza ions is p o ided in Sec ion V-A.
The de ails o he applica ion o ou me hodology o iden i y
con ex a e desc ibed in Sec ion V-B. Expe imen se up o
he applica ion o ou p edic i e moni o ing echnique is
de ined in Sec ion V-C1 and he applica ion o ou p edic i e
moni o ing echnique is desc ibed in Sec ion V-C.
A. SCENARIO ANALYSIS
Two eal-li e o ganiza ions we e conside ed in ou expe i-
men s: Techmas e (TM) and he IT Depa men o a Spanish
Heal hca e P o ide (SHP).
TechMas e is a B azilian company which p o ides IT
in as uc u e and managemen o IT en i onmen s.1The
s udied business p ocess models he Techmas e IT inciden
managemen . This p ocess s o es he in o ma ion o he man-
agemen o inciden s egis e ed a his company. A solu ion
should be es ablished o each inciden in o de o es o e
he se ice wi h minimum dis up ion o he business. A e
p o iding a solu ion o he p oblem and e i ying ha he
1h p:// echmas e .com.b /
222056 VOLUME 8, 2020
A. E. MÁRQUEZ-CHAMORRO e al.: CAP3
TABLE 4. PPI-a ibu e ma ix o Techmas e p ocess.
se ice is es o ed, he inciden is closed. Each inciden is
ep esen ed as a icke which e lec s a p ocess ins ance, and
each icke is composed by di e en a icles ha ep esen
p ocess e en s. A icke can ha e ze o o many a icles. Each
icke has a Techmas e employee assigned. Each inciden is
classi ied wi h a p io i y le el (1-5).
The IT Depa men o he Spanish Heal hca e P o ide
unde s udy (SHP)2p o ides IT se ices and suppo o he
di e en heal h cen e s associa ed. The s udied business p o-
cess ep esen s he SHP IT inciden managemen as i was
pe o med be ween 2014 and 2016. The p ocess is composed
by di e en e en s om he s a o he esolu ion o he
inciden . Inciden s can occu in any heal h cen e o hospi al
associa ed o his p o ide and can be a ended by phone, mail
o in ane . The inciden s a e classi ied in o h ee ca ego ies:
Ha dwa e, Sys em and O he . Each e en o he p ocess has
a ce ain le el o p io i y (low, medium and high). In his
scena io, a se ice le el ag eemen (SLA) is es ablished
conside ing ce ain PPIs. This SLA de e mines he penal ies
de i ed om he unde - ul illmen o a h eshold o each o
he PPIs. Thus, p edic i e moni o ing is necessa y o wa n
he possibili y o iola ion o he SLA. In his case, h ee
PPIs a e conside ed: K01, which de e mines i an inciden
was sol ed in a longe ime han expec ed (du a ion ime >
17h); K06, which de e mines i an inciden has been eopened
because i was no co ec ly sol ed; and K20, which indica es
an abuse o he s opping ime (idle ime >0). Idle ime is
he unp oduc i e ime on he pa o employees as a esul o
ac o s beyond hei con ol.
B. IDENTIFYING CONTEXT
This sec ion de ails he applica ion o ou me hodology
desc ibed in Sec ion IV-A.
1) EXPERIMENTAL SCENARIO: TECHMASTER
The p oposed me hodology was applied o he p ocess o
esolu ion o inciden s o Techmas e . Fi s ly, we analyze
he p ocess model and de e mine he PPIs o be p edic ed
(Figu e 1, S ep 1). We ha e de e mined ha he du a ion o he
p ocess and he numbe o inciden s di ided by he numbe
o se ice eques s (R) a e good candida es o be conside ed
as p edic ed PPI. This second PPI is an agg ega ed indica o .
2The name o he o ganiza ion can no be p o ided due o con iden iali y
issues.
Two expe s in he p ocess ha e pa icipa ed in he ul illmen
o he semi-s uc u ed guideline3(Figu e 1, S ep 2). The wo
p ocess analys s who collabo a ed wi h us we e: he p ocess
manage , esponsible o he echnical ope a ions, 15 yea s
wo king in his p ocess and he di ec o o he company,
esponsible o he ela ionships wi h clien s, 25 yea s wo k-
ing in his p ocess. Acco ding o he in e iew, we ha e ound
he lis o i ems (EBE) associa ed wi h he essen ial ac i i ies.
This lis consis s o he icke s and he epo s (i.e. so wa e
in en o y, ha dwa e in en o y and high impac inciden s).
La e , we ha e iden i ied a se o essen ial ac i i ies o he
p ocess (Figu e 1, Subp ocess in S ep 3) such as Open icke ,
Upda e icke , Communica e wi h clien , Discussion abou he
icke , Build epo s, Send epo s o he cus ome and Send
in oice ( ela ed o he icke ). Finally, acco ding o Figu e 1,
S eps 3, 4 and 5 a e ca ied ou and he con ex a ibu es iden-
i ied a e ollowing he guidelines a e: he human esou ce
in cha ge, he echnical cha ac e is ic o he equipmen , ma u-
i y le el o cus ome s’ in as uc u e, he emo e suppo and
he p io i y. All hese elemen s and he PPIs which a e ela ed
wi h hem a e shown in Table 4(Figu e 1, S ep 6).
2) EXPERIMENTAL SCENARIO: SHP
Fi s , we ha e analyzed he model and iden i y he di e en
PPI o he p ocess (Figu e 1, S ep 1). In his case, he ime
du a ion o an execu ed ins ance o he p ocess has been
selec ed as PPI o he p edic ion. This is due o he ac ha
se e al SLAs de ined in he p e ious sec ion o his p ocess
a e ime- ela ed. A SHP p ocess expe has also ul illed
he semi-s uc u ed guideline 4(Figu e 1, S ep 2). He is
esponsible o he Depa men o IT Se ice Managemen ,
wi h 13 yea s o expe ience and a high knowledge o he
p ocess since he has been he p ocess owne se e al yea s
and has been in ol ed in i s con inuous imp o emen since
he beginning o his wo k he e. Acco ding o he in e iew,
we ha e ound he EBE lis which is o med by he icke ,
he epo s and all he in o ma ion abou he sys em h ough-
ou he p ocess, such as in e ac ions o commen s. Associa ed
wi h he EBE lis , we ha e iden i ied as essen ial ac i i-
ies o he p ocess (Figu e 1, subp ocess e lec ed in S ep
3): Regis a ion o he inciden , De e mina ion o p io i y,
3In e iew wi h Techmas e expe s: h ps://bi .ly/2U n8uk
4In e iew wi h SHP expe : h ps://bi .ly/2VzTIL6
VOLUME 8, 2020 222057
A. E. MÁRQUEZ-CHAMORRO e al.: CAP3
TABLE 5. PPI-a ibu e ma ix o SHP p ocess.
Diagnosis and esolu ion. Finally, he con ex a ibu es iden-
i ied (Figu e 1, S eps 3, 4 and 5) a e ollowing he guide-
lines a e: he p io i y le el and he cen e ela ed o he
inciden . This a ibu e de ines he ype o he cen e (heal h
cen e o hospi al) and i s loca ion. The con ex a ibu es
ex ac ed and hei ela ed PPIs (Figu e 1, S ep 6) a e shown
in Table 5.
C. PREDICTIVE MONITORING EXPERIMENTS
1) EXPERIMENTAL SETUP
This sec ion de ails he se up o ou expe imen add essing
he s eps o ou p edic i e moni o ing echnique desc ibed in
Sec ion IV-B.
a: ENCODING
We ha e selec ed a ypical agg ega ion encoding desc ibed in
[36] as one o he mos used in li e a u e o encode he p ocess
cases. Thus, all e en s since he beginning o he case a e
conside ed. An agg ega ion unc ion is applied o he alues
aken by a speci ic a ibu e h oughou he case li e ime.
In ou case, his unc ion is he numbe o imes ha each
speci ic a ibu e appea s in he case ( equency encoding).
We ha e no di ided he cases in he e en log in o di e en
bucke s. This echnique is named Ze o bucke ing as de ined
in [36]. We ha e also inco po a ed he o de o he e en s as
a new a ibu e in all he logs (i.e. he ela i e posi ion o he
e en in he case), as well as he elapsed ime be ween he
e en and he beginning o he case and he ime be ween
he p e ious e en and he cu en one.
To selec ele an ea u es om he da ase s, ee-based
es ima o s a e employed. They can be used o compu e
impu i y-based ea u e impo ance, which in u n can be used
o disca d i ele an ea u es. In ou case ea u e impo ances
a e ob ained using Ex aT eesClassi ie om Sciki -lea n
lib a y [37].
b: BUILDING THE MODEL
As p edic i e algo i hm we ha e used andom o es [38] and
ex eme g adien boos ing [39] as seen in p e ious wo ks
in he li e a u e [3], [36]. Random o es (RF) is a combi-
na ion o p edic o ees such ha each ee depends on he
alues o a andom ec o es ed independen ly and wi h
he same dis ibu ion o each o hem. G adien Boos ing
is based on he combina ion o weak lea ne s, such as deci-
sion ees, o c ea e a s ong p edic i e model. The gen-
e a ion o he weak decision ees is done in a sequen ial
way, each ee being c ea ed in such a way ha i co -
ec s he e o s o he p e ious ee. One o he pa ame e s
o he algo i hm is he lea ning a e, which con ols he
deg ee o imp o emen o a ee wi h espec o he p e ious
one. We ha e employed Ex eme g adien boos ing (XGB)
[39] which is a g adien boos ing implemen a ion especially
no ewo hy. In [36], au ho s highligh XGBoos and RF as
wo o he bes echniques in p edic i e moni o ing. Fo
he implemen a ion, we ha e used RandomFo es Reg esso
me hod om Sciki -lea n [37] lib a y o machine lea ning in
Py hon and XGBReg esso me hod om he Xgboos Py hon
lib a y. We ha e applied an op imisa ion echnique o he
hype pa ame e s uning o bo h algo i hms. We ha e pe -
o med a andomized sea ch on hype pa ame e s using Ran-
domizedSea chCV om Sciki -lea n lib a y. The pa ame e s
a e op imized by c oss- alida ed sea ch o e pa ame e
se ings. The selec ed RF pa ame e s a e he execu ion
o he me hod a e: n_es ima o s=100, max_ ea u es=au o
and max_dep h=12. Fo XGB, he selec ed pa ame e s
a e: colsample_by ee=1, lea ning_ a e=0.3, max_dep h=6,
alpha=0 and n_es ima o s=100. We ha e spli he aces o
ou da ase in 80% o aining and 20% o es o alida e
he p edic i e algo i hms.
c: EVALUATION
We ha e used Mean Absolu e E o (MAE) and Roo Mean
Squa ed E o (RMSE) as e alua ion measu es since we
a e going o p edic a nume ic alue wi h eg ession mod-
els and hese a e e alua ion measu es commonly used in
he li e a u e o his pu pose [12], [36]. MAE is a isk
me ic co esponding o he expec ed alue o he absolu e
e o loss and RMSE is a isk me ic co esponding o he
expec ed alue o he squa e oo o he squa ed (quad a ic)
e o .
We also p o ide an expe imen al analysis o he ele ance
o he de ini ion and inclusion o con ex a ibu es o he
p edic i e moni o ing o business p ocesses. In o de o do so
e ec i ely, we ha e ob ained six di e en da ase s ex ac ed
om he e en logs o s udied p ocesses, which conside
he con ex in o ma ion ob ained a e he applica ion o ou
me hodology, acco ding o he ollowing classi ica ion:
•All (ALL): all a ibu es o he o iginal log a e
conside ed (excep iden i ie a ibu es).
•None (NONE): jus he basic a ibu es o he log, such
as name o ac i i y and imes amp (da e and ime when
he ac i i y was pe o med) a e conside ed.
•Au oma ic (AUTO): a ibu es de ec ed as ele an o a
decision ee algo i hms a e selec ed o he expe imen-
a ion.
•Random (RND): a se o andomly selec ed a ibu es a e
aken in o accoun .
•Con ex (CTXT): de ec ed con ex a ibu es by ou
me hod a e included.
•Wi hou con ex (WCTXT): de ec ed con ex a ibu es
a e excluded om he log.
222058 VOLUME 8, 2020
