Neu oFlinkCEP: Neu osymbolic Complex E en Recogni ion
Op imized ac oss IoT Pla o ms
Ou ania N ouni
Technical Uni e si y o C e e
[email p o ec ed]
Dimi ios Banelas
Technical Uni e si y o C e e
[email p o ec ed]
Nikos Gia akos
Technical Uni e si y o C e e
[email p o ec ed]
ABSTRACT
We demons a e Neu oFlinkCEP, he i s amewo k ha in e-
g a es neu al and symbolic Complex E en Recogni ion (CER) o e
a s a e-o - he-a Big Da a pla o m, also op imizing neu osymbolic
CER upon ope a ing o e IoT se ings. Neu oFlinkCEP ecei es ex-
p essed pa e ns as ex ended egula exp essions and au oma ically
ans o ms hem o FlinkCEP jobs pe de ice. To enable de ec ion
o simple e en s in ol ed in CER pa e ns, Neu oFlinkCEP can
in eg a e any neu al model in FlinkCEP jobs. To op imally assign
ope a o execu ion in-ne wo k, we inco po a e and ex end a s a e-
o - he-a IoT op imize .
PVLDB Re e ence Fo ma :
Ou ania N ouni, Dimi ios Banelas, and Nikos Gia akos. Neu oFlinkCEP:
Neu osymbolic Complex E en Recogni ion Op imized ac oss IoT
Pla o ms. PVLDB, 18(12): 5355 - 5358, 2025.
doi:10.14778/3750601.3750670
PVLDB A i ac A ailabili y:
The sou ce code, da a, and/o o he a i ac s ha e been made a ailable a
h ps://neu o linkcep.gi hub.io/.
1 INTRODUCTION
Complex E en Recogni ion (CER) [
6
] di e en ia es i sel om con-
en ional s eam p ocessing. Ins ead o lea ing o clien applica ions
he esponsibili y o assigning a meaning o he aw ou pu s o a
que y, CER Engines inges s eams o Simple E en s (SEs) and a se
o pa e ns co esponding o Complex E en s (CEs), i.e., in e es ing
phenomena ha should be acked by he applica ion; and moni o
i such CEs a e p og essi ely e ealed by combining inges ed SEs.
Conside he ollowing pa e n, s emming om a sma ac o y
scena io, whe e aw obo na iga ion da a a e moni o ed:
𝑅success ul_deli e y :=(¬S a ionDe ec ed)∗· (¬S a ionDe ec ed)·
(S a ionDe ec ed ∧ ¬Deli e yManeu e )∗·
(S a ionDe ec ed ∧Deli e yManeu e )
The SE S a ionDe ec ed occu s when a obo de ec s a s a ion in
he sma ac o y. The Deli e yManeu e SE occu s when he obo
is mo ing a a ce ain speed, changing di ec ions while maneu e -
ing o app oach he de ec ed ac o y s a ion. The CE
𝑅successul_deli e y
is sa is ied when ini ially a obo has no de ec ed a s a ion, hen
This wo k is licensed unde he C ea i e Commons BY-NC-ND 4.0 In e na ional
License. Visi h ps://c ea i ecommons.o g/licenses/by-nc-nd/4.0/ o iew a copy o
his license. Fo any use beyond hose co e ed by his license, ob ain pe mission by
emailing [email p o ec ed]. Copy igh is held by he owne /au ho (s). Publica ion igh s
licensed o he VLDB Endowmen .
P oceedings o he VLDB Endowmen , Vol. 18, No. 12 ISSN 2150-8097.
doi:10.14778/3750601.3750670
i de ec s one and — ha ing de ec ed he s a ion — i a ains he
equi ed speed and epea ed change o mo emen di ec ion o ap-
p oach i . The CER sys em con inuously e alua es he apidly in-
ges ed obo s eams, con e s hem o he a o emen ioned SEs and
deduces a success ul deli e y CE.
In such a scena io, each SE ep esen s he de ec ion o a be-
ha io ha can only be deduced by a machine o neu al lea n-
ing model. The ole o he neu al model is o ecei e s eams o
ames (e.g., om ision, LIDAR, o o he con ex ual cues) and he
maneu e ing beha io o deli e y (which can be qui e nuanced
o deduce only om posi ional s eams) and p o ide classi ica-
ion ou comes, i.e., class/symbol
𝐴=S a ionDec ec ed
and symbol
𝐵=Deli e yManeu e
, o CER o be possible. Then, a CER engine
will inges he a o emen ioned SEs (symbols A, B) and e alua e he
occu ence o he in ol ed CEs. E iden ly, such scena ios call o
bo h neu al in e ence and symbolic CER o ope a e syne gis ically.
FlinkCEP is he CER API o a s a e-o - he-a Big Da a pla o m,
namely Apache Flink. FlinkCEP ocuses on scaling-ou he com-
pu a ion o a numbe o machines in a compu e clus e /cloud,
wo king in pa allel on pa i ions o he s eams, o speed up con in-
uous analy ic ou comes. FlinkCEP p o ides a CER language o high
exp essi e powe [
1
,
6
] in e ms o o mula ing pa e ns o CEs.
Howe e , he e a e ce ain ba ie s o he adop ion o FlinkCEP.
Fi s , FlinkCEP equi es business analys s, who a e no necessa ily
expe p og amme s, o w i e unc ional p og amming code. Sec-
ond, pa e n exp ession and pa ame e iza ion in ol e cumbe some
no a ion, making he whole code w i ing p ocess e o -p one [
6
,
7
].
Thi d, wi h he p oli e a ion o IoT de ices as SE p oduce s, he
classic pa adigm in which we i s accumula e aw da a a he cloud
and hen submi a FlinkCEP job is se e ely subop imal [
8
]. Fo in-
s ance, in ou unning example, sending ideo ames om obo s
o he cloud and hen pe o ming CER would deple e he a ail-
able bandwid h, causing ne wo k la encies ha would p e en he
eal- ime cha ac e o he in ol ed applica ions. Wha should be
done ins ead, is o ship ained neu al models and FlinkCEP jobs o
ne wo k de ices, assign pa s o he SE and/o CER p ocess di ec ly
on hem (e.g. wi h Raspbe y Pi boa ds which nowadays ha e bo h
GPU and CPU capaci y), and only a subse o SEs and/o CEs should
be deli e ed o he cloud, ale ing o he occu ed e en s.
Despi e he ac ha ew p e ious e o s ha e in eg a ed neu al
and symbolic CER [
9
], no exis ing app oach has enabled nei he pa -
allel p ocessing o neu osymbolic CER no op imized, dis ibu ed
neu osymbolic CER o e IoT se ings. This wo k con ibu es ad-
ancing he s a e-o - he-a by ackling all he a o emen ioned
challenges. We demons a e Neu oFlinkCEP, he i s amewo k
ha in eg a es neu al (aka sub-symbolic) and symbolic CER o e
a s a e-o - he a Big Da a pla o m o pa allel p ocessing ha is
also op imized o ope a e dis ibu edly o e IoT se ings composed
Inpu SEs
Pa e n
FlinkCEP
synapSEFlow
RegEx2Neu oFlinkCEP
S
e
l
e
c
i
o
n
s
a
e
g
y
C
o
n
s
u
m
p
i
o
n
P
o
l
i
c
y
Window
Figu e 1: Ana omy o a Neu oFlink-
CEP Ope a o .
Job Dispa che
DAG*4CEPNeu al Ne Repo
RapidMine S udio GUI
Logical Wo k low Pa se
Logical Plan
Physical Plan
.pb ile
.pb pa h De ice
Regis y
S a is ics
Collec o ⑦①
②
③
④
⑤
⑥
Figu e 2: Neu oFlinkCEP Wo k low Design, IoT Op imiza ion & Dis ibu ed Execu ion.
o a ious de ices. To alle ia e business analys s om he bu den
o w i ing FlinkCEP p og ams, Neu oFlinkCEP ecei es exp essed
pa e ns in he o m o ex ended egula exp essions (RegEx) and
ans o ms hem o FlinkCEP jobs. To enable de ec ion o SEs and
CEs, Neu oFlinkCEP in eg a es any chosen, domain-speci ic neu al
model inside he FlinkCEP job deployed pe de ice. To op imally
decide whe he ope a o s o he CER wo k low should be execu ed
a he cloud o he de ice ne wo k side, Neu oFlinkCEP enhances
a s a e-o - he-a IoT op imize [
8
] wi h CER-speci ic op imiza-
ions [4].
2 ARCHITECTURE
Neu oFlinkCEP’s key a chi ec u al componen s a e: (i) he RegEx2-
Neu oFlinkCEP ope a o , (ii) he synapSE low ope a o and (iii) he
DAG*4CER Op imize . RegEx2Neu oFlinkCEP and synapSE low op-
e a o s a e nes ed in o a newly in oduced Neu oFlinkCEP ope a o .
We ha e de eloped a Neu oFlinkCEP GUI o g aphical wo k low
design using Neu oFlinkCEP ope a o s and we ha e inco po a ed
i as an ex ension o a comme cial pla o m, RapidMine S udio.
The RegEx2Neu oFlinkCEP ope a o : ecei es as inpu Ex-
ended Regula Exp essions desc ibing he pa e n based on which
a CE would be de ec ed, i.e. his nes ed ope a o desc ibes he sym-
bolic pa o a Neu oFlinkCEP ope a o . As shown in Figu e 1, each
such pa e n can be pa ame e ized wi h ime windowing cons ain s
as well as selec ion s a egies and consump ion policies suppo ed
by FlinkCEP. FlinkCEP suppo s he ollowing SE selec ion s a e-
gies (i) S ic Con igui y: whe e ma ching e en s appea s ic ly one
a e he o he , (ii) Relaxed Con igui y: whe e non-ma ching e en s
appea ing in-be ween he ma ching ones a e igno ed, and (iii) Non-
De e minis ic Relaxed Con igui y: ha allows non-de e minis ic
ac ions be ween ma ching e en s. Fo e en consump ion policies,
FlinkCEP p o ides he op ions: (i) NO_SKIP: p oduce all ma ches (ii)
SKIP_TO_NEXT: disca d pa ial ma ches ha s a ed wi h he same
CE e en (iii) SKIP_PAST_LAST_EVENT: disca d pa ial ma ches
a e CE ma ch s a ed bu be o e i ends. SKIP_TO_FIRST[p] and
SKIP_TO_LAST[p] a e simila o SKIP_TO_NEXT and SKIP_PAST_-
LAST_EVENT, espec i ely, bu hey use a pa e n
𝑝
o dic a e he
s a ( esp. las ) e en in he CE.
The synapSE low ope a o : nes s he Tenso Flow Ja a API wi hin
he Neu oFlinkCEP ope a o . I ecei es as inpu he Tenso low
(
.pb
) ile wi h a ained neu al model o he neu al p edic ion pa
o he Neu oFlinkCEP ope a o . The synapSE low ope a o loads
he ained neu al model p esc ibed and unde akes he espon-
sibili y o inges ing he incoming aw s eams (e.g, ideo ames,
posi ional s eams in ou unning example) composing ea u e ec-
o s. I hen eeds hese ea u es h ough he loaded neu al model
and de i es co esponding labels/symbols. synapSE low also has
wo mo e impo an esponsibili ies: (i) i di ec s he simple e en
ou pu s o he neu al model (depic ed as geome ic shapes in Fig-
u e 1) o he co e o FlinkCEP o he pa e n ma ching p ocess o
p oceed based on he pinpoin ed SEs, (ii) i lis ens o a b oadcas
s eam o model upda es h ough Ka ka. When a new model is
ecei ed, i is eco ded in ha cloud o de ice side, local B oadcas
S a e. synapSE low ensu es ha local Flink asks on he cloud o
de ice ecei e he model and ins all i , so ha all p edic ions o
synapSE low use he la es e sion o he ained neu al ne wo k.
An impo an obse a ion is ha , in case no
.pb
ile is speci ied, he
inpu o a downs eam Neu oFlinkCEP ope a o should be ano he ,
ups eam Neu oFlinkCEP ope a o eeding SEs eadily a ailable
o pa e n ma ching.
The DAG*4CER Op imize : inco po a es a s a e-o - he-a IoT op-
imiza ion algo i hm, namely DAG* [
8
] and ex ends i o DAG*4CER,
a a ia ion wi h plausible CER, FlinkCEP-speci ic op imiza ions [
4
].
We choose o in eg a e and ex end DAG* because i p unes la ge
amoun o he ope a o placemen sea ch space, gua an ees op imal-
i y o he IoT execu ion plan and has been shown o ou pe o m [
8
]
ecen , ele an app oaches in he ield [2, 3].
The DAG*4CER Op imize ecei es as inpu a logical wo k low
designed in RapidMine S udio. This logical wo k low inco po a es
he applica ion logic, bu i is dep i ed o any physical execu ion
de ails engaging he ne wo k o de ices. I is composed o a ious
Neu oFlinkCEP ope a o s (each po en ially equipped wi h a di e -
en synapSE low neu al model, alid o he SEs o he in ol ed
pa e n) exp essing he ules/pa e ns o be moni o ed. The ou pu
o he DAG*4CER Op imize is a physical plan p esc ibing he ne -
wo k si e (cloud o ne wo k de ice) each Neu oFlinkCEP ope a o
should be assigned o execu ion.
DAG* [
8
] is an A*-alike algo i hm ha swi ly ou pu s a phys-
ical execu ion plan ha is gua an eed o be op imal, wi hou ex-
haus i ely examining he en i e sea ch space. To achie e ha , i
opologically so s he logical wo k low and p og essi ely exam-
ines physical ins an ia ions o ne wo k de ices o he in ol ed
ope a o s. A each s ep
𝑛
, he algo i hm adds a new ope a o o he
pa ial physical plan ha has been buil so a and compu es an
es ima ed cos compu ed as
𝑔(𝑛)=𝑓(𝑛) + ℎ(𝑛)
, whe e
𝑓(𝑛)
is he
cu en cos o he pa ial physical plan and
ℎ(𝑛)
is a heu is ic cos
exp essing an unde -es ima ion o how much cos will be added o
he pa ial plan o make a ull plan, i.e, wi h all ope a o s assigned
o de ices. Then, he pa ial plan is inse ed in o a p io i y queue,
so ed on
𝑔(𝑛)
. The i s pa ial plan in he queue is hen dequeued
and he algo i hm a emp s o expand i by adding a new ope a o
as desc ibed abo e. When he dequeued plan is a ull physical plan,
he algo i hm concludes and ou pu s he op imal plan ound. To
ou pu a physical wo k low execu ion g aph, ins ead o a pa h ( ha
is he ou pu o he o iginal A* algo i hm), DAG* imposes wo im-
po an ules: (i) a each s ep, only one ope a o can be examined
o physical ins an ia ions o expand a dequeued pa ial plan, and
(ii) no logical ope a o can be ins an ia ed o any physical imple-
men a ion unless all o i s ups eam ope a o s ha e been included
in he cu en ly examined pa ial plan. Please e e o [
8
] o de ails.
Howe e e icien DAG* is in p uning he sea ch space o possible
Neu oFlinkCEP ope a o assignmen s, i is designed o gene ic
IoT s eam p ocessing scena ios. The e o e, i does no inco po a e
CER-o ien ed que y ew i ings ha can boos he pe o mance
o candida e physical CER plans. In he DAG*4CER Op imize we
ex end DAG* wi h he ollowing CER, FlinkCEP-speci ic que y
ew i ings:
•
Pa e n Decomposi ion. In case a CE in ol es SEs sensed on di -
e en de ices, he exis ing DAG* would assign he en i e pa e n
e alua ion a one de ice. This may be subop imal when all o he
de ices will ha e o communica e hei aw da a o ha de ice.
Ou DAG*4CEP p o ides, o he op imiza ion p ocess, he op ion
o decompose a pa e n o i s SEs and assign he e alua ion o SEs
a di e en de ices a he han assigning he e alua ion o an en-
i e pa e n o one single de ice. Conside , o ins ance a pa e n
AB+C
. Ins ead o deploying a Neu oFlinkCEP ope a o o
AB+C
a one de ice, DAG*4CEP may decide o assign a Neu oFlinkCEP
ope a o a De ice 0 wi h a neu al ne a he synapSE low spe-
cialized o de ec SEs o ype
A
and a Neu oFlinkCEP ope a o a
De ice 1, wi h a neu al ne a synapSE low specialized o de ec
SEs o ype Band C.
•
Ea ly Fil e ing employs selec i e c i e ia using he Flink ope a-
o
Da aS eam. il e ()
, igh a e SEs labeled by synapSE-
Flow become a ailable in Flink, bu be o e hey a e used in Flink-
CEP o pa e n ma ching. I he eby op imizes esou ce usage by
educing i ele an e en olume, be o e eaching CER ope a o s
like CEP.pa e n(), i allowed by he pa e n.
•
Reo de ing in ol es ea anging pa e n e alua ion condi ions
(
.whe e()
ope ands) wi hin he CER pa e n de ini ion in Flink-
CEP i sel . This eo de ing checks less complex, e y selec i e
p edica es i s o p omp ly ule ou non-ma ching e en s wi hin
he CER engine. Unlike Ea ly Fil e ing, Reo de ing does no p une
e en s be o e hey en e FlinkCEP, bu op imizes he p ocessing
sequence in FlinkCEP. This educes compu a ional o e head and
s a e complexi y in
Pa e nS eam
o FlinkCEP and imp o es
ma ching pe o mance.
•
Pushing P edica es Ups eam ope a es like Ea ly Fil e ing, bu
o connec ed Neu oFlinkCEP ope a o s assigned o di e en de-
ices, i.e., one ups eam’s ope a o CEs a one de ice a e he SEs
o a connec ed downs eam Neu oFlinkCEP ope a o a ano he
si e. To educe ne wo k communica ion cos s, pushing p edica es
ups eam mo es il e ing logic di ec ly o sou ce connec o s like
Ka kaSou ce
a he ups eam ope a o s. I hus elimina es i -
ele an e en s om going downs eam, sa ing conside able
ne wo k esou ces and la encies.
3
USER EXPERIENCE AND DEMO SCENARIOS
Sma Fac o y, Robo ic Scena io: This scena io in ol es obo s
mo ing in a sma ac o y e ain including physical obs acles and
10 p oduc ion s a ions. The mission o he obo s is o deli e pa -
icles om one p oduc ion s a ion o he o he ill ull p oduc
i em compila ion. Se e al GBs o da a a e p o ided by DFKI Kaise -
lau e n in he scope o he EVENFLOW p ojec acknowledged in
his wo k. Neu al models a e ained ia ROS simula ions. Exam-
ples o Complex E en s o in e es include: (i) Success ul Deli e y
CE: as in he example o Sec ion 1., (ii) Collision Reco e y CE:
o collisions occu ing be ween obo s o be ween a obo and
physical obs acles, whe e he obo manages o eco e and make
i s way o a p oduc ion s a ion, (iii) S a ion Skipping CE: a obo
de ec s and mo es owa ds a s a ion, bu hen heads o ano he
s a ion, missing deli e y (iii) P olonged S op a a S a ion CE: a o-
bo de ec s a s a ion and maneu e s o app oach i , bu i akes oo
much ime wi hou deli e ing, he e o e abo ing he app oach, (i )
Round-T ip Comple ion CE: a obo success ully pe o ms a ound
ip de ec ing, app oaching and success ully deli e ing p oduc ion
pa icles ac oss all p oduc ion s a ions.
Telecom Scena io: we will use Call De ail Reco d (CDR) [
5
] da a.
A si e in he ne wo k is ei he a co po a e da a cen e o edge com-
pu ing nodes (Raspbe y Pi, Je son Nano, Mobile Edge Compu ing
se e ) a BTS (Base T anscei e S a ion) nea he communica ion
an ennas. These cap u e call me ada a om Radio Access Ne wo k
(RAN) logs and can p ocess e en da a locally be o e o wa ding i
o he co e ne wo k. Examples o ele an CEs include: (i) Long Call
A Nigh CE: epo s long calls o p emium loca ions du ing nigh
hou s, (ii) F equen Long Calls A Nigh CE: aises an ala m upon
he occu ence o a numbe o long-las ing calls made o p emium
loca ions du ing nigh hou s pe Calle ID, (iii) F equen Long Calls
CE: igge s an ala m when mul iple calls made o a p emium loca-
ion sum up o a p olonged du a ion in a day, (i ) F equen Each
Long Call CE: No i ies o a high numbe o long-las ing calls made
o a p emium loca ion in a day. No ice ha hese do no in ol e
p ede ined alues o anges o "p emium loca ions", "long-las ing",
"nigh hou s", "p olonged du a ion" which depend on a ious ac-
o s such as calle loca ion and pas calling beha io . The e o e,
in ol ed SEs should be a ibu ed by synapSE low neu al models.
Also no e ha some o he CEs in he abo e scena ios inges
o he CEs as hei own SEs, i.e., in e connec ing Neu oFlinkCEP
ope a o s, in CER wo k lows. The DAG*4CER op imize will ha e
o decide whe he each Neu oFlinkCEP ope a o o he abo e CER
wo k lows will be placed a some edge de ice o a he cloud side.
Figu e 3: Neu oFlinkCEP Ope a o Pa-
ame e iza ion.
Figu e 4: Dashboa d o he Sma Fac o y, Robo ic Scena io.
Use Expe ience: Du ing he demons a ion, use s will be able o
choose ei he scena io and pe o m he numbe ed asks in Figu e 2:
1
○
The use in e ac s wi h he Neu oFlinkCEP GUI in Rapidmine
S udio o design and pa ame e ize hei own logical wo k lows,
besides some p ebuild ones p o ided by us, o each applica ion
scena io. The use d ags and d ops each Neu oFlinkCEP ope a o
on a can as and connec s Neu oFlinkCEP ope a o s and Ka ka
Sou ce/Sinks o de ine he da a low as shown on he uppe igh
pa o Figu e 2. As shown in Figu e 3, o each Neu oFLinkCEP
ope a o , he use g aphically de ines he pa e n o in e es , se-
lec ion s a egy, consump ion policy, ime window o he nes ed
RegEx2Neu oFlinkCEP ope a o . Also hey speci y he
.pb
ilepa h
o he nes ed synapSE low ope a o .
2
○
When he use submi s he logical wo k low, a Logical Wo k low
Pa se checks i s alidi y. I hen con e s his logical plan o a JSON
ile ha is ed o DAG*4CEP op imize and o a Neu al Ne Repo.
3
○
The DAG*4CER op imize de ec s he a ailable ne wo k de ices
ia a De ice Regis y and examines physical assignmen s o de ices
o each Neu oFLinkCEP ope a o , ou pu ing he op imal physical
plan. I also p ojec s he op imized physical plan back o he GUI
o RapidMine S udio. The e a e 3 op ions o he use o in e ac
wi h he DAG*4CER Op imize : (a)
op imize and deploy
: which
ins uc s he Op imize o di ec ly eed he op imal plan o he Job
Dispa che , (b)
only op imize
: which ins uc s he Op imize o
show he sugges ed physical plan in he GUI o he use o inspec
i o change i , be o e deploying i , (c)
only deploy
: which will
eed he wo k low, a e (b), o he Dispa che .
4
○
In
3
○
(a),
3
○
(c), he DAG*4CEP op imize eeds he physical plan
o he Job Dispa che , while he Neu al Ne Repo p o ides he
.pb
iles o he neu al ne s engaged in he CER wo k low.
5
○
The Job Dispa che submi s Flink jobs o he ne wo k si es based
on he assignmen o Neu oFlinkCEP ope a o s by DAG*4CEP.
6
○
De ec ed CEs a e con inuously isualized in an in e ac i e dash-
boa d. Figu e 4 shows he dashboa d o he obo ic scena io.
7
○
The deployed plan is moni o ed and s a is ics including p ocess-
ing and ne wo k la ency, h oughpu and o he ele an me ics
a e collec ed o u u e DAG*4CEP plan cos es ima ions.
The use will also be able o bypass DAG*4CER and manually
assign ope a o s o ne wo k si es (e.g. assign all Neu oFLinkCEP
ope a o s a he cloud side) so ha he audience can expe ience he
di e ence be ween he high la ency CE deli e y upon cen alized
p ocessing, e sus he op imal execu ion sugges ed by DAG*4CER.
Finally, he use will be able o ain a neu al model and deploy
i on cu en ly unning Neu oFlinkCEP ope a o s (in hei co e-
sponding synapSE low nes ed ope a o s, in pa icula ) a un ime.
ACKNOWLEDGMENTS
O. N ouni and N, Gia akos we e suppo ed by he EU p ojec EVEN-
FLOW unde Ho izon Eu ope ag eemen No.101070430. D. Banelas
was suppo ed by he EU p ojec CREXDATA unde Ho izon Eu ope
ag eemen No.101092749.
REFERENCES
[1]
A. A ikis, A. Ma ga a, M. Uga e, S. Vansumme en, and M. Weidlich. 2017. Com-
plex E en Recogni ion Languages: Tu o ial. In DEBS. 7–10.
[2]
X. Cha ziliadis, E. Zacha a ou, A. E aca , S. Zeuch, and V. Ma kl. 2024. E icien
Placemen o Decomposable Agg ega ion Func ions o S eam P ocessing o e
La ge Geo-Dis ibu ed Topologies. P oc. VLDB Endow. 17, 6 (2024), 1501–1514.
[3]
A. Chaudha y, S. Zeuch, and V. Ma kl. 2020. Go e no : Ope a o Placemen o a
Uni ied Fog-Cloud En i onmen . In EDBT. 631–634.
[4]
I. Flou is, N. Gia akos, A. Deligiannakis, M. Ga o alakis, M. Kamp, and M. Mock.
2017. Issues in complex e en p ocessing: S a us and p ospec s in he Big Da a
e a. J. Sys . So w. 127 (2017), 217–236.
[5]
Ioannis Flou is, Vasiliki Manikaki, Nikos Gia akos, An onios Deligiannakis, Mi-
nos N. Ga o alakis, Michael Mock, Sebas ian Bo he, Inna Ska bo sky, Fabiana
Fou nie , Ma ko S ajce , Tomisla K izan, Jona han Yom-To , and Taji Cu in. 2016.
FERARI: A P o o ype o Complex E en P ocessing o e S eaming Mul i-cloud
Pla o ms. In SIGMOD. 2093–2096.
[6]
N. Gia akos, E. Ale izos, A. A ikis, A. Deligiannakis, and M. Ga o alakis. 2020.
Complex e en ecogni ion in he Big Da a e a: a su ey. VLDB J. 29, 1 (2020),
313–352.
[7]
Nikos Gia akos, Eleni Kougioum zi, An onios Kon axakis, An onios Deligian-
nakis, and Yannis Ko idis. 2021. EasyFlinkCEP: Big E en Da a Analy ics o E e y-
one. In CIKM, Gianluca Dema ini, Guido Zuccon, J. Shane Culpeppe , Zi Huang,
and Hanghang Tong (Eds.). 3029–3033.
[8]
E. S e inio is, D. Banelas, N. Gia akos, and A. Deligiannakis. 2025. DAG*: A
No el A*-like Algo i hm o Op imal Wo k low Execu ion ac oss IoT Pla o ms.
In ICDE. 807–820.
[9]
M. Vilamala, T. Xing, H. Taylo , L. Ga cia, M. S i as a a, L. Kaplan, A. P eece,
A. Kimmig, and F. Ce u i. 2023. DeepP obCEP: A neu o-symbolic app oach o
complex e en p ocessing in ad e sa ial se ings. Expe Sys . Appl. 215 (2023),
119376.
