I
AE
S In
t
er
na
t
io
na
l J
o
urna
l o
f
Art
if
icia
l In
t
ellig
ence
(
I
J
-
AI
)
Vo
l.
1
5
,
No
.
3
,
J
u
n
e
2
0
2
6
,
p
p
.
2761
~
2
7
7
3
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
15
.i
3
.
p
p
2
7
6
1
-
2
7
7
3
2761
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Effec
ts of spa
rse
da
tas
ets on time
i
nterva
l
-
a
wa
re sel
f
-
a
tt
e
ntion
sequentia
l r
ecom
menda
tion mo
del
s
Weish
a
n O
o
i,
L
ee
Yeng
O
ng
,
M
eng
-
Chew
L
eo
w
F
a
c
u
l
t
y
o
f
I
n
f
o
r
ma
t
i
o
n
S
c
i
e
n
c
e
a
n
d
T
e
c
h
n
o
l
o
g
y
,
M
u
l
t
i
me
d
i
a
U
n
i
v
e
r
s
i
t
y
,
M
e
l
a
k
a
,
M
a
l
a
y
si
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Dec
3
0
,
2
0
2
4
R
ev
is
ed
Ma
r
1
9
,
2
0
2
6
Acc
ep
ted
Ap
r
2
2
,
2
0
2
6
Re
c
o
m
m
e
n
d
a
ti
o
n
m
o
d
e
ls
se
r
v
e
a
s
c
ru
c
ial
fil
ters
in
m
a
n
a
g
in
g
in
fo
rm
a
ti
o
n
,
y
e
t
th
e
y
fa
c
e
a
fe
w
c
ru
c
ial
c
h
a
ll
e
n
g
e
s
,
su
c
h
a
s
c
a
p
tu
rin
g
u
se
r
-
it
e
m
in
tera
c
ti
o
n
b
e
h
a
v
io
rs
in
s
p
a
rse
d
a
tas
e
ts.
Da
ta
sp
a
rsity
re
fe
rs
to
a
n
issu
e
wh
e
re
th
e
re
is
a
lac
k
o
f
i
n
tera
c
ti
o
n
s
o
r
m
issin
g
v
a
lu
e
s
in
t
h
e
re
c
o
m
m
e
n
d
a
ti
o
n
d
a
tas
e
t.
A
sp
a
rse
d
a
tas
e
t
with
a
m
a
ss
iv
e
n
u
m
b
e
r
o
f
m
issin
g
v
a
lu
e
s
a
n
d
in
tera
c
ti
o
n
s
lea
d
s
to
m
o
re
d
y
n
a
m
ic
u
se
r
b
e
h
a
v
i
o
rs
,
w
h
ich
su
ffe
rs
a
p
o
o
r
re
c
o
m
m
e
n
d
a
ti
o
n
q
u
a
li
t
y
.
Th
e
se
lf
-
a
tt
e
n
ti
o
n
m
e
c
h
a
n
i
sm
fro
m
Tran
sfo
rm
e
r
c
a
n
a
ll
e
v
iate
t
h
e
e
ffe
c
ts
o
f
d
a
ta
s
p
a
rsity
i
n
d
a
tas
e
ts
b
y
a
ss
ig
n
in
g
we
ig
h
ts
t
o
i
tem
s
o
f
i
n
tera
c
ti
o
n
b
e
h
a
v
io
rs
.
Th
is
a
ll
o
ws
th
e
m
o
d
e
l
to
c
a
p
t
u
re
th
e
u
se
r
d
e
p
e
n
d
e
n
c
ies
in
c
o
m
p
l
e
x
u
se
r
b
e
h
a
v
i
o
r
,
w
h
ich
is
b
e
n
e
ficia
l
fo
r
sp
a
rse
d
a
tas
e
ts
with
p
a
tt
e
r
n
s
t
h
a
t
a
re
n
o
t
imm
e
d
iate
ly
a
p
p
a
r
e
n
t.
T
h
i
s
a
p
p
ro
a
c
h
h
a
s
s
h
o
w
n
i
ts
c
a
p
a
b
i
li
ty
t
o
h
a
n
d
le
larg
e
a
n
d
sp
a
rse
d
a
tas
e
ts,
a
s
se
e
n
in
ti
m
e
in
terv
a
l
-
a
wa
re
se
lf
-
a
tt
e
n
ti
o
n
se
q
u
e
n
ti
a
l
re
c
o
m
m
e
n
d
a
ti
o
n
m
o
d
e
l
(Ti
S
ASRec
).
It
u
ti
li
z
e
d
th
e
s
e
lf
-
a
tt
e
n
ti
o
n
m
e
c
h
a
n
ism
,
c
o
n
si
d
e
rin
g
t
h
e
ti
m
e
sta
m
p
a
n
d
a
b
s
o
lu
te
p
o
siti
o
n
s
o
f
i
tem
s
to
e
stim
a
te
t
h
e
h
ig
h
e
r
a
tt
e
n
ti
o
n
we
ig
h
ts
to
sh
o
w
th
e
imp
o
rtan
c
e
o
f
re
c
e
n
t
it
e
m
s.
Th
u
s,
t
h
is
stu
d
y
a
ims
to
in
v
e
stig
a
te
th
e
e
ffe
c
ts
o
f
sp
a
rse
d
a
tas
e
ts
b
y
c
o
m
p
a
ri
n
g
t
h
e
p
e
rf
o
r
m
a
n
c
e
o
f
Ti
S
ASRec
m
o
d
e
l
wit
h
se
lf
-
a
tt
e
n
ti
o
n
b
a
se
d
se
q
u
e
n
ti
a
l
re
c
o
m
m
e
n
d
a
ti
o
n
m
o
d
e
l
(S
ASRec
)
,
wh
ich
e
x
c
l
u
d
e
s
ti
m
e
in
terv
a
l
-
a
wa
re
n
e
ss
.
K
ey
w
o
r
d
s
:
SASR
ec
Seq
u
en
tial r
ec
o
m
m
e
n
d
atio
n
Sp
ar
s
e
d
ataset
T
ao
b
ao
T
iSAS
R
ec
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
:
L
ee
Yen
g
On
g
Facu
lty
o
f
I
n
f
o
r
m
atio
n
Scien
c
e
an
d
T
ec
h
n
o
lo
g
y
,
M
u
ltime
d
ia
Un
iv
er
s
ity
Ay
er
Ker
o
h
L
am
a
R
o
ad
,
7
5
4
5
0
Me
lak
a,
Ma
lay
s
ia
E
m
ail: ly
o
n
g
@
m
m
u
.
e
d
u
.
m
y
1.
I
NT
RO
D
UCT
I
O
N
T
h
e
p
r
o
d
u
ct
r
ec
o
m
m
en
d
atio
n
is
co
n
s
id
er
ed
a
well
-
k
n
o
wn
tech
n
iq
u
e
f
o
r
b
o
n
d
in
g
cu
s
to
m
er
s
an
d
p
r
o
d
u
cts
b
y
d
eliv
er
in
g
in
s
ig
h
t
f
u
l
s
u
g
g
esti
o
n
s
[
1
]
.
Mo
d
er
n
r
ec
o
m
m
en
d
atio
n
m
o
d
els
s
tan
d
as
o
n
e
o
f
th
e
m
o
s
t
o
r
d
in
ar
y
ap
p
licatio
n
s
in
th
e
b
ig
d
ata
d
o
m
ain
[
2
]
as
it
c
o
u
ld
p
e
r
f
o
r
m
p
r
e
d
ictio
n
s
o
n
u
s
er
s
’
p
r
ef
er
e
n
ce
s
b
y
r
etr
ie
v
in
g
v
ar
io
u
s
v
o
l
u
m
e
s
o
f
d
ata
f
o
r
p
u
r
ch
ase
h
is
to
r
y
,
click
t
h
r
o
u
g
h
r
ate
(
C
T
R
)
,
r
atin
g
s
,
o
r
o
n
lin
e
r
ev
iews
[
3
]
.
Dee
p
lea
r
n
in
g
h
as
b
ee
n
co
n
tr
ib
u
tin
g
to
r
ec
o
m
m
en
d
atio
n
m
o
d
els
as
it
h
a
s
th
e
ca
p
ab
ilit
y
of
h
an
d
lin
g
s
eq
u
en
tial
d
ata
s
u
c
h
as
u
s
er
’
s
b
r
o
wsi
n
g
o
r
p
u
r
ch
ase
h
is
to
r
y
,
w
h
ich
ca
n
ca
p
tu
r
e
th
e
tem
p
o
r
al
d
y
n
am
ics an
d
ev
o
lu
tio
n
o
f
u
s
e
r
s
’
p
r
ef
er
e
n
ce
s
an
d
in
te
r
ests
[
4
]
.
Alth
o
u
g
h
r
ec
en
t
r
ec
o
m
m
en
d
atio
n
m
o
d
els
ar
e
a
b
le
to
r
et
u
r
n
r
e
m
ar
k
ab
le
r
ec
o
m
m
e
n
d
at
io
n
s
,
th
er
e
ex
is
ts
a
cr
u
cial
ch
allen
g
e
in
ca
p
tu
r
in
g
u
s
er
-
item
in
ter
ac
t
io
n
b
e
h
av
io
r
s
i
n
s
p
ar
s
e
d
atasets
.
Data
s
p
ar
s
ity
co
m
m
o
n
l
y
h
a
p
p
en
s
in
lar
g
e
-
s
ca
le
d
ata,
wh
er
e
m
an
y
ex
p
e
cted
v
alu
es,
s
u
ch
as
r
atin
g
s
o
r
in
ter
ac
tio
n
s
,
ar
e
m
is
s
in
g
o
r
u
n
av
ailab
le
in
th
e
d
ataset.
T
h
is
is
p
ar
ticu
lar
ly
p
r
o
b
lem
atic
f
o
r
co
llab
o
r
ativ
e
f
il
ter
in
g
alg
o
r
ith
m
s
,
wh
ich
r
ely
o
n
t
h
ese
r
atin
g
s
to
u
n
co
v
e
r
b
en
ef
icial
r
elatio
n
s
h
ip
s
am
o
n
g
u
s
er
s
an
d
item
s
[
5
]
a
n
d
th
is
is
also
f
u
r
th
er
m
en
tio
n
ed
in
th
e
r
e
ce
n
t
s
tu
d
ies
[
6
]
wh
er
e
c
o
llab
o
r
ativ
e
f
ilter
in
g
alg
o
r
ith
m
s
ass
u
m
e
all
p
ast
in
ter
ac
tio
n
s
ar
e
eq
u
ally
im
p
o
r
tan
t
an
d
th
is
ass
u
m
p
tio
n
o
f
ten
f
ails
in
r
ea
l
-
wo
r
ld
s
ce
n
ar
io
s
,
as
th
e
u
s
er
in
ten
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
3
,
J
u
n
e
2
0
2
6
:
2
7
6
1
-
2
7
7
3
2762
ca
n
s
h
if
t
a
n
d
s
h
o
r
t
-
ter
m
b
eh
a
v
io
r
s
m
atter
m
o
r
e
th
an
s
tatic
lo
n
g
-
ter
m
r
atin
g
s
.
Fu
r
th
er
m
o
r
e,
a
s
p
ar
s
e
d
ataset
with
a
m
ass
iv
e
n
u
m
b
e
r
o
f
m
i
s
s
in
g
v
alu
es
an
d
in
te
r
ac
tio
n
s
l
ea
d
s
to
m
o
r
e
d
y
n
am
ic
u
s
er
b
e
h
av
io
r
.
Hen
ce
,
th
e
r
ec
o
m
m
en
d
atio
n
m
o
d
els
ca
n
n
o
t
id
e
n
tify
t
h
ese
r
elatio
n
s
h
i
p
s
wh
en
d
atasets
ar
e
s
p
ar
s
e,
r
esu
ltin
g
i
n
lo
wer
ac
cu
r
ac
y
an
d
s
u
b
o
p
tim
al
r
ec
o
m
m
en
d
atio
n
lis
ts
[
7
]
.
Nex
t,
th
e
co
llab
o
r
ativ
e
r
ec
o
m
m
en
d
atio
n
m
o
d
el
o
u
g
h
t
to
ca
p
tu
r
e
th
e
u
s
er
b
eh
a
v
io
r
s
f
r
o
m
th
e
u
s
er
-
item
in
ter
ac
tio
n
s
to
p
er
f
o
r
m
r
ec
o
m
m
e
n
d
atio
n
s
.
I
n
u
s
er
b
eh
a
v
io
r
,
it
ca
n
b
e
id
en
tifie
d
as
lo
n
g
-
ter
m
an
d
s
h
o
r
t
-
ter
m
u
s
er
b
e
h
av
io
r
[
8
]
.
I
n
lo
n
g
-
ter
m
u
s
er
b
eh
av
i
o
r
,
it
r
ef
er
s
to
a
u
s
er
’
s
co
n
s
is
ten
t
in
ter
est
o
v
er
an
ex
ten
d
ed
p
er
i
o
d
a
n
d
r
ef
lects
as
th
e
g
en
er
al
p
r
e
f
er
en
ce
s
.
Fo
r
s
h
o
r
t
-
ter
m
b
e
h
av
io
r
,
it
r
ef
lects
th
e
u
s
er
’
s
im
m
ed
iate
o
r
r
ec
en
t
i
n
ter
ests
,
wh
ich
wer
e
in
f
lu
en
ce
d
b
y
te
m
p
o
r
al
n
ee
d
s
o
r
s
p
ec
if
ic
cir
c
u
m
s
tan
ce
s
at
a
g
iv
en
tim
e.
Ho
wev
er
,
th
e
u
s
er
b
e
h
a
v
io
r
s
ar
e
d
y
n
a
m
ic
,
an
d
it
te
n
d
to
ch
a
n
g
e
o
v
er
tim
e
[
9
]
.
T
h
is
ev
en
tu
ally
r
ed
u
ce
s
th
e
ca
p
ab
ilit
y
o
f
co
llab
o
r
ativ
e
r
ec
o
m
m
en
d
atio
n
m
o
d
els
to
ca
p
tu
r
e
th
e
lo
n
g
-
ter
m
d
ep
en
d
en
c
ies
f
r
o
m
u
s
er
-
ite
m
in
ter
ac
tio
n
s
f
o
r
m
o
d
ellin
g
o
u
t
th
e
u
s
er
b
eh
a
v
io
r
s
.
T
h
er
ef
o
r
e,
th
e
s
eq
u
en
tial
r
ec
o
m
m
en
d
atio
n
(
SR
)
m
o
d
el
is
em
p
lo
y
e
d
f
o
r
ca
p
tu
r
in
g
th
e
lo
n
g
-
te
r
m
d
ep
en
d
e
n
cies
to
p
r
o
v
i
d
e
in
s
i
g
h
ts
f
o
r
th
e
u
s
er
b
e
h
av
io
r
s
in
u
s
er
-
item
i
n
ter
ac
tio
n
s
.
T
h
is
o
f
ten
o
cc
u
r
s
in
e
-
co
m
m
er
ce
c
o
n
tex
t
wh
er
e
t
h
e
u
s
er
b
eh
av
i
o
r
in
u
s
er
-
ite
m
in
ter
ac
tio
n
s
u
s
u
ally
h
ap
p
e
n
s
s
u
cc
ess
iv
ely
in
a
s
eq
u
en
ce
,
r
at
h
er
t
h
an
in
an
is
o
lated
m
an
n
e
r
[
1
0
]
.
SR
is
co
m
m
o
n
ly
r
ef
er
r
e
d
to
s
eq
u
en
ce
-
a
war
e
co
m
m
e
n
d
atio
n
wh
er
e
it
r
ec
eiv
es
s
eq
u
en
tially
o
r
d
er
e
d
u
s
er
in
ter
ac
ti
o
n
s
as
in
p
u
t
co
m
p
ar
ed
t
o
a
tr
ad
itio
n
al
u
s
er
-
item
r
atin
g
m
atr
ix
.
I
t
also
ca
p
tu
r
es
c
o
m
p
lex
u
s
er
p
r
ef
er
en
ce
s
f
r
o
m
u
s
e
r
s
’
ch
r
o
n
o
lo
g
ical
b
eh
av
io
r
a
n
d
p
r
ed
icts
th
e
n
ex
t
m
o
s
t
lik
ely
item
f
o
r
r
ec
o
m
m
en
d
atio
n
[
1
1
]
.
I
n
SR
,
th
e
ite
m
s
in
th
e
u
s
er
-
item
in
ter
ac
ti
o
n
s
ar
e
co
n
s
id
er
e
d
co
r
r
elate
d
to
ea
c
h
o
th
e
r
.
T
h
is
i
s
b
ec
au
s
e
th
e
ch
o
ice
o
f
th
e
n
e
x
t item
d
ep
en
d
s
o
n
th
e
p
r
ev
i
o
u
s
item
s
,
as it in
f
er
s
th
e
u
s
er
s
’
s
elec
tio
n
b
ased
o
n
th
eir
p
r
io
r
s
elec
tio
n
s
.
T
h
e
f
o
ll
o
win
g
item
o
f
a
s
eq
u
en
ce
is
p
r
ed
icted
b
ased
o
n
th
e
h
is
to
r
ical
elem
en
ts
f
r
o
m
th
e
u
s
er
-
item
in
ter
ac
tio
n
s
eq
u
en
ce
[
1
2
]
.
I
n
o
r
d
er
to
ca
p
tu
r
e
th
e
u
s
er
b
eh
av
io
r
f
r
o
m
u
s
er
-
item
in
ter
ac
tio
n
s
e
q
u
en
ce
s
,
v
a
r
io
u
s
SR
ap
p
r
o
ac
h
es
in
clu
d
e
r
ec
u
r
r
e
n
t
n
e
u
r
al
n
etwo
r
k
s
(
R
NN)
,
co
n
v
o
l
u
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NN)
,
s
elf
-
atten
tio
n
,
m
eta
tr
an
s
itio
n
al
lear
n
in
g
(
Me
taT
L
)
,
tem
p
o
r
al
g
r
ap
h
tr
an
s
f
o
r
m
er
(
T
GT
)
,
a
n
d
g
r
ap
h
n
eu
r
al
n
etwo
r
k
s
m
o
d
els
ha
v
e
b
ee
n
in
tr
o
d
u
ce
d
[
1
2
]
–
[
1
8
]
.
I
n
SR
,
tr
ad
itio
n
al
m
o
d
els
lik
e
R
NN
s
h
av
e
an
is
s
u
e
o
f
v
a
n
is
h
in
g
g
r
ad
ien
t
wh
ich
lead
s
to
lo
wer
p
er
f
o
r
m
an
ce
an
d
r
elativ
ely
lo
n
g
e
r
tr
ai
n
in
g
tim
e
b
ec
a
u
s
e
o
f
its
n
atu
r
e
wh
ile
c
ap
tu
r
in
g
lo
n
g
-
ter
m
d
ep
en
d
e
n
cies.
T
h
e
em
p
lo
y
m
e
n
t
o
f
s
elf
-
atten
tio
n
m
ec
h
an
is
m
h
elp
s
in
m
o
d
elin
g
lo
n
g
-
ter
m
d
ep
en
d
en
cies
b
y
ca
p
tu
r
in
g
co
n
n
ec
tio
n
s
b
etwe
e
n
an
y
p
air
s
o
f
item
s
r
e
g
ar
d
les
s
o
f
th
eir
d
is
tan
ce
in
t
h
e
u
s
er
-
item
s
eq
u
en
ce
.
I
n
ad
d
itio
n
,
it
also
ass
ig
n
s
atten
tio
n
weig
h
ts
to
m
ea
s
u
r
e
th
e
im
p
o
r
tan
ce
o
f
d
if
f
er
en
t
item
s
to
h
ig
h
lig
h
t
th
eir
r
elev
an
cy
an
d
ca
r
r
y
o
u
t
r
ec
o
m
m
en
d
atio
n
s
.
Fro
m
th
e
p
r
e
v
io
u
s
wo
r
k
s
[
1
2
]
–
[
1
8
]
,
r
ec
o
m
m
en
d
atio
n
s
ar
e
m
o
s
tly
ca
r
r
ied
o
u
t
b
y
en
co
d
in
g
u
s
er
s
’
p
ast
item
in
ter
ac
tio
n
s
in
to
a
v
ec
to
r
with
a
lef
t
-
to
-
r
ig
h
t
s
eq
u
en
tial
m
o
d
el
b
ased
o
n
th
is
h
id
d
en
r
e
p
r
esen
tatio
n
.
I
n
s
o
m
e
wo
r
k
s
,
it
em
p
lo
y
s
an
atten
tio
n
m
ec
h
an
is
m
to
p
r
io
r
itize
in
f
lu
en
tial
u
s
er
-
item
in
ter
ac
tio
n
s
b
y
wei
g
h
in
g
d
if
f
e
r
en
t
item
s
.
B
y
r
ef
e
r
r
in
g
to
ea
ch
u
s
er
-
item
in
ter
a
ctio
n
s
eq
u
en
ce
,
th
e
atten
tio
n
m
ec
h
a
n
is
m
is
ab
le
to
ca
p
tu
r
e
th
e
u
s
er
b
eh
av
i
o
r
f
o
r
th
e
tr
ain
in
g
m
o
d
el.
Alth
o
u
g
h
s
o
m
e
ap
p
r
o
ac
h
es
in
clu
d
e
an
atten
tio
n
m
ec
h
an
is
m
,
th
ey
ar
e
n
o
t
tim
e
-
in
ter
v
al
a
war
e.
I
n
SR
,
as
th
e
in
ter
ac
ted
item
s
ar
e
ar
r
an
g
ed
in
a
s
eq
u
en
ce
,
th
e
tim
e
in
ter
v
als
b
etwe
en
in
ter
ac
tio
n
s
ar
e
ig
n
o
r
ed
,
wh
ic
h
is
cr
u
cial
f
o
r
ca
p
tu
r
in
g
th
e
u
s
er
p
r
ef
er
en
ce
s
.
SR
m
o
d
els
th
at
lack
c
o
n
s
id
er
atio
n
o
f
tim
e
in
ter
v
als
h
av
e
less
in
s
ig
h
t
a
b
o
u
t
u
s
er
b
eh
av
io
r
,
lead
in
g
to
lo
wer
p
er
f
o
r
m
a
n
ce
o
f
r
ec
o
m
m
en
d
atio
n
s
,
p
ar
ticu
la
r
ly
in
a
s
p
ar
s
e
d
ataset.
I
n
th
is
s
tu
d
y
,
th
e
im
p
lem
en
ta
tio
n
o
f
tim
e
in
ter
v
al
-
awa
r
e
s
e
lf
-
atten
tio
n
s
eq
u
en
tial
r
ec
o
m
m
en
d
atio
n
m
o
d
el
(
T
iSAS
R
ec
)
aim
s
to
c
ap
tu
r
e
th
e
tem
p
o
r
al
d
e
p
en
d
e
n
cies
in
u
s
er
-
item
i
n
ter
ac
tio
n
s
f
o
r
d
atasets
with
d
if
f
er
en
t le
v
els o
f
s
p
ar
s
ity
.
T
iSAS
R
ec
em
p
lo
y
s
tim
e
in
ter
v
al
-
awa
r
e
b
y
co
n
s
id
er
in
g
th
e
tim
e
in
ter
v
als b
etwe
en
ea
ch
item
t
h
at
ar
e
n
o
t
eq
u
ally
s
p
ac
ed
in
tim
e.
T
h
is
in
d
ica
tes
th
at
th
e
r
ec
o
m
m
e
n
d
atio
n
p
er
f
o
r
m
an
ce
v
ar
ies
b
ased
o
n
th
e
v
alu
es
o
f
tim
e
g
ap
s
.
T
h
is
d
if
f
er
s
f
r
o
m
o
th
e
r
s
elf
-
atten
tio
n
-
b
ased
SR
m
o
d
els
,
s
u
ch
as
SASR
ec
,
wh
ich
ex
clu
d
es
tim
e
in
ter
v
al
-
awa
r
en
ess
.
B
y
u
tili
zin
g
th
e
ti
m
e
in
ter
v
als
b
etwe
en
ea
ch
ite
m
,
th
e
s
elf
-
atten
tio
n
m
ec
h
an
is
m
in
T
iSAS
R
ec
in
co
r
p
o
r
ates
th
e
ab
s
o
lu
te
item
p
o
s
itio
n
s
.
T
h
is
allo
ws
T
iSAS
R
ec
to
ass
ig
n
d
if
f
e
r
en
t
atten
tio
n
weig
h
ts
b
ased
o
n
ite
m
s
’
ab
s
o
lu
te
p
o
s
itio
n
an
d
r
el
ativ
e
tim
e
in
ter
v
als,
ef
f
ec
tiv
el
y
m
ap
p
in
g
o
u
t
th
e
im
p
o
r
tan
ce
o
f
d
if
f
er
e
n
t
item
s
in
v
ar
y
i
n
g
tim
estam
p
s
an
d
ca
p
tu
r
in
g
th
e
tem
p
o
r
al
u
s
er
p
r
ef
er
en
ce
s
in
d
atasets
.
I
n
th
is
wo
r
k
,
t
h
e
p
er
f
o
r
m
a
n
ce
co
m
p
ar
is
o
n
o
f
T
iSAS
R
ec
an
d
SASR
ec
is
in
v
esti
g
ated
to
s
h
o
wca
s
e
th
e
ef
f
ec
ts
o
f
d
i
f
f
er
en
t
p
ar
a
m
eter
v
alu
es
o
f
r
ec
o
m
m
en
d
atio
n
m
o
d
els
o
n
d
atasets
with
v
ar
y
in
g
lev
els
o
f
s
p
ar
s
ity
,
wh
ic
h
is
n
o
t
ex
p
licitly
d
is
cu
s
s
ed
in
th
e
ex
is
tin
g
wo
r
k
s
.
T
h
er
e
a
r
e
two
m
ain
co
n
tr
ib
u
tio
n
s
in
th
is
s
tu
d
y
.
Firstl
y
,
th
e
ap
p
licatio
n
o
f
T
iSAS
R
ec
an
d
SAS
R
ec
m
o
d
els
ac
r
o
s
s
d
atasets
w
ith
d
if
f
er
en
t
lev
els
o
f
s
p
ar
s
ity
,
wh
ich
ar
e
Mo
v
ieL
en
s
-
1
M
a
n
d
T
a
o
b
ao
d
atasets
.
Seco
n
d
ly
,
a
co
m
p
a
r
ativ
e
an
aly
s
is
o
f
th
e
p
er
f
o
r
m
an
ce
o
f
d
if
f
er
en
t
p
ar
am
eter
v
al
u
es in
b
o
th
T
iSAS
R
ec
an
d
SASR
ec
m
o
d
els ac
r
o
s
s
two
d
atasets
.
T
h
e
s
tr
u
ctu
r
e
o
f
th
is
p
ap
er
i
s
as
f
o
llo
ws
.
Sectio
n
2
d
is
cu
s
s
es
th
e
d
ev
elo
p
m
en
t
o
f
SR
an
d
th
e
ap
p
licatio
n
s
o
f
co
n
tr
asti
v
e
lear
n
in
g
,
wh
ile
s
ec
t
io
n
3
ex
p
lain
s
th
e
m
eth
o
d
o
lo
g
y
u
s
ed
in
th
is
s
tu
d
y
.
I
n
s
ec
tio
n
4
,
th
e
ex
p
er
im
en
tal
r
esu
lts
f
r
o
m
b
o
th
m
o
d
els
a
r
e
c
o
m
p
ar
e
d
ac
r
o
s
s
two
d
atasets
with
d
if
f
er
en
t
lev
els
o
f
s
p
ar
s
ity
.
L
astl
y
,
th
e
ef
f
ec
t o
f
s
p
ar
s
e
d
at
asets
o
n
tim
e
in
ter
v
al
-
awa
r
e
s
elf
-
atten
tio
n
SR
m
o
d
els
is
co
n
clu
d
ed
in
s
ec
tio
n
5
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
E
ffects o
f sp
a
r
s
e
d
a
ta
s
ets o
n
time
in
terva
l
-
a
w
a
r
e
s
elf
-
a
tten
ti
o
n
s
eq
u
en
tia
l
…
(
Weis
h
a
n
Oo
i
)
2763
2.
L
I
T
E
R
AT
U
RE
R
E
VI
E
W
T
h
e
is
s
u
e
o
f
d
ata
s
p
a
r
s
ity
in
r
ec
o
m
m
en
d
atio
n
s
y
s
tem
s
o
cc
u
r
s
wh
en
th
e
r
e
is
in
s
u
f
f
icien
t
i
n
ter
ac
tio
n
d
ata
av
ailab
le
to
ac
cu
r
ately
d
e
ter
m
in
e
u
s
er
s
im
ilar
ities
an
d
p
r
ef
er
en
ce
s
.
Hig
h
s
p
ar
s
ity
o
f
th
e
d
ataset
m
ak
es
it
d
if
f
icu
lt
to
p
r
ed
ict
u
s
er
p
r
ef
er
en
ce
s
d
u
e
to
th
e
lack
o
f
r
eliab
le
p
atter
n
s
f
r
o
m
d
y
n
a
m
ic
u
s
er
b
eh
av
i
o
r
.
C
o
n
s
eq
u
en
tly
,
th
is
r
esu
lts
in
p
o
o
r
p
er
f
o
r
m
an
ce
o
f
r
ec
o
m
m
en
d
atio
n
ac
cu
r
ac
y
an
d
ef
f
icie
n
cy
.
Fo
r
in
s
tan
ce
,
Mo
v
ieL
en
s
-
1
M
an
d
Netf
lix
h
av
e
s
p
ar
s
ity
r
ates
o
f
9
5
.
1
6
%
an
d
9
8
.
8
2
%
,
r
esp
ec
tiv
ely
[
1
9
]
,
wh
ile
th
e
e
-
co
m
m
er
ce
d
ataset
n
am
ely
T
ao
b
ao
h
as
a
s
p
ar
s
ity
r
ate
o
f
9
9
.
8
5
%,
m
ak
in
g
it
ch
al
len
g
in
g
to
r
etr
iev
e
s
eq
u
en
tial
p
atter
n
s
lik
e
lo
n
g
-
ter
m
d
ep
en
d
en
cies
f
r
o
m
th
e
d
ataset
[
2
0
]
.
Alth
o
u
g
h
in
cr
e
asin
g
th
e
s
ca
le
o
f
d
atasets
an
d
th
eir
s
p
ar
s
ity
lev
el
en
h
an
ce
s
d
ata
av
ailab
ilit
y
an
d
p
r
o
v
id
es
m
o
r
e
d
ata
p
o
in
ts
f
o
r
id
en
tify
i
n
g
u
s
er
p
r
ef
er
en
ce
s
,
s
p
a
r
s
e
d
atasets
al
s
o
r
eq
u
ir
e
m
o
r
e
c
o
m
p
u
tatio
n
al
r
eso
u
r
ce
s
to
p
e
r
f
o
r
m
r
ec
o
m
m
en
d
atio
n
s
.
SR
p
er
f
o
r
m
s
r
ec
o
m
m
en
d
atio
n
s
b
ased
o
n
s
eq
u
en
tial
o
r
d
e
r
o
f
u
s
er
in
ter
ac
tio
n
s
,
h
av
e
em
p
lo
y
ed
v
ar
io
u
s
ap
p
r
o
ac
h
es,
in
clu
d
in
g
R
NN,
C
NN,
Me
taT
L
,
T
G
T
,
an
d
g
r
ap
h
n
eu
r
al
n
etwo
r
k
s
.
T
h
e
GR
U4
R
ec
+
m
o
d
el
ap
p
lied
with
R
NN
[
1
4
]
an
d
th
e
C
aser
m
o
d
el
u
s
ed
C
NN
[
1
5
]
ac
h
iev
ed
th
e
NDCG
@
1
0
v
alu
es
o
f
0
.
5
5
1
3
an
d
0
.
5
5
3
8
,
r
esp
ec
ti
v
ely
,
o
n
Mo
v
ieL
en
s
-
1
M
d
a
taset,
in
d
icatin
g
th
eir
ca
p
ab
ilit
ies
in
r
an
k
in
g
r
ec
o
m
m
en
d
atio
n
item
s
.
B
o
th
o
f
th
em
r
ef
er
r
e
d
to
a
n
ea
r
ly
at
tem
p
t
in
SR
s
tu
d
ies
an
d
u
s
ed
i
t
f
o
r
co
m
p
ar
is
o
n
i
n
th
e
f
u
r
th
er
s
tu
d
y
[
2
1
]
.
Ho
we
v
er
,
b
o
th
m
o
d
els
s
tr
u
g
g
le
to
ca
p
tu
r
e
t
h
e
c
o
r
e
u
s
er
p
r
ef
er
e
n
ce
s
f
r
o
m
lo
n
g
-
ter
m
d
ep
en
d
e
n
cies
in
s
p
ar
s
e
d
atas
et,
wh
ich
p
r
o
v
e
n
th
at
d
ata
s
p
ar
s
ity
ca
n
af
f
ec
t
r
ec
o
m
m
en
d
atio
n
q
u
ality
.
T
h
e
ch
allen
g
es
o
f
co
ld
-
s
tar
t
u
s
er
s
with
m
in
im
al
in
ter
ac
tio
n
s
in
SR
[
1
8
]
is
ex
p
er
im
en
ted
o
n
th
r
ee
d
atasets
f
r
o
m
Am
az
o
n
E
lectr
o
n
ics,
Mo
v
ie,
an
d
B
o
o
k
.
T
h
e
r
esu
lts
s
h
o
w
th
e
HR
@
1
v
alu
es
o
f
0
.
2
2
4
,
0
.
2
5
8
,
a
n
d
0
.
4
2
0
f
o
r
ea
ch
d
ataset,
r
esp
ec
tiv
el
y
.
T
o
im
p
r
o
v
e
o
v
e
r
co
m
e
th
is
ch
alle
n
g
e,
SR
m
o
d
el
m
a
y
u
tili
ze
a
d
d
itio
n
al
u
s
er
o
r
item
in
f
o
r
m
atio
n
f
o
r
c
o
ld
-
s
tar
t
u
s
er
s
.
An
o
th
er
m
o
d
el
n
am
ed
s
eq
u
en
tial
r
ec
o
m
m
en
d
atio
n
with
g
r
ap
h
n
e
u
r
al
n
etwo
r
k
s
(
SUR
GE
)
[
2
2
]
in
te
g
r
ates
d
if
f
er
en
t
t
y
p
es
o
f
u
s
er
p
r
ef
er
e
n
ce
s
in
to
clu
s
ter
s
,
f
o
r
m
in
g
d
en
s
e
in
ter
es
t
g
r
ap
h
s
.
T
h
is
h
elp
s
d
is
tin
g
u
is
h
co
r
e
in
ter
ests
f
r
o
m
n
o
is
y
b
e
h
av
io
r
s
.
I
t
ac
h
iev
e
d
NDCG@
2
v
alu
es
o
f
0
.
3
6
2
5
an
d
0
.
9
4
9
5
f
o
r
th
e
T
a
o
b
ao
an
d
Ku
aish
o
u
d
atasets
,
r
esp
ec
tiv
ely
.
Ho
wev
er
,
its
n
ar
r
o
w
f
o
c
u
s
o
n
NDCG
@
2
to
s
h
o
w
th
e
to
p
two
r
ec
o
m
m
en
d
atio
n
s
m
ay
n
o
t
f
u
lly
r
ef
lect
its
p
er
f
o
r
m
an
ce
in
r
an
k
in
g
a
b
r
o
ad
er
s
et
o
f
r
ele
v
an
t
item
s
,
wh
er
ea
s
NDCG
@
1
0
wo
u
ld
p
r
o
v
id
e
a
m
o
r
e
co
m
p
r
eh
e
n
s
iv
e
ev
alu
atio
n
o
f
th
e
r
ec
o
m
m
en
d
atio
n
q
u
ality
.
On
th
e
o
th
er
h
an
d
,
th
e
s
elf
-
at
ten
tio
n
m
ec
h
a
n
is
m
in
s
elf
-
atten
tio
n
b
ased
s
eq
u
e
n
tial
r
ec
o
m
m
en
d
atio
n
m
o
d
el
(
SASR
ec
)
ca
n
ca
p
tu
r
e
th
e
e
n
tire
u
s
er
s
eq
u
e
n
ce
wi
th
o
u
t
r
ec
u
r
r
e
n
t
o
r
c
o
n
v
o
lu
tio
n
al
o
p
e
r
atio
n
s
a
n
d
ad
ap
tiv
ely
ass
ig
n
weig
h
ts
to
p
r
ev
io
u
s
item
s
f
o
r
p
r
ed
ictio
n
s
.
T
h
e
s
elf
-
atten
tio
n
c
o
m
p
o
n
en
t
tak
es
u
s
er
-
item
in
ter
ac
tio
n
s
eq
u
en
ce
s
as
in
p
u
t
,
with
ea
ch
item
r
ep
r
esen
ted
b
y
an
em
b
ed
d
in
g
.
Po
s
itio
n
al
en
co
d
in
g
is
ad
d
ed
to
ca
p
tu
r
e
t
h
e
o
r
d
er
o
f
in
ter
ac
tio
n
s
.
T
h
e
s
elf
-
atten
tio
n
co
m
p
o
n
en
t
co
m
p
u
tes
th
e
r
elatio
n
s
h
ip
b
etwe
en
all
p
air
s
o
f
item
s
in
th
e
s
eq
u
en
ce
s
with
t
h
e
q
u
e
r
y
(
Q)
,
k
e
y
(
K)
,
an
d
v
alu
e
(
V)
v
ec
t
o
r
s
f
o
r
ea
ch
ite
m
.
T
h
ese
v
ec
to
r
s
ar
e
th
en
u
s
ed
to
ca
lcu
late
th
e
atte
n
tio
n
s
co
r
es,
wh
ich
ar
e
ass
ig
n
ed
to
ea
ch
item
to
weig
h
its
co
n
tr
ib
u
tio
n
in
th
e
s
eq
u
en
ce
s
f
o
r
n
ex
t
item
p
r
ed
i
ctio
n
.
T
h
e
SASR
ec
h
as
s
h
o
w
n
its
ca
p
a
b
ilit
y
to
ca
p
tu
r
e
lo
n
g
-
ter
m
d
ep
e
n
d
en
cies
with
h
ig
h
ef
f
icie
n
cy
a
n
d
s
ca
lab
ilit
y
o
n
s
p
ar
s
e
d
atasets
,
ac
h
iev
in
g
an
NDCG@
1
0
v
alu
e
o
f
0
.
5
9
0
5
o
n
th
e
Mo
v
ieL
en
s
-
1
M
d
ataset
[
1
6
]
.
T
h
e
SASR
ec
m
o
d
el
is
o
f
ten
r
ef
er
r
ed
as
th
e
b
aselin
e
m
o
d
el
in
atten
tio
n
-
b
ased
SR
r
esear
ch
.
T
h
is
is
b
ec
a
u
s
e
i
t
was
o
n
e
o
f
th
e
p
io
n
ee
r
s
elf
-
atten
tio
n
ar
c
h
itectu
r
es
f
o
r
SR
an
d
it
is
also
u
s
ed
f
o
r
co
m
p
ar
is
o
n
i
n
r
ec
en
t stu
d
i
es a
s
well
[
2
3
]
.
T
h
e
T
GT
f
r
am
ewo
r
k
[
1
7
]
wh
i
ch
aim
s
to
a
d
d
r
ess
th
e
ch
allen
g
e
o
f
m
o
d
ellin
g
th
e
tim
e
-
ev
o
lv
in
g
u
s
er
p
r
ef
er
en
ce
s
,
p
ar
ticu
lar
ly
o
n
m
u
lti
-
b
eh
av
io
r
in
ter
ac
tio
n
s
in
S
R
s
y
s
tem
s
.
T
h
e
T
GT
m
o
d
el
ac
h
iev
ed
NDCG@
1
0
v
alu
es
o
f
0
.
2
6
3
an
d
0
.
3
3
0
o
n
T
ao
b
ao
-
d
ata
an
d
t
h
e
I
J
C
AI
C
o
n
test
d
ataset,
r
esp
ec
tiv
ely
,
an
d
HR
@
1
0
v
alu
es
o
f
0
.
4
5
2
an
d
0
.
5
1
9
o
n
th
e
s
am
e
d
atasets
.
T
h
e
T
GT
m
o
d
el
h
as
d
em
o
n
s
tr
ated
its
ab
ilit
y
to
d
y
n
am
ically
f
u
s
e
an
d
ex
tr
ac
t
u
s
er
s
ac
tiv
ated
c
o
r
e
in
ter
ests
f
r
o
m
n
o
is
y
u
s
er
b
e
h
av
io
r
s
eq
u
en
ce
s
b
y
ap
p
ly
in
g
th
e
s
elf
-
atten
tio
n
m
ec
h
an
is
m
.
Ho
wev
er
,
m
o
s
t
o
f
th
e
SR
m
o
d
els
o
n
ly
r
etr
ie
v
e
th
e
o
r
d
er
s
o
f
u
s
er
-
item
in
ter
ac
tio
n
s
,
wh
ich
is
in
s
u
f
f
icien
t
f
o
r
ca
p
tu
r
i
n
g
f
u
lly
d
y
n
a
m
ic
u
s
er
b
e
h
av
io
r
.
T
h
is
in
co
m
p
lete
u
n
d
er
s
tan
d
in
g
o
f
th
e
c
o
r
e
u
s
er
p
r
ef
er
en
ce
s
lead
s
to
p
o
o
r
r
ec
o
m
m
en
d
atio
n
q
u
ality
.
C
o
m
p
ar
ed
to
SASR
ec
,
T
iSAS
R
ec
[
2
3
]
i
n
tr
o
d
u
ce
s
tim
e
in
ter
v
al
em
b
ed
d
in
g
s
to
ca
p
tu
r
e
th
e
tem
p
o
r
al
g
ap
s
b
etwe
en
ea
c
h
u
s
er
-
item
in
ter
ac
tio
n
in
th
e
s
eq
u
e
n
ce
s
,
wh
ich
ar
e
th
en
u
s
ed
in
t
h
e
co
m
p
u
tatio
n
s
o
f
atten
tio
n
s
co
r
es,
p
r
o
v
i
d
in
g
a
m
o
r
e
n
u
an
ce
d
u
n
d
er
s
tan
d
i
n
g
o
f
u
s
er
b
eh
a
v
io
r
.
T
h
is
en
h
an
c
es
th
e
d
is
tr
ib
u
tio
n
o
f
atten
tio
n
s
co
r
es
f
r
o
m
th
e
q
u
e
r
y
,
k
ey
,
an
d
v
alu
e
(
QKV
)
co
m
p
o
n
e
n
ts
in
th
e
s
elf
-
atten
tio
n
m
ec
h
an
is
m
s
.
T
h
e
co
m
p
ar
ativ
e
s
tu
d
ies
b
etwe
en
SASR
ec
an
d
T
iSAS
R
ec
o
n
m
u
ltip
le
h
ig
h
l
y
s
p
ar
s
e
d
atasets
[
2
1
]
,
[
2
4
]
.
T
h
ey
ar
e
m
ain
ly
f
o
cu
s
ed
o
n
NDCG@
1
0
a
n
d
HR
@
1
0
v
alu
es
with
o
u
t
f
u
r
th
er
ex
p
lo
r
in
g
th
e
ef
f
ec
ts
o
f
o
th
er
m
o
d
el
p
ar
am
eter
s
.
Hen
ce
,
f
u
r
th
e
r
in
v
esti
g
atio
n
s
ar
e
n
ee
d
ed
to
f
in
d
o
u
t
th
e
ef
f
ec
ts
o
f
m
o
d
el
p
ar
am
eter
s
,
s
u
ch
as
atten
tio
n
b
lo
ck
,
d
r
o
p
o
u
t r
ates
,
a
n
d
t
i
m
e
i
n
t
e
r
v
a
l
a
c
r
o
s
s
d
i
f
f
e
r
en
t
s
p
a
r
s
e
d
at
a
s
e
ts
.
S
e
v
e
r
a
l
p
r
i
o
r
s
t
u
d
ie
s
r
e
p
o
r
t
e
d
e
x
p
e
r
i
m
e
n
t
a
l
c
o
n
f
i
g
u
r
a
t
i
o
n
s
a
n
d
t
r
a
i
n
i
n
g
e
f
f
i
ci
e
n
c
y
u
n
d
e
r
d
i
f
f
e
r
e
n
t
c
o
m
p
u
t
a
t
i
o
n
a
l
s
e
tt
i
n
g
s
,
a
l
t
h
o
u
g
h
t
r
a
i
n
i
n
g
ti
m
e
w
as
n
o
t
c
o
n
s
is
t
e
n
tl
y
d
o
c
u
m
e
n
t
e
d
a
c
r
o
s
s
a
l
l
w
o
r
k
s
.
S
p
ec
i
f
i
c
al
l
y
,
t
h
e
r
e
s
ea
r
c
h
i
n
[
1
4
]
,
[
1
8
]
d
i
d
n
o
t
e
x
p
l
i
ci
t
l
y
r
e
p
o
r
t
t
r
a
i
n
i
n
g
t
i
m
e
,
d
es
p
i
te
t
h
e
s
t
u
d
ie
s
f
r
o
m
[
1
8
]
i
n
c
l
u
d
e
d
t
h
e
e
x
p
e
r
i
m
e
n
t
a
l
e
v
al
u
a
t
i
o
n
s
i
n
v
o
l
v
i
n
g
t
h
e
S
ASR
e
c
m
o
d
e
l
.
I
n
c
o
n
t
r
a
s
t
,
i
n
t
h
e
s
t
u
d
i
es
f
r
o
m
[
1
5
]
d
i
d
p
r
o
v
i
d
e
d
d
e
t
a
i
l
e
d
t
r
a
i
n
i
n
g
d
u
r
a
t
i
o
n
u
s
i
n
g
a
4
-
c
o
r
e
I
n
t
e
l
i
7
C
P
U
w
i
t
h
3
2
G
B
R
AM
,
r
e
p
o
r
t
i
n
g
a
p
p
r
o
x
im
a
t
e
l
y
o
n
e
h
o
u
r
f
o
r
t
h
e
M
o
v
i
e
L
e
n
s
d
a
t
as
e
t
,
t
w
o
h
o
u
r
s
f
o
r
t
h
e
G
o
w
al
l
a
a
n
d
F
o
u
r
s
q
u
a
r
e
d
a
t
a
s
et
s
a
n
d
o
n
e
h
o
u
r
f
o
r
t
h
e
T
m
a
l
l
d
at
as
e
t
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
3
,
J
u
n
e
2
0
2
6
:
2
7
6
1
-
2
7
7
3
2764
N
e
x
t
,
s
t
u
d
y
f
r
o
m
[
1
6
]
h
a
s
c
o
n
d
u
c
t
e
d
e
x
p
e
r
i
m
e
n
t
s
c
o
m
p
a
r
i
n
g
SR
m
o
d
el
s
a
n
d
r
e
p
o
r
t
e
d
p
e
r
-
ep
o
c
h
t
r
a
i
n
i
n
g
t
i
m
es
,
w
h
e
r
e
SA
SR
e
c
r
e
q
u
i
r
e
d
o
n
l
y
1
.
7
s
e
c
o
n
d
s
p
e
r
e
p
o
c
h
,
c
o
m
p
a
r
e
d
t
o
1
9
.
1
s
e
c
o
n
d
s
f
o
r
C
as
e
r
an
d
3
0
.
7
s
e
c
o
n
d
s
f
o
r
G
R
U
4
R
e
c+
.
F
u
r
t
h
e
r
m
o
r
e
,
S
AS
R
ec
c
o
n
v
e
r
g
e
d
t
o
o
p
t
i
m
a
l
p
e
r
f
o
r
m
a
n
c
e
w
i
t
h
i
n
a
p
p
r
o
x
i
m
a
t
e
ly
3
5
0
s
e
c
o
n
d
s
o
n
t
h
e
M
o
v
i
e
L
e
n
s
-
1
M
d
a
t
as
e
t
.
X
i
a
et
a
l
.
[
1
7
]
c
o
n
d
u
c
t
e
d
e
x
p
e
r
i
m
e
n
t
s
u
s
i
n
g
a
n
N
V
I
D
I
A
T
I
T
A
N
R
T
X
GP
U
a
n
d
an
X
e
o
n
W
-
2
1
3
3
C
P
U
;
w
h
i
l
e
S
AS
R
ec
w
as
e
v
a
l
u
at
e
d
o
n
t
h
e
T
ao
b
a
o
d
a
t
a
s
e
t
b
u
t
T
i
SA
SR
e
c
w
as
n
o
t
t
es
t
e
d
,
a
n
d
t
h
e
p
r
o
p
o
s
e
d
m
o
d
e
l
r
e
q
u
i
r
e
d
a
p
p
r
o
x
i
m
a
t
e
l
y
3
2
.
0
s
ec
o
n
d
s
p
e
r
e
p
o
c
h
.
C
h
a
n
g
e
t
a
l
.
[
22
]
h
a
s
r
e
p
o
r
te
d
t
h
a
t
t
h
e
S
UR
G
E
m
o
d
e
l
c
o
n
v
e
r
g
e
d
o
n
t
h
e
T
a
o
b
a
o
d
a
t
a
s
et
i
n
1
8
.
7
4
m
i
n
u
t
e
s
wh
e
n
t
r
a
i
n
e
d
o
n
a
n
N
V
I
D
I
A
T
es
l
a
V
1
0
0
GP
U
w
i
t
h
3
2
G
B
m
e
m
o
r
y
.
S
i
m
i
l
a
r
l
y
,
L
i
et
a
l
.
[
2
5
]
w
as
t
r
a
i
n
e
d
a
n
d
e
v
al
u
a
t
e
d
u
s
i
n
g
a
s
i
n
g
l
e
T
e
s
la
V
1
0
0
G
P
U
,
a
l
t
h
o
u
g
h
t
h
e
t
o
t
a
l
t
r
a
i
n
i
n
g
t
i
m
e
w
as
n
o
t
e
x
p
l
i
c
i
tl
y
s
t
a
te
d
.
Ul
t
i
m
at
e
l
y
,
m
o
s
t
o
f
s
t
u
d
i
e
s
a
r
e
t
es
t
e
d
w
i
t
h
t
h
e
M
o
v
i
e
L
e
n
s
d
a
t
as
et
a
l
o
n
g
w
i
t
h
t
h
e
SA
SR
e
c
m
o
d
e
l
i
n
t
h
e
s
t
u
d
i
es
o
f
SR
,
y
et
t
h
e
r
e
h
a
s
n
o
t
b
e
e
n
a
s
t
u
d
y
f
o
r
t
h
e
t
r
a
i
n
i
n
g
ti
m
e
t
a
k
en
a
c
r
o
s
s
d
i
f
f
e
r
e
n
t
p
a
r
a
m
e
t
e
r
s
o
f
S
A
SR
e
c
a
n
d
T
i
SA
SR
e
c
u
n
d
e
r
a
n
e
x
t
r
e
m
e
l
y
s
p
a
r
s
e
d
a
t
as
e
t
o
f
th
e
T
a
o
b
a
o
d
a
t
a
s
e
t
.
3.
M
E
T
H
O
D
T
h
is
s
tu
d
y
im
p
lem
e
n
ted
t
h
e
co
m
p
ar
is
o
n
b
etwe
en
S
ASR
ec
an
d
T
iSAS
R
ec
m
o
d
els
with
Mo
v
ieL
en
s
-
1
M
d
ataset
an
d
T
ao
b
ao
u
s
er
-
b
e
h
av
io
r
d
ataset
s
.
C
o
m
p
ar
ed
to
SASR
ec
m
o
d
el,
th
e
T
iSAS
R
e
c
m
o
d
el
m
a
p
s
o
u
t
th
e
ab
s
o
lu
te
p
o
s
itio
n
s
an
d
r
elativ
e
tim
e
i
n
ter
v
als
o
f
ea
ch
item
in
d
ex
f
r
o
m
th
e
u
s
er
-
item
in
ter
ac
tio
n
s
to
em
p
h
asize
th
e
im
p
o
r
tan
ce
an
d
c
o
r
r
elatio
n
s
b
etwe
en
item
s
.
T
h
is
h
elp
s
t
h
e
atten
tio
n
b
lo
ck
s
ca
lcu
late
atten
tio
n
weig
h
ts
b
ased
o
n
th
ese
em
b
ed
d
in
g
s
an
d
co
m
p
u
te
th
e
u
s
er
s
’
p
r
ef
e
r
en
ce
s
co
r
es
f
o
r
ea
ch
item
.
T
h
u
s
,
th
e
T
iSAS
R
ec
m
o
d
el
d
is
ce
r
n
s
th
e
v
ar
y
in
g
im
p
ac
t
(
n
o
is
es)
o
f
p
ast
in
te
r
ac
tio
n
s
o
n
f
u
t
u
r
e
in
ter
ac
tio
n
s
,
r
eg
ar
d
less
o
f
th
e
s
p
ar
s
ity
o
f
th
e
u
s
er
-
item
i
n
ter
a
ctio
n
s
eq
u
en
ce
s
.
T
h
e
f
ir
s
t
e
x
p
er
im
e
n
t
is
ca
r
r
ied
o
u
t
b
y
ap
p
ly
i
n
g
b
o
th
SASR
ec
an
d
T
iSAS
R
ec
m
o
d
els
in
Mo
v
ieL
en
s
-
1
M
d
ataset
an
d
T
ao
b
ao
u
s
er
-
b
eh
a
v
io
r
d
ataset
,
r
esp
ec
tiv
ely
with
d
ef
a
u
lt
m
o
d
el
p
ar
am
eter
s
to
co
m
p
ar
e
th
eir
p
er
f
o
r
m
an
ce
s
in
s
p
ar
s
e
d
atasets
.
T
h
en
,
f
u
r
th
er
ex
p
er
im
e
n
ts
ar
e
ca
r
r
ied
o
u
t
to
p
er
f
o
r
m
a
co
m
p
ar
ativ
e
an
aly
s
is
o
n
th
e
p
er
f
o
r
m
an
ce
o
f
b
o
th
m
o
d
els
with
d
if
f
er
en
t
p
ar
am
eter
v
alu
es.
B
o
th
m
o
d
els
ar
e
tr
ain
ed
an
d
r
u
n
in
th
e
en
v
ir
o
n
m
en
t
o
f
NVI
DI
A
GeFo
r
ce
GT
X
1
0
8
0
.
T
h
e
d
escr
ip
tio
n
s
o
f
th
e
d
atasets
ar
e
s
u
m
m
ar
ized
in
T
a
b
le
1
an
d
el
ab
o
r
ated
as f
o
llo
ws:
–
Mo
v
ieL
en
s
[
2
6
]
:
a
d
ataset
co
l
lecte
d
b
y
Gr
o
u
p
L
en
s
wh
ic
h
c
o
n
tain
s
1
,
0
0
0
,
2
0
9
an
o
n
y
m
o
u
s
r
atin
g
s
(
1
~5
)
o
f
ap
p
r
o
x
im
ately
3
,
9
0
0
m
o
v
ies
m
ad
e
b
y
6
,
0
4
0
Mo
v
ieL
e
n
s
u
s
er
s
in
th
e
y
ea
r
2
0
0
0
.
T
h
e
u
s
er
-
item
in
ter
ac
tio
n
r
ec
o
r
d
s
ar
e
s
to
r
ed
in
‘
r
atin
g
s
.
d
at’
,
wh
e
r
e
it
co
n
s
is
ts
o
f
m
o
v
ie
r
atin
g
s
with
f
o
u
r
d
if
f
e
r
en
t
lab
els o
f
‘
User
I
D’
,
‘
M
o
v
ieI
D’
‘
R
atin
g
’
,
an
d
‘
T
im
estam
p
’
.
–
T
ao
b
ao
[
2
7
]
:
a
d
ataset
co
llected
an
d
o
f
f
er
e
d
f
r
o
m
Alib
a
b
a
f
o
r
a
n
o
n
lin
e
s
h
o
p
p
in
g
p
l
atf
o
r
m
,
wh
ic
h
co
n
tain
s
1
0
0
,
1
5
0
,
8
0
7
an
o
n
y
m
o
u
s
in
ter
ac
tio
n
s
f
r
o
m
9
8
7
,
9
9
4
u
s
er
s
f
o
r
4
,
1
6
2
,
0
2
4
item
s
with
9
,
4
3
9
ca
teg
o
r
ies.
T
h
e
u
s
er
-
item
in
t
er
ac
tio
n
r
ec
o
r
d
s
ar
e
s
to
r
ed
i
n
‘
User
B
eh
av
io
r
.
csv
’
,
wh
e
r
e
it
co
n
s
is
ts
o
f
u
s
er
b
eh
av
io
r
d
ata
with
f
iv
e
d
if
f
er
en
t
lab
els
o
f
‘
User
I
D
’
,
‘
I
tem
I
D
’
,
‘
C
ateg
o
r
y
I
D
’
,
‘
B
eh
av
io
r
t
ype
’,
an
d
'
T
im
estam
p
'
.
T
ab
le
1
.
Statis
tics
o
f
T
ao
b
ao
u
s
er
-
b
eh
av
io
r
d
ataset
D
a
t
a
s
e
t
s
N
u
mb
e
r
s
o
f
u
n
i
q
u
e
u
ser
N
u
mb
e
r
s
o
f
u
n
i
q
u
e
i
t
e
m
To
t
a
l
i
n
t
e
r
a
c
t
i
o
n
s
D
a
t
a
sp
a
r
si
t
y
r
a
t
e
(
%)
D
a
t
a
d
e
n
s
i
t
y
r
a
t
e
(
%)
A
v
e
r
a
g
e
seq
u
e
n
c
e
l
e
n
g
t
h
Ta
o
b
a
o
u
s
e
r
-
b
e
h
a
v
i
o
r
9
8
7
,
9
9
4
4
,
1
6
2
,
0
2
4
1
0
0
,
1
5
0
,
8
0
7
9
9
.
8
5
0
.
1
5
5
.
5
0
M
o
v
i
e
L
e
n
s
-
1M
6
,
0
4
0
3
,
4
1
6
9
9
9
,
6
1
1
9
5
.
1
6
4
.
8
4
1
6
3
.
5
Data
s
et
s
p
ar
s
ity
r
ef
er
s
to
th
e
p
r
o
p
o
r
tio
n
o
f
m
is
s
in
g
o
r
u
n
a
v
ailab
le
in
ter
ac
tio
n
s
b
etwe
en
u
s
er
s
an
d
item
s
o
v
er
tim
e.
A
s
p
ar
s
e
d
at
aset
h
as
a
lar
g
e
n
u
m
b
e
r
o
f
u
n
o
b
s
er
v
ed
in
ter
ac
tio
n
s
r
elativ
e
to
th
e
p
o
s
s
ib
le
to
tal
in
ter
ac
tio
n
s
,
m
ea
n
in
g
th
at
u
s
e
r
s
in
ter
ac
t
with
o
n
ly
a
s
m
all
f
r
ac
tio
n
o
f
th
e
a
v
ailab
le
item
s
.
C
o
n
v
er
s
ely
,
a
d
en
s
e
d
ataset
h
as
m
o
r
e
in
ter
ac
tio
n
s
an
d
f
ewe
r
g
ap
s
,
in
d
icatin
g
a
h
ig
h
er
c
o
v
er
a
g
e
o
f
th
e
p
o
ten
t
ial
u
s
er
-
item
p
air
s
.
T
h
e
r
elatio
n
s
h
ip
b
etwe
en
s
p
ar
s
ity
an
d
d
en
s
ity
is
in
v
er
s
e
as t
h
e
d
en
s
ity
in
cr
ea
s
es,
th
e
s
p
ar
s
ity
d
ec
r
ea
s
es.
T
h
e
T
ao
b
ao
u
s
er
-
b
e
h
av
io
r
d
a
taset
h
as
a
h
ig
h
er
d
ata
s
p
a
r
s
ity
r
ate
co
m
p
a
r
ed
to
t
h
e
Mo
v
i
eL
en
s
-
1
M
d
ataset
d
esp
ite
h
o
ld
in
g
lar
g
e
r
n
u
m
b
er
s
o
f
u
n
iq
u
e
u
s
er
s
an
d
item
s
.
I
n
g
en
er
al,
a
d
ataset
with
h
ig
h
er
d
e
n
s
ity
h
o
ld
s
m
o
r
e
u
s
er
-
item
in
te
r
ac
tio
n
h
is
to
r
y
an
d
v
ice
v
er
s
a.
T
h
is
lead
s
to
b
etter
tr
ain
in
g
f
o
r
m
ac
h
in
e
lear
n
in
g
m
o
d
els,
as
it
h
as
m
o
r
e
d
ata
to
lear
n
p
atter
n
s
f
r
o
m
.
R
ich
er
c
o
n
ten
t
h
el
p
s
th
e
r
ec
o
m
m
en
d
atio
n
m
o
d
el
ca
p
tu
r
e
u
s
er
p
r
ef
er
en
ce
s
m
o
r
e
ef
f
ec
tiv
ely
,
r
esu
ltin
g
in
im
p
r
o
v
ed
p
er
f
o
r
m
a
n
ce
an
d
m
o
r
e
ac
cu
r
at
e
r
ec
o
m
m
en
d
atio
n
s
.
I
n
co
n
tr
ast,
a
s
p
ar
s
e
d
ataset
ca
n
lead
to
u
n
d
er
f
itti
n
g
o
r
less
r
eliab
le
r
ec
o
m
m
en
d
atio
n
s
,
as
th
e
m
o
d
el
h
as
lim
ited
d
ata
to
g
en
e
r
alize
u
s
er
p
r
ef
er
e
n
ce
s
o
r
in
f
er
p
atter
n
s
.
B
ased
o
n
T
ab
le
1
,
b
o
t
h
d
atas
ets
ar
e
s
h
o
wn
to
h
av
e
d
if
f
er
e
n
t
len
g
th
s
o
f
u
s
er
-
item
in
te
r
ac
tio
n
len
g
t
h
f
o
r
ea
c
h
u
s
er
.
T
h
e
d
atasets
with
lo
n
g
e
r
av
e
r
ag
e
s
eq
u
en
ce
l
en
g
th
h
av
e
r
elativ
ely
r
ic
h
er
i
n
ter
ac
tio
n
d
ata
an
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
E
ffects o
f sp
a
r
s
e
d
a
ta
s
ets o
n
time
in
terva
l
-
a
w
a
r
e
s
elf
-
a
tten
ti
o
n
s
eq
u
en
tia
l
…
(
Weis
h
a
n
Oo
i
)
2765
p
o
ten
tial
to
r
etu
r
n
b
etter
v
alu
es
in
ter
m
s
o
f
p
er
f
o
r
m
an
ce
m
etr
ics.
T
h
e
∑
(
_
[
]
)
_
is
r
ef
er
r
ed
as
th
e
to
tal
n
u
m
b
er
o
f
in
ter
ac
tio
n
s
ac
r
o
s
s
all
u
s
er
s
wh
ile
th
e
|
_
|
is
th
e
to
tal
n
u
m
b
er
o
f
u
s
er
s
in
th
e
tr
ain
in
g
s
et.
T
h
e
eq
u
atio
n
o
f
ca
lcu
latio
n
s
o
n
a
v
er
ag
e
s
eq
u
en
ce
len
g
th
o
f
a
u
s
er
in
a
d
ataset
is
d
ef
in
ed
in
(
1
)
.
ℎ
=
∑
(
_
[
]
)
_
|
_
|
(
1
)
3
.
1
.
Da
t
a
pre
-
pro
ce
s
s
ing
,
pa
dd
ing
,
a
nd
t
ra
in
-
t
est
-
v
a
lid
a
t
io
n sp
lit
T
h
e
d
ata
p
r
e
-
p
r
o
ce
s
s
in
g
s
tep
s
ar
e
co
n
d
u
cte
d
b
ased
o
n
[
1
6
]
an
d
[
2
3
]
,
wh
ic
h
in
clu
d
e
d
ata
clea
n
in
g
,
d
ata
f
ilter
in
g
with
f
iv
e
c
o
r
e
u
s
er
in
ter
ac
tio
n
s
,
a
n
d
a
r
r
an
g
em
e
n
ts
b
ased
o
n
tim
estam
p
s
.
T
h
e
p
r
o
ce
s
s
o
f
f
ilter
in
g
o
u
t
at
least
f
iv
e
co
r
e
u
s
er
in
te
r
ac
tio
n
s
f
r
o
m
ea
ch
u
s
er
is
to
e
n
s
u
r
e
th
at
t
h
e
tr
ain
in
g
,
test
in
g
an
d
v
alid
atio
n
d
ata
ar
e
s
u
f
f
icien
t.
U
p
o
n
f
ilter
in
g
o
n
5
co
r
e
u
s
er
in
ter
ac
tio
n
s
,
ea
ch
u
s
er
is
p
r
e
-
p
r
o
ce
s
s
ed
with
p
ad
d
in
g
with
n
u
m
b
er
o
f
2
0
0
a
n
d
1
0
r
esp
ec
tiv
ely
f
o
r
M
o
v
ieL
en
s
-
1
M
a
n
d
T
ao
b
ao
d
ataset.
Pad
d
in
g
is
a
tech
n
iq
u
e
t
h
at
is
wid
ely
ap
p
lied
in
SR
m
o
d
els
to
h
an
d
le
th
e
s
eq
u
en
ce
s
o
f
d
if
f
er
en
t
len
g
th
s
.
Fo
r
SR
,
ea
ch
u
s
er
h
as
d
if
f
er
en
t
len
g
th
s
o
f
u
s
er
-
item
in
ter
ac
tio
n
s
,
wh
er
e
it
r
eq
u
ir
es
a
f
ix
ed
len
g
th
o
f
u
s
er
-
item
in
ter
ac
ti
o
n
s
ets
f
o
r
ef
f
ec
tiv
e
tr
ain
in
g
o
f
th
e
SR
m
o
d
el.
Pad
d
in
g
ad
d
s
s
o
m
e
d
u
m
m
y
item
s
at
th
e
en
d
o
f
ea
c
h
u
s
er
’
s
item
in
ter
ac
tio
n
s
eq
u
en
ce
s
u
n
til
it
r
ea
c
h
es
th
e
s
am
e
len
g
th
as
th
e
lo
n
g
est
s
eq
u
en
ce
.
Fo
r
ex
am
p
le,
w
h
en
th
e
u
s
er
-
item
in
ter
ac
tio
n
s
eq
u
e
n
ce
o
f
(
1
,
2
,
…
,
15
,
16
)
h
as
a
m
ax
im
u
m
in
ter
ac
tio
n
u
n
til
16
,
th
e
p
a
d
d
in
g
tech
n
i
q
u
e
is
ap
p
lied
to
ex
p
an
d
th
e
u
s
er
-
ite
m
in
ter
ac
tio
n
s
eq
u
en
ce
u
n
til
50
to
f
u
lf
il th
e
m
ax
im
u
m
len
g
th
o
f
SASR
ec
th
at
i
s
f
ix
ed
to
5
0
in
d
ef
a
u
lt.
I
n
SASR
ec
,
it
h
an
d
les
v
a
r
iab
le
-
len
g
t
h
s
eq
u
en
ce
s
b
y
f
illi
n
g
u
p
d
u
m
m
y
d
ata
with
v
alu
e
‘
0
’
.
I
n
th
is
way
,
th
e
m
o
d
el
ca
n
p
r
o
ce
s
s
th
e
s
eq
u
en
ce
s
in
b
atc
h
es.
I
n
SASR
e
c
an
d
T
iSA
Sr
ec
,
e
v
er
y
u
s
e
r
-
it
em
s
e
q
u
e
n
ce
is
s
p
l
it
in
to
t
r
ai
n
i
n
g
,
tes
ti
n
g
a
n
d
v
al
i
d
at
io
n
s
ets
.
Fo
r
t
r
ai
n
i
n
g
s
et
,
it
h
o
l
d
s
t
h
e
wh
o
le
u
s
e
r
-
ite
m
i
n
t
er
ac
ti
o
n
s
e
q
u
e
n
c
e
o
f
ea
c
h
u
s
e
r
b
u
t
ex
cl
u
d
i
n
g
t
h
e
tw
o
l
at
est
ite
m
s
a
t th
e
en
d
o
f
t
h
e
s
e
q
u
e
n
c
e
i
n
p
u
t.
O
n
th
e
o
t
h
er
h
a
n
d
,
test
in
g
s
et
is
r
e
tr
ie
v
e
d
f
r
o
m
t
h
e
la
t
est
i
n
t
e
r
ac
te
d
ite
m
o
f
ea
ch
u
s
e
r
,
wh
ile
t
h
e
s
ec
o
n
d
m
o
s
t
r
e
ce
n
t
i
n
t
er
ac
te
d
it
em
s
ar
e
tr
ea
t
e
d
as
v
ali
d
ati
o
n
s
et
.
T
h
e
v
al
id
ati
o
n
s
et
is
u
s
e
d
t
o
v
e
r
i
f
y
th
e
t
r
ai
n
i
n
g
p
e
r
f
o
r
m
a
n
c
e
wit
h
2
0
e
p
o
c
h
s
.
Up
o
n
r
u
n
n
in
g
t
h
e
las
t
e
p
o
c
h
,
t
h
e
m
o
d
el
is
v
e
r
i
f
ie
d
wit
h
t
h
e
t
esti
n
g
s
et
.
F
in
all
y
,
t
h
e
p
e
r
f
o
r
m
a
n
c
es
o
f
b
o
t
h
m
o
d
els
a
r
e
e
v
a
lu
ate
d
wit
h
NDCG
a
n
d
h
it
r
ate
m
et
r
ics
.
3
.
2
.
M
o
del im
plem
ent
a
t
io
n
B
o
th
m
o
d
els
ar
e
im
p
lem
en
te
d
to
c
o
m
p
ar
e
th
eir
p
er
f
o
r
m
a
n
ce
o
n
two
d
atasets
with
th
e
m
ax
im
u
m
len
g
th
o
f
s
eq
u
e
n
ce
2
0
0
a
n
d
1
0
f
o
r
Mo
v
ieL
e
n
s
-
1
M
an
d
T
a
o
b
ao
d
ataset
r
esp
ec
tiv
ely
.
T
h
e
m
o
d
el
p
ar
am
eter
s
an
d
h
y
p
er
p
a
r
am
eter
s
o
f
SASR
ec
an
d
T
iSAS
R
ec
ar
e
co
n
f
ig
u
r
ed
to
d
ef
au
lt
v
alu
es
in
th
e
f
ir
s
t
co
m
p
ar
is
o
n
.
T
h
en
,
b
o
t
h
m
o
d
els
ar
e
in
v
est
ig
ated
with
d
if
f
er
en
t
co
m
b
in
atio
n
s
o
f
p
ar
am
eter
s
f
o
r
b
o
th
d
atasets
.
T
h
e
k
ey
p
ar
am
eter
s
in
th
e
SASR
ec
m
o
d
el
in
clu
d
e:
h
id
d
en
_
u
n
its
’
,
‘
n
u
m
_
b
l
o
ck
s
’
,
‘
n
u
m
_
h
ea
d
s
’
,
a
n
d
‘
d
r
o
p
o
u
t_
r
ate’
.
T
h
e
k
ey
p
ar
am
eter
s
in
T
iSAS
R
ec
ar
e
s
im
ilar
w
ith
SA
SR
ec
m
o
d
el
with
an
ad
d
iti
o
n
al
p
ar
am
eter
o
f
‘
tim
e_
s
p
an
’
to
in
d
icate
d
if
f
er
en
t
v
alu
es
o
f
tim
e
-
i
n
ter
v
al
.
T
h
e
d
ef
au
lt
p
ar
am
eter
v
al
u
es
ar
e
s
tated
as:
h
id
d
en
_
u
n
its
=
5
0
’
,
‘
n
u
m
_
b
lo
c
k
s
=
2
’
,
‘
n
u
m
_
h
ea
d
s
=
1
’
,
‘
d
r
o
p
o
u
t_
r
ate
=
0
.
2
’
,
an
d
‘
tim
e_
s
p
a
n
=
2
5
6
’
.
3.
3
.
E
v
a
lua
t
i
o
n m
et
rics
T
o
ev
alu
ate
th
e
p
er
f
o
r
m
a
n
ce
o
f
th
e
m
o
d
els
with
ex
is
tin
g
wo
r
k
s
,
th
is
s
tu
d
y
f
o
cu
s
es
o
n
im
p
lem
en
tin
g
n
o
r
m
alize
d
d
is
co
u
n
ted
cu
m
u
lativ
e
g
ain
(
NDCG@
1
0
)
an
d
h
it
r
ate
(
HR
@
1
0
)
ev
alu
atio
n
m
etr
ics
[
1
0
]
.
T
h
e
in
(
2
)
is
a
d
ef
i
n
ed
v
alu
e
o
f
a
cu
to
f
f
p
o
in
t
w
h
ile
ev
alu
atin
g
th
e
r
an
k
i
n
g
q
u
ality
.
Fo
r
ex
am
p
le,
th
e
c
u
to
f
f
p
o
in
t
in
th
is
s
tu
d
y
f
o
r
b
o
th
d
atasets
ar
e
1
0
,
wh
ich
o
n
ly
l
o
o
k
s
at
t
h
e
to
p
1
0
item
s
f
o
r
r
ec
o
m
m
e
n
d
atio
n
.
NDCG
is
in
d
icate
d
as
a
p
o
s
itio
n
-
awa
r
e
m
etr
ic,
wh
ic
h
m
ea
s
u
r
es
th
e
p
er
f
o
r
m
an
ce
o
f
a
r
ec
o
m
m
en
d
atio
n
m
o
d
el
b
y
co
n
s
id
er
in
g
th
e
r
an
k
in
g
s
o
f
i
tem
s
an
d
ass
ig
n
in
g
s
co
r
es
to
th
e
h
ig
h
e
r
r
an
k
s
o
f
item
s
.
C
o
n
tr
ar
y
to
NDCG,
(
HR
@
1
0
)
f
o
cu
s
es
m
o
r
e
o
n
th
e
p
r
esen
ce
o
f
th
e
c
o
r
r
ec
t
item
in
th
e
to
p
1
0
lis
t,
wh
ich
is
d
e
d
u
ce
d
in
(
3
)
.
ℎ
@
r
ef
er
s
to
th
e
n
u
m
b
er
o
f
h
its
f
o
r
th
e
u
s
er
an
d
#I
(
)
d
en
o
tes
th
e
n
u
m
b
er
o
f
in
te
r
ac
tio
n
s
o
f
u
s
er
.
T
o
p
r
ev
en
t
a
s
ev
er
e
co
m
p
u
tatio
n
b
u
r
d
e
n
o
n
all
u
s
er
-
item
p
air
s
,
th
e
s
tu
d
y
u
s
es
1
0
0
r
an
d
o
m
ly
ch
o
s
en
n
eg
ativ
e
s
am
p
les an
d
m
ix
es with
th
e
g
r
o
u
n
d
-
tr
u
th
item
s
(
p
o
s
itiv
e
s
am
p
les)
to
r
an
k
th
ese
item
s
f
o
r
ea
ch
u
s
er
u
.
@
=
∑
2
−
1
(
+
1
)
=
1
(
2
)
@
=
1
|
|
∑
#ℎ
@
#
(
)
|
|
=
1
(
3
)
B
o
th
@
an
d
@
v
alu
es
r
an
g
e
f
r
o
m
0
to
1
.
T
h
e
h
ig
h
er
v
alu
e
o
f
N
DC
G
an
d
HR
in
d
icate
s
b
etter
r
ec
o
m
m
e
n
d
atio
n
q
u
alit
y
.
NDCG@
k
an
d
HR
@
k
v
al
u
es
th
at
eq
u
al
to
1
in
th
e
ca
s
e
o
f
id
ea
l
r
a
n
k
in
g
,
wh
er
e
all
th
e
item
s
ar
e
p
er
f
e
ctly
in
clu
d
e
d
in
th
e
to
p
-
k
lis
t
an
d
ar
e
s
o
r
te
d
b
y
r
ele
v
an
ce
.
On
th
e
o
th
er
h
a
n
d
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
3
,
J
u
n
e
2
0
2
6
:
2
7
6
1
-
2
7
7
3
2766
@
an
d
@
v
alu
e
th
at
eq
u
al
0
in
d
icate
s
th
at
th
er
e
ar
e
n
o
r
elev
an
t
item
s
in
clu
d
ed
in
th
e
to
p
-
k
li
s
t.
T
h
er
ef
o
r
e,
th
e
h
ig
h
e
r
th
e
v
alu
es
o
f
@
an
d
@
,
th
e
m
o
r
e
r
elev
a
n
t
item
s
ar
e
in
clu
d
ed
in
th
e
to
p
-
k
lis
t a
n
d
s
o
r
ted
b
y
r
elev
an
c
y
.
4.
E
XP
E
R
I
M
E
N
T
A
L
RE
SUL
T
S AN
D
A
NALY
SI
S
T
h
e
ex
p
er
im
en
ts
ar
e
co
n
d
u
ct
ed
with
SAS
R
ec
an
d
T
iSA
S
R
ec
m
o
d
els
o
n
two
s
p
ar
s
e
d
a
tasets
.
T
h
e
ex
p
er
im
en
tal
r
esu
lts
h
av
e
r
ec
o
r
d
ed
th
e
s
co
r
i
n
g
i
n
NDCG@
1
0
an
d
HR
@
1
0
alo
n
g
with
th
eir
tr
ain
i
n
g
tim
e.
T
h
e
ex
p
er
im
e
n
tal
r
esu
lts
ar
e
s
h
o
wn
in
T
ab
les
2
to
5
to
in
v
esti
g
ate
th
e
ef
f
ec
ts
o
f
v
ar
y
i
n
g
v
alu
es
o
f
m
o
d
el
p
ar
am
eter
s
o
n
d
if
f
er
e
n
t sp
ar
s
ity
o
f
d
atasets
.
4
.
1
.
E
f
f
ec
t
s
o
f
m
o
del pa
ra
met
er
s
o
n
t
he
lo
wer
s
pa
rsity
da
t
a
s
et
T
ab
les
2
a
n
d
3
d
e
p
ict
th
e
co
m
p
ar
is
o
n
o
f
SASR
ec
an
d
T
i
SASR
ec
m
o
d
els
with
v
ar
y
in
g
p
ar
am
eter
v
alu
es
an
d
th
e
tr
ain
in
g
tim
e
f
o
r
p
r
o
ce
s
s
in
g
t
h
e
Mo
v
ieL
e
n
s
-
1
M
d
ataset.
E
ac
h
ex
p
er
im
en
t u
s
es
d
if
f
er
en
t
v
alu
es
f
o
r
atten
tio
n
b
lo
c
k
s
an
d
d
r
o
p
o
u
t
r
ates,
in
o
r
d
er
to
co
m
p
ar
e
with
th
e
d
ef
a
u
lt
p
ar
am
eter
s
(
atten
tio
n
b
lo
ck
s
=
2
,
d
r
o
p
o
u
t
r
ate
=
0
.
2
)
.
T
h
e
p
er
f
o
r
m
an
ce
m
etr
ics
NDCG@
1
0
a
n
d
R
@
1
0
ar
e
u
s
ed
to
ass
ess
t
h
e
p
er
f
o
r
m
an
ce
o
f
d
if
f
er
en
t
p
ar
am
eter
c
o
n
f
ig
u
r
at
io
n
s
,
with
ch
an
g
es c
o
m
p
ar
e
d
t
o
th
e
d
ef
a
u
lt p
ar
am
ete
r
s
r
ef
lec
ted
as p
er
ce
n
tag
es.
T
ab
le
2
.
NDCG@
1
0
an
d
tr
ain
in
g
tim
e
co
m
p
ar
is
o
n
o
f
SASR
ec
an
d
T
iSAS
R
ec
in
Mo
v
ieL
e
n
s
-
1
M
d
ataset
D
a
t
a
s
e
t
M
o
d
e
l
p
a
r
a
me
t
e
r
s
S
A
S
R
e
c
Ti
S
A
S
R
e
c
N
D
C
G
@
10
C
h
a
n
g
e
s
(
%)
Tr
a
i
n
i
n
g
t
i
me
N
D
C
G
@
10
C
h
a
n
g
e
s
(
%)
Tr
a
i
n
i
n
g
t
i
me
M
o
v
i
e
L
e
n
s
-
1M
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=2
D
r
o
p
o
u
t
r
a
t
e
=
0
.
2
1
0
.
5
8
7
1
-
0
h
2
3
m
3
5
s
0
.
6
0
3
0
-
3
h
5
7
m
4
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
4
D
r
o
p
o
u
t
r
a
t
e
=
0
.
2
0
.
5
9
7
7
+
1
.
8
0
0
h
5
5
m
2
s
0
.
6
0
9
1
2
+
1
.
0
1
1
7
h
2
6
m
5
6
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
8
D
r
o
p
o
u
t
r
a
t
e
=
0
.
2
0
.
5
9
8
1
+
1
.
8
7
1
h
2
9
m
2
9
s
0
.
5
9
6
5
-
1
.
0
8
2
3
h
2
8
m
5
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
2
D
r
o
p
o
u
t
r
a
t
e
=
0
.
3
0
.
5
8
5
4
-
0
.
2
9
0
h
3
7
m
3
1
s
0
.
5
9
9
7
-
0
.
5
5
4
h
3
2
m
3
3
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
2
D
r
o
p
o
u
t
r
a
t
e
=
0
.
4
0
.
5
7
8
6
3
-
1
.
4
5
0
h
4
3
m
4
8
s
0
.
5
9
8
7
-
0
.
7
1
4
h
3
2
m
2
5
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
2
D
r
o
p
o
u
t
r
a
t
e
=
0
.
2
Ti
me
-
s
p
a
n
=
1
,
0
2
4
-
-
-
0
.
6
0
2
8
-
0
.
0
3
3
h
5
0
m
3
9
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
2
D
r
o
p
o
u
t
r
a
t
e
=
0
.
2
Ti
me
-
s
p
a
n
=
2
,
0
4
8
-
-
-
0
.
6
0
2
3
--
0
.
1
2
3
h
4
6
m
1
s
*
N
o
t
e
:
1
D
e
f
a
u
l
t
p
a
r
a
me
t
e
r
v
a
l
u
e
s (
a
d
d
i
t
i
o
n
a
l
t
i
m
e
-
sp
a
n
p
a
r
a
me
t
e
r
v
a
l
u
e
i
s
2
5
6
f
o
r
Ti
S
A
S
R
e
c
2
I
n
d
i
c
a
t
e
s a
s
t
h
e
h
i
g
h
e
s
t
p
e
r
f
o
r
ma
n
c
e
v
a
l
u
e
i
n
t
h
e
d
a
t
a
s
e
t
3
I
n
d
i
c
a
t
e
s a
s
t
h
e
l
o
w
e
st
p
e
r
f
o
r
ma
n
c
e
v
a
l
u
e
i
n
t
h
e
d
a
t
a
s
e
t
T
ab
le
3
.
HR
@
1
0
an
d
tr
ai
n
in
g
tim
e
co
m
p
ar
is
o
n
o
f
SASR
ec
an
d
T
iSAS
R
ec
in
Mo
v
ieL
en
s
-
1
M
d
ataset
D
a
t
a
s
e
t
M
o
d
e
l
p
a
r
a
me
t
e
rs
S
A
S
R
e
c
Ti
S
A
S
R
e
c
H
R
@
1
0
C
h
a
n
g
e
s
(
%)
Tr
a
i
n
i
n
g
t
i
me
H
R
@
1
0
C
h
a
n
g
e
s
(
%)
Tr
a
i
n
i
n
g
t
i
m
e
M
o
v
i
e
L
e
n
s
-
1M
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
2
D
r
o
p
o
u
t
r
a
t
e
=
0
.
2
1
0
.
8
2
1
4
-
0
h
2
3
m
3
5
s
0
.
8
2
6
3
-
3
h
5
7
m
4
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
4
D
r
o
p
o
u
t
r
a
t
e
=
0
.
2
0
.
8
2
5
7
+
0
.
5
2
0
h
5
5
m
2
s
0
.
8
3
0
8
2
+
0
.
5
5
17
h
2
6
m
5
6
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
8
D
r
o
p
o
u
t
r
a
t
e
=
0
.
2
0
.
8
2
2
7
+
0
.
1
6
1
h
2
9
m
2
9
s
0
.
8
2
2
4
-
0
.
4
7
23
h
2
8
m
5
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
2
D
r
o
p
o
u
t
r
a
t
e
=
0
.
3
0
.
8
2
1
7
+
0
.
0
4
0
h
3
7
m
3
1
s
0
.
8
2
3
3
-
0
.
3
6
4
h
3
2
m
3
3
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
2
D
r
o
p
o
u
t
r
a
t
e
=
0
.
4
0
.
8
1
6
6
3
-
0
.
5
9
0
h
4
3
m
4
8
s
0
.
8
2
4
2
-
0
.
2
5
4
h
3
2
m
2
5
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
2
D
r
o
p
o
u
t
r
a
t
e
=
0
.
2
Ti
me
-
s
p
a
n
=
1
,
0
2
4
-
-
-
0
.
8
2
3
7
-
0
.
3
1
3
h
5
0
m
3
9
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
2
D
r
o
p
o
u
t
r
a
t
e
=
0
.
2
Ti
me
-
s
p
a
n
=
2
,
0
4
8
-
-
-
0
.
8
2
6
3
-
3
h
4
6
m
1
s
*
N
o
t
e
:
1
D
e
f
a
u
l
t
p
a
r
a
me
t
e
r
v
a
l
u
e
s (
a
d
d
t
i
o
n
a
l
t
i
me
-
s
p
a
n
p
a
r
a
m
e
t
e
r
v
a
l
u
e
i
s
2
5
6
f
o
r
T
i
S
A
S
R
e
c
)
2
I
n
d
i
c
a
t
e
s a
s
t
h
e
h
i
g
h
e
s
t
p
e
r
f
o
r
ma
n
c
e
v
a
l
u
e
i
n
t
h
e
d
a
t
a
s
e
t
3
I
n
d
i
c
a
t
e
s a
s
t
h
e
l
o
w
e
st
p
e
r
f
o
r
ma
n
c
e
v
a
l
u
e
i
n
t
h
e
d
a
t
a
s
e
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
E
ffects o
f sp
a
r
s
e
d
a
ta
s
ets o
n
time
in
terva
l
-
a
w
a
r
e
s
elf
-
a
tten
ti
o
n
s
eq
u
en
tia
l
…
(
Weis
h
a
n
Oo
i
)
2767
T
iSAS
R
ec
m
o
d
el
ac
h
iev
es
th
e
h
ig
h
est
v
alu
es
o
f
NDCG
@
1
0
an
d
HR
@
1
0
in
Mo
v
ieL
en
s
-
1
M
d
ataset
with
0
.
6
0
9
1
a
n
d
0
.
8
3
0
8
,
r
esp
e
ctiv
ely
f
o
r
th
e
co
n
f
i
g
u
r
atio
n
s
o
f
4
atten
tio
n
b
lo
ck
s
an
d
0
.
2
d
r
o
p
o
u
t
r
ate.
On
th
e
o
th
er
h
a
n
d
,
SASR
ec
m
o
d
el
e
x
h
ib
its
th
e
lo
west
v
alu
es
o
f
NDCG
@
1
0
an
d
HR
@
1
0
with
0
.
5
7
8
6
an
d
0
.
8
1
6
6
r
esp
ec
tiv
ely
f
o
r
th
e
c
o
n
f
i
g
u
r
a
tio
n
s
o
f
2
atten
tio
n
b
lo
ck
s
a
n
d
0
.
4
d
r
o
p
o
u
t
r
ate.
Alth
o
u
g
h
T
iSAS
R
ec
h
as
a
g
en
er
ally
l
o
n
g
e
r
tr
ain
i
n
g
tim
e
co
m
p
ar
ed
to
SASR
ec
f
o
r
Mo
v
ieL
en
s
-
1
M,
it
is
wo
r
th
n
o
tin
g
th
at
th
e
T
iSAS
R
ec
p
r
o
v
id
es
m
o
r
e
ac
cu
r
ate
r
ec
o
m
m
en
d
atio
n
r
a
n
k
in
g
with
h
ig
h
er
NDCG
@
1
0
an
d
HR
@
1
0
v
alu
es
co
m
p
ar
ed
to
SASR
ec
.
A
r
ec
o
m
m
en
d
atio
n
m
o
d
el
th
at
d
eliv
er
s
s
u
p
er
io
r
r
ec
o
m
m
e
n
d
atio
n
q
u
ality
ca
n
lead
to
b
etter
cu
s
to
m
er
s
atis
f
ac
tio
n
.
B
o
th
SASR
ec
an
d
T
iSAS
R
e
c
m
o
d
els
ar
e
test
ed
with
d
if
f
er
en
t
n
u
m
b
er
s
o
f
atten
tio
n
b
lo
ck
s
to
in
v
esti
g
ate
th
e
ef
f
ec
ts
o
n
th
eir
r
ec
o
m
m
e
n
d
atio
n
p
er
f
o
r
m
an
ce
.
I
n
th
e
M
o
v
ieL
en
s
-
1
M
d
ataset,
T
iSAS
R
ec
p
er
f
o
r
m
ed
b
est
with
th
ese
m
o
d
if
icatio
n
s
.
Atten
tio
n
b
lo
ck
s
co
n
s
is
t
o
f
s
elf
-
atten
tio
n
la
y
er
s
th
at
h
elp
p
r
eser
v
e
s
eq
u
en
ce
o
r
d
er
an
d
im
p
r
o
v
e
tr
ain
in
g
.
Ad
d
in
g
m
o
r
e
atten
ti
o
n
b
lo
c
k
s
in
cr
ea
s
es
th
e
m
o
d
e
ls
'
co
m
p
lex
ity
an
d
tr
ain
in
g
tim
e
f
o
r
d
ee
p
er
r
ep
r
esen
tatio
n
lear
n
in
g
.
Fo
r
SASR
ec
,
in
cr
ea
s
in
g
atten
tio
n
b
lo
ck
s
f
r
o
m
2
t
o
8
im
p
r
o
v
e
d
NDCG@
1
0
f
r
o
m
0
.
5
8
7
1
to
0
.
5
9
8
1
(
a
1
.
8
7
%
im
p
r
o
v
em
e
n
t)
.
Ho
wev
e
r
,
HR
@
1
0
d
ec
r
ea
s
ed
f
r
o
m
0
.
5
2
%
to
0
.
1
6
%
wh
en
in
cr
ea
s
in
g
f
r
o
m
4
to
8
atte
n
tio
n
b
lo
c
k
s
.
T
h
is
m
ea
n
s
th
at
wh
ile
m
o
r
e
atten
tio
n
b
lo
ck
s
im
p
r
o
v
e
t
h
e
r
an
k
in
g
o
f
r
elev
a
n
t
item
s
,
th
ey
m
ay
also
r
esu
lt in
m
is
s
in
g
s
o
m
e
item
s
in
th
e
t
o
p
1
0
lis
t.
I
n
d
e
n
s
er
d
atasets
lik
e
Mo
v
ieL
en
s
-
1
M,
i
n
cr
ea
s
in
g
atten
tio
n
b
lo
ck
s
h
el
p
s
ac
cu
r
ately
r
an
k
r
elev
an
t ite
m
s
b
u
t m
ay
n
ar
r
o
w
th
e
s
co
p
e
o
f
th
e
t
o
p
1
0
r
ec
o
m
m
en
d
atio
n
s
,
f
o
cu
s
in
g
o
n
a
f
e
w
h
ig
h
ly
r
an
k
e
d
item
s
r
ath
er
t
h
an
a
b
r
o
a
d
er
r
a
n
g
e.
T
h
is
ca
n
m
ak
e
th
e
m
o
d
el
less
ef
f
ec
tiv
e
at
id
en
tif
y
in
g
m
o
r
e
r
elev
an
t item
s
.
NDCG
@
1
0
o
f
T
iSAS
R
ec
s
l
ig
h
tly
in
cr
ea
s
ed
f
r
o
m
0
.
6
0
3
0
to
0
.
6
0
9
1
wh
en
th
e
atten
tio
n
b
l
o
ck
s
wer
e
in
cr
ea
s
ed
to
4
,
b
u
t
d
r
o
p
p
ed
b
y
1
.
0
8
%
to
0
.
5
9
6
5
wh
e
n
in
cr
ea
s
ed
to
8
.
As
t
h
e
m
o
d
el
co
m
p
le
x
ity
in
cr
ea
s
ed
with
m
o
r
e
atten
tio
n
b
lo
c
k
s
,
th
e
tr
ain
in
g
tim
e
also
in
c
r
ea
s
ed
,
b
u
t
with
o
u
t
s
ig
n
if
ica
n
t
p
er
f
o
r
m
an
ce
g
ain
s
.
T
h
is
s
u
g
g
ests
th
at
8
atten
tio
n
b
lo
ck
s
ar
e
ex
ce
s
s
iv
e
f
o
r
th
is
d
atas
et,
wh
ich
h
as
a
s
p
ar
s
ity
o
f
9
5
.
1
6
%,
an
d
m
ay
lead
to
p
o
o
r
g
en
e
r
aliza
tio
n
.
Desp
ite
u
s
in
g
th
e
s
am
e
n
u
m
b
e
r
o
f
atten
tio
n
b
lo
ck
s
,
T
iSA
SR
ec
co
n
s
is
ten
tly
o
u
tp
er
f
o
r
m
ed
SASR
ec
in
NDCG
@
1
0
an
d
HR
@
1
0
.
T
h
is
is
b
ec
au
s
e
T
iSA
SR
ec
in
co
r
p
o
r
ates
tim
e
in
ter
v
al
in
f
o
r
m
atio
n
in
to
its
s
elf
-
atten
tio
n
m
ec
h
an
is
m
,
ca
p
tu
r
in
g
tem
p
o
r
al
d
y
n
am
ics
an
d
ev
o
lv
in
g
u
s
er
p
r
ef
er
en
ce
s
m
o
r
e
ef
f
ec
tiv
ely
.
T
h
is
tem
p
o
r
al
awa
r
en
ess
allo
ws
T
iSAS
R
e
c
to
p
r
o
v
id
e
b
etter
r
ec
o
m
m
e
n
d
atio
n
s
.
Ho
wev
er
,
T
iSAS
R
ec
g
en
er
ally
r
eq
u
ir
es
m
o
r
e
tr
ain
in
g
tim
e
d
u
e
to
th
e
ad
d
e
d
c
o
m
p
lex
ity
o
f
p
r
o
ce
s
s
in
g
b
o
th
item
s
eq
u
en
ce
s
an
d
tem
p
o
r
al
in
f
o
r
m
atio
n
.
I
n
c
r
ea
s
in
g
th
e
tim
e
-
s
p
an
v
alu
e
f
o
r
th
e
T
iSAS
R
ec
m
o
d
el
in
th
e
Mo
v
ieL
en
s
-
1
M
d
ataset
s
lig
h
tly
wo
r
s
en
e
d
th
e
NDCG@
1
0
p
er
f
o
r
m
a
n
ce
.
T
h
e
NDCG@
1
0
v
alu
e
d
r
o
p
p
ed
f
r
o
m
0
.
6
0
3
0
to
0
.
6
0
2
8
an
d
0
.
6
0
2
3
f
o
r
tim
e
-
s
p
an
v
alu
es
o
f
1
,
0
2
4
an
d
2
,
0
4
8
,
r
esp
ec
tiv
ely
.
Fo
r
HR
@
1
0
,
th
e
v
alu
e
d
r
o
p
p
ed
f
r
o
m
0
.
8
2
6
3
t
o
0
.
8
2
3
7
at
a
tim
e
s
p
an
o
f
1
,
0
2
4
b
u
t
r
em
ain
ed
th
e
s
am
e
at
2
,
0
4
8
.
T
h
is
in
d
icate
s
th
at
in
cr
ea
s
in
g
th
e
tim
e
s
p
an
in
th
e
d
en
s
er
Mo
v
ieL
en
s
-
1
M
d
ataset
d
o
es
n
o
t
im
p
r
o
v
e
NDCG@
1
0
p
er
f
o
r
m
a
n
ce
d
u
e
to
ad
d
e
d
n
o
is
e
in
th
e
tr
ai
n
in
g
d
ata.
As
th
e
tim
e
s
p
an
in
cr
ea
s
es,
ea
r
lier
in
ter
ac
tio
n
s
th
at
a
r
e
less
r
elev
an
t
to
th
e
cu
r
r
en
t
r
ec
o
m
m
en
d
atio
n
task
ar
e
i
n
clu
d
ed
,
tu
r
n
in
g
in
to
n
o
is
e
as
u
s
er
p
r
ef
er
e
n
ce
s
ch
an
g
e
o
v
er
tim
e.
T
h
is
p
r
ev
en
ts
th
e
m
o
d
el
f
r
o
m
ac
c
u
r
ately
p
r
ed
ictin
g
t
h
e
n
e
x
t r
elev
an
t item
.
Fo
r
SASR
ec
,
th
e
d
r
o
p
o
u
t
r
ate
h
elp
s
p
r
ev
en
t
o
v
e
r
f
itti
n
g
b
y
r
a
n
d
o
m
ly
s
ettin
g
a
f
r
ac
tio
n
o
f
th
e
in
p
u
t
to
ze
r
o
d
u
r
i
n
g
tr
ain
in
g
.
W
h
ile
HR
@
1
0
s
lig
h
tly
im
p
r
o
v
es
b
y
0
.
0
4
%
at
a
d
r
o
p
o
u
t
r
ate
o
f
0
.
3
,
in
cr
ea
s
in
g
th
e
d
r
o
p
o
u
t
r
ate
g
en
er
ally
h
as
a
m
in
im
al
p
o
s
itiv
e
im
p
ac
t
o
n
N
DC
G
@
1
0
an
d
HR
@
1
0
.
At
a
d
r
o
p
o
u
t
r
ate
o
f
0
.
4
,
NDCG
@
1
0
an
d
HR
@
1
0
d
r
o
p
to
0
.
5
7
8
6
an
d
0
.
8
1
6
6
,
r
esp
ec
tiv
ely
,
in
d
icatin
g
wo
r
s
e
r
ec
o
m
m
en
d
atio
n
q
u
ality
an
d
less
ef
f
ec
tiv
e
t
o
p
-
1
0
it
em
in
clu
s
io
n
.
As
th
e
d
r
o
p
o
u
t
r
ate
in
c
r
ea
s
es,
SASR
ec
's
p
er
f
o
r
m
an
ce
s
to
p
s
im
p
r
o
v
in
g
,
lea
d
in
g
to
u
n
d
e
r
f
it
tin
g
an
d
an
in
ab
ilit
y
to
lear
n
m
ea
n
in
g
f
u
l
p
atter
n
s
f
r
o
m
th
e
d
ata.
Fo
r
T
iSAS
R
ec
,
in
cr
ea
s
in
g
th
e
d
r
o
p
o
u
t
r
ate
d
o
es
n
o
t
im
p
r
o
v
e
NDCG@
1
0
a
n
d
HR
@
1
0
v
al
u
es.
A
h
ig
h
er
d
r
o
p
o
u
t
r
ate
ca
u
s
es
T
iSAS
R
ec
to
lo
s
e
to
o
m
u
c
h
in
f
o
r
m
atio
n
d
u
r
i
n
g
tr
ain
i
n
g
,
im
p
air
in
g
its
ab
ilit
y
to
ca
p
tu
r
e
co
m
p
lex
p
atter
n
s
an
d
in
ter
ac
tio
n
s
,
an
d
u
ltima
tely
f
ai
lin
g
to
r
an
k
r
elev
a
n
t item
s
ef
f
ec
tiv
ely
.
4
.
2
.
E
f
f
ec
t
s
o
f
m
o
del pa
ra
met
er
s
o
n hig
her
s
pa
rsity
da
t
a
s
et
Similar
to
th
e
p
r
ev
i
o
u
s
s
ec
tio
n
,
T
ab
le
s
4
a
n
d
5
d
ep
ict
th
e
c
o
m
p
ar
is
o
n
o
f
b
o
th
m
o
d
els
with
v
ar
y
in
g
v
alu
es
o
f
m
o
d
el
p
ar
am
eter
s
an
d
th
e
tr
ain
in
g
tim
e
f
o
r
p
r
o
c
e
s
s
in
g
th
e
T
ao
b
ao
d
ataset.
T
h
e
d
ef
au
lt
p
ar
am
eter
s
f
o
r
atten
tio
n
b
lo
c
k
s
,
d
r
o
p
o
u
t
r
ate
an
d
tim
e
s
p
an
r
em
ain
th
e
s
am
e
as
in
th
e
p
r
e
v
io
u
s
s
ec
tio
n
.
SASR
ec
m
o
d
el
ac
h
iev
es
th
e
h
ig
h
est
v
alu
e
o
f
NDCG
@
1
0
an
d
HR
@
1
0
wi
th
2
atten
tio
n
b
lo
c
k
s
an
d
a
0
.
3
d
r
o
p
o
u
t
r
ate.
I
n
co
n
tr
ast,
th
e
T
iSAS
R
ec
m
o
d
el
h
as
th
e
lo
west
ND
C
G
@
1
0
an
d
HR
@
1
0
with
8
atten
tio
n
b
lo
ck
s
an
d
a
0
.
2
d
r
o
p
o
u
t
r
ate.
T
h
is
s
h
o
ws
th
at
SASR
ec
p
r
o
v
id
es
m
o
r
e
a
cc
u
r
ate
r
ec
o
m
m
en
d
atio
n
s
o
n
th
e
h
ig
h
er
s
p
ar
s
ity
T
ao
b
ao
d
ataset.
I
n
co
n
tr
ast
to
M
o
v
ieL
en
s
-
1
M
d
ataset,
SASR
ec
s
h
o
ws
b
etter
p
er
f
o
r
m
a
n
ce
with
c
h
an
g
es
i
n
atten
tio
n
b
lo
ck
s
in
s
tead
o
f
T
iSAS
R
ec
.
I
n
cr
ea
s
in
g
th
e
atten
tio
n
b
l
o
ck
s
in
SASR
ec
o
n
t
h
e
T
a
o
b
ao
d
ataset
led
to
s
ig
n
if
ican
t
im
p
r
o
v
em
en
ts
,
esp
ec
ially
with
4
atten
tio
n
b
l
o
ck
s
,
s
h
o
win
g
a
9
.
0
5
%
in
cr
e
m
en
t
in
NDCG
@
1
0
an
d
1
1
.
2
3
%
in
HR
@
1
0
.
W
ith
4
at
ten
tio
n
b
lo
ck
s
an
d
a
0
.
2
d
r
o
p
o
u
t
r
ate,
SASR
ec
ac
h
iev
ed
N
DC
G
@
1
0
o
f
0
.
2
5
0
6
an
d
HR
@
1
0
o
f
0
.
3
8
2
3
.
Ho
w
ev
er
,
p
er
f
o
r
m
a
n
ce
d
r
o
p
p
ed
t
o
0
.
2
2
0
7
an
d
0
.
3
4
4
2
w
h
en
t
h
e
atten
tio
n
b
lo
ck
s
in
cr
ea
s
ed
to
8
.
Desp
ite
th
e
in
c
r
ea
s
ed
co
m
p
lex
ity
an
d
tr
ain
in
g
tim
e,
SASR
ec
'
s
p
er
f
o
r
m
an
c
e
s
lig
h
tly
d
ec
r
ea
s
ed
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
3
,
J
u
n
e
2
0
2
6
:
2
7
6
1
-
2
7
7
3
2768
af
ter
4
atten
tio
n
b
l
o
ck
s
d
u
e
t
o
o
v
e
r
f
itti
n
g
o
n
th
e
h
ig
h
e
r
s
p
ar
s
ity
d
ataset.
T
h
e
T
ao
b
ao
d
a
taset,
b
ein
g
s
p
ar
s
er
th
an
Mo
v
ieL
en
s
-
1
M
,
h
as
a
lo
wer
d
en
s
ity
,
lead
in
g
to
m
o
r
e
m
is
s
in
g
u
s
er
-
item
in
ter
ac
tio
n
s
an
d
in
ef
f
icien
t
tr
ain
in
g
.
T
h
u
s
,
th
e
tim
e
in
ter
v
al
-
awa
r
e
in
f
o
r
m
atio
n
u
s
ed
b
y
T
iSAS
R
ec
is
less
ef
f
ec
tiv
e
in
h
ig
h
er
s
p
ar
s
ity
d
atasets
,
r
ed
u
cin
g
its
ad
v
a
n
tag
es o
v
er
SASR
ec
.
T
ab
le
4
.
NDCG@
1
0
an
d
t
r
ain
in
g
tim
e
co
m
p
ar
is
o
n
o
f
SASR
ec
an
d
T
iSAS
R
ec
in
T
ao
b
ao
d
ataset
D
a
t
a
s
e
t
M
o
d
e
l
p
a
r
a
me
t
e
r
s
S
A
S
R
e
c
Ti
S
A
S
R
e
c
N
D
C
G
@
1
0
C
h
a
n
g
e
s
(
%)
Tr
a
i
n
i
n
g
t
i
me
N
D
C
G
@
1
0
C
h
a
n
g
e
s
(
%)
Tr
a
i
n
i
n
g
t
i
me
Ta
o
b
a
o
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
2
D
r
o
p
o
u
t
r
a
t
e
=
0
.
2
1
0
.
2
2
9
8
-
3
h
5
7
m
4
s
0
.
2
3
4
6
-
4
h
2
8
m
4
3
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
4
D
r
o
p
o
u
t
r
a
t
e
=
0
.
2
0
.
2
5
0
6
+
9
.
0
5
10
h
2
9
m
4
2
s
0
.
2
3
5
9
+
0
.
5
5
1
h
3
9
m
2
3
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
8
D
r
o
p
o
u
t
r
a
t
e
=
0
.
2
0
.
2
2
0
7
-
3
.
9
6
9
h
3
7
m
6
s
0
.
2
2
0
3
3
-
6
.
0
9
4
h
3
9
m
3
5
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
2
D
r
o
p
o
u
t
r
a
t
e
=
0
.
3
0
.
2
6
4
6
2
+
1
5
.
1
4
4
h
12
m
2
6
s
0
.
2
6
2
0
+
1
1
.
6
8
1
h
1
8
m
3
3
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
2
D
r
o
p
o
u
t
r
a
t
e
=
0
.
4
0
.
2
6
2
3
+
1
4
.
1
4
5
h
1
3
m
3
3
s
0
.
2
5
6
1
+
9
.
1
7
3
h
39
m
5
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
2
D
r
o
p
o
u
t
r
a
t
e
=
0
.
2
Ti
me
-
s
p
a
n
=
1
0
2
4
-
-
-
0
.
2
5
2
7
+
7
.
7
2
%
3
h
1
3
m
2
5
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
2
D
r
o
p
o
u
t
r
a
t
e
=
0
.
2
Ti
me
-
s
p
a
n
=
2
0
4
8
-
-
-
0
.
2
5
3
9
+
7
.
8
1
%
3
h
3
0
m
3
5
s
*
N
o
t
e
:
1
D
e
f
a
u
l
t
p
a
r
a
me
t
e
r
v
a
l
u
e
s (
a
d
d
i
t
i
o
n
a
l
t
i
m
e
-
sp
a
n
p
a
r
a
me
t
e
r
v
a
l
u
e
i
s
2
5
6
f
o
r
Ti
S
A
S
R
e
c
)
2
I
n
d
i
c
a
t
e
s a
s
t
h
e
h
i
g
h
e
s
t
p
e
r
f
o
r
ma
n
c
e
v
a
l
u
e
i
n
t
h
e
d
a
t
a
s
e
t
3
I
n
d
i
c
a
t
e
s a
s
t
h
e
l
o
w
e
st
p
e
r
f
o
r
ma
n
c
e
v
a
l
u
e
i
n
t
h
e
d
a
t
a
s
e
t
T
ab
le
5
.
HR
@
1
0
an
d
tr
ai
n
in
g
tim
e
co
m
p
ar
is
o
n
o
f
SASR
ec
an
d
T
iSAS
R
ec
in
T
ao
b
ao
d
atas
et
D
a
t
a
s
e
t
M
o
d
e
l
p
a
r
a
me
t
e
r
s
S
A
S
R
e
c
Ti
S
A
S
R
e
c
H
R
@
1
0
C
h
a
n
g
e
s
(
%)
Tr
a
i
n
i
n
g
t
i
me
H
R
@
1
0
C
h
a
n
g
e
s
(
%)
Tr
a
i
n
i
n
g
t
i
me
Ta
o
b
a
o
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=2
D
r
o
p
o
u
t
r
a
t
e
=
0
.
2
1
0
.
3
4
3
7
-
3
h
5
7
m
4
s
0
.
3
4
0
7
-
4
h
2
8
m
4
3
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
4
D
r
o
p
o
u
t
r
a
t
e
=
0
.
2
0
.
3
8
2
3
+
1
1
.
2
3
1
0
h
2
9
m
4
2
s
0
.
3
4
5
8
+
1
.
5
0
1
h
3
9
m
2
3
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
8
D
r
o
p
o
u
t
r
a
t
e
=
0
.
2
0
.
3
4
4
2
+
0
.
2
5
9
h
3
7
m
6
s
0
.
3
3
1
5
3
-
2
.
7
0
4
h
3
9
m
3
5
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
2
D
r
o
p
o
u
t
r
a
t
e
=
0
.
3
0
.
3
9
4
7
2
+
1
4
.
8
5
4
h
1
2
m
2
6
s
0
.
3
7
5
2
+
1
0
.
1
4
1
h
1
8
m
3
3
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
2
D
r
o
p
o
u
t
r
a
t
e
=
0
.
4
0
.
3
8
8
2
+
1
2
.
9
5
5
h
1
3
m
3
3
s
0
.
3
7
5
5
+
1
0
.
2
1
3
h
3
9
m
5
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
2
D
r
o
p
o
u
t
r
a
t
e
=
0
.
2
Ti
me
-
s
p
a
n
=
1
0
2
4
-
-
-
0
.
3
5
6
5
+
4
.
6
4
3
h
1
3
m
2
5
s
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
2
D
r
o
p
o
u
t
r
a
t
e
=
0
.
2
Ti
me
-
s
p
a
n
=
2
0
4
8
-
-
-
0
.
3
6
0
1
+
5
.
7
0
3
h
3
0
m
3
5
s
*
N
o
t
e
:
1
D
e
f
a
u
l
t
p
a
r
a
me
t
e
r
v
a
l
u
e
s (
a
d
d
i
t
i
o
n
a
l
t
i
m
e
-
sp
a
n
p
a
r
a
me
t
e
r
v
a
l
u
e
i
s
2
5
6
f
o
r
Ti
S
A
S
R
e
c
)
2
I
n
d
i
c
a
t
e
s a
s
t
h
e
h
i
g
h
e
s
t
p
e
r
f
o
r
ma
n
c
e
v
a
l
u
e
i
n
t
h
e
d
a
t
a
s
e
t
3
I
n
d
i
c
a
t
e
s a
s
t
h
e
l
o
w
e
st
p
e
r
f
o
r
ma
n
c
e
v
a
l
u
e
i
n
t
h
e
d
a
t
a
s
e
t
NDCG
@
1
0
o
f
SASR
ec
im
p
r
o
v
ed
b
y
1
5
.
1
4
% f
r
o
m
0
.
2
2
9
8
to
0
.
2
6
4
6
wh
e
n
th
e
d
r
o
p
o
u
t r
ate
in
cr
ea
s
ed
f
r
o
m
0
.
2
to
0
.
3
.
Dr
o
p
o
u
t
r
ate
h
elp
s
p
r
ev
en
t
o
v
er
f
itti
n
g
an
d
im
p
r
o
v
e
g
en
er
aliza
tio
n
,
wh
ic
h
is
wh
y
SASR
ec
s
h
o
wed
s
u
ch
im
p
r
o
v
em
e
n
t
o
n
th
e
s
p
a
r
s
e
T
ao
b
ao
d
ataset.
C
o
m
p
ar
ed
to
th
e
Mo
v
ieL
en
s
-
1
M
d
ataset,
SASR
ec
h
ad
m
o
r
e
s
ig
n
if
ican
t
im
p
r
o
v
e
m
en
ts
with
d
if
f
er
en
t
d
r
o
p
o
u
t
r
ate
o
n
T
ao
b
ao
.
A
h
ig
h
s
p
a
r
s
ity
o
f
9
9
.
8
5
%
in
T
ao
b
ao
in
d
icate
s
f
ewe
r
in
ter
a
ctio
n
s
p
er
u
s
er
,
w
h
ich
in
c
r
ea
s
es
d
ata
co
m
p
lex
ity
f
o
r
r
ec
o
m
m
en
d
atio
n
m
o
d
els.
T
h
is
co
m
p
lex
ity
ar
is
es
f
r
o
m
s
h
o
r
ter
u
s
er
-
item
s
eq
u
en
ce
s
a
n
d
n
o
is
e
s
en
s
itiv
ity
,
m
ak
in
g
it
h
ar
d
f
o
r
m
o
d
els
to
id
en
tify
u
s
er
p
r
ef
e
r
en
ce
s
.
Mo
d
els
s
tr
u
g
g
le
with
p
atter
n
r
ec
o
g
n
itio
n
d
u
e
to
lim
ited
d
ata
an
d
in
cr
ea
s
ed
n
o
is
e
as
r
an
d
o
m
i
n
ter
ac
tio
n
s
m
ay
b
e
m
is
in
ter
p
r
eted
as
s
ig
n
if
ican
t
p
atter
n
s
.
As
a
r
esu
lt,
m
o
d
els
m
ay
lear
n
n
o
is
e
p
atter
n
s
in
s
tead
o
f
tr
u
e
p
r
e
f
er
e
n
ce
s
,
r
ed
u
cin
g
th
eir
r
ec
o
m
m
e
n
d
atio
n
p
er
f
o
r
m
an
ce
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
E
ffects o
f sp
a
r
s
e
d
a
ta
s
ets o
n
time
in
terva
l
-
a
w
a
r
e
s
elf
-
a
tten
ti
o
n
s
eq
u
en
tia
l
…
(
Weis
h
a
n
Oo
i
)
2769
SASR
ec
h
as
a
s
im
p
ler
ar
c
h
itectu
r
e
with
o
u
t
tim
e
i
n
ter
v
al
-
aw
ar
e
em
b
e
d
d
in
g
s
co
m
p
ar
ed
to
T
iSAS
R
ec
.
I
n
cr
ea
s
in
g
d
r
o
p
o
u
t
r
ates
in
S
ASR
ec
h
elp
s
b
y
r
an
d
o
m
ly
d
r
o
p
p
in
g
n
eu
r
o
n
s
d
u
r
in
g
tr
ain
i
n
g
,
lead
in
g
to
b
etter
g
en
er
aliza
tio
n
,
an
d
im
p
r
o
v
ed
r
an
k
in
g
ac
c
u
r
ac
y
.
Fo
r
ex
am
p
le,
with
tu
n
ed
m
o
d
el
p
ar
am
eter
s
o
f
2
atten
tio
n
b
lo
ck
s
an
d
0
.
3
d
r
o
p
o
u
t
r
ates
f
o
r
SASR
ec
,
it
ac
h
iev
ed
0
.
2
6
4
6
f
o
r
th
e
NDCG@
1
0
s
co
r
in
g
o
n
T
ao
b
a
o
,
s
lig
h
tly
h
ig
h
er
th
an
T
iSAS
R
ec
’
s
b
est
NDCG
@
1
0
s
co
r
in
g
o
f
0
.
2
5
3
9
.
T
h
is
in
d
icate
s
th
at
in
an
ex
tr
em
ely
s
p
ar
s
e
s
ettin
g
lik
e
T
ao
b
ao
d
ataset,
in
co
r
p
o
r
atin
g
tim
e
in
ter
v
als
is
y
et
t
o
y
ield
a
lar
g
e
ac
c
u
r
ac
y
g
ai
n
o
v
e
r
th
e
s
im
p
ler
SASR
ec
.
Ho
wev
er
,
h
ig
h
er
d
r
o
p
o
u
t
r
ates
r
e
q
u
ir
e
m
o
r
e
ep
o
ch
s
to
co
n
v
er
g
e,
ex
te
n
d
in
g
tr
ain
in
g
tim
es.
T
h
e
tr
ain
in
g
tim
e
o
f
SASR
ec
in
cr
ea
s
ed
f
o
r
ap
p
r
o
x
im
ately
3
2
.
2
6
%
f
r
o
m
1
4
,
2
2
4
s
at
a
0
.
2
d
r
o
p
o
u
t
r
ate
to
1
8
,
8
1
3
s
at
a
0
.
4
d
r
o
p
o
u
t
r
ate
b
y
r
ef
er
r
in
g
t
o
T
ab
le
6
.
C
o
n
v
e
r
s
ely
,
T
iSAS
R
ec
's
tr
ain
in
g
tim
e
d
ec
r
ea
s
ed
f
o
r
ap
p
r
o
x
im
ately
7
0
.
7
7
%
f
r
o
m
1
6
,
1
2
3
s
to
4
,
7
1
3
s
wh
e
n
th
e
d
r
o
p
o
u
t
r
ate
in
c
r
ea
s
ed
to
0
.
3
,
h
itti
n
g
a
s
wee
t
s
p
o
t
f
o
r
h
ig
h
er
s
p
ar
s
ity
d
ataset,
w
h
er
e
it
ef
f
ec
tiv
ely
r
e
g
u
lar
ized
th
e
m
o
d
el,
allo
win
g
it
to
co
n
v
er
g
e
m
o
r
e
q
u
ick
l
y
b
y
p
r
e
v
en
tin
g
o
v
e
r
f
itti
n
g
wit
h
o
u
t
s
ev
er
ely
im
p
ac
tin
g
its
ca
p
ac
ity
to
lear
n
f
r
o
m
th
e
s
p
a
r
s
e
d
ata.
Ho
wev
er
,
f
u
r
th
er
in
cr
em
en
t
o
f
d
r
o
p
o
u
t
r
ate
to
0
.
4
led
to
lo
n
g
er
tr
ai
n
in
g
tim
es
d
u
e
to
u
n
d
er
f
itti
n
g
an
d
o
p
tim
izatio
n
d
if
f
icu
lties
,
wh
ich
p
r
o
lo
n
g
ed
th
e
tr
ain
in
g
p
r
o
ce
s
s
as
th
e
m
o
d
el
s
tr
u
g
g
led
t
o
ca
p
t
u
r
e
t
h
e
lim
ited
p
atter
n
s
av
ailab
le
in
th
e
s
p
ar
s
e
d
atase
t.
T
h
er
ef
o
r
e,
a
n
o
tab
le
d
i
f
f
er
en
ce
was
o
b
s
er
v
ed
i
n
tr
ain
in
g
ef
f
icien
cy
wh
er
e
T
iSAS
R
ec
o
f
ten
co
n
v
er
g
e
d
f
aster
with
th
e
ex
i
s
ten
ce
o
f
t
h
e
tim
e
-
s
p
an
in
ter
v
al.
I
n
s
ev
er
al
co
n
f
ig
u
r
atio
n
s
,
T
iSAS
R
ec
’
s
tr
ain
in
g
tim
e
was
s
ig
n
if
ican
tly
s
h
o
r
ter
th
an
SASR
ec
’
s
o
n
th
e
s
am
e
d
ata.
Fo
r
in
s
tan
ce
,
th
e
SASR
ec
with
4
s
elf
-
atten
tio
n
b
lo
c
k
s
o
v
er
3
7
,
7
8
2
s
to
tr
ain
o
n
T
a
o
b
ao
,
wh
er
ea
s
T
iSAS
R
ec
f
in
is
h
ed
in
ab
o
u
t
5
,
9
6
3
s
with
th
e
s
am
e
n
u
m
b
er
o
f
b
lo
ck
s
p
lu
s
with
t
h
e
tim
e
-
s
p
an
o
f
2
5
6
.
Ho
wev
e
r
,
f
u
r
th
er
i
n
cr
em
en
t
o
f
d
r
o
p
o
u
t
r
ate
to
0
.
4
led
to
lo
n
g
er
tr
ain
in
g
tim
es
d
u
e
to
u
n
d
er
f
itti
n
g
an
d
o
p
tim
i
za
tio
n
d
if
f
icu
lties
,
wh
ich
p
r
o
lo
n
g
ed
t
h
e
tr
ain
in
g
p
r
o
ce
s
s
as
th
e
m
o
d
el
s
tr
u
g
g
le
d
to
ca
p
tu
r
e
th
e
lim
ited
p
atter
n
s
av
ailab
le
in
th
e
s
p
ar
s
e
d
ataset.
T
ab
le
6
.
T
r
ai
n
in
g
tim
e
c
o
m
p
ar
is
o
n
tab
le
b
etwe
en
SASR
ec
an
d
T
iSAS
R
ec
f
o
r
T
ao
b
ao
d
atase
t
D
a
t
a
s
e
t
M
o
d
e
l
p
a
r
a
me
t
e
r
s
S
A
S
R
e
c
t
r
a
i
n
i
n
g
t
i
m
e
Ti
S
A
S
R
e
c
t
r
a
i
n
i
n
g
t
i
m
e
Ta
o
b
a
o
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
2
D
r
o
p
o
u
t
r
a
t
e
=
0
.
2
Ti
me
-
s
p
a
n
=
2
5
6
(
N
o
t
a
p
p
l
i
c
a
b
l
e
f
o
r
S
A
S
R
e
c
)
3
h
5
7
m
4
s
(
1
4
,
2
2
4
s)
4
h
2
8
m
4
3
s (
1
6
,
1
2
3
s)
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
4
D
r
o
p
o
u
t
r
a
t
e
=
0
.
2
Ti
me
-
s
p
a
n
=
2
5
6
(
N
o
t
a
p
p
l
i
c
a
b
l
e
f
o
r
S
A
S
R
e
c
)
10
h
2
9
m
4
2
s (3
7
,
7
8
2
s)
1
h
3
9
m
2
3
s (
5
,
9
6
3
s)
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
8
D
r
o
p
o
u
t
r
a
t
e
=
0
.
2
Ti
me
-
s
p
a
n
=
2
5
6
(
N
o
t
a
p
p
l
i
c
a
b
l
e
f
o
r
S
A
S
R
e
c
)
9
h
3
7
m
6
s
(
3
4
,
6
2
6
s)
4
h
3
9
m
3
5
s (
1
6
,
7
7
5
s)
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
2
D
r
o
p
o
u
t
r
a
t
e
=
0
.
3
Ti
me
-
s
p
a
n
=
2
5
6
(
N
o
t
a
p
p
l
i
c
a
b
l
e
f
o
r
S
A
S
R
e
c
)
4
h
1
2
m
2
6
s (
1
5
,
1
4
6
s)
1
h
1
8
m
3
3
s (
4
,
7
1
3
s)
A
t
t
e
n
t
i
o
n
b
l
o
c
k
=
2
D
r
o
p
o
u
t
r
a
t
e
=
0
.
4
Ti
me
-
s
p
a
n
=
2
5
6
(
N
o
t
a
p
p
l
i
c
a
b
l
e
f
o
r
S
A
S
R
e
c
)
5
h
1
3
m
3
3
s (
1
8
,
8
1
3
s)
3
h
3
9
m
5
s
(
1
3
,
1
4
5
s)
B
ased
o
n
th
e
p
er
f
o
r
m
a
n
ce
o
f
T
iSAS
R
ec
in
T
ao
b
ao
d
ataset,
in
cr
ea
s
in
g
th
e
tim
e
-
s
p
an
v
alu
e
im
p
r
o
v
es
NDCG
@
1
0
an
d
HR
@
1
0
.
W
ith
a
tim
e,
s
p
an
o
f
1
,
0
2
4
,
ND
C
G
@
1
0
an
d
HR
@
1
0
in
cr
ea
s
ed
f
r
o
m
0
.
2
3
4
6
an
d
0
.
3
4
0
7
to
0
.
2
5
2
7
an
d
0
.
3
5
6
5
.
W
ith
a
tim
e,
s
p
an
o
f
2
,
0
4
8
,
t
h
ey
f
u
r
t
h
er
in
cr
ea
s
ed
to
0
.
2
5
3
9
an
d
0
.
3
6
0
1
.
T
h
e
h
ig
h
er
tim
e
-
s
p
a
n
v
alu
e
ca
p
tu
r
es
lo
n
g
-
ter
m
d
ep
en
d
en
cies
an
d
ass
ig
n
s
m
o
r
e
ac
cu
r
ate
atten
tio
n
weig
h
ts
b
ased
o
n
tem
p
o
r
al
d
is
tan
ce
,
wh
ich
i
s
esp
ec
ially
ef
f
ec
tiv
e
f
o
r
a
s
p
ar
s
e
d
ataset.
Ho
wev
er
,
th
e
tr
a
in
in
g
tim
e
s
lig
h
tly
in
cr
ea
s
ed
b
y
a
b
o
u
t
1
7
m
in
u
tes
wh
en
th
e
tim
e
-
s
p
an
v
alu
e
is
r
aised
f
r
o
m
1
,
0
2
4
to
2
,
0
4
8
.
T
h
is
is
b
ec
au
s
e
th
e
in
cr
em
en
t o
f
th
e
m
ax
im
u
m
r
a
n
g
e
o
f
tim
e
in
ter
v
als b
etwe
en
u
s
er
-
item
in
ter
ac
tio
n
s
r
eq
u
ir
es
a
b
r
o
ad
er
atten
tio
n
s
co
p
e
in
th
e
s
p
ar
s
e
d
ataset
an
d
m
o
r
e
co
m
p
le
x
ca
lcu
latio
n
s
f
o
r
atten
tio
n
weig
h
ts
,
as
it
in
tr
o
d
u
ce
s
m
o
r
e
d
is
tan
t
item
s
in
to
a
s
in
g
le
u
s
er
-
item
in
ter
ac
tio
n
v
ec
to
r
.
C
o
n
tr
ar
y
to
Mo
v
ieL
en
s
-
1
M
,
th
e
lack
o
f
tr
ain
in
g
d
ata
in
h
ig
h
er
s
p
ar
s
ity
d
ataset
lik
e
T
ao
b
ao
lim
its
th
e
p
er
f
o
r
m
a
n
ce
o
f
th
e
r
ec
o
m
m
en
d
atio
n
m
o
d
el
d
u
e
to
th
e
s
h
o
r
te
r
len
g
th
o
f
u
s
er
-
item
in
ter
ac
tio
n
s
.
As
th
e
tim
e
-
s
p
an
v
alu
e
in
cr
ea
s
es,
m
o
r
e
d
ata
ca
n
b
e
i
n
clu
d
ed
f
o
r
tr
ain
in
g
T
iSAS
R
ec
,
allo
win
g
it
to
lear
n
co
r
e
p
atter
n
s
th
at
r
e
p
r
esen
t
u
s
er
p
r
ef
er
e
n
ce
s
in
s
p
ar
s
e
d
at
asets
,
wh
ich
is
les
s
lik
ely
to
b
e
m
is
d
ir
ec
ted
in
d
en
s
er
d
atasets
.
B
a
s
ed
o
n
th
e
ex
p
er
im
en
tal
r
esu
lts
,
T
iSAS
R
ec
h
as
a
g
en
er
ally
lo
wer
tr
ain
in
g
tim
e
co
m
p
ar
e
d
to
th
e
r
esu
lts
f
r
o
m
SAS
R
e
c
m
o
d
el.
Ho
wev
er
,
th
e
o
v
er
a
ll
tr
ain
in
g
tim
e
f
o
r
T
iSAS
R
ec
is
s
h
o
r
ter
th
an
SASR
ec
ac
r
o
s
s
d
if
f
er
en
t
p
ar
am
et
er
co
n
f
ig
u
r
atio
n
s
,
m
a
k
in
g
T
iS
ASR
ec
s
u
itab
le
f
o
r
h
ig
h
er
s
p
ar
s
ity
d
ataset
with
less
er
p
r
o
ce
s
s
in
g
tim
e
d
esp
ite
its
s
lig
h
tly
lo
wer
ac
cu
r
ac
y
f
r
o
m
th
e
r
esu
lt.
I
n
T
iSAS
R
ec
,
i
t
in
tr
o
d
u
ce
d
th
e
tim
e
-
s
p
an
in
ter
v
al
as
o
n
e
o
f
th
e
attem
p
ts
o
n
im
p
r
o
v
em
en
t
o
f
atten
tio
n
ef
f
icien
cy
.
Similar
ly
in
s
tu
d
ies
o
f
[
2
4
]
,
[
2
5
]
,
th
ey
b
o
th
im
p
lem
en
t
d
is
tin
ct
s
tr
ateg
ies
o
f
s
p
ar
s
e
cr
o
s
s
-
atten
tio
n
m
ec
h
a
n
is
m
s
an
d
lev
er
a
g
e
cr
o
s
s
-
atten
tio
n
to
in
co
r
p
o
r
ate
c
o
n
tex
tu
al
attr
ib
u
tes
,
r
esp
ec
tiv
ely
to
im
p
r
o
v
e
atten
tio
n
ef
f
icien
cy
i
n
s
p
ar
s
e
d
atasets
.
Gen
er
ally
,
it
s
er
v
es
as
an
en
r
ich
in
g
ap
p
r
o
ac
h
o
n
th
e
s
p
ar
s
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
3
,
J
u
n
e
2
0
2
6
:
2
7
6
1
-
2
7
7
3
2770
in
ter
ac
tio
n
d
ata
with
au
x
iliar
y
s
ig
n
als.
T
h
u
s
,
it
ca
n
ac
h
iev
e
f
aster
tr
ain
in
g
an
d
m
ain
tain
r
el
ativ
ely
r
em
ar
k
a
b
le
p
er
f
o
r
m
an
ce
ev
e
n
in
h
ig
h
ly
s
p
ar
s
e
en
v
ir
o
n
m
en
ts
,
esp
ec
ially
in
ex
tr
em
ely
s
p
ar
s
e
d
atasets
with
a
s
p
ar
s
ity
r
ate
m
o
r
e
th
an
9
5
%
an
d
a
s
h
o
r
ter
av
er
ag
e
s
eq
u
e
n
ce
len
g
th
.
T
h
i
s
s
u
g
g
ests
b
etter
tr
ain
in
g
ef
f
i
cien
cy
b
y
r
etain
i
n
g
lo
wer
v
alu
e
o
f
atten
tio
n
b
lo
c
k
s
an
d
co
n
s
id
er
in
g
o
th
er
co
n
t
ex
tu
al
v
ar
ia
b
les
th
at
h
elp
in
i
m
p
r
o
v
i
n
g
atten
ti
o
n
ef
f
icien
cy
wh
e
n
tr
ain
in
g
with
an
ex
tr
em
ely
s
p
a
r
s
e
d
ataset
in
s
tead
o
f
test
in
g
with
m
o
r
e
atten
tio
n
b
lo
c
k
s
.
4
.
3
.
At
t
ent
io
n
weig
ht
dis
t
rib
utio
n o
n bo
t
h m
o
dels
in T
a
o
ba
o
da
t
a
s
et
T
h
e
v
alu
es
o
f
NDCG@
1
0
an
d
HR
@
1
0
g
r
ad
u
ally
im
p
r
o
v
e
as
th
e
v
alu
e
o
f
tim
e
-
s
p
an
m
o
d
el
p
ar
am
eter
s
in
cr
ea
s
es.
T
ab
le
s
2
-
5
s
h
o
w
th
at
b
o
th
SASR
ec
an
d
T
iSAS
R
ec
m
o
d
el
h
ad
s
ig
n
if
i
ca
n
t im
p
r
o
v
em
en
ts
in
p
er
f
o
r
m
an
ce
ev
alu
atio
n
m
etr
ics
f
o
r
T
ao
b
ao
d
ataset.
Ho
wev
er
,
th
e
v
alu
es
o
f
NDCG@
1
0
an
d
HR
@
1
0
in
T
iSAS
R
ec
ar
e
r
elativ
ely
lo
we
r
th
an
th
e
v
alu
es
o
f
NDCG@
1
0
an
d
HR
@
1
0
in
SASR
ec
f
o
r
T
ao
b
ao
d
ataset.
T
h
is
is
f
u
r
th
er
an
aly
s
ed
in
Fi
g
u
r
e
1
,
wh
ich
illu
s
tr
ates
th
e
atten
tio
n
weig
h
t
d
is
tr
ib
u
tio
n
m
ap
o
f
th
e
last
1
0
u
s
e
r
-
i
t
e
m
s
e
q
u
e
n
c
e
s
i
n
b
o
t
h
m
o
d
e
l
s
.
I
n
F
i
g
u
r
e
s
1
(
a
)
a
n
d
(
b
)
,
t
h
e
x
-
a
x
i
s
r
e
p
r
e
s
e
n
t
s
t
h
e
p
o
s
i
t
i
o
n
o
f
i
t
e
m
s
i
n
s
e
q
u
e
n
c
e
,
a
n
d
e
a
c
h
p
o
i
n
t
o
n
t
h
e
y
-
a
x
i
s
c
o
r
r
e
s
p
o
n
d
s
t
o
a
s
p
e
ci
f
i
c
t
i
m
e
s
t
e
p
i
n
s
e
q
u
e
n
c
e
o
f
u
s
e
r
-
i
t
e
m
i
n
t
e
r
ac
t
i
o
n
s
.
T
h
e
n
u
m
b
e
r
s
o
f
s
tar
ted
f
r
o
m
0
to
9
wh
ich
s
tatin
g
f
r
o
m
th
e
m
o
s
t
r
ec
en
t
to
th
e
o
ld
est
ite
m
p
o
s
itio
n
.
On
th
e
o
th
er
h
an
d
,
th
e
y
-
ax
is
r
ep
r
esen
ts
th
e
tim
e
-
s
tep
s
at
w
h
ich
th
e
m
o
d
el
m
ak
es
r
ec
o
m
m
en
d
atio
n
s
,
wh
er
e
0
in
d
icate
s
as
th
e
f
ir
s
t
item
a
n
d
9
r
ef
e
r
s
as
th
e
last
item
to
b
e
r
ec
o
m
m
en
d
ed
in
a
tim
e
-
s
tep
.
I
n
Fig
u
r
e
1
(
a)
,
th
e
p
ar
am
eter
s
ettin
g
s
o
f
th
e
SA
SR
ec
atten
tio
n
weig
h
t
d
is
tr
ib
u
tio
n
m
a
p
ar
e
c
o
n
f
i
g
u
r
ed
with
d
ef
a
u
lt
p
ar
am
eter
v
alu
es
(
atten
tio
n
b
lo
ck
s
=2
,
d
r
o
p
o
u
t
r
ate
=0
.
2
’
)
wh
ile
Fig
u
r
e
1
(
b
)
is
co
n
f
i
g
u
r
e
d
with
‘
att
en
tio
n
b
lo
c
k
s
=2
’
,
‘
d
r
o
p
o
u
t r
ate
=
0
.
2
’
an
d
‘
tim
e
-
s
p
an
=2
5
6
’
f
o
r
th
e
T
iSAS
R
ec
m
o
d
el.
(
a)
(
b
)
Fig
u
r
e
1
.
Atten
tio
n
weig
h
t d
is
tr
ib
u
tio
n
m
a
p
in
l
ast 1
0
u
s
er
-
item
s
eq
u
en
ce
of
(
a)
SASR
ec
at
ten
tio
n
h
ea
tm
ap
o
n
T
ao
b
ao
an
d
(
b
)
T
iSAS
R
ec
atten
tio
n
h
ea
tm
ap
o
n
T
a
o
b
ao
Evaluation Warning : The document was created with Spire.PDF for Python.