I
nte
rna
t
io
na
l J
o
urna
l o
f
E
lect
rica
l a
nd
Co
m
pu
t
er
E
ng
ineering
(
I
J
E
CE
)
Vo
l.
1
6
,
No
.
3
,
J
u
n
e
20
2
6
,
p
p
.
1518
~
1
5
3
0
I
SS
N:
2088
-
8
7
0
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijece.
v
1
6
i
3
.
pp
1
5
1
8
-
1
5
3
0
1518
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ec
e.
ia
esco
r
e.
co
m
A
surv
ey
of r
etrie
v
a
l alg
o
rithms in ad a
nd conten
t
recom
menda
tion
sy
stems
Yu Zh
a
o
1
,
F
a
ng
L
iu
2
,
Yua
n
Yua
n
3
,
Yif
a
n Da
ng
4
1
R
o
t
ma
n
S
c
h
o
o
l
o
f
M
a
n
a
g
e
me
n
t
,
U
n
i
v
e
r
si
t
y
o
f
T
o
r
o
n
t
o
,
To
r
o
n
t
o
,
C
a
n
a
d
a
2
Y
a
l
e
S
c
h
o
o
l
o
f
M
a
n
a
g
e
me
n
t
,
Y
a
l
e
U
n
i
v
e
r
s
i
t
y
,
N
e
w
H
a
v
e
n
,
U
n
i
t
e
d
S
t
a
t
e
s
o
f
A
meri
c
a
3
S
c
h
o
o
l
o
f
E
n
g
i
n
e
e
r
i
n
g
a
n
d
S
c
i
e
n
c
e
,
S
t
e
v
e
n
s I
n
st
i
t
u
t
e
o
f
Te
c
h
n
o
l
o
g
y
,
H
o
b
o
k
e
n
,
U
n
i
t
e
d
S
t
a
t
e
s
o
f
A
meri
c
a
4
R
o
b
e
r
t
H
.
S
mi
t
h
S
c
h
o
o
l
o
f
B
u
si
n
e
ss
,
U
n
i
v
e
r
si
t
y
o
f
M
a
r
y
l
a
n
d
,
C
o
l
l
e
g
e
P
a
r
k
,
U
n
i
t
e
d
S
t
a
t
e
s
o
f
A
m
e
r
i
c
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Feb
1
1
,
2
0
2
6
R
ev
is
ed
Ma
r
2
8
,
2
0
2
6
Acc
ep
ted
Ap
r
2
7
,
2
0
2
6
Th
is
p
a
p
e
r
p
re
se
n
ts
a
s
u
rv
e
y
o
f
re
tri
e
v
a
l
a
l
g
o
ri
th
m
s
u
se
d
i
n
a
d
v
e
rti
si
n
g
re
c
o
m
m
e
n
d
a
ti
o
n
a
n
d
o
rg
a
n
ic
c
o
n
ten
t
re
c
o
m
m
e
n
d
a
ti
o
n
sy
ste
m
s.
M
o
d
e
r
n
d
ig
it
a
l
p
latfo
rm
s
re
ly
o
n
re
tri
e
v
a
l
-
b
a
se
d
m
o
d
e
ls
t
o
e
fficie
n
tl
y
m
a
tch
u
se
rs
with
re
lev
a
n
t
a
d
v
e
rti
se
m
e
n
ts
o
r
p
e
rso
n
a
li
z
e
d
c
o
n
te
n
t.
Th
is
su
r
v
e
y
re
v
iew
s
k
e
y
tec
h
n
i
q
u
e
s
i
n
c
lu
d
in
g
in
v
e
rt
e
d
in
d
e
x
m
e
th
o
d
s,
c
o
ll
a
b
o
ra
ti
v
e
fil
terin
g
,
c
o
n
ten
t
-
b
a
se
d
fil
teri
n
g
,
h
y
b
ri
d
re
c
o
m
m
e
n
d
a
ti
o
n
m
o
d
e
ls,
a
n
d
t
h
e
t
wo
-
to
we
r
n
e
u
ra
l
n
e
tw
o
rk
a
rc
h
i
tec
tu
re
wi
d
e
ly
u
se
d
i
n
larg
e
-
sc
a
le
re
c
o
m
m
e
n
d
a
ti
o
n
sy
ste
m
s.
Th
e
p
a
p
e
r
c
o
m
p
a
re
s
th
e
o
b
jec
ti
v
e
s,
d
a
ta
u
ti
li
z
a
ti
o
n
stra
teg
ies
,
a
n
d
e
v
a
lu
a
ti
o
n
m
e
tri
c
s
o
f
a
d
targ
e
ti
n
g
a
n
d
o
r
g
a
n
ic
re
tri
e
v
a
l
sy
ste
m
s.
P
ra
c
ti
c
a
l
c
h
a
ll
e
n
g
e
s
su
c
h
a
s
c
o
l
d
-
sta
rt
p
ro
b
lem
s,
d
a
ta
q
u
a
li
ty
,
sc
a
lab
il
it
y
,
a
n
d
p
riv
a
c
y
c
o
n
sid
e
ra
ti
o
n
s
a
re
a
lso
d
isc
u
ss
e
d
.
Th
is
su
r
v
e
y
f
u
rth
e
r
h
i
g
h
li
g
h
ts
t
h
e
g
ro
wi
n
g
c
o
n
n
e
c
ti
o
n
b
e
twe
e
n
in
d
u
strial
re
c
o
m
m
e
n
d
a
ti
o
n
p
ip
e
li
n
e
s
a
n
d
e
m
e
rg
in
g
re
tri
e
v
a
l
m
e
c
h
a
n
ism
s
u
se
d
i
n
larg
e
lan
g
u
a
g
e
m
o
d
e
l
(
LL
M
)
sy
st
e
m
s.
Th
is
su
rv
e
y
p
r
o
v
i
d
e
s
in
si
g
h
ts
i
n
to
t
h
e
d
e
sig
n
p
ri
n
c
ip
les
o
f
m
o
d
e
r
n
re
tri
e
v
a
l
sy
ste
m
s
a
n
d
o
u
tl
in
e
s
f
u
tu
re
re
se
a
rc
h
d
irec
ti
o
n
s
a
t
t
h
e
in
ters
e
c
ti
o
n
o
f
re
c
o
m
m
e
n
d
a
ti
o
n
s
y
ste
m
s a
n
d
LL
M
.
K
ey
w
o
r
d
s
:
Ad
tar
g
etin
g
I
n
d
u
s
tr
ial
r
ec
o
m
m
en
d
atio
n
s
y
s
tem
s
I
n
f
o
r
m
atio
n
r
etr
ie
v
al
L
ar
g
e
lan
g
u
ag
e
m
o
d
els
R
ec
o
m
m
en
d
atio
n
s
y
s
tem
s
R
etr
iev
al
alg
o
r
ith
m
s
T
wo
-
to
wer
n
e
u
r
al
n
etwo
r
k
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
Fan
g
L
iu
Yale
Sch
o
o
l o
f
Ma
n
a
g
em
en
t,
Yale
Un
iv
er
s
ity
New
Hav
en
,
C
T
,
Un
ited
States
o
f
Am
er
ica
E
m
ail: f
an
g
liu
4
3
5
@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
W
ith
th
e
r
ap
id
g
r
o
wth
o
f
d
ig
ital
p
latf
o
r
m
s
,
d
eliv
er
in
g
p
er
s
o
n
alize
d
co
n
ten
t
an
d
ad
v
er
tis
em
en
ts
h
as
b
ec
o
m
e
ess
en
tial
f
o
r
im
p
r
o
v
i
n
g
u
s
er
en
g
ag
em
e
n
t
an
d
p
latf
o
r
m
r
ev
en
u
e.
R
etr
iev
al
alg
o
r
i
th
m
s
p
lay
a
ce
n
tr
al
r
o
le
in
r
ec
o
m
m
en
d
atio
n
s
y
s
tem
s
b
y
m
atch
in
g
u
s
er
s
with
r
elev
an
t
ad
v
er
tis
em
en
ts
an
d
o
r
g
a
n
ic
co
n
ten
t
[
1
]
,
[
2
]
.
Ad
tar
g
etin
g
f
o
cu
s
es
o
n
m
a
x
i
m
izin
g
en
g
a
g
em
en
t
a
n
d
co
n
v
er
s
io
n
r
ates
th
r
o
u
g
h
p
er
s
o
n
ali
ze
d
ad
v
er
tis
em
en
ts
b
ased
o
n
u
s
er
p
r
o
f
iles
an
d
b
eh
av
io
r
al
d
ata,
wh
er
ea
s
o
r
g
a
n
ic
r
etr
iev
al
s
y
s
tem
s
aim
to
im
p
r
o
v
e
u
s
er
ex
p
er
ien
ce
b
y
r
ec
o
m
m
en
d
i
n
g
co
n
ten
t a
lig
n
ed
with
u
s
er
s
’
in
ter
ests
an
d
p
r
ef
er
en
ce
s
.
I
n
r
ec
en
t
y
ea
r
s
,
r
ec
o
m
m
en
d
atio
n
s
y
s
tem
s
h
av
e
ev
o
lv
ed
f
r
o
m
tr
ad
itio
n
al
in
f
o
r
m
atio
n
r
etr
iev
al
tech
n
iq
u
es
to
d
ee
p
lear
n
in
g
–
b
ased
r
etr
iev
al
a
r
ch
itectu
r
es u
s
ed
in
lar
g
e
-
s
ca
le
in
d
u
s
tr
ial
p
la
tf
o
r
m
s
[
3
]
,
[
4
]
,
with
ad
d
itio
n
al
r
ec
e
n
t
s
u
r
v
ey
s
[
5
]
.
Var
io
u
s
r
ec
o
m
m
en
d
atio
n
ap
p
r
o
ac
h
es
h
av
e
b
ee
n
d
e
v
elo
p
e
d
t
o
im
p
r
o
v
e
ac
cu
r
ac
y
an
d
s
ca
lab
ilit
y
,
in
clu
d
in
g
c
o
n
ten
t
-
b
ased
f
ilter
in
g
,
c
o
llab
o
r
ativ
e
f
ilter
in
g
,
an
d
h
y
b
r
id
r
ec
o
m
m
en
d
atio
n
m
eth
o
d
s
[
6
]
,
[
7
]
.
M
o
r
e
r
ec
en
tl
y
,
n
eu
r
al
r
etr
ie
v
al
ar
ch
itectu
r
e
s
s
u
ch
as
th
e
two
-
to
wer
m
o
d
el
h
av
e
b
ee
n
wid
ely
ad
o
p
ted
in
i
n
d
u
s
tr
ial
r
ec
o
m
m
en
d
atio
n
p
i
p
elin
es
d
u
e
to
th
eir
ef
f
ec
tiv
en
ess
in
lar
g
e
-
s
ca
le
ca
n
d
id
ate
g
en
e
r
atio
n
an
d
r
ep
r
esen
tatio
n
lear
n
i
n
g
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
A
s
u
r
ve
y
o
f retr
ieva
l a
lg
o
r
ith
ms in
a
d
a
n
d
co
n
ten
t reco
mme
n
d
a
tio
n
s
ystems
(
Yu
Zh
a
o
)
1519
T
h
is
p
ap
er
s
u
r
v
ey
s
r
et
r
iev
al
alg
o
r
ith
m
s
u
s
ed
i
n
ad
v
er
tis
in
g
an
d
o
r
g
an
ic
r
ec
o
m
m
en
d
atio
n
s
y
s
tem
s
,
co
v
er
in
g
r
ep
r
esen
tativ
e
a
p
p
r
o
ac
h
es su
ch
as c
o
n
ten
t
-
b
ased
f
ilter
in
g
,
co
llab
o
r
ativ
e
f
ilter
in
g
,
h
y
b
r
id
m
o
d
els,
an
d
n
eu
r
al
r
etr
iev
al
ar
c
h
itectu
r
es
in
clu
d
in
g
th
e
two
-
t
o
wer
m
o
d
el.
Pra
ctica
l
ch
allen
g
es
s
u
ch
as
co
ld
s
tar
t,
s
ca
lab
ilit
y
,
d
ata
q
u
ality
,
an
d
p
r
iv
ac
y
co
n
ce
r
n
s
ar
e
also
d
is
cu
s
s
ed
.
Un
lik
e
m
an
y
ex
is
tin
g
s
u
r
v
ey
s
th
at
b
r
o
ad
ly
r
ev
iew
r
ec
o
m
m
en
d
atio
n
alg
o
r
ith
m
s
,
th
is
p
ap
er
f
o
cu
s
es
s
p
ec
if
ically
o
n
r
etr
iev
al
-
s
tag
e
alg
o
r
ith
m
s
in
m
o
d
e
r
n
in
d
u
s
tr
ial
r
ec
o
m
m
e
n
d
atio
n
p
ip
e
lin
es.
T
h
e
s
u
r
v
e
y
ex
am
in
es
r
etr
iev
al
tech
n
iq
u
es
u
s
ed
i
n
b
o
th
ad
v
er
tis
in
g
an
d
o
r
g
an
ic
r
ec
o
m
m
en
d
atio
n
s
y
s
t
em
s
an
d
h
ig
h
lig
h
ts
h
o
w
ca
n
d
id
ate
g
en
er
atio
n
o
p
e
r
ates
in
lar
g
e
-
s
ca
le
p
latf
o
r
m
s
.
I
t
also
d
is
cu
s
s
es
p
o
ten
tial
co
n
n
ec
tio
n
s
b
etwe
en
r
etr
iev
al
-
b
ased
r
ec
o
m
m
en
d
atio
n
s
y
s
tem
s
an
d
lar
g
e
la
n
g
u
a
g
e
m
o
d
el
(
L
L
M
)
–
en
h
an
ce
d
r
etr
i
ev
al
ar
ch
itectu
r
es.
Alth
o
u
g
h
th
is
s
u
r
v
e
y
d
o
es
n
o
t
ex
p
lo
r
e
th
is
t
o
p
ic
i
n
d
ep
th
,
t
h
e
r
etr
ie
v
al
m
ec
h
an
is
m
s
d
is
cu
s
s
ed
h
er
e
s
h
ar
e
co
n
ce
p
t
u
al
s
im
ilar
ities
with
r
etr
iev
al
-
au
g
m
e
n
ted
g
e
n
er
atio
n
(
R
AG)
f
r
am
ewo
r
k
s
u
s
ed
in
LLM
[
8
]
–
[
1
3
]
,
wh
er
e
r
etr
iev
al
s
er
v
es
as
an
in
ter
m
ed
iate
s
tep
to
id
en
tify
r
ele
v
an
t
in
f
o
r
m
atio
n
p
r
io
r
t
o
d
o
wn
s
tr
ea
m
g
en
e
r
atio
n
o
r
r
an
k
in
g
,
th
er
eb
y
im
p
r
o
v
in
g
r
elev
an
ce
,
co
n
te
x
tu
al
g
r
o
u
n
d
in
g
,
an
d
o
v
er
all
s
y
s
tem
p
er
f
o
r
m
an
ce
.
I
n
R
AG
s
y
s
tem
s
,
r
elev
an
t
in
f
o
r
m
atio
n
is
f
ir
s
t
r
etr
ie
v
ed
f
r
o
m
e
x
ter
n
al
k
n
o
wled
g
e
s
o
u
r
ce
s
an
d
t
h
e
n
in
co
r
p
o
r
ated
i
n
to
lan
g
u
ag
e
m
o
d
els
to
im
p
r
o
v
e
th
e
q
u
ality
an
d
f
ac
tu
al
g
r
o
u
n
d
in
g
o
f
g
en
e
r
ated
r
esp
o
n
s
es.
Similar
ly
,
r
ec
o
m
m
en
d
atio
n
s
y
s
tem
s
r
ely
o
n
r
etr
iev
al
alg
o
r
ith
m
s
to
id
en
tify
r
elev
an
t u
s
er
an
d
item
r
e
p
r
esen
tatio
n
s
b
ef
o
r
e
d
o
wn
s
tr
ea
m
r
an
k
in
g
o
r
p
r
ed
ic
tio
n
task
s
.
Ov
er
all,
th
is
s
u
r
v
ey
p
r
o
v
id
es
a
s
tr
u
ctu
r
ed
o
v
er
v
iew
o
f
r
etr
ie
v
al
alg
o
r
ith
m
s
with
a
p
ar
ticu
la
r
f
o
cu
s
o
n
in
d
u
s
tr
ial
d
ep
lo
y
m
en
t
s
ce
n
ar
io
s
.
B
y
an
aly
zin
g
b
o
th
tr
ad
i
tio
n
al
r
etr
iev
al
tech
n
i
q
u
es
a
n
d
m
o
d
er
n
n
eu
r
al
ar
ch
itectu
r
es,
th
e
p
a
p
er
h
ig
h
lig
h
ts
k
ey
d
esig
n
p
r
in
ci
p
les
an
d
s
y
s
tem
c
h
allen
g
es
in
lar
g
e
-
s
ca
le
in
d
u
s
tr
ial
r
ec
o
m
m
en
d
atio
n
p
i
p
elin
es.
T
h
e
r
e
m
a
i
n
d
e
r
o
f
t
h
is
p
a
p
e
r
i
s
o
r
g
a
n
i
z
e
d
a
s
f
o
ll
o
ws
.
S
ec
t
i
o
n
2
i
n
t
r
o
d
u
c
e
s
m
a
j
o
r
a
d
ta
r
g
e
t
i
n
g
te
c
h
n
i
q
u
e
s
a
n
d
i
n
d
e
x
i
n
g
s
t
r
at
e
g
i
es
.
S
e
c
ti
o
n
3
r
e
v
i
e
ws
r
e
t
r
i
e
v
a
l
a
r
c
h
i
t
e
ct
u
r
e
s
u
s
e
d
i
n
r
e
c
o
m
m
e
n
d
a
ti
o
n
s
y
s
t
e
m
s
,
i
n
cl
u
d
i
n
g
t
h
e
t
w
o
-
t
o
w
e
r
m
o
d
el
.
S
e
c
ti
o
n
4
co
m
p
a
r
e
s
a
d
v
e
r
t
is
i
n
g
r
e
c
o
m
m
en
d
a
t
i
o
n
a
n
d
o
r
g
a
n
i
c
r
et
r
i
e
v
a
l
p
i
p
e
l
i
n
e
s
.
S
ec
t
i
o
n
5
d
i
s
c
u
s
s
es
e
v
a
l
u
a
ti
o
n
m
e
t
r
i
cs
a
n
d
e
x
p
e
r
i
m
e
n
t
a
t
i
o
n
f
r
a
m
e
w
o
r
k
s
,
f
o
l
l
o
w
e
d
b
y
t
h
e
c
o
n
c
l
u
s
i
o
n
i
n
s
e
c
ti
o
n
6
.
I
n
th
is
s
u
r
v
ey
,
r
ep
r
esen
tativ
e
r
etr
iev
al
alg
o
r
ith
m
s
u
s
ed
i
n
a
d
v
er
tis
in
g
an
d
c
o
n
ten
t
r
ec
o
m
m
en
d
atio
n
s
y
s
tem
s
ar
e
r
ev
iewe
d
.
R
elev
an
t
s
tu
d
ies
wer
e
id
en
tifie
d
t
h
r
o
u
g
h
m
ajo
r
ac
ad
e
m
ic
d
atab
ases
s
u
ch
as
I
E
E
E
Xp
lo
r
e,
AC
M
Dig
ital
L
ib
r
ar
y
,
an
d
Go
o
g
le
Sch
o
la
r
,
f
o
c
u
s
in
g
o
n
p
u
b
licatio
n
s
f
r
o
m
2
0
1
5
to
2
0
2
4
.
Selectio
n
cr
iter
ia
in
clu
d
ed
r
elev
a
n
ce
to
r
etr
iev
al
-
s
tag
e
r
ec
o
m
m
en
d
atio
n
s
y
s
tem
s
,
citatio
n
im
p
ac
t,
an
d
in
d
u
s
tr
ial
ap
p
licab
ilit
y
.
Stu
d
ies
wer
e
s
elec
ted
b
ased
o
n
th
eir
in
f
l
u
en
ce
in
b
o
th
ac
a
d
em
ic
r
esear
ch
an
d
in
d
u
s
tr
ial
p
r
ac
tice.
T
h
e
r
ev
iewe
d
m
eth
o
d
s
ar
e
ca
teg
o
r
ized
in
to
tr
ad
itio
n
al
r
etr
iev
al
tec
h
n
iq
u
es
a
n
d
n
eu
r
al
r
etr
iev
al
ar
ch
itectu
r
es,
an
d
ea
c
h
m
eth
o
d
is
an
aly
ze
d
in
ter
m
s
o
f
its
m
ec
h
an
is
m
,
s
ca
lab
ilit
y
,
in
d
u
s
tr
i
al
ap
p
licab
ilit
y
,
an
d
u
s
ag
e
s
ce
n
ar
io
s
in
m
o
d
er
n
r
ec
o
m
m
en
d
atio
n
p
ip
elin
es.
T
h
e
m
ain
co
n
tr
ib
u
tio
n
s
o
f
th
is
p
ap
er
a
r
e
s
u
m
m
ar
ized
as f
o
ll
o
ws:
a.
T
h
is
p
ap
er
p
r
o
v
id
es
a
f
o
c
u
s
ed
s
u
r
v
ey
o
n
r
etr
iev
al
-
s
tag
e
al
g
o
r
ith
m
s
in
m
o
d
er
n
r
ec
o
m
m
e
n
d
atio
n
s
y
s
tem
s
,
wh
ich
ar
e
less
em
p
h
asized
in
ex
is
tin
g
liter
atu
r
e.
b.
I
t
p
r
esen
ts
a
co
m
p
ar
ativ
e
an
a
ly
s
is
b
etwe
en
ad
tar
g
etin
g
s
y
s
tem
s
an
d
o
r
g
a
n
ic
r
ec
o
m
m
en
d
atio
n
s
y
s
tem
s
f
r
o
m
a
u
n
if
ied
r
etr
iev
al
p
e
r
s
p
ec
tiv
e.
c.
I
t a
n
aly
ze
s
tr
ad
e
-
o
f
f
s
am
o
n
g
d
if
f
er
en
t r
etr
iev
al
ap
p
r
o
ac
h
es i
n
in
d
u
s
tr
ial
d
e
p
lo
y
m
e
n
t scen
ar
io
s
.
d.
I
t
d
i
s
c
u
s
s
es
e
m
e
r
g
i
n
g
c
o
n
n
e
c
tio
n
s
b
e
t
w
e
e
n
r
e
c
o
m
m
e
n
d
a
ti
o
n
r
e
t
r
i
e
v
a
l
a
n
d
L
L
M
-
b
as
e
d
r
e
t
r
i
e
v
a
l
f
r
a
m
e
w
o
r
k
s
.
2.
AD
T
ARG
E
T
I
NG
M
O
D
E
L
S
Ad
ta
r
g
eti
n
g
is
wi
d
e
ly
u
s
e
d
i
n
m
o
d
er
n
d
i
g
i
tal
m
ar
k
eti
n
g
t
o
d
e
li
v
e
r
p
e
r
s
o
n
ali
ze
d
ad
v
er
t
i
s
em
e
n
ts
t
o
s
p
e
ci
f
ic
a
u
d
ie
n
c
es
a
n
d
m
a
x
i
m
iz
e
e
n
g
a
g
e
m
e
n
t
a
n
d
c
o
n
v
er
s
io
n
r
at
es.
On
e
c
o
m
m
o
n
t
ec
h
n
i
q
u
e
is
t
h
e
u
s
e
o
f
in
v
e
r
t
ed
in
d
e
x
es,
w
h
i
c
h
e
f
f
ici
e
n
tl
y
m
at
ch
u
s
er
p
r
o
f
i
les
w
it
h
r
ele
v
a
n
t
ad
v
er
tis
e
m
e
n
ts
in
la
r
g
e
-
s
ca
l
e
i
n
f
o
r
m
ati
o
n
r
et
r
i
ev
al
s
y
s
t
em
s
[
1
4
]
–
[
1
6
]
.
S
ev
er
al
t
a
r
g
eti
n
g
s
t
r
a
te
g
ies
a
r
e
co
m
m
o
n
l
y
u
s
e
d
,
i
n
cl
u
d
i
n
g
d
e
m
o
g
r
a
p
h
ic
ta
r
g
et
in
g
,
r
et
ar
g
eti
n
g
,
k
e
y
w
o
r
d
ta
r
g
et
in
g
,
b
e
h
a
v
i
o
r
al
ta
r
g
et
in
g
,
a
n
d
c
o
n
t
ex
t
u
al
ta
r
g
et
in
g
[
1
7
]
,
[
1
8
]
.
2
.
1
.
I
nv
er
t
ed
ind
ex
Ad
tar
g
etin
g
co
m
m
o
n
ly
lev
e
r
ag
es
in
v
er
ted
in
d
e
x
es
to
im
p
r
o
v
e
th
e
ef
f
icien
cy
an
d
ac
cu
r
ac
y
o
f
d
eliv
er
in
g
p
er
s
o
n
alize
d
ad
v
e
r
tis
em
en
ts
[
1
4
]
.
An
in
v
e
r
ted
i
n
d
ex
is
a
d
ata
s
tr
u
ctu
r
e
th
at
m
ap
s
item
s
(
s
u
ch
as
ad
v
er
tis
em
en
ts
)
to
ass
o
ciate
d
k
ey
wo
r
d
s
o
r
attr
i
b
u
tes,
en
a
b
lin
g
f
ast
s
ea
r
ch
a
n
d
r
etr
iev
al.
As
illu
s
tr
ated
i
n
Fig
u
r
e
1
,
th
e
i
n
v
er
ted
in
d
e
x
li
n
k
s
k
ey
wo
r
d
s
to
co
r
r
esp
o
n
d
i
n
g
item
s
,
allo
win
g
th
e
s
y
s
tem
to
q
u
ick
l
y
r
etr
ie
v
e
r
elev
an
t
ad
v
e
r
tis
em
en
ts
b
ased
o
n
u
s
er
q
u
er
ies
o
r
p
r
o
f
ile
attr
i
b
u
tes.
I
n
ad
tar
g
etin
g
s
y
s
tem
s
,
th
e
in
v
er
ted
in
d
ex
ty
p
ically
o
p
e
r
ates
th
r
o
u
g
h
s
ev
er
al
s
tag
es.
First,
an
in
d
ex
is
c
o
n
s
tr
u
cted
f
r
o
m
a
v
ailab
le
ad
v
er
tis
em
en
ts
,
wh
er
e
ea
ch
ad
v
er
tis
em
en
t
is
r
ep
r
es
en
ted
b
y
a
s
et
o
f
k
ey
wo
r
d
s
o
r
attr
ib
u
tes
d
escr
ib
in
g
its
co
n
ten
t
a
n
d
tar
g
et
au
d
ien
ce
.
Nex
t,
u
s
er
s
ar
e
p
r
o
f
iled
b
ased
o
n
th
eir
o
n
lin
e
ac
ti
v
ities
,
s
u
ch
as
s
ea
r
c
h
q
u
er
ies,
b
r
o
wsi
n
g
h
is
to
r
y
,
s
o
cial
m
ed
ia
in
ter
ac
tio
n
s
,
an
d
p
u
r
ch
ase
b
eh
a
v
io
r
.
E
ac
h
u
s
er
p
r
o
f
ile
is
r
ep
r
esen
ted
b
y
a
s
et
o
f
k
ey
wo
r
d
s
o
r
attr
ib
u
tes r
ef
lectin
g
th
e
u
s
er
’
s
in
ter
ests
an
d
p
r
ef
er
e
n
ce
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
3
,
J
u
n
e
20
2
6
:
1
5
1
8
-
1
5
3
0
1520
Fin
ally
,
wh
en
a
u
s
er
in
ter
ac
ts
with
a
p
latf
o
r
m
,
th
e
s
y
s
tem
r
etr
iev
es
th
e
u
s
er
p
r
o
f
ile
an
d
m
atch
es
it
ag
ain
s
t
th
e
in
v
er
ted
in
d
ex
.
B
y
lo
ca
tin
g
k
e
y
wo
r
d
s
ass
o
ciate
d
with
th
e
u
s
er
,
th
e
s
y
s
tem
ca
n
ef
f
icien
tly
r
etr
iev
e
ad
v
er
tis
em
en
ts
with
m
atch
in
g
o
r
r
elate
d
attr
i
b
u
tes.
T
h
is
m
e
ch
an
is
m
en
ab
les
ef
f
icien
t
r
ea
l
-
tim
e
ad
tar
g
etin
g
b
y
allo
win
g
p
latf
o
r
m
s
to
q
u
ick
ly
r
etr
iev
e
ad
v
e
r
tis
em
en
ts
th
at
m
atch
u
s
er
s
’
in
ter
ests
e
v
en
in
lar
g
e
-
s
ca
le
d
atasets
.
Fig
u
r
e
1
.
I
n
v
er
te
d
in
d
e
x
s
tr
u
ct
u
r
e
f
o
r
ad
v
e
r
tis
em
en
t r
etr
iev
al
2
.
2
.
De
m
o
g
r
a
ph
ic
Dem
o
g
r
ap
h
ic
tar
g
etin
g
d
eliv
e
r
s
ad
v
er
tis
em
en
ts
to
u
s
er
s
b
as
ed
o
n
d
em
o
g
r
ap
h
ic
attr
ib
u
tes s
u
ch
as
ag
e
an
d
g
en
d
er
.
Fo
r
ex
am
p
le,
to
y
s
ar
e
p
r
im
ar
ily
m
ar
k
eted
to
c
h
ild
r
en
,
ed
u
ca
tio
n
al
r
eso
u
r
ce
s
to
teen
ag
er
s
,
an
d
r
etir
em
en
t
o
r
f
in
a
n
cial
p
lan
n
i
n
g
s
er
v
ices
to
o
ld
er
u
s
er
s
.
Dem
o
g
r
ap
h
ic
in
f
o
r
m
atio
n
is
ty
p
ically
o
b
tain
ed
f
r
o
m
u
s
er
r
eg
is
tr
atio
n
d
ata,
p
r
o
f
ile
in
f
o
r
m
atio
n
,
o
r
in
f
er
r
e
d
f
r
o
m
b
eh
av
io
r
al
s
ig
n
als
s
u
ch
as
b
r
o
wsi
n
g
h
is
to
r
y
an
d
co
n
ten
t
co
n
s
u
m
p
tio
n
p
atter
n
s
.
T
h
ese
attr
ib
u
tes
ar
e
th
en
in
co
r
p
o
r
ated
in
t
o
th
e
ad
v
e
r
tis
em
en
t
in
d
ex
in
g
p
r
o
ce
s
s
,
wh
er
e
ad
v
e
r
tis
em
en
ts
ar
e
ass
o
ciate
d
with
d
e
m
o
g
r
a
p
h
ic
-
r
el
ated
k
ey
wo
r
d
s
o
r
attr
ib
u
tes.
Du
r
in
g
th
e
r
etr
iev
a
l
p
r
o
ce
s
s
,
th
e
s
y
s
tem
c
o
m
p
ar
es
u
s
er
d
em
o
g
r
ap
h
ic
ch
a
r
ac
ter
is
tics
with
th
e
in
d
ex
ed
a
d
v
er
tis
em
en
ts
.
B
y
m
atch
in
g
u
s
er
attr
ib
u
tes
with
r
elev
an
t
d
em
o
g
r
ap
h
ic
ca
teg
o
r
i
es,
th
e
ad
v
er
tis
in
g
s
y
s
tem
ca
n
ef
f
icien
tly
r
etr
iev
e
ad
v
er
tis
em
en
ts
th
at
ar
e
m
o
r
e
l
ik
ely
to
b
e
r
elev
an
t to
th
e
u
s
er
[
1
7
]
,
[
1
8
]
.
2
.
3
.
Re
-
t
a
rg
e
t
ing
R
etar
g
etin
g
,
also
k
n
o
wn
as
r
em
ar
k
etin
g
,
f
o
cu
s
es
o
n
u
s
er
s
wh
o
h
av
e
p
r
ev
i
o
u
s
ly
in
ter
ac
ted
with
a
web
s
ite
o
r
ap
p
licatio
n
b
u
t
h
a
v
e
n
o
t
co
m
p
leted
a
d
esire
d
a
ctio
n
,
s
u
ch
as
m
ak
in
g
a
p
u
r
c
h
ase.
T
h
is
tech
n
iq
u
e
aim
s
to
r
e
-
en
g
a
g
e
p
o
ten
tial
cu
s
to
m
er
s
b
y
r
em
in
d
in
g
th
e
m
o
f
p
r
o
d
u
cts
o
r
s
er
v
ices
th
ey
h
av
e
p
r
ev
i
o
u
s
ly
v
iewe
d
.
As
illu
s
tr
ated
in
Fig
u
r
e
2
,
th
e
r
etar
g
etin
g
p
r
o
ce
s
s
g
en
er
ally
in
v
o
lv
es
co
llectin
g
u
s
er
in
ter
ac
tio
n
d
ata,
tr
ac
k
in
g
b
r
o
wsi
n
g
b
eh
av
i
o
r
,
an
d
d
eliv
e
r
in
g
ta
r
g
eted
a
d
v
e
r
tis
em
en
ts
o
n
ex
ter
n
al
p
latf
o
r
m
s
s
u
ch
as
s
o
cial
m
ed
ia
o
r
s
ea
r
ch
en
g
in
es to
en
co
u
r
ag
e
u
s
er
s
to
r
etu
r
n
an
d
co
m
p
lete
th
e
d
esire
d
ac
tio
n
.
Fig
u
r
e
2
.
R
etar
g
etin
g
wo
r
k
f
lo
w
in
d
ig
ital a
d
v
er
tis
in
g
s
y
s
tem
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
A
s
u
r
ve
y
o
f retr
ieva
l a
lg
o
r
ith
ms in
a
d
a
n
d
co
n
ten
t reco
mme
n
d
a
tio
n
s
ystems
(
Yu
Zh
a
o
)
1521
I
n
p
r
ac
tice,
r
eta
r
g
etin
g
s
y
s
tem
s
r
ely
o
n
s
ev
e
r
al
s
tag
es
o
f
d
at
a
p
r
o
ce
s
s
in
g
.
First,
u
s
er
in
ter
a
ctio
n
d
ata
is
co
llected
th
r
o
u
g
h
co
o
k
ies,
tr
ac
k
in
g
p
ix
els,
an
d
o
th
er
tr
ac
k
in
g
tech
n
o
l
o
g
ies.
T
h
is
in
f
o
r
m
atio
n
m
ay
in
clu
d
e
v
is
ited
p
ag
es,
v
iewe
d
p
r
o
d
u
ct
s
,
an
d
u
s
er
ac
tio
n
s
o
n
t
h
e
p
latf
o
r
m
.
Ne
x
t,
ad
v
er
tis
em
en
ts
ar
e
in
d
ex
e
d
ac
co
r
d
in
g
to
th
e
s
p
ec
if
ic
ac
tio
n
s
p
er
f
o
r
m
ed
b
y
u
s
er
s
.
Fo
r
in
s
tan
ce
,
a
p
r
o
d
u
ct
lef
t
in
a
s
h
o
p
p
in
g
ca
r
t
m
ay
tr
ig
g
er
ad
v
er
tis
em
en
ts
ass
o
ciate
d
with
th
at
p
r
o
d
u
ct
ca
teg
o
r
y
o
r
r
el
ated
item
s
.
Fin
ally
,
wh
en
th
e
u
s
er
r
etu
r
n
s
to
th
e
web
s
ite
o
r
co
n
tin
u
es
b
r
o
wsi
n
g
o
th
er
p
latf
o
r
m
s
,
t
h
e
s
y
s
tem
m
atch
es
th
e
u
s
er
’
s
p
r
ev
io
u
s
in
ter
ac
tio
n
s
with
r
elev
an
t
ad
v
er
tis
em
en
ts
,
in
cr
e
asin
g
th
e
lik
elih
o
o
d
th
at
th
e
u
s
er
will
r
etu
r
n
an
d
co
m
p
lete
t
h
e
tr
an
s
ac
tio
n
[
1
7
]
,
[
1
8
]
.
2
.
4
.
K
ey
wo
rd
t
a
r
g
et
ing
Key
wo
r
d
tar
g
eti
n
g
d
eliv
e
r
s
ad
v
er
tis
em
en
ts
b
ased
o
n
k
ey
w
o
r
d
s
th
at
u
s
er
s
en
ter
i
n
s
ea
r
ch
q
u
er
ies
o
r
th
at
ap
p
ea
r
in
th
e
co
n
ten
t
th
ey
ar
e
cu
r
r
en
tly
v
iewin
g
.
B
y
an
aly
zin
g
th
ese
tex
tu
al
s
ig
n
als,
ad
v
er
tis
in
g
s
y
s
tem
s
ca
n
m
atch
u
s
er
i
n
ten
t w
ith
r
el
ev
an
t a
d
v
e
r
tis
em
en
ts
.
R
ec
en
t
s
tu
d
ies
s
u
g
g
est
th
at
L
L
Ms
ca
n
f
u
r
th
er
en
h
an
ce
k
e
y
wo
r
d
tar
g
etin
g
b
y
ex
p
an
d
in
g
an
d
r
ef
in
in
g
k
ey
wo
r
d
s
ets.
Mo
d
els
s
u
c
h
as
GPT
-
4
[
1
0
]
a
n
d
o
th
er
tr
an
s
f
o
r
m
er
-
b
ased
ar
ch
itectu
r
es
[
9
]
,
[
1
9
]
ar
e
ca
p
ab
le
o
f
g
en
er
atin
g
r
elate
d
k
ey
wo
r
d
v
ar
iatio
n
s
th
at
ca
p
tu
r
e
d
if
f
e
r
en
t
way
s
u
s
er
s
ex
p
r
ess
th
eir
s
ea
r
ch
in
ten
t.
Fo
r
ex
am
p
le,
a
k
e
y
wo
r
d
s
u
ch
as
“r
u
n
n
i
n
g
s
h
o
es”
m
ay
b
e
ex
p
an
d
ed
i
n
to
r
elate
d
p
h
r
ases
in
clu
d
in
g
“a
th
leti
c
f
o
o
twea
r
,
”
“jo
g
g
in
g
s
n
ea
k
er
s
,
”
o
r
“m
ar
ath
o
n
tr
ain
e
r
s
.
”
Su
ch
s
em
an
tic
ex
p
an
s
io
n
in
cr
ea
s
es
th
e
lik
elih
o
o
d
o
f
m
atch
in
g
r
elev
a
n
t q
u
er
ies an
d
im
p
r
o
v
es a
d
v
er
tis
em
en
t r
elev
a
n
ce
.
I
n
p
r
ac
tice,
k
ey
wo
r
d
s
ar
e
ty
p
ically
co
llected
f
r
o
m
s
ea
r
c
h
q
u
er
ies,
web
p
ag
e
c
o
n
ten
t,
an
d
u
s
er
-
g
en
er
ated
te
x
t
s
u
ch
as
co
m
m
e
n
ts
o
r
r
ev
iews.
Ad
v
er
tis
em
en
ts
ar
e
in
d
ex
e
d
u
s
in
g
t
h
ese
k
ey
wo
r
d
s
s
o
th
ey
ca
n
b
e
ef
f
icien
tly
r
etr
iev
ed
wh
e
n
r
elate
d
q
u
er
ies
o
cc
u
r
.
W
h
en
u
s
er
s
s
ea
r
ch
f
o
r
o
r
in
ter
ac
t
wit
h
r
elev
an
t
co
n
ten
t
,
th
e
s
y
s
tem
co
m
p
ar
es
th
e
ass
o
ciate
d
k
ey
wo
r
d
s
with
in
d
ex
e
d
ad
v
er
tis
em
en
ts
an
d
r
etr
iev
es
ad
s
th
at
m
atch
th
e
u
s
er
’
s
cu
r
r
en
t c
o
n
tex
t
o
r
s
ea
r
c
h
in
ten
t.
2
.
5
.
B
eha
v
io
r
a
l
t
a
rg
et
ing
B
eh
av
io
r
al
tar
g
etin
g
is
a
wid
e
ly
u
s
ed
tech
n
iq
u
e
i
n
ad
v
e
r
tis
em
en
t
r
ec
o
m
m
e
n
d
atio
n
s
y
s
tem
s
.
I
t
r
elie
s
o
n
a
n
aly
zin
g
u
s
er
s
’
o
n
lin
e
ac
tiv
ities
in
o
r
d
er
t
o
d
eliv
er
ad
v
e
r
tis
em
en
ts
th
at
r
e
f
lect
th
eir
in
ter
ests
an
d
p
r
ef
er
en
ce
s
.
User
b
eh
av
i
o
r
al
d
ata
co
llected
f
r
o
m
b
r
o
wsi
n
g
h
is
to
r
y
,
s
ea
r
ch
q
u
er
ies,
s
o
cial
m
ed
ia
in
ter
ac
tio
n
s
,
an
d
p
u
r
ch
ase
b
eh
a
v
io
r
ar
e
a
g
g
r
eg
ated
o
v
er
tim
e
to
b
u
ild
u
s
er
p
r
o
f
iles
th
at
ca
p
tu
r
e
l
o
n
g
-
ter
m
in
ter
ests
an
d
en
g
ag
em
e
n
t p
atter
n
s
.
Ad
v
er
tis
em
en
ts
ar
e
th
e
n
in
d
ex
ed
ac
co
r
d
in
g
to
b
eh
a
v
io
r
al
p
atter
n
s
.
Fo
r
ex
am
p
le,
ad
v
e
r
tis
em
en
ts
r
elate
d
to
o
u
td
o
o
r
eq
u
ip
m
en
t
m
ay
b
e
ass
o
ciate
d
with
b
eh
av
io
r
al
in
d
icato
r
s
s
u
ch
as
h
i
k
in
g
,
ca
m
p
i
n
g
,
o
r
ad
v
en
tu
r
e
-
r
elate
d
s
ea
r
ch
es.
D
u
r
in
g
t
h
e
r
etr
iev
al
p
r
o
ce
s
s
,
th
e
s
y
s
tem
an
aly
ze
s
u
s
er
b
e
h
av
io
r
an
d
m
atch
es
it
with
ad
v
er
tis
em
en
ts
th
at
co
r
r
e
s
p
o
n
d
to
th
ese
o
b
s
er
v
ed
i
n
ter
e
s
ts
.
Fo
r
in
s
tan
ce
,
u
s
er
s
wh
o
f
r
eq
u
en
tly
s
ea
r
ch
f
o
r
h
ik
in
g
g
ea
r
m
ay
b
e
p
r
esen
ted
with
ad
v
er
tis
em
en
ts
f
o
r
o
u
td
o
o
r
eq
u
i
p
m
en
t
[
1
7
]
,
[
1
8
]
.
2
.
6
.
Co
nte
x
t
ua
l
t
a
rg
et
ing
C
o
n
tex
tu
al
tar
g
etin
g
d
eliv
e
r
s
ad
v
er
tis
em
en
ts
b
ased
o
n
th
e
co
n
ten
t
o
f
th
e
we
b
p
ag
e
th
at
a
u
s
er
is
cu
r
r
en
tly
v
iewin
g
.
I
n
s
tead
o
f
r
ely
in
g
p
r
i
m
ar
ily
o
n
u
s
er
p
r
o
f
iles
,
th
is
ap
p
r
o
ac
h
an
aly
z
es
th
e
s
u
r
r
o
u
n
d
i
n
g
co
n
ten
t to
d
eter
m
in
e
wh
ic
h
ad
v
er
tis
em
en
ts
ar
e
m
o
s
t r
elev
an
t
in
a
g
iv
e
n
co
n
te
x
t.
C
o
n
tex
tu
al
in
f
o
r
m
atio
n
is
ty
p
ically
ex
tr
ac
ted
f
r
o
m
web
p
a
g
e
tex
t,
im
a
g
es,
an
d
m
etad
at
a.
Natu
r
al
lan
g
u
ag
e
p
r
o
ce
s
s
in
g
(
NL
P)
a
n
d
co
n
ten
t
an
aly
s
is
tech
n
iq
u
e
s
ar
e
co
m
m
o
n
ly
u
s
ed
to
u
n
d
e
r
s
tan
d
th
e
s
em
an
tic
m
ea
n
in
g
o
f
th
e
p
ag
e
an
d
id
en
t
if
y
r
elev
an
t t
o
p
ics
[
1
9
]
,
[
2
0
]
.
Ad
v
er
tis
em
en
ts
ar
e
i
n
d
ex
e
d
with
k
ey
wo
r
d
s
th
at
co
r
r
esp
o
n
d
to
s
p
ec
if
ic
c
o
n
tex
tu
al
th
em
es.
Fo
r
ex
am
p
le,
a
d
v
er
tis
em
en
ts
f
o
r
g
y
m
m
e
m
b
er
s
h
ip
s
o
r
s
p
o
r
ts
wea
r
m
ay
b
e
ass
o
ciate
d
with
k
ey
wo
r
d
s
r
elate
d
to
f
itn
ess
,
h
ea
lth
,
an
d
ex
er
cise.
W
h
en
a
u
s
er
v
is
its
a
web
p
ag
e,
th
e
s
y
s
tem
an
al
y
ze
s
th
e
p
ag
e
c
o
n
ten
t
a
n
d
r
etr
iev
es a
d
v
er
tis
em
en
ts
th
at
m
atch
th
e
d
etec
ted
c
o
n
tex
t
[
1
8
]
.
3.
R
E
T
RIE
V
AL
S
Y
ST
E
M
S
R
ec
o
m
m
en
d
atio
n
s
y
s
tem
s
h
a
v
e
b
ee
n
wid
ely
s
tu
d
ied
i
n
th
e
liter
atu
r
e
[
1
]
,
[
2
0
]
.
Ad
tar
g
eti
n
g
f
o
c
u
s
es
o
n
d
eliv
er
in
g
p
e
r
s
o
n
alize
d
ad
v
er
tis
em
en
ts
to
u
s
er
s
b
ased
o
n
th
eir
p
r
o
f
iles
an
d
b
e
h
av
io
r
s
,
a
im
in
g
to
m
ax
i
m
ize
en
g
ag
em
e
n
t
an
d
co
n
v
er
s
io
n
r
ates.
I
n
co
n
tr
ast,
o
r
g
an
ic
r
etr
iev
al
em
p
h
asizes
im
p
r
o
v
in
g
u
s
er
ex
p
er
ien
ce
b
y
r
ec
o
m
m
en
d
in
g
c
o
n
ten
t o
r
p
r
o
d
u
cts th
at
alig
n
with
u
s
er
s
’
p
r
ef
er
en
ce
s
with
o
u
t d
i
r
ec
t m
o
n
e
tar
y
in
f
lu
e
n
ce
.
3
.
1
.
O
rg
a
nic
re
t
riev
a
l
Or
g
an
ic
r
etr
iev
al,
also
k
n
o
w
n
as
o
r
g
an
ic
r
ec
o
m
m
en
d
atio
n
,
aim
s
t
o
e
n
h
an
ce
u
s
er
ex
p
er
ien
ce
b
y
p
r
o
v
id
i
n
g
p
er
s
o
n
alize
d
co
n
te
n
t
o
r
p
r
o
d
u
ct
s
u
g
g
esti
o
n
s
b
ased
o
n
u
s
er
p
r
ef
er
en
ce
s
an
d
b
eh
av
io
r
.
T
h
ese
s
y
s
tem
s
an
aly
ze
h
is
to
r
ical
in
ter
ac
tio
n
s
to
id
en
tify
item
s
th
at
ar
e
m
o
s
t r
elev
an
t to
ea
ch
u
s
er
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
3
,
J
u
n
e
20
2
6
:
1
5
1
8
-
1
5
3
0
1522
Or
g
an
ic
r
etr
iev
al
s
y
s
tem
s
ar
e
wid
ely
u
s
ed
in
d
o
m
ain
s
s
u
ch
as
e
-
co
m
m
er
ce
,
s
tr
ea
m
in
g
p
lat
f
o
r
m
s
,
an
d
s
o
cial
m
ed
ia
[
1
]
,
[
2
0
]
.
L
ar
g
e
-
s
ca
le
in
d
u
s
tr
ial
s
y
s
tem
s
s
u
ch
as
Yo
u
T
u
b
e
r
ely
h
ea
v
ily
o
n
n
eu
r
al
r
et
r
iev
al
ar
ch
itectu
r
es
f
o
r
ca
n
d
id
ate
g
en
er
atio
n
[
2
1
]
.
Fo
r
ex
am
p
le,
Netf
lix
r
ec
o
m
m
en
d
s
m
o
v
ie
s
b
ased
o
n
u
s
er
s
’
v
iewin
g
h
is
to
r
y
,
w
h
ile
Sp
o
tify
’
s
Dis
co
v
er
W
ee
k
ly
in
tr
o
d
u
ce
s
n
ew
m
u
s
ic
ac
co
r
d
in
g
to
lis
ten
in
g
b
e
h
av
io
r
.
Desp
ite
th
eir
s
u
cc
es
s
,
o
r
g
an
ic
r
etr
iev
al
s
y
s
tem
s
f
ac
e
s
ev
er
al
ch
allen
g
es,
in
clu
d
in
g
d
ata
s
p
ar
s
ity
,
th
e
co
ld
s
tar
t
p
r
o
b
lem
,
an
d
m
ain
t
ain
in
g
r
ec
o
m
m
e
n
d
atio
n
d
i
v
er
s
ity
[
2
2
]
.
Data
s
p
ar
s
ity
o
cc
u
r
s
wh
en
in
ter
ac
tio
n
d
ata
is
lim
ited
,
wh
ile
th
e
c
o
ld
s
tar
t
p
r
o
b
lem
ar
is
es
wh
en
n
ew
u
s
er
s
o
r
item
s
lack
h
is
to
r
ical
r
ec
o
r
d
s
.
Ma
in
tain
in
g
d
iv
e
r
s
ity
is
also
im
p
o
r
tan
t to
p
r
e
v
en
t r
ec
o
m
m
e
n
d
atio
n
s
f
r
o
m
b
ec
o
m
in
g
o
v
er
l
y
n
ar
r
o
w.
T
r
ad
itio
n
al
o
r
g
a
n
ic
r
etr
iev
al
m
eth
o
d
s
in
clu
d
e
co
n
te
n
t
-
b
as
ed
f
ilter
in
g
,
co
llab
o
r
ati
v
e
f
ilter
in
g
,
an
d
h
y
b
r
id
a
p
p
r
o
ac
h
es.
C
o
n
ten
t
-
b
ased
f
ilter
in
g
r
ec
o
m
m
en
d
s
item
s
with
s
im
ilar
ch
ar
ac
ter
is
t
ics
to
th
o
s
e
p
r
ev
io
u
s
ly
p
r
ef
er
r
e
d
b
y
th
e
u
s
er
.
C
o
llab
o
r
ativ
e
f
ilter
in
g
id
en
tifie
s
p
atte
r
n
s
am
o
n
g
u
s
er
s
with
s
im
ilar
p
r
ef
er
en
ce
s
an
d
h
as
ev
o
lv
ed
in
to
n
e
u
r
al
co
llab
o
r
ativ
e
f
ilter
in
g
ar
ch
itectu
r
es
f
o
r
lar
g
e
-
s
ca
le
s
y
s
tem
s
[
7
]
,
[
2
3
]
.
Ma
n
y
m
o
d
er
n
p
latf
o
r
m
s
co
m
b
i
n
e
b
o
th
a
p
p
r
o
ac
h
es
to
im
p
r
o
v
e
r
ec
o
m
m
e
n
d
atio
n
ac
cu
r
ac
y
an
d
r
o
b
u
s
tn
ess
.
T
o
p
r
o
v
id
e
a
clea
r
er
o
v
er
v
iew
o
f
d
if
f
er
e
n
t
r
ec
o
m
m
en
d
atio
n
ap
p
r
o
ac
h
es
,
T
ab
le
1
s
u
m
m
ar
izes
th
e
k
ey
ch
ar
ac
ter
is
tics
,
ad
v
an
tag
es,
an
d
lim
itatio
n
s
o
f
s
ev
er
al
co
m
m
o
n
ly
u
s
ed
r
etr
iev
al
m
eth
o
d
s
.
As
s
h
o
wn
in
T
ab
le
1
,
n
eu
r
al
r
etr
iev
al
m
eth
o
d
s
p
r
o
v
id
e
b
ett
er
s
ca
lab
ilit
y
b
u
t r
eq
u
ir
e
lar
g
e
-
s
ca
le
tr
ain
in
g
d
ata.
T
ab
le
1
.
C
o
m
p
a
r
is
o
n
o
f
m
ajo
r
r
ec
o
m
m
en
d
atio
n
r
etr
iev
al
ap
p
r
o
ac
h
es
M
e
t
h
o
d
K
e
y
I
d
e
a
S
t
r
e
n
g
t
h
Li
mi
t
a
t
i
o
n
C
o
n
t
e
n
t
-
b
a
se
d
f
i
l
t
e
r
i
n
g
R
e
c
o
mm
e
n
d
s
i
t
e
ms w
i
t
h
s
i
mi
l
a
r
a
t
t
r
i
b
u
t
e
s
S
i
mp
l
e
a
n
d
i
n
t
e
r
p
r
e
t
a
b
l
e
Li
mi
t
e
d
d
i
s
c
o
v
e
r
y
o
f
n
e
w
i
n
t
e
r
e
s
t
s
C
o
l
l
a
b
o
r
a
t
i
v
e
f
i
l
t
e
r
i
n
g
U
ses
b
e
h
a
v
i
o
r
o
f
si
m
i
l
a
r
u
s
e
r
s
C
a
p
t
u
r
e
s
i
m
p
l
i
c
i
t
p
r
e
f
e
r
e
n
c
e
s
C
o
l
d
st
a
r
t
a
n
d
s
p
a
r
s
i
t
y
i
ssu
e
s
H
y
b
r
i
d
met
h
o
d
s
C
o
m
b
i
n
e
s m
u
l
t
i
p
l
e
r
e
c
o
m
me
n
d
a
t
i
o
n
s
t
r
a
t
e
g
i
e
s
I
mp
r
o
v
e
s
a
c
c
u
r
a
c
y
a
n
d
r
o
b
u
st
n
e
s
s
H
i
g
h
e
r
s
y
st
e
m
c
o
mp
l
e
x
i
t
y
N
e
u
r
a
l
r
e
t
r
i
e
v
a
l
(
t
w
o
-
t
o
w
e
r
)
Le
a
r
n
s
u
ser/
i
t
e
m
e
mb
e
d
d
i
n
g
s f
o
r
si
mi
l
a
r
i
t
y
se
a
r
c
h
S
c
a
l
a
b
l
e
f
o
r
l
a
r
g
e
s
y
st
e
ms
R
e
q
u
i
r
e
s
l
a
r
g
e
t
r
a
i
n
i
n
g
d
a
t
a
se
t
s
3
.
2
.
T
wo
-
t
o
wer
net
wo
rk
f
o
r
re
t
riev
a
l
T
h
e
two
-
to
wer
m
o
d
el,
also
k
n
o
wn
as
th
e
d
u
al
-
to
wer
m
o
d
e
l,
is
a
d
ee
p
lear
n
i
n
g
a
r
ch
itectu
r
e
wid
ely
u
s
ed
in
r
ec
o
m
m
e
n
d
atio
n
s
y
s
t
em
s
f
o
r
co
n
ten
t
r
etr
iev
al
[
3
]
,
[
4
]
,
[
2
1
]
,
[
2
4
]
,
[
2
5
]
.
I
t
ex
te
n
d
s
tr
ad
itio
n
al
m
atr
ix
f
ac
to
r
izatio
n
a
p
p
r
o
ac
h
es
[
6
]
b
y
lear
n
i
n
g
em
b
ed
d
in
g
r
e
p
r
e
s
en
tatio
n
s
f
o
r
u
s
er
s
an
d
ite
m
s
th
r
o
u
g
h
n
e
u
r
al
n
etwo
r
k
s
.
T
h
e
ar
ch
itectu
r
e
c
o
n
s
is
ts
o
f
two
in
d
ep
en
d
en
t
n
e
u
r
al
n
etwo
r
k
s
:
a
u
s
er
to
wer
an
d
an
item
to
wer
.
T
h
e
u
s
er
to
wer
en
co
d
es
u
s
er
f
ea
tu
r
es
s
u
ch
as
d
em
o
g
r
ap
h
ic
attr
i
b
u
tes,
b
r
o
wsi
n
g
h
is
to
r
y
,
an
d
p
ast
in
ter
ac
tio
n
s
in
to
a
laten
t
r
e
p
r
esen
tatio
n
.
T
h
e
item
to
wer
en
co
d
es
item
f
ea
t
u
r
e
s
s
u
ch
as
m
etad
ata,
tex
tu
al
d
e
s
cr
ip
tio
n
s
,
o
r
v
is
u
al
co
n
ten
t
in
to
item
em
b
e
d
d
in
g
s
.
B
o
th
r
ep
r
esen
tatio
n
s
ar
e
p
r
o
jecte
d
in
to
a
s
h
a
r
ed
laten
t
s
p
a
ce
wh
er
e
s
im
ilar
ity
b
etwe
en
u
s
er
s
an
d
item
s
ca
n
b
e
co
m
p
u
te
d
.
L
et
d
en
o
te
th
e
in
p
u
t
f
ea
tu
r
es
o
f
a
u
s
er
an
d
d
en
o
te
th
e
in
p
u
t
f
ea
tu
r
es
o
f
an
item
.
T
h
e
u
s
er
to
wer
m
ap
s
th
e
u
s
er
f
ea
tu
r
es
in
to
a
laten
t
r
ep
r
esen
tatio
n
th
r
o
u
g
h
a
n
eu
r
al
m
ap
p
i
n
g
f
u
n
ctio
n
(
⋅
)
,
wh
ile
th
e
item
to
wer
m
ap
s
item
f
ea
tu
r
es in
to
an
em
b
ed
d
in
g
v
ec
to
r
th
r
o
u
g
h
an
o
th
e
r
n
eu
r
al
f
u
n
ctio
n
(
⋅
)
:
=
(
)
=
(
)
T
h
e
r
elev
an
ce
b
etwe
en
u
s
er
s
an
d
item
s
is
ty
p
ically
m
ea
s
u
r
ed
u
s
in
g
s
im
ilar
ity
f
u
n
ctio
n
s
s
u
ch
as
th
e
d
o
t
p
r
o
d
u
ct:
(
,
)
=
o
r
co
s
in
e
s
im
ilar
ity
(
,
)
=
|
|
u
|
|
|
|
v
|
|
As
illu
s
tr
ated
in
Fig
u
r
e
3
,
th
is
ar
ch
itectu
r
e
allo
ws
u
s
er
an
d
i
tem
r
ep
r
esen
tatio
n
s
to
b
e
lear
n
ed
in
d
e
p
en
d
e
n
tly
wh
ile
en
ab
lin
g
ef
f
icien
t la
r
g
e
-
s
ca
le
r
etr
iev
al.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
A
s
u
r
ve
y
o
f retr
ieva
l a
lg
o
r
ith
ms in
a
d
a
n
d
co
n
ten
t reco
mme
n
d
a
tio
n
s
ystems
(
Yu
Zh
a
o
)
1523
Fig
u
r
e
3
.
Ar
c
h
itectu
r
e
o
f
th
e
t
wo
-
to
wer
n
eu
r
al
r
etr
iev
al
m
o
d
el
3
.
3
.
T
ra
ini
ng
a
nd
infe
re
nce
T
r
ain
in
g
t
h
e
two
-
to
wer
m
o
d
el
in
v
o
lv
es
lear
n
in
g
e
m
b
ed
d
in
g
s
th
at
ca
p
tu
r
e
co
m
p
atib
ilit
y
b
etwe
en
u
s
er
s
an
d
item
s
.
Du
r
in
g
tr
ain
i
n
g
,
u
s
er
–
item
p
air
s
ar
e
s
am
p
led
f
r
o
m
h
is
to
r
ical
in
ter
ac
tio
n
d
ata.
Po
s
itiv
e
p
air
s
r
ep
r
esen
t
o
b
s
er
v
ed
in
ter
ac
tio
n
s
,
wh
ile
n
eg
ativ
e
s
am
p
les
r
ep
r
esen
t
n
o
n
-
in
ter
ac
tio
n
s
.
T
h
is
s
tr
ateg
y
en
ab
les
th
e
m
o
d
el
to
d
is
tin
g
u
is
h
r
elev
a
n
t
item
s
f
r
o
m
ir
r
ele
v
an
t
o
n
es.
N
eg
ativ
e
s
am
p
lin
g
tec
h
n
iq
u
es
ar
e
o
f
ten
ap
p
lied
to
ad
d
r
ess
th
e
im
b
alan
ce
b
etwe
en
p
o
s
itiv
e
an
d
n
eg
ativ
e
e
x
am
p
les
[
1
1
]
,
[
2
6
]
.
T
h
e
m
o
d
el
is
ty
p
ically
o
p
tim
ized
u
s
in
g
lo
s
s
f
u
n
ctio
n
s
s
u
ch
as
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
[
2
3
]
,
[
2
7
]
,
[
2
8
]
o
r
p
air
wis
e
r
an
k
in
g
lo
s
s
,
w
h
ich
ar
e
c
o
m
m
o
n
ly
u
s
ed
i
n
im
p
licit
f
ee
d
b
ac
k
r
ec
o
m
m
en
d
atio
n
s
y
s
tem
s
.
Mo
d
el
p
ar
am
eter
s
ar
e
u
p
d
ated
th
r
o
u
g
h
g
r
a
d
ien
t
-
b
ased
o
p
tim
izatio
n
alg
o
r
ith
m
s
s
u
ch
as
s
to
ch
asti
c
g
r
ad
ien
t
d
escen
t
(
SGD)
o
r
Ad
am
.
Du
r
in
g
i
n
f
er
en
ce
,
th
e
tr
ain
e
d
m
o
d
el
g
en
e
r
ates
em
b
e
d
d
in
g
v
ec
to
r
s
f
o
r
u
s
er
s
an
d
item
s
.
T
h
e
u
s
er
to
wer
p
r
o
d
u
ce
s
a
v
ec
to
r
r
ep
r
esen
tatio
n
f
o
r
a
g
iv
e
n
u
s
er
,
wh
ile
th
e
item
to
wer
g
en
er
at
es
em
b
ed
d
in
g
s
f
o
r
ca
n
d
id
ate
item
s
.
T
h
e
s
im
ilar
ity
b
etwe
en
u
s
er
an
d
item
v
ec
to
r
s
is
th
en
co
m
p
u
ted
u
s
in
g
s
im
ilar
ity
m
etr
ics
s
u
ch
as
d
o
t
p
r
o
d
u
cts
o
r
co
s
in
e
s
im
ilar
ity
.
T
h
e
s
y
s
tem
r
etr
iev
es
th
e
to
p
-
N
item
s
with
th
e
h
ig
h
est
s
im
ilar
ity
s
co
r
es
an
d
r
ec
o
m
m
en
d
s
th
em
to
th
e
u
s
er
.
I
n
lar
g
e
-
s
ca
le
in
d
u
s
tr
ial
s
y
s
tem
s
,
th
is
r
etr
iev
al
p
r
o
ce
s
s
is
o
f
ten
ac
ce
ler
ated
u
s
in
g
ef
f
icien
t
in
d
ex
in
g
tech
n
iq
u
es
s
u
ch
as
ap
p
r
o
x
im
ate
n
ea
r
est
n
eig
h
b
o
r
(
ANN)
s
ea
r
ch
an
d
d
is
tr
ib
u
ted
r
etr
iev
al
in
f
r
astru
ctu
r
e
[
3
]
,
[
1
6
]
.
T
h
e
o
v
er
all
r
etr
iev
al
p
r
o
ce
d
u
r
e
o
f
a
two
-
to
wer
r
ec
o
m
m
en
d
atio
n
s
y
s
tem
ca
n
b
e
s
u
m
m
ar
ized
as
f
o
llo
ws.
First,
th
e
u
s
er
to
wer
en
co
d
es
u
s
er
f
ea
t
u
r
es
to
g
e
n
er
ate
a
u
s
er
em
b
e
d
d
in
g
v
ec
to
r
.
Seco
n
d
,
th
e
item
to
wer
p
r
o
d
u
ce
s
em
b
ed
d
in
g
r
e
p
r
esen
tatio
n
s
f
o
r
ca
n
d
id
ate
item
s
,
wh
ich
a
r
e
o
f
ten
p
r
ec
o
m
p
u
ted
an
d
s
to
r
ed
in
a
n
in
d
ex
in
g
s
tr
u
ctu
r
e
.
T
h
ir
d
,
a
s
im
ilar
ity
f
u
n
ctio
n
s
u
ch
as
d
o
t
p
r
o
d
u
ct
o
r
c
o
s
in
e
s
im
ilar
ity
is
u
s
ed
to
m
ea
s
u
r
e
t
h
e
r
elev
an
ce
b
etwe
en
th
e
u
s
er
e
m
b
ed
d
in
g
an
d
item
em
b
e
d
d
in
g
s
.
Fin
ally
,
th
e
s
y
s
tem
r
etr
iev
es
th
e
to
p
-
N
item
s
with
th
e
h
ig
h
est s
im
ilar
ity
s
co
r
es a
n
d
r
etu
r
n
s
th
em
as r
ec
o
m
m
en
d
atio
n
ca
n
d
id
ates.
3
.
4
.
Cha
lleng
es
a
nd
ex
t
ens
io
ns
Mo
d
er
n
r
e
co
m
m
e
n
d
atio
n
s
y
s
tem
s
em
p
lo
y
a
v
a
r
iety
o
f
r
etr
iev
al
m
o
d
els
b
ey
o
n
d
t
h
e
two
-
to
wer
ar
ch
itectu
r
e.
Den
s
e
v
ec
to
r
r
et
r
iev
al
m
eth
o
d
s
ar
e
wid
ely
ad
o
p
ted
in
lar
g
e
-
s
ca
le
s
y
s
tem
s
[
1
6
]
,
wh
er
e
u
s
er
an
d
item
r
ep
r
esen
tatio
n
s
ar
e
lear
n
ed
in
a
s
h
ar
ed
em
b
ed
d
i
n
g
s
p
ac
e
an
d
ef
f
icien
t
n
ea
r
est
-
n
eig
h
b
o
r
s
ea
r
ch
is
u
s
ed
f
o
r
ca
n
d
id
ate
g
e
n
er
atio
n
.
I
n
ad
d
itio
n
,
g
r
ap
h
n
e
u
r
al
n
et
wo
r
k
(
GNN)
–
b
ased
m
o
d
els
h
av
e
b
ee
n
in
tr
o
d
u
ce
d
to
ca
p
tu
r
e
co
m
p
lex
u
s
er
–
item
r
elatio
n
s
h
ip
s
th
r
o
u
g
h
in
ter
ac
tio
n
g
r
ap
h
s
[
2
9
]
–
[
3
1
]
.
R
ep
r
esen
tativ
e
ex
ten
s
io
n
s
in
clu
d
e
s
o
cial
r
ec
o
m
m
en
d
atio
n
m
o
d
els
[
3
2
]
an
d
lig
h
tweig
h
t
g
r
ap
h
c
o
n
v
o
lu
tio
n
ap
p
r
o
ac
h
es
s
u
ch
as
L
ig
h
tGC
N
[
3
3
]
,
wh
ile
s
u
r
v
ey
s
tu
d
ies p
r
o
v
id
e
b
r
o
ad
e
r
o
v
er
v
iews o
f
th
is
r
esear
ch
ar
ea
[
3
4
]
–
[
3
6
]
.
Kn
o
wled
g
e
g
r
a
p
h
–
b
ased
r
ec
o
m
m
en
d
atio
n
m
eth
o
d
s
f
u
r
th
er
en
h
an
ce
r
etr
ie
v
al
b
y
in
c
o
r
p
o
r
ati
n
g
s
tr
u
ctu
r
ed
r
elatio
n
al
in
f
o
r
m
at
io
n
b
etwe
en
u
s
er
s
an
d
item
s
,
im
p
r
o
v
in
g
b
o
th
ac
cu
r
ac
y
a
n
d
ex
p
lain
a
b
ilit
y
in
co
m
p
lex
r
ec
o
m
m
en
d
atio
n
s
ce
n
ar
io
s
[
3
7
]
–
[
3
9
]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
3
,
J
u
n
e
20
2
6
:
1
5
1
8
-
1
5
3
0
1524
Mu
ltimo
d
al
r
etr
ie
v
al
ap
p
r
o
a
ch
es
f
u
r
th
er
e
n
h
an
ce
r
ec
o
m
m
en
d
atio
n
p
er
f
o
r
m
a
n
ce
b
y
in
teg
r
atin
g
in
f
o
r
m
atio
n
f
r
o
m
m
u
ltip
le
m
o
d
alities
s
u
ch
as
tex
t,
im
a
g
es,
an
d
au
d
i
o
[
2
2
]
,
[
4
0
]
.
A
k
ey
ch
allen
g
e
in
r
ec
o
m
m
en
d
atio
n
s
y
s
tem
s
is
th
e
co
ld
s
tar
t
p
r
o
b
lem
[
4
1
]
–
[
4
3
]
,
wh
e
r
e
n
ew
u
s
er
s
o
r
ite
m
s
lack
s
u
f
f
icien
t
in
ter
ac
tio
n
d
ata.
T
o
ad
d
r
ess
th
is
is
s
u
e,
s
ev
er
al
s
tr
ateg
ies
h
av
e
b
ee
n
ex
p
lo
r
e
d
,
in
clu
d
i
n
g
tr
an
s
f
er
lear
n
in
g
,
wh
ich
lev
er
a
g
es
k
n
o
wled
g
e
f
r
o
m
lar
g
e
-
s
ca
le
d
atasets
,
an
d
s
y
n
th
etic
d
ata
g
en
e
r
atio
n
,
wh
ic
h
au
g
m
en
ts
s
p
ar
s
e
in
ter
ac
tio
n
d
ata.
Fed
er
ated
le
ar
n
in
g
h
as
also
b
ee
n
p
r
o
p
o
s
ed
to
en
ab
le
co
llab
o
r
ativ
e
m
o
d
el
tr
ain
in
g
ac
r
o
s
s
d
is
tr
ib
u
ted
d
ata
s
o
u
r
ce
s
wh
ile
p
r
eser
v
in
g
u
s
er
p
r
iv
a
cy
[
4
4
]
.
An
o
th
er
im
p
o
r
tan
t
f
ac
to
r
af
f
ec
tin
g
s
y
s
tem
p
er
f
o
r
m
an
ce
i
s
d
ata
q
u
ality
[
4
2
]
.
R
ec
o
m
m
en
d
atio
n
ef
f
ec
tiv
en
ess
r
elies
h
ea
v
ily
o
n
r
eliab
le
i
n
ter
ac
tio
n
d
ata,
an
d
n
o
is
y
o
r
s
p
ar
s
e
d
atasets
ca
n
s
i
g
n
if
ican
tly
d
eg
r
a
d
e
m
o
d
el
p
er
f
o
r
m
a
n
ce
.
R
ec
en
t
wo
r
k
h
as
p
r
o
p
o
s
ed
s
ev
er
al
e
x
ten
s
io
n
s
to
en
h
an
ce
t
h
e
tr
a
d
itio
n
al
two
-
to
wer
ar
ch
itectu
r
e.
Mu
lti
-
task
lear
n
in
g
allo
ws
th
e
m
o
d
el
to
jo
in
tly
o
p
tim
ize
m
u
ltip
le
o
b
jectiv
es
s
u
ch
as
cl
ick
-
th
r
o
u
g
h
r
ate
p
r
e
d
ictio
n
an
d
u
s
er
en
g
ag
em
e
n
t
[
2
6
]
,
[
4
5
]
.
T
h
e
th
r
ee
-
t
o
wer
ar
c
h
itectu
r
e
f
u
r
th
er
e
x
ten
d
s
th
i
s
d
esig
n
b
y
i
n
tr
o
d
u
cin
g
a
n
ad
d
itio
n
al
n
etwo
r
k
to
m
o
d
el
co
n
tex
tu
al
in
f
o
r
m
atio
n
,
in
cl
u
d
in
g
tem
p
o
r
al
s
ig
n
als,
lo
ca
tio
n
d
ata,
a
n
d
cr
o
s
s
-
f
ea
tu
r
e
in
ter
ac
tio
n
s
[
4
6
]
,
[
4
7
]
.
As
illu
s
tr
ated
in
Fig
u
r
e
4
,
t
h
e
th
r
ee
-
t
o
wer
ar
c
h
itectu
r
e
p
r
o
ce
s
s
es
u
s
er
f
ea
tu
r
es,
item
f
ea
tu
r
es,
an
d
co
n
tex
tu
al
f
ea
tu
r
es th
r
o
u
g
h
s
e
p
ar
ate
n
eu
r
al
to
wer
s
b
ef
o
r
e
c
o
m
b
in
in
g
th
em
in
a
s
h
ar
ed
r
e
p
r
esen
tatio
n
lay
er
f
o
r
f
in
al
p
r
ed
ictio
n
.
T
h
is
d
esig
n
en
ab
les
th
e
m
o
d
el
to
ca
p
tu
r
e
r
ich
er
in
ter
ac
tio
n
s
an
d
im
p
r
o
v
e
r
ec
o
m
m
en
d
atio
n
p
er
f
o
r
m
an
ce
in
c
o
m
p
lex
s
ce
n
a
r
io
s
.
R
ec
en
t
s
tu
d
ies
h
av
e
also
ex
p
lo
r
ed
in
teg
r
atin
g
L
L
Ms
in
t
o
r
ec
o
m
m
e
n
d
atio
n
p
ip
elin
es
[
1
1
]
–
[
1
3
]
.
L
L
Ms
ca
n
g
en
er
ate
s
y
n
th
eti
c
in
ter
ac
tio
n
d
ata
an
d
en
h
an
ce
s
em
an
tic
r
e
p
r
esen
tatio
n
s
o
f
tex
t
u
al
f
ea
tu
r
es,
h
elp
in
g
allev
iate
co
ld
s
tar
t
is
s
u
es
an
d
im
p
r
o
v
in
g
th
e
r
o
b
u
s
tn
ess
o
f
r
ec
o
m
m
en
d
atio
n
s
y
s
tem
s
.
I
n
ad
d
itio
n
,
L
L
Ms
ca
n
b
e
u
s
ed
to
en
r
i
ch
u
s
er
,
item
,
an
d
c
o
n
tex
tu
al
r
ep
r
esen
tatio
n
s
,
co
m
p
lem
en
tin
g
m
u
lti
-
to
we
r
ar
ch
itectu
r
es a
n
d
im
p
r
o
v
in
g
r
e
co
m
m
en
d
atio
n
p
er
f
o
r
m
a
n
ce
in
s
p
ar
s
e
an
d
co
m
p
lex
d
ata
s
ce
n
ar
io
s
.
Fig
u
r
e
4
.
Ar
c
h
itectu
r
e
o
f
th
e
t
h
r
ee
-
to
wer
n
eu
r
al
r
ec
o
m
m
en
d
atio
n
m
o
d
el
4.
CO
M
P
ARI
NG
AD
T
ARG
E
T
I
NG
AND
O
RG
AN
I
C
R
E
T
RI
E
VA
L
Fro
m
a
s
y
s
tem
d
esig
n
p
er
s
p
ec
tiv
e,
b
o
th
p
a
r
ad
ig
m
s
s
h
ar
e
s
im
ilar
r
etr
iev
al
m
ec
h
an
is
m
s
b
u
t
d
iv
er
g
e
in
o
p
tim
izatio
n
o
b
jectiv
es
an
d
e
v
alu
atio
n
cr
iter
ia.
Mo
d
e
r
n
ad
v
er
tis
in
g
s
y
s
tem
s
o
f
ten
r
ely
o
n
lar
g
e
-
s
ca
le
click
p
r
ed
ictio
n
m
o
d
els
[
4
5
]
.
At
a
f
u
n
d
am
e
n
tal
lev
el,
ad
tar
g
etin
g
an
d
o
r
g
an
ic
r
ec
o
m
m
en
d
atio
n
r
etr
iev
al
s
h
ar
e
a
s
im
ilar
o
b
jectiv
e:
m
atch
i
n
g
u
s
er
in
ten
t
with
r
elev
a
n
t
co
n
ten
t
.
Ad
v
er
tis
in
g
s
y
s
tem
s
an
aly
ze
u
s
er
b
eh
av
io
r
a
n
d
d
em
o
g
r
a
p
h
ic
in
f
o
r
m
atio
n
to
p
r
ed
ict
wh
ic
h
a
d
v
er
tis
em
en
ts
ar
e
m
o
s
t
lik
ely
to
g
en
e
r
ate
en
g
ag
em
e
n
t,
wh
ile
s
ea
r
ch
en
g
in
es
a
n
d
r
ec
o
m
m
e
n
d
atio
n
s
y
s
tem
s
an
aly
ze
q
u
er
ies,
in
ter
ac
tio
n
h
is
to
r
ies,
an
d
co
n
ten
t
f
ea
tu
r
es
to
id
en
tify
r
elev
an
t
item
s
.
I
n
b
o
t
h
ca
s
es,
m
ac
h
in
e
lear
n
in
g
an
d
lar
g
e
-
s
ca
le
d
ata
p
r
o
ce
s
s
in
g
p
l
ay
a
ce
n
tr
al
r
o
le
in
im
p
r
o
v
in
g
th
e
ac
cu
r
ac
y
o
f
co
n
ten
t
d
eliv
er
y
.
T
h
is
m
et
h
o
d
o
lo
g
ical
s
im
ilar
ity
h
ig
h
lig
h
ts
th
e
in
cr
ea
s
in
g
co
n
v
er
g
en
ce
b
etwe
en
tar
g
eted
ad
v
er
tis
in
g
tech
n
o
lo
g
ies an
d
co
n
ten
t r
etr
iev
al
s
y
s
tem
s
[
1
]
,
[
2
0
]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
A
s
u
r
ve
y
o
f retr
ieva
l a
lg
o
r
ith
ms in
a
d
a
n
d
co
n
ten
t reco
mme
n
d
a
tio
n
s
ystems
(
Yu
Zh
a
o
)
1525
I
n
d
ex
-
b
ased
a
d
tar
g
etin
g
is
co
n
ce
p
tu
ally
s
im
ilar
to
t
h
e
i
n
v
er
ted
in
d
ex
wid
ely
u
s
ed
i
n
o
r
g
an
ic
in
f
o
r
m
atio
n
r
etr
iev
al
[
1
5
]
.
B
o
th
ap
p
r
o
ac
h
es
r
ely
o
n
s
tr
u
c
tu
r
ed
in
d
ex
i
n
g
m
ec
h
a
n
is
m
s
to
ef
f
icien
tly
m
atch
u
s
er
s
with
r
elev
an
t c
o
n
ten
t.
I
n
ad
v
er
tis
in
g
s
y
s
tem
s
,
u
s
er
p
r
o
f
iles
ar
e
ca
teg
o
r
ized
b
ased
o
n
b
eh
av
io
r
al
p
atter
n
s
,
in
ter
ests
,
an
d
d
em
o
g
r
a
p
h
ic
attr
ib
u
tes,
allo
win
g
ad
v
er
tis
er
s
to
id
en
tif
y
p
o
ten
tial
au
d
i
en
ce
s
f
o
r
s
p
ec
if
ic
ca
m
p
aig
n
s
.
I
n
o
r
g
an
ic
r
etr
iev
al
s
y
s
tem
s
,
in
v
er
ted
in
d
ices
m
ap
k
ey
wo
r
d
s
o
r
f
ea
tu
r
es
to
d
o
cu
m
e
n
ts
o
r
item
s
,
en
ab
lin
g
s
ea
r
ch
en
g
in
es to
r
ap
id
ly
r
etr
iev
e
r
ele
v
an
t r
esu
lts
in
r
esp
o
n
s
e
to
u
s
er
q
u
er
ies.
T
ar
g
etin
g
s
tr
ateg
ies
in
a
d
v
er
tis
in
g
also
r
esem
b
le
r
etr
iev
a
l
m
ec
h
an
is
m
s
u
s
ed
in
o
r
g
an
ic
s
ea
r
ch
s
y
s
tem
s
.
I
n
ter
est
-
b
ased
tar
g
etin
g
d
is
p
lay
s
ad
v
er
tis
em
en
ts
to
u
s
er
s
ac
co
r
d
in
g
t
o
th
eir
in
f
er
r
ed
in
ter
ests
an
d
b
r
o
wsi
n
g
b
eh
a
v
io
r
s
,
wh
ich
p
ar
allels
r
ec
o
m
m
en
d
atio
n
s
y
s
tem
s
th
at
r
etr
iev
e
co
n
ten
t
alig
n
ed
with
u
s
er
p
r
ef
er
en
ce
s
.
Key
w
o
r
d
tar
g
etin
g
f
o
c
u
s
es
o
n
s
p
ec
if
ic
wo
r
d
s
o
r
p
h
r
ases
an
d
f
u
n
ctio
n
s
s
im
ilar
ly
to
tr
ad
itio
n
a
l
s
ea
r
ch
r
etr
iev
al,
wh
er
e
u
s
er
q
u
er
ies
ar
e
m
atc
h
ed
a
g
ain
s
t
in
d
ex
ed
co
n
ten
t
co
n
tain
i
n
g
r
elev
an
t
ter
m
s
.
Desp
ite
th
ese
s
im
ilar
itie
s
,
th
e
two
s
y
s
tem
s
d
if
f
er
in
th
eir
o
p
tim
iz
atio
n
g
o
als,
d
ata
u
s
ag
e,
an
d
ev
alu
atio
n
m
etr
ics.
T
ab
le
2
s
u
m
m
ar
izes
th
e
k
ey
d
if
f
er
en
ce
s
b
etwe
en
ad
tar
g
etin
g
s
y
s
tem
s
an
d
o
r
g
an
ic
r
etr
iev
al
s
y
s
tem
s
in
ter
m
s
o
f
o
b
jectiv
es,
d
ata
s
o
u
r
ce
s
,
o
p
tim
izatio
n
tar
g
ets,
an
d
e
v
alu
atio
n
m
etr
ics.
T
h
ese
d
if
f
e
r
en
ce
s
ar
e
f
u
r
th
e
r
r
ef
lecte
d
in
th
eir
d
ata
u
s
ag
e
an
d
ev
alu
atio
n
s
tr
ateg
ies.
Ad
tar
g
etin
g
r
elies
h
ea
v
ily
o
n
d
em
o
g
r
ap
h
ic
an
d
b
eh
av
io
r
al
d
ata
f
o
r
p
r
ec
is
e
ta
r
g
etin
g
,
wh
ile
o
r
g
an
ic
r
etr
iev
al
em
p
h
asizes
h
is
to
r
ical
in
ter
ac
tio
n
s
to
im
p
r
o
v
e
r
elev
an
ce
an
d
lo
n
g
-
ter
m
e
n
g
ag
em
en
t.
I
n
ad
d
itio
n
,
ad
s
y
s
tem
s
o
f
ten
r
aise
g
r
ea
ter
p
r
iv
a
cy
co
n
ce
r
n
s
d
u
e
t
o
ex
ten
s
iv
e
u
s
er
p
r
o
f
ilin
g
[
1
7
]
,
[
1
8
]
.
T
ab
le
2
.
C
o
m
p
a
r
is
o
n
o
f
ad
tar
g
etin
g
an
d
o
r
g
an
ic
r
etr
iev
al
s
y
s
tem
s
A
sp
e
c
t
A
d
T
a
r
g
e
t
i
n
g
S
y
st
e
ms
O
r
g
a
n
i
c
R
e
t
r
i
e
v
a
l
S
y
s
t
e
ms
O
b
j
e
c
t
i
v
e
M
a
x
i
m
i
z
e
a
d
v
e
r
t
i
si
n
g
r
e
v
e
n
u
e
a
n
d
c
o
n
v
e
r
si
o
n
r
a
t
e
s
I
mp
r
o
v
e
u
s
e
r
sa
t
i
sf
a
c
t
i
o
n
a
n
d
e
n
g
a
g
e
men
t
P
r
i
mary
d
a
t
a
s
o
u
r
c
e
U
ser d
e
mo
g
r
a
p
h
i
c
s
,
b
r
o
w
s
i
n
g
h
i
s
t
o
r
y
,
b
e
h
a
v
i
o
r
a
l
t
r
a
c
k
i
n
g
H
i
st
o
r
i
c
a
l
i
n
t
e
r
a
c
t
i
o
n
s
,
p
r
e
f
e
r
e
n
c
e
s,
a
n
d
c
o
n
su
mp
t
i
o
n
p
a
t
t
e
r
n
s
O
p
t
i
mi
z
a
t
i
o
n
t
a
r
g
e
t
C
l
i
c
k
-
t
h
r
o
u
g
h
r
a
t
e
(
C
T
R
)
,
c
o
n
v
e
r
si
o
n
r
a
t
e
,
a
d
v
e
r
t
i
si
n
g
R
O
I
R
e
l
e
v
a
n
c
e
,
d
i
v
e
r
s
i
t
y
,
l
o
n
g
-
t
e
r
m u
ser
e
n
g
a
g
e
m
e
n
t
Ev
a
l
u
a
t
i
o
n
me
t
r
i
c
s
C
o
s
t
p
e
r
c
l
i
c
k
(
C
P
C
)
,
C
T
R
,
c
o
n
v
e
r
si
o
n
r
a
t
e
,
r
e
v
e
n
u
e
D
A
U
/
M
A
U
,
r
e
t
e
n
t
i
o
n
r
a
t
e
,
u
ser
e
n
g
a
g
e
m
e
n
t
met
r
i
c
s
B
u
s
i
n
e
ss r
o
l
e
D
i
r
e
c
t
m
o
n
e
t
i
z
a
t
i
o
n
t
h
r
o
u
g
h
a
d
v
e
r
t
i
si
n
g
I
n
d
i
r
e
c
t
r
e
v
e
n
u
e
t
h
r
o
u
g
h
u
ser
r
e
t
e
n
t
i
o
n
a
n
d
p
l
a
t
f
o
r
m
g
r
o
w
t
h
P
r
i
v
a
c
y
c
o
n
c
e
r
n
s
H
i
g
h
d
u
e
t
o
e
x
t
e
n
s
i
v
e
u
ser
p
r
o
f
i
l
i
n
g
M
o
d
e
r
a
t
e
,
p
r
i
m
a
r
i
l
y
f
o
c
u
se
d
o
n
i
n
t
e
r
a
c
t
i
o
n
d
a
t
a
Ty
p
i
c
a
l
i
n
d
u
s
t
r
i
a
l
e
x
a
mp
l
e
s
G
o
o
g
l
e
A
d
s,
F
a
c
e
b
o
o
k
A
d
s,
p
r
o
g
r
a
m
mat
i
c
a
d
v
e
r
t
i
si
n
g
N
e
t
f
l
i
x
r
e
c
o
mm
e
n
d
a
t
i
o
n
s,
Y
o
u
Tu
b
e
r
e
c
o
m
me
n
d
a
t
i
o
n
s,
S
p
o
t
i
f
y
p
l
a
y
l
i
st
s
As
s
h
o
wn
in
T
ab
le
2
,
alth
o
u
g
h
ad
tar
g
etin
g
a
n
d
o
r
g
an
ic
r
et
r
iev
al
s
y
s
tem
s
s
h
ar
e
s
im
ilar
u
n
d
er
ly
in
g
r
etr
iev
al
m
ec
h
an
is
m
s
,
th
ey
d
if
f
er
s
ig
n
if
ican
tly
in
th
eir
o
b
ject
iv
es,
d
ata
u
s
ag
e,
an
d
ev
alu
atio
n
m
etr
ics,
lead
in
g
to
d
if
f
e
r
en
t
o
p
tim
izatio
n
s
tr
at
eg
ies
in
r
ea
l
-
wo
r
ld
s
y
s
tem
s
.
A
k
ey
i
n
s
ig
h
t
is
th
at
alth
o
u
g
h
ad
tar
g
etin
g
an
d
o
r
g
an
ic
r
et
r
iev
al
s
y
s
tem
s
s
h
ar
e
s
im
ilar
r
etr
iev
al
m
ec
h
an
is
m
s
,
th
eir
o
p
tim
izatio
n
o
b
jectiv
es
lead
to
f
u
n
d
am
e
n
tally
d
if
f
er
en
t
s
y
s
te
m
d
esig
n
s
.
Ad
s
y
s
tem
s
p
r
io
r
it
ize
s
h
o
r
t
-
ter
m
r
ev
en
u
e
m
etr
ic
s
s
u
ch
as
C
T
R
an
d
co
n
v
er
s
io
n
r
ate,
wh
ile
o
r
g
an
i
c
s
y
s
tem
s
em
p
h
asize
lo
n
g
-
ter
m
u
s
er
en
g
a
g
em
en
t,
d
i
v
er
s
ity
,
an
d
r
eten
tio
n
.
T
h
is
d
if
f
er
en
ce
s
ig
n
if
ica
n
tly
in
f
lu
en
ce
s
h
o
w
r
etr
iev
al
m
o
d
els
ar
e
tr
ain
ed
,
tu
n
e
d
,
an
d
d
ep
l
o
y
ed
in
r
ea
l
-
wo
r
ld
s
y
s
tem
s
.
5.
M
E
T
RI
CS A
ND
E
XP
E
RI
M
E
NT
AT
I
O
N
5
.
1
.
M
et
rics
E
v
alu
atio
n
m
etr
ics
an
d
u
s
er
ex
p
er
im
en
ts
ar
e
wid
ely
u
s
ed
to
ass
es
s
r
ec
o
m
m
en
d
atio
n
s
y
s
tem
p
er
f
o
r
m
an
ce
[
2
1
]
,
[
2
6
]
.
I
n
la
r
g
e
-
s
ca
le
r
ec
o
m
m
en
d
atio
n
a
n
d
ad
v
er
tis
in
g
s
y
s
tem
s
,
co
m
m
o
n
ly
u
s
ed
m
etr
ics
in
clu
d
e
click
-
th
r
o
u
g
h
r
ate
(
C
T
R
)
,
ar
ea
u
n
d
e
r
th
e
r
ec
eiv
e
r
o
p
er
atin
g
c
h
ar
ac
ter
is
tic
cu
r
v
e
(
AUC),
p
r
ec
is
io
n
,
an
d
r
ec
all
[
2
6
]
,
[
4
5
]
.
C
T
R
m
ea
s
u
r
es th
e
p
r
o
p
o
r
ti
o
n
o
f
im
p
r
e
s
s
io
n
s
th
at
r
esu
lt in
u
s
er
click
s
an
d
is
wid
ely
u
s
ed
in
ad
v
er
tis
in
g
s
y
s
tem
s
.
Pre
ci
s
io
n
an
d
r
ec
all
ev
alu
ate
th
e
r
elev
an
ce
o
f
r
ec
o
m
m
e
n
d
ed
item
s
,
wh
ile
A
UC
r
ef
lects th
e
m
o
d
el’
s
ab
ilit
y
to
co
r
r
ec
tly
r
a
n
k
p
o
s
itiv
e
in
ter
ac
tio
n
s
h
ig
h
er
t
h
an
n
e
g
ativ
e
o
n
es
.
T
h
ese
m
etr
ics ar
e
ty
p
ically
u
s
ed
in
o
f
f
lin
e
e
v
alu
atio
n
b
ef
o
r
e
m
o
d
els ar
e
d
ep
l
o
y
ed
to
o
n
lin
e
s
y
s
tem
s
[
2
6
]
,
[
4
5
]
.
C
o
n
ten
t
r
ec
o
m
m
e
n
d
atio
n
s
y
s
tem
s
p
r
im
ar
ily
f
o
c
u
s
o
n
m
e
tr
ics
th
at
r
ef
lect
u
s
er
en
g
a
g
e
m
en
t
an
d
r
eten
tio
n
[
2
1
]
,
[
4
8
]
,
[
4
9
]
,
s
u
c
h
as
m
o
n
th
l
y
ac
tiv
e
u
s
er
s
(
M
AU)
,
d
aily
ac
tiv
e
u
s
er
s
(
DAU
)
,
an
d
u
s
er
r
eten
tio
n
r
ates
[
2
1
]
.
T
h
ese
s
y
s
tem
s
aim
to
in
c
r
ea
s
e
u
s
er
e
n
g
ag
e
m
en
t
an
d
r
ete
n
tio
n
t
h
r
o
u
g
h
p
er
s
o
n
a
lized
co
n
ten
t.
T
h
e
u
ltima
te
g
o
al
is
to
in
cr
ea
s
e
u
s
er
lo
y
alty
an
d
s
atis
f
ac
tio
n
,
lea
d
in
g
t
o
s
u
s
tain
ed
g
r
o
wth
in
u
s
er
b
ase
a
n
d
ac
tiv
ity
lev
els.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
3
,
J
u
n
e
20
2
6
:
1
5
1
8
-
1
5
3
0
1526
I
n
ad
d
itio
n
to
th
ese
h
ig
h
-
lev
el
m
etr
ics,
m
o
r
e
g
r
an
u
lar
m
etr
i
cs
lik
e
click
s
,
lik
es,
an
d
f
o
llo
ws
ar
e
al
s
o
cr
u
cial.
T
h
ese
m
etr
ics
ar
e
p
ar
ticu
lar
ly
s
en
s
itiv
e
to
alg
o
r
ith
m
ic
u
p
d
ates
an
d
p
r
o
v
id
e
im
m
ed
iate
f
ee
d
b
ac
k
o
n
u
s
er
en
g
a
g
em
en
t
with
t
h
e
r
ec
o
m
m
en
d
e
d
c
o
n
ten
t.
T
h
e
y
a
r
e
clo
s
ely
m
o
n
ito
r
e
d
t
o
g
au
g
e
t
h
e
ef
f
ec
tiv
en
ess
o
f
th
e
r
ec
o
m
m
en
d
atio
n
alg
o
r
it
h
m
s
an
d
to
m
ak
e
tim
ely
ad
ju
s
tm
en
ts
th
at
ca
n
f
u
r
th
er
en
h
a
n
c
e
u
s
er
in
ter
ac
tio
n
an
d
s
atis
f
ac
tio
n
.
I
n
co
n
tr
ast,
ad
s
y
s
tem
s
o
p
er
ate
with
in
a
th
r
ee
-
p
ar
ty
ec
o
s
y
s
tem
in
v
o
lv
in
g
a
d
v
er
tis
er
s
,
th
e
p
latf
o
r
m
,
an
d
u
s
er
s
.
E
ac
h
p
ar
ty
h
as
d
is
tin
ct
in
ter
ests
an
d
m
etr
ics
f
o
r
s
u
cc
ess
.
Ad
v
er
tis
er
s
ar
e
p
r
im
ar
ily
co
n
ce
r
n
e
d
with
m
etr
ics
s
u
ch
as
co
s
t
p
er
click
(
C
PC
)
an
d
co
n
v
er
s
io
n
r
ates
[
1
7
]
,
[
1
8
]
,
[
4
5
]
,
[
4
8
]
,
wh
ich
in
d
icate
th
e
ef
f
ec
tiv
en
ess
an
d
ef
f
icien
cy
o
f
th
eir
a
d
s
p
en
d
.
T
h
e
p
latf
o
r
m
,
o
n
th
e
o
th
er
h
an
d
,
m
ea
s
u
r
es
r
ev
en
u
e
g
en
e
r
ated
f
r
o
m
a
d
p
lace
m
en
ts
,
aim
in
g
t
o
m
ax
im
ize
o
v
e
r
all
p
r
o
f
itab
ilit
y
.
User
s
’
in
ter
ests
lie
in
th
e
r
elev
an
ce
a
n
d
n
o
n
-
in
tr
u
s
iv
en
ess
o
f
th
e
ad
s
th
ey
en
co
u
n
ter
;
h
i
g
h
r
ele
v
an
ce
c
an
en
h
a
n
ce
u
s
er
ex
p
e
r
ien
ce
,
wh
ile
ir
r
elev
an
t
o
r
ex
ce
s
s
iv
e
ad
s
ca
n
d
etr
ac
t
f
r
o
m
it.
B
alan
cin
g
th
ese
th
r
ee
s
ets
o
f
in
ter
ests
i
s
cr
u
cial
f
o
r
th
e
s
u
cc
ess
o
f
ad
s
y
s
tem
s
.
5
.
2
.
Co
ntr
o
lled
ex
perim
ent
a
t
io
n f
ra
m
ew
o
rk
T
o
ev
alu
ate
an
d
o
p
tim
ize
b
o
th
co
n
ten
t
r
ec
o
m
m
en
d
a
tio
n
an
d
ad
s
y
s
tem
s
,
co
n
tr
o
lled
A/B
ex
p
er
im
en
tatio
n
f
r
am
ew
o
r
k
s
[
2
6
]
,
[
3
4
]
,
[
4
5
]
a
r
e
co
m
m
o
n
l
y
u
s
ed
.
I
n
th
is
s
etu
p
,
u
s
er
s
ar
e
r
an
d
o
m
ly
d
iv
i
d
ed
in
to
co
n
tr
o
l
an
d
tr
ea
tm
en
t
g
r
o
u
p
s
.
T
h
e
co
n
tr
o
l
g
r
o
u
p
co
n
tin
u
es
to
e
x
p
er
ien
ce
th
e
s
y
s
tem
as
u
s
u
al,
wh
ile
th
e
tr
ea
tm
en
t g
r
o
u
p
is
ex
p
o
s
ed
to
a
n
ew
f
ea
tu
r
e
o
r
ch
an
g
e
b
ein
g
test
ed
.
B
y
co
m
p
ar
in
g
th
e
b
eh
av
io
r
an
d
o
u
tco
m
es
o
f
th
ese
two
g
r
o
u
p
s
,
th
e
im
p
a
ct
o
f
t
h
e
n
ew
f
ea
tu
r
e
ca
n
b
e
a
cc
u
r
ately
ass
ess
ed
.
As
illu
s
tr
ated
in
Fig
u
r
e
5
,
th
e
A/B
test
in
g
wo
r
k
f
lo
w
i
n
v
o
l
v
es
s
p
litt
in
g
u
s
er
s
in
to
ex
p
er
i
m
en
tal
g
r
o
u
p
s
,
co
llectin
g
b
e
h
av
io
r
al
d
ata,
an
d
an
aly
zin
g
th
e
r
esu
lts
to
g
u
id
e
s
y
s
tem
im
p
r
o
v
em
en
ts
[
2
7
]
.
Fig
u
r
e
5
.
A/B
ex
p
er
im
e
n
tatio
n
wo
r
k
f
lo
w
Ho
wev
er
,
A/B
ex
p
er
im
en
ts
in
th
ese
s
y
s
tem
s
co
m
e
wi
th
u
n
iq
u
e
ch
allen
g
es.
Fo
r
u
s
er
ex
p
e
r
im
en
ts
in
co
n
ten
t
r
ec
o
m
m
en
d
atio
n
,
th
e
r
e
is
a
r
is
k
o
f
tr
a
f
f
ic
s
tealin
g
,
wh
er
e
t
h
e
tr
ea
tm
en
t
g
r
o
u
p
m
ay
in
ad
v
er
ten
tl
y
attr
ac
t
u
s
er
s
wh
o
wo
u
ld
o
th
er
wis
e
b
elo
n
g
to
th
e
co
n
tr
o
l
g
r
o
u
p
.
T
h
is
ca
n
s
k
ew
th
e
r
esu
lts
,
m
ak
in
g
it
d
if
f
icu
lt
to
m
ea
s
u
r
e
th
e
tr
u
e
ef
f
ec
t
o
f
th
e
ex
p
er
im
en
t.
Me
tr
ics
s
u
c
h
as
click
s
,
lik
es,
an
d
f
o
llo
ws
ar
e
p
ar
ticu
lar
ly
s
en
s
itiv
e
to
alg
o
r
ith
m
ic
u
p
d
a
tes
an
d
n
ee
d
to
b
e
cl
o
s
ely
m
o
n
ito
r
ed
d
u
r
in
g
th
ese
e
x
p
er
i
m
en
ts
to
ass
ess
th
e
im
m
ed
iate
im
p
ac
t o
n
u
s
er
en
g
ag
em
en
t
[
2
6
]
.
W
h
en
ex
p
e
r
im
en
tin
g
with
c
o
n
ten
t
cr
ea
to
r
s
o
r
ad
v
e
r
tis
er
s
,
it’s
ess
en
tial
to
s
ep
ar
ate
u
s
er
s
an
d
cr
ea
to
r
s
to
av
o
id
“g
h
o
s
t
ex
p
e
r
im
en
ta
tio
n
”
d
if
f
e
r
en
ce
s
.
T
h
is
m
ea
n
s
en
s
u
r
in
g
th
at
u
s
er
s
an
d
c
o
n
ten
t
cr
ea
t
o
r
s
(
o
r
ad
v
er
tis
er
s
)
d
o
n
o
t
c
r
o
s
s
-
co
n
t
am
in
ate
th
e
co
n
tr
o
l
a
n
d
tr
ea
t
m
en
t
g
r
o
u
p
s
.
Fo
r
in
s
tan
ce
,
if
an
ex
p
er
im
e
n
t
is
r
u
n
with
a
s
u
b
s
et
o
f
co
n
ten
t
cr
ea
to
r
s
,
th
e
au
d
ien
ce
f
o
r
t
h
ese
cr
ea
t
o
r
s
s
h
o
u
ld
b
e
e
v
en
ly
d
is
tr
ib
u
t
ed
b
etwe
en
c
o
n
tr
o
l
an
d
tr
ea
tm
en
t
g
r
o
u
p
s
to
en
s
u
r
e
ac
cu
r
ate
m
ea
s
u
r
em
e
n
t o
f
t
h
e
ex
p
er
im
en
t
’
s
im
p
ac
t.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
A
s
u
r
ve
y
o
f retr
ieva
l a
lg
o
r
ith
ms in
a
d
a
n
d
co
n
ten
t reco
mme
n
d
a
tio
n
s
ystems
(
Yu
Zh
a
o
)
1527
Fo
r
ad
ex
p
er
im
en
ts
,
it
is
also
cr
itical
to
en
s
u
r
e
th
at
b
o
th
th
e
tr
ea
tm
en
t
an
d
co
n
tr
o
l
g
r
o
u
p
s
h
av
e
th
e
s
am
e
b
u
d
g
et.
T
h
is
h
elp
s
to
m
ain
tain
a
f
air
co
m
p
ar
is
o
n
b
etwe
en
th
e
two
g
r
o
u
p
s
an
d
en
s
u
r
es
th
at
an
y
d
if
f
er
en
ce
s
in
p
er
f
o
r
m
an
ce
ca
n
b
e
attr
ib
u
te
d
to
ex
p
er
im
en
tal
ch
an
g
es r
ath
e
r
th
an
b
u
d
g
et
d
i
s
cr
ep
an
cies.
6.
CO
NCLU
SI
O
N
R
etr
iev
al
alg
o
r
ith
m
s
p
lay
a
ce
n
tr
al
r
o
le
in
m
o
d
er
n
ad
v
er
tis
in
g
an
d
c
o
n
ten
t
r
ec
o
m
m
en
d
atio
n
s
y
s
tem
s
b
y
e
n
ab
lin
g
ef
f
icien
t
m
atch
i
n
g
b
etwe
en
u
s
er
s
a
n
d
r
elev
an
t
item
s
.
T
h
is
s
u
r
v
ey
r
ev
ie
wed
r
ep
r
esen
tativ
e
r
etr
iev
al
ap
p
r
o
ac
h
es
u
s
ed
in
b
o
th
ad
v
er
tis
in
g
r
ec
o
m
m
en
d
atio
n
an
d
o
r
g
a
n
ic
co
n
ten
t
r
ec
o
m
m
en
d
atio
n
s
y
s
tem
s
,
in
clu
d
in
g
tr
a
d
itio
n
al
r
etr
iev
al
tech
n
iq
u
es
an
d
n
eu
r
al
r
etr
ie
v
al
ar
ch
itectu
r
es
s
u
ch
as
th
e
two
-
to
wer
m
o
d
el.
T
h
ese
m
eth
o
d
s
f
o
r
m
th
e
f
o
u
n
d
atio
n
o
f
lar
g
e
-
s
ca
le
r
ec
o
m
m
e
n
d
atio
n
p
ip
elin
es
b
y
e
n
ab
lin
g
ef
f
icien
t
ca
n
d
id
ate
g
en
er
atio
n
.
Desp
ite
s
ig
n
if
ican
t
p
r
o
g
r
ess
,
s
ev
er
al
ch
allen
g
es
r
em
ai
n
in
r
ea
l
-
wo
r
ld
s
y
s
tem
s
,
in
clu
d
in
g
d
ata
s
p
ar
s
ity
,
co
ld
s
tar
t
is
s
u
es,
d
at
a
q
u
ality
,
an
d
p
r
iv
ac
y
co
n
ce
r
n
s
.
Pra
ctica
l
s
o
lu
tio
n
s
s
u
ch
as
tr
an
s
f
er
lear
n
in
g
,
s
y
n
th
etic
d
ata
g
en
er
atio
n
,
a
n
d
f
ed
e
r
ated
lear
n
in
g
h
av
e
b
ee
n
ex
p
l
o
r
ed
to
m
itig
ate
th
ese
lim
itatio
n
s
.
I
n
ad
d
itio
n
,
i
n
f
r
astru
ctu
r
e
s
ca
lab
i
lity
is
a
k
ey
co
n
s
id
er
atio
n
in
i
n
d
u
s
tr
ial
d
ep
l
o
y
m
en
ts
,
wh
er
e
tech
n
iq
u
es
s
u
ch
as
ap
p
r
o
x
im
ate
n
ea
r
est
n
eig
h
b
o
r
s
ea
r
ch
,
d
is
tr
ib
u
ted
r
etr
ie
v
al
p
ip
elin
es,
an
d
h
ar
d
wa
r
e
ac
ce
ler
atio
n
(
e.
g
.
,
GPU/TPU
-
b
ased
s
y
s
tem
s
)
ar
e
wid
ely
ad
o
p
te
d
to
s
u
p
p
o
r
t e
f
f
i
cien
t la
r
g
e
-
s
ca
le
r
etr
iev
al.
L
o
o
k
in
g
f
o
r
war
d
,
r
etr
ie
v
al
al
g
o
r
ith
m
s
ar
e
ex
p
ec
ted
to
co
n
tin
u
e
e
v
o
lv
in
g
alo
n
g
s
id
e
ad
v
an
ce
s
in
m
ac
h
in
e
lear
n
in
g
an
d
lar
g
e
-
s
ca
le
co
m
p
u
tin
g
.
E
m
e
r
g
in
g
r
esear
ch
d
ir
ec
tio
n
s
in
clu
d
e
d
ee
p
er
i
n
teg
r
atio
n
b
etwe
en
r
etr
iev
al
ar
ch
itectu
r
e
s
an
d
LLM
s
,
as
well
a
s
im
p
r
o
v
ed
s
y
s
tem
d
esig
n
s
th
at
b
alan
ce
p
er
s
o
n
aliza
tio
n
,
s
ca
lab
ilit
y
,
an
d
p
r
iv
ac
y
.
O
v
er
all,
r
etr
iev
al
-
b
ased
ar
ch
itectu
r
e
is
ex
p
ec
ted
to
r
em
ain
a
f
u
n
d
am
en
tal
co
m
p
o
n
en
t
o
f
m
o
d
e
r
n
r
ec
o
m
m
en
d
atio
n
s
y
s
tem
s
.
R
ec
en
t
ad
v
an
ce
s
in
n
eu
r
al
r
an
k
in
g
an
d
d
en
s
e
r
etr
ie
v
al
f
u
r
th
er
h
ig
h
lig
h
t
th
e
r
ap
id
ev
o
lu
tio
n
o
f
r
etr
iev
a
l
ar
ch
itectu
r
es,
e
m
p
h
asizin
g
t
h
e
im
p
o
r
tan
ce
o
f
s
ca
lab
le
s
em
an
tic
m
atch
in
g
an
d
ef
f
icien
t
r
e
p
r
esen
tatio
n
lea
r
n
in
g
i
n
m
o
d
er
n
in
f
o
r
m
atio
n
s
y
s
tem
s
.
T
h
ese
tr
e
n
d
s
s
u
g
g
est
th
at
f
u
tu
r
e
r
ec
o
m
m
en
d
atio
n
s
y
s
tem
s
will
in
cr
ea
s
in
g
ly
r
ely
o
n
u
n
if
ie
d
r
etr
iev
al
a
r
ch
itectu
r
es
th
at
in
teg
r
ate
s
em
an
tic
u
n
d
er
s
tan
d
i
n
g
,
s
ca
lab
ilit
y
,
an
d
r
ea
l
-
tim
e
d
ec
is
io
n
-
m
a
k
in
g
.
T
h
is
wo
r
k
p
r
o
v
id
es
a
p
r
ac
tical
r
ef
er
en
ce
f
o
r
b
o
th
ac
ad
em
ic
r
esear
ch
a
n
d
in
d
u
s
tr
ial
d
ep
lo
y
m
e
n
t o
f
r
etr
iev
al
-
b
a
s
ed
r
ec
o
m
m
en
d
atio
n
s
y
s
tem
s
.
F
UNDING
I
NF
O
R
M
A
T
I
O
N
T
h
is
r
esear
ch
r
ec
eiv
ed
n
o
e
x
te
r
n
al
f
u
n
d
in
g
.
AUTHO
R
CO
NT
RI
B
UT
I
O
NS ST
A
T
E
M
E
N
T
T
h
is
jo
u
r
n
al
u
s
es
th
e
C
o
n
tr
ib
u
to
r
R
o
les
T
ax
o
n
o
m
y
(
C
R
ed
iT)
to
r
ec
o
g
n
ize
in
d
iv
id
u
al
au
th
o
r
co
n
tr
ib
u
tio
n
s
,
r
ed
u
ce
au
th
o
r
s
h
ip
d
is
p
u
tes,
an
d
f
ac
ilit
ate
co
llab
o
r
atio
n
.
Na
m
e
o
f
Aut
ho
r
C
M
So
Va
Fo
I
R
D
O
E
Vi
Su
P
Fu
Yu
Z
h
ao
✓
✓
Fan
g
L
iu
✓
✓
✓
Yu
an
Yu
an
✓
Yif
an
Dan
g
✓
C
:
C
o
n
c
e
p
t
u
a
l
i
z
a
t
i
o
n
M
:
M
e
t
h
o
d
o
l
o
g
y
So
:
So
f
t
w
a
r
e
Va
:
Va
l
i
d
a
t
i
o
n
Fo
:
Fo
r
mal
a
n
a
l
y
s
i
s
I
:
I
n
v
e
s
t
i
g
a
t
i
o
n
R
:
R
e
so
u
r
c
e
s
D
:
D
a
t
a
C
u
r
a
t
i
o
n
O
:
W
r
i
t
i
n
g
-
O
r
i
g
i
n
a
l
D
r
a
f
t
E
:
W
r
i
t
i
n
g
-
R
e
v
i
e
w
&
E
d
i
t
i
n
g
Vi
:
Vi
su
a
l
i
z
a
t
i
o
n
Su
:
Su
p
e
r
v
i
s
i
o
n
P
:
P
r
o
j
e
c
t
a
d
mi
n
i
st
r
a
t
i
o
n
Fu
:
Fu
n
d
i
n
g
a
c
q
u
i
si
t
i
o
n
CO
NF
L
I
C
T
O
F
I
N
T
E
R
E
S
T
ST
A
T
E
M
E
NT
Au
th
o
r
s
s
tate
n
o
co
n
f
lict o
f
in
t
er
est.
DATA AV
AI
L
AB
I
L
I
T
Y
Data
av
ailab
ilit
y
is
n
o
t
ap
p
li
ca
b
le
to
th
is
p
ap
er
as
n
o
n
e
w
d
ata
wer
e
cr
ea
ted
o
r
an
aly
ze
d
in
th
is
s
tu
d
y
.
Evaluation Warning : The document was created with Spire.PDF for Python.