Recei ed Ma ch 30, 2022, accep ed Ap il 15, 2022, da e o publica ion Ap il 20, 2022, da e o cu en e sion Ap il 27, 2022.
Digi al Objec Iden i ie 10.1109/ACCESS.2022.3168977
Real-Time Focused Ex ac ion
o Social Media Use s
RODRIGO MARTÍNEZ-CASTAÑO, DAVID E. LOSADA , AND JUAN C. PICHEL
CiTIUS, Uni e sidade de San iago de Compos ela, 15782 San iago de Compos ela, Spain
Co esponding au ho : Juan C. Pichel ([email p o ec ed])
This wo k was suppo ed in pa by he Minis e io de Ciencia e Inno ación (MICINN) unde G an RTI2018-093336-B-C21 and G an
PLEC2021-007662; in pa by Xun a de Galicia unde G an ED431G/08, G an ED431G-2019/04, G an ED431C 2018/19, and G an
ED431F 2020/08; and in pa by he Eu opean Regional De elopmen Fund (ERDF).
ABSTRACT In his pape , we explo e a eal- ime au oma ion challenge: he p oblem o ocused ex ac ion o
Social Media use s. This challenge can be seen as a special o m o ocused c awling whe e he main a ge
is o de ec use s wi h ce ain pa e ns. Gi en a speci ic use p o ile, he ask consis s o apidly inges ing
Social Media da a and ea ly de ec ing a ge use s. This is a eal- ime in elligen au oma ion ask ha has
nume ous applica ions in domains such as sa e y, heal h o ma ke ing. The olume and dynamics o Social
Media con en s demand e icien eal- ime solu ions able o p edic which use s a e wo h o explo e. To mee
his aim, we p opose and e alua e se e al me hods ha e ec i ely allow us o ha es ele an use s. E en
wi h li le con ex ual in o ma ion (e.g., a single use submission), ou me hods quickly ocus on he mos
p omising use s. We also de eloped a dis ibu ed mic ose ice a chi ec u e ha suppo s eal- ime pa allel
ex ac ion o Social Media use s. This modula a chi ec u e scales up in clus e s o compu e s and i can be
easily adap ed o use ex ac ion in mul iple domains and Social Media sou ces. Ou expe imen s sugges
ha some o he p oposed p io i isa ion me hods, which wo k wi h minimal use con ex , a e e ec i e a
apidly ocusing on he mos ele an use s. These me hods pe o m sa is ac o ily wi h huge olumes o use s
and in e ac ions and lead o ha es a ios 2 o 9 imes highe han hose achie ed by andom p io i isa ion.
INDEX TERMS Big da a, dis ibu ed sys ems, ocused use ex ac ion, supe ised lea ning, in o ma ion
e ie al, eal- ime p ocessing, social media.
I. INTRODUCTION
Focused C awling echniques a e o ien ed o ex ac ing web
pages ha sa is y a gi en p ope y o opic o in e es . This
is ypically suppo ed by supe ised lea ning echnology
(e.g., ex classi ie s) ha help o decide wha links a e
wo h o explo e. Nowadays, i is easible o un c awling
ools ha e icien ly ex ac on opic con en s om massi e
eposi o ies such as he web o ce ain Social Media (SM)
sou ces. Focused C awling was i s in oduced in [1] and
coined in [2]. I has been in ensely discussed in he li e a u e
[3], [4] and i plays a undamen al ole, o example, in build-
ing e ical sea ch engines [5]. Ne e heless, o he bes o
ou knowledge, he eal- ime ex ac ion o SM use s ha a e
ele an o a a ge opic has ecei ed li le a en ion. In his
pape , we make a i s a emp o ill his gap. Many s udies
in he li e a u e ha e p oposed di e en al e na i es o c awl
The associa e edi o coo dina ing he e iew o his manusc ip and
app o ing i o publica ion was Massimo Ca a o .
o ex ac SM con en s. Howe e , he no ion o use is o en
igno ed and mos exis ing me hods do no y o an icipa e
which use s a e wo h o explo e. We claim ha we need
eal- ime echnology ha apidly inges s SM con en s and,
gi en li le con ex (e.g., he las use ’s SM submission),
mo es quickly owa ds ele an use s. We en isage in elligen
o ms o use c awling ha agg ega e he his o y o he a ge
use s wi h li le delay and do no o e load he sys em wi h
use s ha a e likely non- ele an . As epo ed in ou Rela ed
Wo k sec ion, some use -o ien ed SM me hods exis bu hey
a e o ien ed o ex ac samples o use s ( a he han massi ely
analysing he en i e SM si e in eal- ime) o do no implemen
any o m o use p io i isa ion.
Gi en a ce ain SM pla o m, we o mally de ine he ask o
Focused Use Ex ac ion (FUE) and in es iga e he e iciency
and e ec i eness issues in ol ed in ex ac ing a ge use s.
FUE has impo an applica ions in nume ous domains. The
apid g ow h o SM pla o ms and hei con en p oduc ion
dynamics demand Big Da a solu ions able o (1) e icien ly
VOLUME 10, 2022 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/ 42607
R. Ma ínez-Cas año e al.: Real-Time Focused Ex ac ion o Social Media Use s
inges con en s in eal- ime, (2) quickly eac o changes, and
(3) in elligen ly an icipa e which use s a e wo h o explo e.
This ask equi es me hods able o pe o m use ex ac ion
om la ge eposi o ies o SM da a. By sma ly p io i is-
ing SM use s who a e o po en ial in e es , we can ea ly
iden i y a ge indi iduals and, hus, suppo a numbe o
isk assessmen asks such as g ooming de ec ion, c iminal
ec ui men o ea ly iden i ica ion o psychological p oblems.
This equi es no only la ge-scale echnologies bu also sma
p edic i e componen s ha con inuously ank use s based on
eal- ime e idence (e.g., las SM pos ).
In popula SM sou ces, we cons an ly ob ain new pieces
o e idence. Each new SM pos is ypically sho and gi es
only a pa ial desc ip ion and con ex o he au ho o he pos .
The use c awling p ocess needs o e alua e his eal- ime
e idence and, o example, decide whe he o no o c awl
he en i e use his o y. This p edic ion ask esembles when
ocused web c awle s see a link o a new page and ha e o
decide whe he o no o download he linked page. Howe e ,
in FUE, he ull explo a ion o a gi en use o en equi es ens
o hund eds o eques s o he SM se e s and, hus, e ec i e
p io i isa ion o use s becomes c ucial.
Taking only in o accoun he pa ial use con ex allows us
o op imise he exis ing esou ces and ex ac he maximum
numbe o ele an use s pe uni o ime. In his pape ,
we p opose se e al use p io i isa ion me hods ha guide his
explo a ion p ocess and we s udy which o hese me hods lead
o high-speed FUE.
In o de o ex ac and p ocess huge amoun s o da a in eal
ime, i is necessa y o design and implemen an adequa e
a chi ec u e ha can scale up ho izon ally in a clus e o
compu e s. In his ega d, we con ibu e he e by de eloping
an adap able dis ibu ed c awling a chi ec u e based on Ca e-
nae1[10]–[12], Ka ka [13] and Docke [14]. The a chi ec u e
was de eloped o suppo FUE bu i can be easily adap ed o
o he scena ios such as eal- ime que y-based il e ing, opic
ex ac ion o summa isa ion.
Summing up, he pape desc ibes he p ac ical applica ion
o use p io i isa ion me hods in eal- ime in elligen ex ac-
ion o SM use s. The main con ibu ions o his pape a e:
•A o mal de ini ion o he Focused Use Ex ac-
ion (FUE) ask and a sys ema ic analysis o he main
e ec i eness and e iciency issues in ol ed.
•Se e al heu is ic-based me hods ha consis en ly
inc ease he ha es a io o a ge use s and a com-
p ehensi e s udy o hei ela i e impo ance unde
eal- ime SM expe imen s. To he bes o ou knowledge,
his is he i s s udy ha analyses he abili y o a se o
use - ela ed a iables in use c awling p io i isa ion.
•A modula SM c awle ha pe o ms eal- ime ex ac-
ion and explo a ion o SM communi ies and use p o-
iles. An expe imen al e alua ion pe o med on Reddi
demons a es he eal- ime p ocessing capabili ies o
he c awle . The compe ing me hods a e benchma ked
1h ps://gi hub.com/ca enae
agains each o he unde eal- ime expe imen s whe e
he me hods un in pa allel. In his way, he ela i e
bene i s o each p io i isa ion me hod a e no a ec ed
by seasonal ac o s.
•A Docke -based dis ibu ed a chi ec u e ha can be eas-
ily deployed and scales by inc easing he numbe o
ins ances o each module in a clus e o compu e s (pub-
licly a ailable2unde a F ee So wa e license).
The pape is s uc u ed as ollows. Sec ion II desc ibes he
me hodology ollowed by ou c awle o ex ac SM con en s
and ank use s ollowing di e en s a egies. Sec ion III dis-
cusses he a chi ec u e o he sys em, i s modules, connec-
ions and he scalabili y o he p oposed solu ion. Sec ions IV
and Villus a e how di e en use ex ac ion me hods un and
e ol e unde eal- ime expe imen s. In Sec ion VI he adap -
abili y o ou app oach o o he Social Media is discussed.
Sec ion VII discusses ela ed wo k and, inally, Sec ion VII
con ains he main conclusions o his s udy.
II. METHODOLOGY
A. FOCUSED USER EXTRACTION
Gi en a p o ile o in e es ha de ines a ge use s ( a ge use
p o ile), he ele ance o each newly de ec ed use has o be
es ima ed. I possible, his needs o be done using minimal
con ex ual in o ma ion. Fo example, c awling he en i e his-
o y o use pos s, e en i possible, in oduces a signi ican
o e head on he use c awling p ocess. And many o hose
use s migh be non- ele an . An e ec i e use c awle mus
p ocess ligh olumes o use da a (e.g., las pos ) and only
make a ull ex ac ion o use ’s da a o hose use s ha seem
o be on- a ge . This ep esen s a wo-s age p ocess in which
SM web pages a e cons an ly explo ed o ex ac candida e
use names (e.g., om web pages wi h he newes h eads
o use pos s) and he use c awle p ocess p edic s which
candida e use s a e wo h o explo e (i.e., which use s a e
ully explo ed).
The a ge use p o ile can be ep esen ed wi h a d i ing
que y o wi h a se o examples (use s who ul il he a ge
p o ile and which can be used o build a use classi ie ). Gi en
a Social Media pla o m, he aim o he ask is o ex ac
as many a ge use s pe uni o ime as possible. E ec i e
algo i hms ha sol e his p oblem should somehow p io i ise
hose use s ha a e p esumably on a ge and op imise he use
o a ailable esou ces.
The e a e wo main use ul me ics o e alua e he pe o -
mance o his ask. Fi s , he ha es a io (HR), o p ecision,
as he numbe o ele an use s ex ac ed di ided by he o al
numbe o ex ac ed use s. A use is deemed as ele an i i is
ca aloged as so wi h high con idence by a e e ence classi ie .
This e e ence classi ie ep esen s he a ge p o ile o use s.
This means ha he en i e use p o ile –conca ena ion o all
hei pos s–, when passed o he classi ie , exceeds a ce ain
h eshold o he classi ie (p edic ion p obabili y highe han
0.9 in he case o he logis ic eg ession classi ie ). A second
2h ps://gi hub.com/blind-snipe
42608 VOLUME 10, 2022
R. Ma ínez-Cas año e al.: Real-Time Focused Ex ac ion o Social Media Use s
pe o mance me ic is he ex ac ion speed exp essed as he
numbe o ex ac ed use s pe uni o ime.
The e is a huge numbe o ac i e use s on Social Media.
In ealis ic scena ios, whe e ha dwa e is limi ed, ex ac ing
and p ocessing e e y hing would be un easible due o he
massi e olume o con en s p oduced in eal- ime. Fu he -
mo e, he e a e usually some access es ic ions imposed
by he sou ce (e.g., obo s. x cons ain s). As a con-
sequence, ligh weigh explo a ion needs o be pe o med in
o de o ex ac SM use s. A c ucial issue is how o es ima e
use ’s ele ance om he li le con ex a ailable. No e also
ha , e en i we ha e ex ensi e ha dwa e esou ces and com-
pu a ional powe a ou disposal, a andom o a bi a y explo-
a ion o SM use s is no pe inen . Such näi e ex ac ion
o use s slows down he iden i ica ion o ele an use s and,
hus, canno suppo he ea ly iden i ica ion o a ge use s.
Typically, when explo ing a Social Media si e looking o
new use s, candida e use names can be ound on pages ha
lis use s’ desc ip ions o on pages ha con ain a h ead o
use ’s pos s (e.g., he en y page o a gi en use commu-
ni y o o um). The a ailable pos s o en include a i le o
desc ip ion, possibly a snippe , in o ma ion abou he au ho
and o he ele an social p ope ies (such as he numbe o
eplies, o es, e c.). These ypes o lis ings se e as seed pages
o ex ac ing candida e use s. Among he a ailable e idence
ha can be used o guide he c awle , we ha e elemen s such
as he i le o he ex o he las use ’s pos , he da e and
ime o publica ion, he numbe o commen s, he sco e o he
numbe o imes a pos has been liked, he numbe o imes he
pos has been sha ed o e-pos ed, e c. In his pape , we s udy
which o he a ailable pieces o e idence a e p edic i e o
ele ance in o de o guide he c awling p ocess.
B. FOCUSED USER EXTRACTION ON REDDIT
Reddi is a Web pla o m whe e use s submi con en (pos s)
such as ex , images o links and o he use s can commen
and o e o o agains hem. Commen s, a he same ime,
can also be commen ed and o ed. The pla o m is subdi ided
in o communi ies (sub eddi s) ocused on speci ic opics. I is
cu en ly anked in Alexa [15] as he 7 h websi e wi h mo e
a ic in he Uni ed S a es and 19 h globally. The numbe o
a e age daily ac i e use s is highe han 52 million, and he e
a e mo e han 100,000 ac i e communi ies [16].
We ha e chosen Reddi as ou e e ence pla o m o expe -
imen ing wi h FUE algo i hms. This decision was based on
se e al c i e ia. Fi s , i is one o he h ee la ges Social
Media pla o ms wi h inc easing popula i y. Second, he
di e si y o use s and communi ies make Reddi a pe ec
place o use mining. Thi d, i s e ms o se ice s a e hei
open philosophy and willingness o suppo ex e nal applica-
ions o se ices connec ed o Reddi . This allows us o easily
conduc c awling esea ch on his pla o m. Fu he mo e, he
lessons lea ned om ou s udy can be po en ially ans e ed
o o he social ne wo ks.
The c awling p ocess is composed o wo main s eps. Fi s ,
he c awle has o choose he nex sub eddi om which
o explo e he la es pos s and commen s. This sub eddi
explo a ion s ep cons an ly de ec s new use names ha a e
candida es o be ully explo ed. The second s ep consis s o
anking he candida e use s based on he a ailable e idence.
To mee his aim, a use anking is buil (using a numbe o
use p io i isa ion me hods) and he op use is selec ed and
ully explo ed (all his/he submissions a e collec ed).
The i s s age (sub eddi explo a ion) wo ks wi h a sub-
eddi on ie ha s o es he sub eddi names ha ha e been
ound so a . Ini ially, he main on page o Reddi is used
as a seed o ex ac some sub eddi names ha a e s o ed in
he on ie . This on ie g ows du ing he c awling p ocess
(any web page e ie ed om Reddi po en ially con ains
e e ences o unseen sub eddi s) and he selec ion o he nex
sub eddi is done andomly om he sub eddi s a ailable in
he on ie .
The second s age consis s o choosing he nex use o be
analysed in dep h. When explo ing sub eddi s, new use s a e
disco e ed and anked acco ding o di e en me hods (see
below). A e a use is ully explo ed (all his/he submissions
a e e ie ed), he/she is ma ked as p ocessed on he lis o
known use s and, hus, he/she will no be p ocessed again.
Algo i hm 1desc ibes ou ocused use c awle in pseu-
docode. The e a e ou main p ocedu es. Fi s , du ing he
Ini ialisa ion, he Reddi ’s on page is downloaded. This
page is a lis o he mos ecen pos s on he pla o m om any
sub eddi . The names o he sub eddi s a e collec ed and used
as seeds. In addi ion, he ea u es o he pos s ( o example,
i le o numbe o commen s) a e ex ac ed. These ea u es a e
used o upda e he anking o candida e use s (acco ding o
di e en me hods, as we will explain below). A e ob aining
he seeds, h ee asks a e launched in pa allel:
•Communi y Explo a ion. A sub eddi is picked a an-
dom om he lis o known communi ies. The on page
o he selec ed sub eddi is isi ed, ex ac ing he la es
pos s and hei ea u es. The use anking is upda ed by
agg ega ing he new ea u es wi h he exis ing ones.
•Use Ex ac ion. The op use om he use anking is
selec ed. All he a ailable submissions and commen s
pos ed by his use a e ex ac ed om he SM pla o m.
No e ha his equi es se e al calls o Reddi and, hus,
he ocused use c awle aims a ex ac ing only hose
use s who a e likely ele an .
•E alua ion. All he ex ac ed use s a e queued o e alu-
a e hei ele ance. To mee his aim, we employ a use
classi ie . This classi ica ion ool analyses he ull use
in o ma ion and de e mines whe he o no he use is
on- opic. In his way, his e alua ion componen assesses
whe he o no he ex ac ion o he en i e use his o y
was wo hwhile. This e alua ion s a egy ollows [2],
[17], whe e he ele ance o he ex ac ed pages is e al-
ua ed wi h a classi ie . The main eason o es ima e ele-
ance in his way is ha he e a e p ac ical impedimen s
o pe o m a manual analysis o housands o ex s om
hund eds o housands o use s.
VOLUME 10, 2022 42609
R. Ma ínez-Cas año e al.: Real-Time Focused Ex ac ion o Social Media Use s
Algo i hm 1 Pseudo-Code o he Focused Use C awle o Reddi
1: p ocedu e Ini ialisa ion
2: i leng h o sub eddi _lis == 0 hen
3: h ml_code =load_u l(00h ps:// eddi .com/ /all/new/00)
4: sub eddi s =ex ac _sub eddi s(h ml_code)
5: o each sub eddi in sub eddi s do
6: i sub eddi no in sub eddi _ on ie hen
7: sub eddi _ on ie .add(sub eddi )
8: pos s_ ea u es =ex ac _pos s_ ea u es(h ml_code)
9: o each pos _ ea u es in pos s_ ea u es do
10: use s_ anking.upda e_use _ ea u es(pos _ ea u es)
11: do in pa allel
12: Communi y Explo a ion()
13: Use Ex ac ion()
14: E alua ion()
15:
16: p ocedu e Communi y Explo a ion
17: while ue do
18: selec ed_sub eddi =pick_a _ andom(sub eddi _ on ie )
19: h ml_code =load_u l(00h ps:// eddi .com/ /00 +selec ed_sub eddi +00/new/00)
20: pos s_ ea u es =ex ac _pos s_ ea u es(h ml_code)
21: o each pos _ ea u e in pos s_ ea u es do
22: use s_ anking.upda e_use _ ea u es(pos _ ea u es)
23:
24: p ocedu e Use Ex ac ion
25: while ue do
26: selec ed_use =use s_ anking.ge _ i s _use ()
27: h ml_code =load_u l(00h ps:// eddi .com/use /00 +selec ed_use )
28: use _pos s =ex ac _pos s(h ml_code)
29: o each use _pos in use _pos s do
30: sub eddi _ on ie .add(use _pos [00sub eddi 00])
31: use s_pos s_ o_e alua e.add(use _pos s)
32:
33: p ocedu e E alua ion
34: while ue do
35: use _pos s =use s_pos s_ o_e alua e.ge _ i s ()
36: p obabili y =classi ica ion_model.p edic (use _pos s)
37: i p obabili y >THRESHOLD hen
38: posi i e_use s.add(use _pos s)
39: else
40: nega i e_use s.add(use _pos s)
We now make a b ie discussion on he compu a ional
complexi y in ol ed in he p ocedu es desc ibed abo e:
•Communi y Explo a ion: The communi y explo a ion
p ocedu e is qui e ligh weigh . Fo each sub eddi , i ge s
he sub eddi ’s page o new con en s, ex ac s some
me ada a om he a ailable pos s and upda es some
use ’s ea u es. The ime complexi y o his p ocedu e
is O(s), whe e sis he numbe o communi ies o sub-
eddi s. Wi h p ope da a s o age s uc u es, he cos o
choosing a andom sub eddi om he sub eddi on ie
(line 18) and upda ing use ’s ea u es (lines 21-22) do
no add u he complexi y o he p ocess.
•Use Ex ac ion: The compu a ional load o he use
ex ac ion p ocedu e g ows linea ly wi h he numbe o
use s and, o each use , he use ’s pos s a ailable a he
use ’s on page need o be c awled and added o he
co esponding da a s uc u es. This means ha he ime
complexi y o his p ocedu e is O(u·pa g), whe e uis
he numbe o use s and pa g is he mean numbe o pos s
published in he use ’s on page. Obse e ha , ha ing
p ope da a s o age s uc u es, he cos s o ex ac ing
he nex use om he a ailable lis o use s (line 26),
inco po a ing he pos ’s sub eddi o he sub eddi on-
ie (line 30), and inco po a ing he use ’s pos s o he
42610 VOLUME 10, 2022
R. Ma ínez-Cas año e al.: Real-Time Focused Ex ac ion o Social Media Use s
FIGURE 1. Simpli ied a chi ec u e diag am o he eal- ime ocused use c awle o Reddi .
e alua ion’s da a s uc u e (line 31) can be conside ed
negligible.
•E alua ion: This p ocedu e essen ially consis s o pass-
ing use s h ough he classi ica ion model. The cons uc-
ion o he classi ie om (ex e nal) aining da a is no
pa o he c awling p ocess ( he c awling p ocess simply
eads he ained classi ie and he ex ec o ize om
disk). The ex ec o ize p ocesses he use ’s pos s and
p oduces a nume ical ep esen a ion ha is hen passed
o he classi ie . We wo k wi h a ligh weigh classi ie
(Logis ic eg ession, see sec ion IV-F) whose p edic ion
cos g ows linea ly wi h he numbe o ea u es. The
numbe o ea u es depends on he numbe o unique
wo ds in he aining co pus and i is o en la ge han he
numbe o wo ds in he use ’s pos s. As a consequence,
he ime complexi y essen ially depends on he numbe
o use s and he numbe o classi ica ion ea u es: O(u·
), whe e uis he numbe o use s and is he numbe
o classi ica ion ea u es, which is bounded he size o
he ocabula y o he aining collec ion. Again, wi h
p ope s o age s uc u es, he cos o ex ac ing he i s
use (line 35) and inco po a ing he use o he posi i e
o nega i e lis s (lines 38-40) is insigni ican .
In e ms o space complexi y, he size o sub eddi on-
ie is bounded by he numbe o a ailable sub eddi s
(less han 200k) and, in any case, we only need o s o e a
s ing o each sub eddi ( he name). The use s’ ea u es
and pos s demand mo e space bu he c awling p ocess,
i needed, can emo e he ea u es and pos s o he use s
who ha e been al eady passed o he classi ie .
No e also ha communi y explo a ion, use ex ac-
ion and e alua ion un in pa allel (lines 11-14) and,
u he mo e, he ope a ion o each o hese h ee p o-
cedu es can be easily pa allelized. As a gued in he nex
sec ion, depending on he load o each p ocedu e, we can
add mo e compu a ional esou ces (e.g., o do pa allel
ex ac ion o mul iple use s).
III. REAL-TIME CRAWLING ARCHITECTURE
The ocused use c awle is composed o mul iple mic ose -
ices, designed o acili a e he scalabili y o he sys em
when unning on a clus e o compu e s. Modules a e in e -
connec ed o o m an execu ion g aph buil on Ca enae
[10]–[12], a Py hon lib a y o easy de elopmen o scalable
s eam execu ion g aphs. Ca enae uses he Ka ka message
b oke o in e connec di e en mic ose ices. Ca enae-based
sys ems can scale up ho izon ally by inc easing he numbe o
ins ances o any mic ose ice wi hou u he con igu a ion.
Ou a chi ec u e pe mi s he deploymen o his ocused
c awle in in as uc u es whe e dynamic esou ces can be
con igu ed, such as compu ing clouds (e.g., Amazon Web
Se ices, Mic oso Azu e, Google Cloud Pla o m, Alibaba
Cloud). In addi ion, when ew ha dwa e esou ces a e a ail-
able, he c awle can scale up o down wi hou in e up ing
i s execu ion. Due o his dis ibu ed a chi ec u e, he ocused
use c awle can be simply scaled by augmen ing he numbe
o ins ances (e.g., i he e is a bo leneck in one o he mod-
ules). The main modules and connec ions o he sys em a e
shown in Figu e 1. The modules o he ocused use c awle
o Reddi a e explained below:
•Sub eddi Explo e . A his s age, sub eddi s a e
eques ed o he Sub eddi Choose h ough RPC
(Remo e P ocedu e Call). When a sub eddi is assigned,
his module explo es i by ex ac ing he las pos s and
VOLUME 10, 2022 42611

R. Ma ínez-Cas año e al.: Real-Time Focused Ex ac ion o Social Media Use s
commen s. Fo each submission, i ex ac s he a ailable
ea u es and sends hem o a Use Upda e ins ance
( o example, i he c awle p io i ises use s based on
he numbe o o es o he use submissions, hen any
new upda e on he use ’s o es is handled by he Use
Upda e ). When ini ialising he sys em, he Sub eddi
Explo e module ex ac s he sub eddi s o he ex s
a ailable unde /all, a pseudo-communi y ha holds
oge he he pos s om all communi ies. No e ha , a he
beginning o he p ocess, he e is no knowledge on
which sub eddi s a e mo e p omising (po en ially con-
ain mo e on- opic use s) and, hus, we need o s a om
a gene al page o he pla o m.
•Sub eddi Choose . This module is esponsible o
selec ing a new sub eddi o be sen o he sub eddi
explo e s. In his wo k, we expe imen wi h a andom
selec ion o sub eddi s bu he pla o m can easily imple-
men mo e sophis ica ed selec ion me hods.
•Use Upda e . The Use Upda e is esponsible o
upda ing he use ea u es ha guide he ocused c awl-
ing p ocess. Fo example, i a e ages, o e e y use , he
numbe o commen s in his/he pos s. A ull desc ip ion
o use ea u es and ela ed equi emen s o he Use
Upda e module a e gi en in Sec ion IV.
•Use Ex ac o . The Use Ex ac o ins ances eques
use s o he Use Choose module and ex ac he max-
imum possible numbe o use ex s.3The ex ac ed
con en is sen o he Use Classi ie , which e alua es
he pe o mance. Addi ionally, he sub eddi s disco -
e ed du ing his p ocess a e sen o he Sub eddi S o e .
•Use Choose . This module handles he selec ion o new
use s o be p ocessed (acco ding o he use p io i isa ion
p ocess ha guides he ocused c awle ). Use ex ac o s
in oke he Use Choose module h ough RPC once hey
inish he ex ac ion o a use . Since he new use s a e
eques ed only when needed, he use selec ion is cohe -
en wi h he upda ed anking s a e in e e y momen .
The Use Choose assigns a unique use o each use
ex ac o . I also ma ks he selec ed use s as p ocessed
in o de o a oid epea ed p ocessing.
•Use Classi ie . The main goal o his module is o
e alua e he pe o mance o he ocused use c awle .
I ecei es all he a ailable ex s o e e y ex ac ed use ,
classi ies he use and s o es he p obabili y p o ided by
he use classi ie .
•Sub eddi S o e . In his phase, he p e iously unseen
sub eddi names a e s o ed in he da abase. I ecei es
new sub eddi s om a Sub eddi Explo e ins ance
du ing he ini ialisa ion and om he Use Ex ac o
ins ances.
•S a s Dumpe . This module is connec ed o all he o he
modules and s o es e en s wi h in o ma ion ela ed o
he eal- ime execu ion. E e y e en is imes amped,
3In ou expe imen s, we espec Reddi ’s obo s. x and, hus, each
eques ge s a maximum o wo hund ed ex s (pos s and commen s).
so he beha iou o he sys em in se e al dimensions
can be easily analysed du ing o a e he execu ion. This
module is no ep esen ed in Figu e 1due o i s numbe
o in e connec ions.
•Model T aine . This module is esponsible o aining
he use classi ie . In ou ini ial expe imen s, he use
classi ie is buil once ( om ex e nal da a) and emains
unchanged he ea e .
•Ba ch Upda e . This module is needed o one o he
ad anced p io i isa ion me hods de ailed in Sec ion V.
I upda es use - ela ed da a a e lea ning new use clas-
si ica ion models.
MongoDB [18] is used as a s o age sys em and o manag-
ing he ankings. Ae ospike [19], a memo y-based key- alue
s o e, is used o a oid p ocessing epea ed pos s du ing he
c awling p ocess.
IV. BASIC PRIORITISATION METHODS
In ou a emp o ea ly iden i y a ge use s du ing c awling,
we ha e de ined se e al me hods o guide he FUE p ocess.
S anda d ocused c awle s employ a numbe o s a egies
when, o example, hey ge a new link and hey need o
p edic whe he i is wo h o download he linked web page.
In FUE, he a ailable e idence is di e en and, hus, we need
o design new s a egies o p io i ising he use s ha a e
ound. These me hods wo k wi h a numbe o ea u es ha
a e a ailable when b owsing he SM websi e. The me hods
conside ed a e:
A. RANDOM
A andom (RND) selec ion o he nex use . This is a näi e
baseline used o compa ison agains mo e sophis ica ed
me hods.
B. AVERAGE SCORE OF USER’S TEXTS
In SM pla o ms, use s’ in e ac ions a e o en sco ed, o ed o
liked. The popula i y o use ’s pos s and commen s migh be
a aluable clue o guide FUE. Fo example, a ocused use
ex ac ion p ocess migh be in e es ed in ea ly iden i ying
hose use s emi ing highly in luen ial con en s. In Reddi , he
sco e o a pos e lec s he u ili y and quali y o he con en
wi hin i s con ex ( he sub eddi ). The in e ac ion o he use s
de e mines he o al sco e o a pos (use s can add, up o e,
o sub ac , down o e, a poin o each pos o commen ). The
sco e o a Reddi ’s pos o commen is calcula ed by summing
he posi i e and nega i e o es. Depending on he ocus o he
c awle , his ea u e migh be indica i e o a ge use s (e.g.,
a ocused ex ac ion o o ensi e use s migh bene i om
he exis ence o many nega i e o es). Du ing he c awling
p ocess, his a e age sco e is accumula ed o each use . No e
ha his is compu ed om he use ex s seen so a (web
pages downloaded) and, hus, i is usually an incomple e iew
o he o e all sco e o his use .
Le Ube he se o candida e use s and Su=
{Su 1,Su 2, . . .}be he se o sco es o each obse ed ex o a
42612 VOLUME 10, 2022
R. Ma ínez-Cas año e al.: Real-Time Focused Ex ac ion o Social Media Use s
gi en use u. Le Sube he a e age sco e o he known ex s o
a gi en use . We p opose wo di e en ways o de e mine he
chosen use (cuse ). These wo a ian s selec he nex use
based on he highes o lowes a e age sco e, espec i ely.
The i s app oach builds (and cons an ly upda es) a anking
o use s by dec easing a e age sco e and ex ac s he op use .
This will be e e ed o as Highes A e age Sco e (HAS).
Simila ly, he Lowes A e age Sco e (LAS) builds a anking
o use s by inc easing a e age sco e.
cuse =a g max
u∈U
Su(1)
cuse =a g min
u∈U
Su(2)
C. AVERAGE NUMBER OF COMMENTS IN USER’S TEXTS
This ea u e also weigh s use s based on he impac o hei
ex s. In his case, we employ he a e age numbe o com-
men s associa ed o he pos s o a gi en use . Use s wi h
long h eads o commen s o hei pos s migh be pa icula ly
ele an o a gi en FUE p ocess. Le Cube he numbe o
commen s o each obse ed ex o a gi en use u. Le Cu
be he a e age numbe o commen s o he known ex s o a
gi en use u.
Again, we de ine wo app oaches ha build descending and
ascending ankings wi h he a e age numbe o commen s
o e e y use : Highes A e age Commen s (HAC) and Low-
es A e age Commen s (LAC), espec i ely. The anking is
upda ed e e y ime new da a is a ailable. When he sys em
needs o ex ac a new use , he op use o he anking is
selec ed.
cuse =a g max
u∈U
Cu(3)
cuse =a g min
u∈U
Cu(4)
D. TIME-BASED USER SELECTION
The pos ime o use ’s ex s migh also be use ul o dis-
co e ing a ge use s. Fo example, a FUE p ocess aiming
a ex ac ing use s showing signs o psychological p ob-
lems (e.g., dep ession) could bene i om ime- ela ed ends
(e.g., use s migh make mo e submissions du ing he nigh
because sleeping p oblems is one o he symp oms o dep es-
sion). To encode his ea u e, we ex ac he hou when each
obse ed ex was pos ed, we a e age hese hou s o each
use and we compu e how close he a e age is o ou selec ed
imes (0h, 6h, 12h and 18h). Le T= {0,6,12,18}be he
se o selec ed hou s. Le Hube he se o hou s o he seen
ex s o a gi en use u. Le Hube he a e age hou when he
seen ex s o a gi en use we e pos ed. We ha e de ined ou
ime- ela ed p io i isa ion me hods: Nea 00:00 UTC±00:00
(N00) gi es mo e weigh o use s ha pos nea midnigh ,
Nea 06:00 UTC±00:00 (N06) gi es mo e weigh o use s
ha pos nea 6 in he mo ning, and so on.
cuse =a g min
u∈U
min(| −Hu|,24 − | −Hu|),whe e ∈T
(5)
E. CLASSIFIER-BASED PRIORITISATION
A ocused c awle is o en guided by a d i ing que y o clas-
si ie . Gi en he ex obse ed om each use , a na u al way
o guide he FUE p ocess is ei he o compu e he ma ching
be ween he d i ing que y and he use ’s ex o o pass he
use ’s ex o he classi ie ha guides FUE. A any poin ,
he use ’s ex s a ailable o he c awle a e a subse o he
use ’s his o y o pos s bu , s ill, his pa ial ep esen a ion
o he use migh be highly aluable o ea ly iden i y a ge
use s. Fo example, he i s ime ha he c awle sees a use ,
i p obably eads a single commen o pos om he use . This
esembles when a ocused c awling o web pages sees he i s
e e ence (link) o a new web page. In such a case, a single
ancho ex and he associa ed URL a e he only pieces o
e idence a ailable bu he ocused c awle can s ill es ima e
opicali y wi h espec o he a ge p o ile. In FUE, we can
main ain (and cons an ly augmen ) a ex ual ep esen a ion
o each use . This consis s o he conca ena ion o he i les
and commen s’ bodies o he pos s obse ed o each use .
Le Dbe he se o inc emen al use ’s ep esen a ions. Gi en
D, we can pass i o he classi ie and ob ain a p obabili y
es ima ion o he use being on- a ge , Pu(D).
We de ine wo app oaches o selec a ge use s: Highes
Classi ica ion P obabili y (HCP) builds a anking o use s
by dec easing p obabili y, whe eas Lowes Classi ica ion
P obabili y (LCP) builds a anking o use s by inc easing
p obabili y. LCP is supposed o pe o m poo ly since his
p io i isa ion makes li le sense. Howe e , LCP se es as a
sa e check in ou empi ical s udy.
cuse =a g max
u∈U
Pu(D) (6)
cuse =a g min
u∈U
Pu(D) (7)
Table 1summa ises all p io i isa ion me hods desc ibed
abo e.
F. EXPERIMENTAL RESULTS
In o de o make ai compa isons among he p oposed me h-
ods, we ha e designed an expe imen al amewo k whe e all
c awling a ian s un in pa allel. In his way, we can ai ly
e alua e hei ela i e me i s on iden i ying a ge use s and
we a oid undesi able biases om seasonal e ec s ha migh
a ec he compa ison.
The se e whe e all he expe imen s in he pape we e
conduc ed has he ollowing cha ac e is ics:
•CPU: 2 x In el R
Xeon R
CPU E5-2630 4 @ 2.20 GHz
- 20 co es (40 h eads)
•Memo y: 12 x Hynix 32 GiB DIMM DDR4 @ 2400
MHz (384 GiB)
•Disk: Toshiba MD04ACA400 4 TB @ 7,200 RPM,
64 MB cache
•In e ne bandwid h: ∼400/200 Mbps (download/upload)
The i s expe imen las ed se en days. Du ing his pe iod,
he o al numbe o unique ex ac ed use s was highe
han 110,000 (655 pe hou ) o each p io i isa ion me hod
VOLUME 10, 2022 42613
R. Ma ínez-Cas año e al.: Real-Time Focused Ex ac ion o Social Media Use s
TABLE 1. Basic p io i isa ion me hods.
FIGURE 2. Numbe o ex ac ed use s by each ocused use c awling me hod.
(see Figu e 2). No e ha he ex ac ion pe o mance can be
highly scaled, while main aining he poli eness o he c awle ,
by mainly launching mo e ins ances o he Use Ex ac o
module, which is he main bo leneck.
The c awling a ian s HAC and HAS a e slowe a ex ac -
ing use s. This is due o he way in which use s a e p io i ised.
These wo a ian s ex ac use s ha ha e a high numbe o
commen s o poin s. These use s end o be highly ac i e on
he pla o m and, hus, when he c awle ex ac s hem, he
e ie al o all hei submissions is cos ly.
Le us now e alua e he abili y o hese p io i isa ion s a e-
gies in iden i ying a ge use s. To mee his aim, we use a use
classi ie buil om ex e nal aining da a. Mo e speci ically,
we buil a bina y classi ie om a sample o Reddi use s
ha p edic s i a gi en use is likely o ha e signs o dep es-
sion. We wo ked wi h a collec ion on dep ession and na u al
language use [20]. The classi ie is a Logis ic Reg ession
model wi h L1 egula isa ion implemen ed in Py hon wi h
sciki -lea n. I was buil wi h a aining se o 486 use s (83
posi i es, 403 nega i es) whe e use s we e ep esen ed wi h
a single documen consis ing o he conca ena ion o all hei
w i ings. In o de o decide i a use is on- a ge , we es ab-
lished a h eshold equal o 0.9 since p e ious expe imen s
demons a ed ha his se ing leads o high p ecision [20].
Figu es 3and 4plo he ha es a io o di e en p io i isa-
ion s a egies and, hus, i helps o compa e he ela i e me i s
o he p oposed me ics in ea ly iden i ying he a ge use s.
To acili a e eadabili y, he me hods ha e been sepa a ed in o
wo independen plo s bu he andom s a egy is shown in
bo h g aphs. The highes ha es a io is achie ed by HCP:
2.8 imes be e han he second bes (N06) and 4.9 imes
be e han RND du ing he i s day. On he las day hese
numbe s decay o 1.4 and 1.8, espec i ely.
A e he i s day, ou o he p oposals beha e be e han
andom. This numbe inc eases o i e a e he hi d day.
A his poin , many s a egies beha e simila ly since mos o
hem ha e al eady explo ed a high numbe o use s. The mos
p omising s a egies a e HCP, N06, HAC and N00 since hey
pe o m conside ably be e han andom. On he o he hand,
HAS, N12, N18, LAC, LAS and LCP can be disca ded.
Obse e ha HCP e ec i ely exploi s he li le pieces o
ex ual e idence a ailable. These small ex ac s seem o be
indica i e o use ele ance. No e also ha none o he o he
good pe o me s employ he ex s w i en by use s o p io i-
ise hem.
V. ADVANCED PRIORITISATION METHODS
Gi en he esul s discussed abo e, we designed h ee new
FUE me hods. The i s is based on e- aining he classi ie
ha guides he HCP c awle ( om new agged use s ob ained
wi h pseudo- aining da a). A second me hod pe o ms a
usion o wo use ankings (HCP and HAC). Finally, he hi d
me hod combines HCP, N06 and HAC in a hie a chical way.
This sec ion p esen s hese new me hods and discusses he
associa ed changes in he FUE a chi ec u e.
42614 VOLUME 10, 2022
R. Ma ínez-Cas año e al.: Real-Time Focused Ex ac ion o Social Media Use s
FIGURE 3. Ha es a io o some ocused use c awling me hods.
FIGURE 4. Ha es a io o some (disca ded) ocused use c awling me hods.
HCP Wi h Pseudo-T aining: Wi h HCP, he c awling sys-
em makes wo ypes o use classi ica ions. The e alua ion
classi ie (see Sec ion II) is he co e e alua ion ool o he
FUE p ocess and assesses whe he o no use ex ac ion
was e ec i e. The p edic ion classi ie ca ego ises he use s
based on he pa ial in o ma ion a ailable (e.g., las pos in
a ecen ly c awled page), yielding a con idence sco e ha is
used o use p io i isa ion. Wi h he s anda d HCP me hod
discussed abo e, hese wo classi ie s a e he same and hei
classi ica ion model is buil once ( om he aining se ) and
ne e changes he ea e . Howe e , he classi ica ion model
o he p edic i e classi ie could be upda ed as we ex ac
and p ocess use s. Mo e speci ically, hose use s ha a e
chosen, ully explo ed, and classi ied by he e alua ion clas-
si ie can be inco po a ed as pseudo- aining examples. The
main idea o HCP wi h pseudo- aining consis s o upda ing
he p edic ion classi ie based on augmen ing he aining
examples wi h new examples and hei pseudo-labels. E e y
i e minu es, a new p edic ion classi ie is buil om he
o iginal aining se plus he new pseudo-examples a ailable.
We expe imen ed wi h h ee a ian s. Wi h HCP-0.9-0.1 only
hose use s classi ied wi h a p obabili y highe han 0.9 o
lowe han 0.1 a e inco po a ed (as posi i e o nega i e,
espec i ely).4Wi h HCP-0.5-0.5 we employ he s anda d
h eshold o he classi ie (0.5) and, hus, any use classi-
ied abo e o below he h eshold is inco po a ed in o he
aining se as posi i e o nega i e, espec i ely. We also
4No e ha we wo k wi h Logis ic Reg ession classi ie s and, hus, we ha e
access o p obabili ies associa ed o he es ima ions.
VOLUME 10, 2022 42615
R. Ma ínez-Cas año e al.: Real-Time Focused Ex ac ion o Social Media Use s
[24] R. Guo, H. Wang, M. Chen, J. Li, and H. Gao, ‘‘Pa allelizing he ex ac ion
o esh in o ma ion om online social ne wo ks,’’ Fu u e Gene . Compu .
Sys ., ol. 59, pp. 33–46, Jun. 2016.
[25] F. E landsson, P. B ódka, M. Bold , and H. Johnson, ‘‘Do we eally
need o ca ch hem all? A new use -guided social media c awling
me hod,’’ En opy, ol. 19, no. 12, p. 686, Dec. 2017. [Online]. A ailable:
h ps://www.mdpi.com/1099-4300/19/12/686
[26] P. Wang, J. Zhao, J. C. S. Lui, D. Towsley, and X. Guan, ‘‘Fas c awling
me hods o explo ing con en dis ibu ed o e la ge g aphs,’’ Knowl. In .
Sys ., ol. 59, no. 1, pp. 67–92, Ap . 2019.
[27] Y. Zhou, R. Ji, J. Su, and J. Yao, ‘‘Unco e ing media bias ia social ne wo k
lea ning,’’ ACM T ans. In ell. Sys . Technol., ol. 12, no. 1, pp. 1–12,
Feb. 2021.
[28] X. Wang, L. Toka chuk, F. Cuad ado, and S. Poslad, ‘‘Exploi ing hash ags
o adap i e mic oblog c awling,’’ in P oc. IEEE/ACM In . Con . Ad .
Social Ne w. Anal. Mining, Aug. 2013, pp. 311–315.
[29] Z. Zhang and O. Nas aoui, ‘‘P o ile-based ocused c awle o social
media-sha ing websi es,’’ in P oc. 20 h IEEE In . Con . Tools Wi h A i .
In ell., No . 2008, pp. 317–324.
[30] C. H. C. Leung, A. W. S. Chan, A. Milani, J. Liu, and Y. Li, ‘‘In el-
ligen social media indexing and sha ing using an adap i e indexing
sea ch engine,’’ ACM T ans. In ell. Sys . Technol., ol. 3, no. 3, pp. 1–27,
May 2012, doi: 10.1145/2168752.2168761.
[31] M.-Y. Kan, ‘‘Web page classi ica ion wi hou he web page,’’ in P oc. 13 h
In . Wo ld Wide Web Con . Al e na e T ack Pape s Pos e s (WWW Al .),
2004, pp. 262–263.
[32] R. Meusel, P. Mika, and R. Blanco, ‘‘Focused c awling o s uc u ed
da a,’’ in P oc. 23 d ACM In . Con . Con . In . Knowl. Manage., No . 2014,
pp. 1039–1048.
[33] F. Mencze , G. Pan , and P. S ini asan, ‘‘Topical web c awle s: E alu-
a ing adap i e algo i hms,’’ ACM T ans. In e ne Technol., ol. 4, no. 4,
pp. 378–419, No . 2004.
[34] G. Gossen, E. Demido a, and T. Risse, ‘‘IC awl: Imp o ing he eshness
o web collec ions by in eg a ing social web and ocused web c awling,’’
in P oc. 15 h ACM/IEEE-CS Join Con . Digi . Lib a ies, Jun. 2015,
pp. 75–84.
[35] T. Gisselb ech , L. Denoye , P. Gallina i, and S. Lamp ie , ‘‘Whichs eams:
A dynamic app oach o ocused da a cap u e om la ge social media,’’ in
P oc. 9 h In . AAAI Con . Web Social Media, 2015, pp. 130–139.
[36] S. Aga wal and A. Su eka, ‘‘Spide and he lies : Focused c awling on
Tumbl o de ec ha e p omo ing communi ies,’’ 2016, a Xi :1603.09164.
[37] X. Tang and C. C. Yang, ‘‘Ranking use in luence in heal hca e social
media,’’ ACM T ans. In ell. Sys . Technol., ol. 3, no. 4, pp. 1–21,
Sep. 2012, doi: 10.1145/2337542.2337558.
[38] J. Li, L. Wu, R. Hong, K. Zhang, Y. Ge, and Y. Li, ‘‘A join neu al
model o use beha io p edic ion on social ne wo king pla o ms,’’ ACM
T ans. In ell. Sys . Technol., ol. 11, no. 6, pp. 1–25, Dec. 2020, doi: 10.
1145/3406540.
[39] B.-Y. Hsu, C.-L. Tu, M.-Y. Chang, and C.-Y. Shen, ‘‘On c awling
communi y-awa e online social ne wo k da a,’’ in P oc. 30 h ACM
Con . Hype ex Social Media, Sep. 2019, pp. 265–266, doi: 10.1145/
3342220.3344937.
[40] J. Zhuang, T. Mei, S. C. H. Hoi, X.-S. Hua, and Y. Zhang, ‘‘Communi y
disco e y om social media by low- ank ma ix eco e y,’’ ACM T ans.
In ell. Sys . Technol., ol. 5, no. 4, pp. 1–19, Jan. 2015.
[41] K. Lee, J. Ca e lee, Z. Cheng, and D. Z. Sui, ‘‘Campaign ex ac ion om
social media,’’ ACM T ans. In ell. Sys . Technol., ol. 5, no. 1, pp. 1–28,
Dec. 2013, doi: 10.1145/2542182.2542191.
[42] S. C. Gun uku, D. B. Yaden, M. L. Ke n, L. H. Unga , and J. C. Eichs aed ,
‘‘De ec ing dep ession and men al illness on social media: An in eg a i e
e iew,’’ Cu en Opinion Beha . Sci., ol. 18, pp. 43–49, Dec. 2017.
[43] E. A. Ríssola, D. E. Losada, and F. C es ani, ‘‘A su ey o compu a ional
me hods o online men al s a e assessmen on social media,’’ ACM T ans.
Compu . Heal hca e, ol. 2, no. 2, pp. 1–31, Ma . 2021.
[44] S. Aga wal and A. Su eka, ‘‘Using KNN and SVM based one-class clas-
si ie o de ec ing online adicaliza ion on Twi e ,’’ in P oc. In . Con .
Dis ib. Compu . In e ne Technol., Cham, Swi ze land: Sp inge , 2015,
pp. 431–442.
[45] B. S. Nandhini and J. I. Sheeba, ‘‘Online social ne wo k bullying de ec ion
using in elligence echniques,’’ P oc. Compu . Sci., ol. 45, pp. 485–492,
Jan. 2015.
RODRIGO MARTÍNEZ-CASTAÑO ecei ed he
B.Sc. deg ee in compu e science and he M.Sc.
deg ee in big da a and da a analysis echnologies
om he Uni e si y o San iago de Compos ela,
Spain. He is cu en ly pu suing he Ph.D. deg ee.
His esea ch in e es s include big da a echnolo-
gies, dis ibu ed sys ems, in o ma ion e ie al,
and blockchain.
DAVID E. LOSADA is cu en ly an Associa e
P o esso in compu e science and a i icial in el-
ligence wi h CiTIUS, Uni e si y o San iago de
Compos ela, Spain. His cu en esea ch in e es s
include a wide ange o in o ma ion e ie al (IR)
and ela ed a eas, such as ea ly isk de ec ion, ex
mining, IR e alua ion, IR p obabilis ic models,
summa iza ion, no el y de ec ion, and sen ence
e ie al. He is an Ac i e Membe o he IR Com-
muni y and he egula ly se es on he P og am
Commi ee o p es igious in e na ional con e ences, such as SIGIR o ECIR.
In 2011, he was ecognized wi h he ACM Senio Membe Awa d.
JUAN C. PICHEL ecei ed he B.Sc. and M.Sc.
deg ees in physics om he Uni e si y o San iago
de Compos ela, Spain, and he Ph.D. deg ee in
compu e science om he Uni e si y o San iago
de Compos ela, in 2006. He was a Visi ing Pos -
doc o al Resea che wi h he Uni e si y Ca los III
de Mad id, Spain, and he Uni e si y o Illinois a
U bana–Champaign, USA. He also wo ked as a
Resea che and he P ojec Manage wi h he Gali-
cia Supe compu ing Cen e , Spain. He is cu en ly
an Associa e P o esso wi h CiTIUS, Uni e si y o San iago de Compos ela.
His esea ch in e es s include pa allel and dis ibu ed compu ing, big da a
echnologies, p og amming models, and so wa e op imiza ion echniques
o eme ging a chi ec u es.
42622 VOLUME 10, 2022