I
n
t
e
r
n
at
ion
al
Jou
r
n
al
of
E
lec
t
r
ical
an
d
Com
p
u
t
e
r
E
n
gin
e
e
r
in
g
(
I
JE
CE
)
Vol.
15
,
No.
1
,
F
e
br
ua
r
y
20
25
,
pp.
711
~
718
I
S
S
N:
2088
-
8708
,
DO
I
:
10
.
11591/i
jec
e
.
v
15
i
1
.
pp
7
11
-
718
711
Jou
r
n
al
h
omepage
:
ht
tp:
//
ij
e
c
e
.
iaes
c
or
e
.
c
om
A
n
al
ysi
s of
b
ig
d
a
t
a f
r
o
m
N
e
w
Y
or
k
t
axi
t
r
i
p
2023:
r
e
v
e
n
u
e
p
r
e
d
ic
t
io
n
u
si
n
g
or
d
in
ar
y l
e
a
st
sq
u
a
r
e
s
sol
u
t
io
n
an
d
li
m
ite
d
-
m
e
m
o
r
y
B
r
oyd
e
n
-
Fle
t
c
h
e
r
-
G
ol
d
f
ar
b
-
S
h
a
n
n
o
al
gor
it
h
m
s
S
ar
a
Rhou
as
,
Nor
e
li
s
lam
E
l
Ham
i
E
ngi
ne
e
r
in
g
S
c
ie
nc
e
L
a
bor
a
to
r
y, N
a
ti
ona
l
S
c
hool
of
A
ppl
ie
d S
c
ie
nc
e
s
, I
bn
T
of
a
il
U
ni
ve
r
s
it
y, K
e
ni
tr
a
, M
or
oc
c
o
Ar
t
icle
I
n
f
o
AB
S
T
RA
CT
A
r
ti
c
le
h
is
tor
y
:
R
e
c
e
ived
J
un
14,
2024
R
e
vis
e
d
S
e
p
6,
2024
Ac
c
e
pted
Oc
t
1,
2024
T
h
i
s
s
t
u
d
y
ex
p
l
o
res
t
h
e
p
re
d
i
c
t
i
o
n
o
f
t
ax
i
t
r
i
p
fare
s
u
s
i
n
g
t
w
o
l
i
n
ear
reg
res
s
i
o
n
me
t
h
o
d
s
:
n
o
rma
l
eq
u
at
i
o
n
s
(
o
rd
i
n
ar
y
l
ea
s
t
s
q
u
ar
es
s
o
l
u
t
i
o
n
(O
L
S
)
)
an
d
l
i
m
i
t
e
d
-
memo
r
y
Bro
y
d
en
-
F
l
et
ch
er
-
G
o
l
d
far
b
-
Sh
a
n
n
o
(
L
-
BFG
S
)
.
U
t
i
l
i
zi
n
g
a
d
a
t
as
e
t
o
f
N
e
w
Y
o
r
k
Ci
t
y
y
e
l
l
o
w
t
ax
i
t
ri
p
s
fr
o
m
2
0
2
3
,
t
h
e
an
al
y
s
i
s
i
n
v
o
l
v
e
s
d
a
t
a
c
l
ean
i
n
g
,
feat
u
re
en
g
i
n
eeri
n
g
,
a
n
d
mo
d
e
l
t
ra
i
n
i
n
g
.
T
h
e
d
at
a
co
n
s
i
s
t
s
o
f
o
v
er
1
2
mi
l
l
i
o
n
reco
rd
s
,
man
ag
e
d
,
an
d
p
ro
c
es
s
ed
t
h
at
i
n
v
o
l
v
e
s
co
n
fi
g
u
r
i
n
g
t
h
e
Sp
ar
k
d
r
i
v
er
an
d
ex
ec
u
t
o
r
m
emo
ry
t
o
eff
i
ci
e
n
t
l
y
p
ro
ce
s
s
t
h
e
Par
q
u
e
t
-
fo
rma
t
d
at
a
s
t
o
red
o
n
h
a
d
o
o
p
d
i
s
t
ri
b
u
t
ed
f
i
l
e
s
y
s
t
em
(
H
D
FS
)
.
K
ey
feat
u
re
s
i
n
f
l
u
e
n
ci
n
g
fare
amo
u
n
t
,
s
u
ch
as
p
as
s
en
g
er
co
u
n
t
,
t
r
i
p
d
i
s
t
a
n
ce,
fare
amo
u
n
t
,
an
d
t
i
p
amo
u
n
t
,
w
ere
an
al
y
zed
fo
r
co
rrel
at
i
o
n
.
Mo
d
el
s
w
ere
t
rai
n
ed
o
n
a
n
8
0
-
2
0
t
rai
n
-
t
e
s
t
s
p
l
i
t
,
an
d
t
h
e
i
r
p
erfo
rman
ce
w
as
ev
al
u
at
e
d
u
s
i
n
g
r
o
o
t
-
mea
n
-
s
q
u
are
erro
r
(
RMSE
)
an
d
mean
s
q
u
are
d
erro
r
(
MSE
)
.
Res
u
l
t
s
s
h
o
w
t
h
a
t
b
o
t
h
met
h
o
d
s
p
r
o
v
i
d
e
co
m
p
a
rab
l
e
accu
rac
y
,
w
i
t
h
s
l
i
g
h
t
d
i
fferen
ce
s
i
n
co
eff
i
ci
e
n
t
s
an
d
t
rai
n
i
n
g
t
i
me.
A
d
d
i
t
i
o
n
a
l
l
y
,
v
en
d
o
r
p
erfo
rma
n
ce
met
r
i
cs
,
i
n
c
l
u
d
i
n
g
t
o
t
al
t
r
i
p
s
,
av
era
g
e
t
ri
p
d
i
s
t
an
ce,
fare
amo
u
n
t
,
an
d
t
i
p
am
o
u
n
t
,
w
ere
an
a
l
y
ze
d
t
o
rev
ea
l
t
re
n
d
s
an
d
i
n
f
o
rm
s
t
rat
e
g
i
c
d
eci
s
i
o
n
s
fo
r
fl
eet
ma
n
ag
eme
n
t
.
T
h
i
s
co
mp
re
h
e
n
s
i
v
e
a
n
al
y
s
i
s
d
em
o
n
s
t
ra
t
es
t
h
e
eff
i
cacy
o
f
l
i
n
ear
reg
re
s
s
i
o
n
t
ec
h
n
i
q
u
es
i
n
p
re
d
i
c
t
i
n
g
t
ax
i
fares
an
d
o
ffer
s
v
al
u
ab
l
e
i
n
s
i
g
h
t
s
f
o
r
o
p
t
i
m
i
zi
n
g
t
ax
i
o
p
era
t
i
o
n
s
.
K
e
y
w
o
r
d
s
:
B
ig
da
ta
Da
ta
a
na
lys
is
L
inea
r
r
e
gr
e
s
s
ion
M
a
c
hine
lea
r
ning
S
pa
r
k
Th
i
s
i
s
a
n
o
p
en
a
c
ces
s
a
r
t
i
c
l
e
u
n
d
e
r
t
h
e
CC
B
Y
-
SA
l
i
ce
n
s
e.
C
or
r
e
s
pon
din
g
A
u
th
or
:
R
houa
s
S
a
r
a
E
nginee
r
ing
S
c
ienc
e
L
a
bor
a
tor
y
,
Na
ti
ona
l
S
c
hool
of
Applied
S
c
ienc
e
s
,
I
bn
T
of
a
il
Unive
r
s
it
y
Av.
de
L
'Unive
r
s
it
é
,
Ké
nit
r
a
,
M
a
r
o
cco
E
mail:
r
houa
s
.
s
a
r
a
@gmail.
c
om
1.
I
NT
RODU
C
T
I
ON
P
r
e
dicting
taxi
t
r
ip
f
a
r
e
s
a
c
c
ur
a
tely
is
a
c
r
it
ica
l
tas
k
f
o
r
bo
th
f
lee
t
ope
r
a
tor
s
a
nd
pa
s
s
e
nge
r
s
in
u
r
ba
n
tr
a
ns
por
tation
s
ys
tems
.
W
it
h
the
a
dve
nt
of
big
da
ta
a
nd
a
dva
nc
e
d
a
na
lyt
ica
l
tool
s
,
it
is
now
po
s
s
ibl
e
to
leve
r
a
ge
e
xtens
ive
da
ta
s
e
ts
to
ga
in
in
s
ight
s
int
o
f
a
r
e
de
ter
mi
na
nts
a
nd
im
pr
ove
f
a
r
e
pr
e
diction
mo
de
ls
[
1]
.
T
his
s
tudy
f
oc
us
e
s
on
uti
li
z
ing
li
ne
a
r
r
e
gr
e
s
s
ion
tec
hniques
to
pr
e
dict
taxi
t
r
ip
f
a
r
e
s
us
ing
da
ta
f
r
om
Ne
w
Yor
k
C
it
y's
ye
ll
ow
taxi
f
lee
t
f
o
r
the
e
nti
r
e
ye
a
r
of
2023
[
2]
.
B
y
c
ompar
ing
two
p
r
omi
ne
nt
r
e
gr
e
s
s
ion
methods
,
nor
mal
e
qua
ti
ons
o
r
dinar
y
lea
s
t
s
qua
r
e
s
(
OL
S
)
s
olut
ion
a
nd
li
mi
ted
-
memor
y
B
r
oyde
n
-
F
letc
he
r
-
Goldf
a
r
b
-
S
ha
nno
(
L
-
B
F
GS
)
,
we
a
im
to
identi
f
y
th
e
mos
t
e
f
f
e
c
ti
ve
a
ppr
oa
c
h
f
o
r
f
a
r
e
pr
e
diction.
Ac
c
ur
a
te
f
a
r
e
pr
e
dictions
c
a
n
e
nha
nc
e
ope
r
a
ti
ona
l
e
f
f
icie
nc
y,
opti
mi
z
e
pr
icing
s
tr
a
tegie
s
,
a
nd
im
p
r
ove
c
us
tom
e
r
s
a
ti
s
f
a
c
ti
on
by
pr
ovidi
ng
t
r
a
ns
pa
r
e
nt
a
nd
pr
e
dicta
b
le
f
a
r
e
e
s
ti
mate
s
[
3]
,
[
4]
.
Ne
w
Yor
k
C
it
y’
s
ye
ll
ow
taxi
da
tas
e
t
pr
ovides
a
r
ich
s
our
c
e
of
inf
or
mation,
e
nc
ompas
s
ing
mi
ll
ions
of
t
r
ip
r
e
c
or
ds
with
diver
s
e
a
tt
r
ibut
e
s
s
uc
h
a
s
t
r
i
p
dis
tanc
e
,
pa
s
s
e
nge
r
c
ount,
f
a
r
e
a
mount
,
ti
p
a
m
ount,
a
nd
tempor
a
l
de
tails
[
5]
.
T
he
lar
ge
volum
e
of
da
ta
a
l
lows
f
or
a
de
tailed
a
na
lys
is
of
f
a
r
e
de
ter
mi
na
nts
a
nd
the
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
15
,
No.
1
,
F
e
br
ua
r
y
20
25
:
711
-
718
712
de
ve
lopm
e
nt
of
r
obus
t
pr
e
dictive
models
.
How
e
ve
r
,
the
pr
e
s
e
nc
e
of
nu
ll
va
lues
a
nd
outl
ier
s
ne
c
e
s
s
it
a
te
s
r
igor
ous
da
ta
c
lea
ning
a
nd
p
r
e
pr
oc
e
s
s
ing.
T
h
is
s
tu
dy
s
ys
tema
ti
c
a
ll
y
a
ddr
e
s
s
e
s
thes
e
c
ha
ll
e
nge
s
,
e
ns
ur
ing
the
int
e
gr
it
y
a
nd
r
e
li
a
bil
it
y
o
f
the
da
tas
e
t.
F
e
a
tur
e
e
nginee
r
ing
tec
hniques
a
r
e
e
mpl
oye
d
to
e
xtr
a
c
t
me
a
ningf
ul
ins
ight
s
f
r
om
the
da
ta,
s
uc
h
a
s
tempor
a
l
pa
tt
e
r
n
s
in
tr
ip
f
r
e
que
nc
ies
a
nd
f
a
r
e
va
r
iations
a
c
r
os
s
dif
f
e
r
e
nt
ve
ndor
s
[
6]
,
[
7
]
.
I
n
a
ddit
ion
to
buil
ding
p
r
e
dictive
models
,
thi
s
s
tu
dy
c
onduc
ts
a
c
ompr
e
he
ns
ive
c
or
r
e
lation
a
na
lys
is
to
unde
r
s
tand
the
r
e
lations
hips
be
twe
e
n
va
r
ious
tr
ip
a
tt
r
ibut
e
s
a
nd
the
f
a
r
e
a
mount
.
B
y
e
xa
mi
ni
ng
thes
e
c
or
r
e
lations
,
we
identif
y
the
mos
t
s
igni
f
ica
nt
f
e
a
tur
e
s
inf
luenc
ing
f
a
r
e
pr
e
dictions
.
T
he
pe
r
f
o
r
manc
e
of
the
r
e
gr
e
s
s
ion
models
is
e
va
lu
a
ted
us
ing
metr
ics
s
u
c
h
a
s
r
oot
-
mea
n
-
s
qua
r
e
e
r
r
or
(
R
M
S
E
)
a
nd
mea
n
s
qua
r
e
d
e
r
r
o
r
(
M
S
E
)
,
pr
ovid
ing
a
qua
nti
tative
mea
s
ur
e
o
f
their
a
c
c
ur
a
c
y.
F
ur
ther
mor
e
,
the
s
tudy
de
lves
int
o
ve
ndor
pe
r
f
or
manc
e
a
na
lys
is
,
c
ompar
ing
ke
y
pe
r
f
o
r
man
c
e
indi
c
a
tor
s
li
ke
tot
a
l
tr
ips
,
a
ve
r
a
ge
t
r
ip
dis
tanc
e
,
f
a
r
e
a
mount
,
a
nd
ti
p
a
mount
a
c
r
os
s
di
f
f
e
r
e
nt
ve
ndor
s
.
T
his
holi
s
ti
c
a
ppr
oa
c
h
not
only
high
li
ghts
the
e
f
f
e
c
ti
ve
ne
s
s
of
li
ne
a
r
r
e
gr
e
s
s
ion
tec
hniques
in
f
a
r
e
pr
e
diction
but
a
ls
o
of
f
e
r
s
va
luable
ins
ight
s
int
o
ve
ndor
ope
r
a
ti
ons
,
c
ontr
ibut
ing
to
the
ove
r
a
ll
opti
mi
z
a
ti
on
of
taxi
s
e
r
v
ice
s
in
Ne
w
Yor
k
C
ity
[
8]
.
2.
M
E
T
HO
D
Our
a
ppr
oa
c
h
to
a
na
lyzing
Ne
w
Yor
k
C
ity
taxi
tr
ip
da
ta
in
2023
c
ombi
ne
s
the
Apa
c
he
s
pa
r
k
platf
or
m
a
nd
li
ne
a
r
r
e
gr
e
s
s
ion
models
f
or
f
a
r
e
pr
e
d
iction.
S
pa
r
k
ha
ndles
lar
ge
da
tas
e
ts
,
e
na
bli
ng
e
f
f
ici
e
nt
da
ta
c
lea
ning,
tr
a
ns
f
o
r
mation,
a
nd
a
na
lys
is
.
Af
te
r
loading
the
da
ta
f
r
om
p
a
r
que
t
f
il
e
s
a
nd
f
il
ter
ing
invalid
r
e
c
or
ds
,
we
us
e
li
ne
a
r
r
e
gr
e
s
s
ion
to
pr
e
dict
f
a
r
e
s
ba
s
e
d
on
f
e
a
tur
e
s
li
ke
pa
s
s
e
nge
r
c
ount,
tr
ip
dis
tanc
e
,
f
a
r
e
,
a
nd
ti
ps
.
W
e
im
pleme
nt
two
methods
f
o
r
li
ne
a
r
r
e
gr
e
s
s
ion
,
n
o
r
mal
e
qua
ti
ons
f
or
s
maller
da
ta
a
nd
L
-
B
F
GS
f
o
r
high
-
dim
e
ns
ional
da
ta
[
9]
.
T
he
da
ta
is
s
pli
t
int
o
tr
a
ini
ng
a
nd
tes
t
s
e
ts
to
e
va
luate
pe
r
f
or
manc
e
us
ing
R
M
S
E
a
nd
M
S
E
.
W
e
a
ls
o
a
s
s
e
s
s
taxi
ve
ndor
s
'
pe
r
f
o
r
manc
e
by
a
na
lyzing
met
r
ics
s
uc
h
a
s
tr
ip
c
ounts
,
a
ve
r
a
ge
dis
tanc
e
,
f
a
r
e
,
a
nd
ti
ps
,
vis
ua
li
z
e
d
thr
ough
ba
r
c
h
a
r
ts
to
highl
ight
pe
r
f
o
r
manc
e
dif
f
e
r
e
nc
e
s
.
T
his
in
tegr
a
ted
a
ppr
oa
c
h
e
nha
nc
e
s
taxi
s
e
r
vice
e
f
f
icie
nc
y
a
nd
s
uppor
ts
s
tr
a
tegic
de
c
is
ion
-
making
in
tr
a
ns
por
tation
[
1
0]
.
2.
1.
Wor
k
m
e
t
h
od
ology
I
n
thi
s
s
e
c
ti
on,
we
will
e
xplor
e
the
va
r
ious
meth
odologi
e
s
a
nd
tool
s
e
mpl
oye
d
to
ha
ndle
big
da
ta,
f
oc
us
ing
on
tec
hniques
that
e
na
ble
e
f
f
icie
nt
pr
oc
e
s
s
ing
a
nd
a
na
lys
is
of
lar
ge
da
tas
e
ts
.
W
e
will
de
lve
int
o
the
a
ppli
c
a
ti
on
of
li
ne
a
r
r
e
g
r
e
s
s
ion
in
mac
hine
lea
r
ni
ng,
dis
c
us
s
ing
how
dif
f
e
r
e
nt
a
pp
r
oa
c
he
s
,
s
uc
h
a
s
or
dinar
y
lea
s
t
s
qua
r
e
s
(
OL
S
)
a
nd
li
mi
ted
-
memor
y
B
r
oyde
n
-
F
letc
he
r
-
Goldf
a
r
b
-
S
ha
nno
(L
-
B
F
GS)
,
c
a
n
be
ut
il
ize
d
to
opti
mi
z
e
model
pa
r
a
mete
r
s
f
or
pr
e
dictive
a
c
c
ur
a
c
y.
Additi
ona
ll
y
,
we
wil
l
e
xa
mi
ne
the
metr
ics
us
e
d
f
or
e
va
luating
model
pe
r
f
or
manc
e
,
s
he
dding
li
ght
on
how
they
mea
s
ur
e
the
e
f
f
e
c
ti
ve
ne
s
s
of
pr
e
dictive
models
,
identif
y
a
r
e
a
s
f
or
im
pr
ove
ment,
a
nd
e
ns
ur
e
that
t
he
c
hos
e
n
a
lgor
it
h
ms
a
li
gn
with
the
goa
ls
of
da
t
a
-
dr
iven
de
c
is
ion
-
making.
T
hr
ough
thi
s
e
xplor
a
ti
on
,
we
a
i
m
to
pr
ovide
a
c
ompr
e
he
ns
ive
unde
r
s
tanding
of
how
big
da
ta
pr
oc
e
s
s
ing
tool
s
,
s
uc
h
a
s
Apa
c
he
S
pa
r
k,
a
n
d
li
ne
a
r
r
e
gr
e
s
s
ion
tec
hniques
c
a
n
be
leve
r
a
ge
d
to
buil
d
,
opti
mi
z
e
,
a
nd
e
va
luate
p
r
e
dictive
models
.
2.
1.
1.
T
ools
f
or
h
an
d
li
n
g
b
ig
d
at
a
Our
a
ppr
oa
c
h
to
mana
ging
big
da
ta
r
e
li
e
s
on
Apa
c
he
s
pa
r
k
,
a
dis
tr
ibut
e
d
c
omput
ing
s
ys
tem
known
f
or
it
s
e
f
f
icie
nc
y
a
nd
s
c
a
labili
ty.
Apa
c
he
s
pa
r
k
e
xc
e
ls
in
pr
oc
e
s
s
ing
lar
ge
da
tas
e
ts
by
dis
tr
ibut
ing
tas
k
s
a
c
r
os
s
a
c
lus
ter
of
c
omput
e
r
s
,
whic
h
e
na
bles
pa
r
a
ll
e
l
pr
oc
e
s
s
ing.
T
his
c
a
pa
bil
it
y
s
igni
f
ica
ntl
y
r
e
duc
e
s
da
ta
pr
oc
e
s
s
ing
ti
me
c
ompar
e
d
to
t
r
a
dit
ional
s
ingl
e
-
mac
hine
methods
[
11]
.
Apa
c
he
s
pa
r
k
us
e
s
r
e
s
il
ient
dis
tr
ibut
e
d
da
tas
e
t
s
(
R
DD
s
)
to
e
ns
ur
e
f
a
ult
tol
e
r
a
nc
e
a
nd
e
nha
nc
e
pe
r
f
or
manc
e
.
R
DD
s
a
r
e
c
a
c
he
d
in
memo
r
y,
a
ll
owi
ng
it
e
r
a
ti
ve
a
lgo
r
it
hms
to
r
e
us
e
int
e
r
media
te
r
e
s
ult
s
a
c
r
os
s
mul
ti
ple
c
omput
a
ti
ons
.
T
h
is
f
e
a
tur
e
gr
e
a
tl
y
s
pe
e
ds
up
mac
hine
lea
r
ning
a
lgor
it
hms
a
nd
other
ta
s
ks
that
r
e
quir
e
mul
ti
ple
da
ta
pa
s
s
e
s
[
12]
.
S
pa
r
k's
unif
ied
a
na
lyt
ics
e
ngine
s
uppor
ts
diver
s
e
d
a
ta
pr
oc
e
s
s
ing
ne
e
ds
,
including
ba
tch
p
r
oc
e
s
s
ing,
r
e
a
l
-
ti
me
s
tr
e
a
m
pr
oc
e
s
s
ing,
a
nd
mac
hine
lea
r
nin
g.
I
t
include
s
s
e
ve
r
a
l
s
pe
c
ialize
d
li
br
a
r
ies
,
s
uc
h
a
s
S
pa
r
k
S
QL
f
or
S
QL
que
r
ies
,
s
pa
r
k
s
tr
e
a
mi
ng
f
or
r
e
a
l
-
ti
me
da
ta,
M
L
li
b
f
or
mac
hine
lea
r
ning
,
a
nd
Gr
a
phX
f
or
gr
a
ph
pr
oc
e
s
s
ing.
T
he
s
e
li
br
a
r
ies
e
xtend
S
pa
r
k's
f
unc
ti
on
a
li
ty
a
nd
make
it
ve
r
s
a
ti
le
f
o
r
va
r
ious
da
ta
tas
ks
[
1
3]
.
One
of
s
pa
r
k’
s
notable
a
dva
ntage
s
is
it
s
in
-
memor
y
c
omput
ing
c
a
pa
bil
it
y,
whic
h
a
ll
ows
f
or
r
a
pid
da
ta
pr
oc
e
s
s
ing
by
s
tor
ing
da
ta
in
memor
y
r
a
ther
t
ha
n
on
dis
k
[
14]
.
T
his
f
e
a
tu
r
e
is
pa
r
ti
c
ular
ly
be
ne
f
icia
l
f
or
it
e
r
a
ti
ve
a
lgor
it
hms
a
nd
int
e
r
a
c
ti
ve
da
ta
e
xplor
a
ti
on
[
15]
.
Addit
ionally,
s
pa
r
k’
s
us
e
r
-
f
r
iendly
AP
I
s
in
J
a
va
,
S
c
a
la,
P
ython,
a
nd
R
s
im
pli
f
y
the
c
r
e
a
ti
on
o
f
c
ompl
e
x
wor
kf
lows
a
nd
da
ta
pipelines
while
p
r
ovidi
ng
a
dva
nc
e
d
c
ontr
ols
f
or
e
xpe
r
ienc
e
d
us
e
r
s
.
S
pa
r
k’
s
e
f
f
icie
nt
pr
oc
e
s
s
ing
a
nd
ve
r
s
a
ti
le
c
a
pa
bil
it
ies
make
it
e
s
s
e
nti
a
l
f
or
moder
n
da
ta
a
na
lyt
ics
a
nd
mac
hine
le
a
r
ning
[
16]
,
[
17]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
A
nalys
is
of
big
data
fr
om
N
e
w
Y
or
k
taxi
tr
ip
2023
:
r
e
v
e
nue
pr
e
diction
us
ing
…
(
Sar
a
R
houas
)
713
2.
1.
2.
L
in
e
ar
r
e
gr
e
s
s
ion
in
m
ac
h
i
n
e
lear
n
in
g
L
inea
r
r
e
gr
e
s
s
ion
is
one
of
the
mos
t
f
unda
menta
l
a
nd
wide
ly
us
e
d
tec
hniques
in
mac
hine
lea
r
ning
f
or
pr
e
dicting
a
c
onti
nuous
tar
ge
t
va
r
iable
ba
s
e
d
on
one
or
mor
e
p
r
e
dictor
va
r
iable
s
[
18]
.
At
i
ts
c
o
r
e
,
li
ne
a
r
r
e
gr
e
s
s
ion
a
im
s
to
model
the
r
e
lations
hip
be
twe
e
n
the
de
pe
nde
nt
va
r
iable
(
the
ta
r
ge
t)
a
nd
the
ind
e
pe
nde
nt
va
r
iable
s
(
the
pr
e
dictor
s
)
by
f
it
ti
ng
a
li
ne
a
r
e
qu
a
ti
on
to
obs
e
r
ve
d
da
ta.
T
he
pr
im
a
r
y
objec
ti
ve
of
li
ne
a
r
r
e
gr
e
s
s
ion
is
to
de
ter
mi
ne
the
opti
mal
va
lues
f
or
thes
e
c
oe
f
f
icie
nts
s
uc
h
that
the
s
um
of
the
s
qua
r
e
d
dif
f
e
r
e
nc
e
s
be
twe
e
n
the
obs
e
r
ve
d
a
c
tual
va
lues
a
nd
the
va
lues
pr
e
dicte
d
by
the
li
ne
a
r
model
(
kno
wn
a
s
the
r
e
s
idual
s
um
o
f
s
qua
r
e
s
)
is
mi
nim
ize
d.
I
n
our
s
t
udy,
we
uti
l
ize
d
two
s
pe
c
if
ic
methods
to
pe
r
f
or
m
li
ne
a
r
r
e
gr
e
s
s
ion:
the
OL
S
.
a
nd
the
L
-
B
F
GS
a
lgor
it
hm
[
1
9]
.
T
he
or
dinar
y
lea
s
t
s
qua
r
e
s
(
OL
S
)
method
is
a
f
u
nda
menta
l
a
ppr
oa
c
h
in
li
ne
a
r
r
e
gr
e
s
s
ion
us
e
d
to
e
s
ti
mate
the
c
oe
f
f
icie
nts
that
mi
nim
ize
the
r
e
s
idual
s
um
of
s
qua
r
e
s
be
twe
e
n
the
obs
e
r
ve
d
va
lues
a
nd
the
va
lues
pr
e
dicte
d
by
the
model
.
T
he
goa
l
is
to
f
ind
the
be
s
t
-
f
it
li
ne
that
c
a
ptur
e
s
the
r
e
lations
hip
be
t
we
e
n
the
indepe
nde
nt
va
r
iable
s
(
pr
e
dictor
s
)
a
nd
the
de
pe
n
de
nt
va
r
iable
(
tar
ge
t
)
[
20]
.
T
he
OL
S
s
olut
ion
is
de
r
ived
us
ing
the
n
or
mal
(
1)
:
=
(
)
−
1
(
1
)
w
he
r
e
r
e
pr
e
s
e
nts
the
ve
c
tor
of
c
oe
f
f
icie
nts
,
X
is
the
matr
ix
of
input
f
e
a
tur
e
s
(
including
a
c
olum
n
of
one
s
f
or
the
int
e
r
c
e
pt
ter
m)
,
y
is
the
ve
c
tor
of
obs
e
r
ve
d
va
lues
,
is
the
tr
a
ns
pos
e
of
the
mat
r
ix
.
T
his
m
e
thod
pr
ovides
a
n
e
xa
c
t
s
olut
ion
by
s
olvi
ng
the
a
bove
e
qua
ti
on,
making
it
s
tr
a
ight
f
o
r
wa
r
d
a
nd
c
omput
a
ti
ona
ll
y
e
f
f
icie
nt
f
or
s
maller
da
tas
e
ts
.
How
e
ve
r
,
f
o
r
ve
r
y
lar
ge
da
tas
e
ts
,
the
matr
ix
inve
r
s
ion
c
a
n
be
c
ome
c
omput
a
ti
ona
ll
y
e
xpe
ns
ive,
whic
h
is
a
li
mi
tation
o
f
thi
s
a
ppr
oa
c
h
[
21]
.
T
he
L
-
B
F
GS
a
lgor
it
hm
is
a
n
it
e
r
a
ti
ve
opti
mi
z
a
ti
on
tec
hnique
pa
r
ti
c
ular
ly
we
ll
-
s
uit
e
d
f
or
lar
ge
-
s
c
a
le
a
nd
high
-
dim
e
ns
ional
da
ta
s
e
ts
[
22]
.
I
t
is
a
va
r
ia
nt
of
the
B
F
GS
a
lgor
it
h
m
that
us
e
s
li
mi
ted
memor
y
to
a
ppr
oxim
a
te
the
inve
r
s
e
He
s
s
ian
matr
ix,
whic
h
is
e
s
s
e
nti
a
l
f
or
de
ter
mi
n
ing
the
dir
e
c
ti
on
of
the
s
tee
pe
s
t
de
s
c
e
nt
in
opti
mi
z
a
ti
on
pr
oblems
.
T
he
it
e
r
a
ti
ve
p
r
o
c
e
s
s
f
oll
ows
thes
e
s
tep
s
(
2)
:
+
1
=
−
−
1
∇
(
)
(
2)
w
he
r
e
is
the
c
oe
f
f
icie
nt
ve
c
tor
a
t
it
e
r
a
ti
on
k,
is
t
he
s
tep
s
ize
(
lea
r
ning
r
a
te)
,
−
1
is
the
inver
s
e
he
s
s
ian
matr
ix
a
ppr
oxi
mation
a
t
it
e
r
a
ti
on
k
,
a
nd
∇
(
)
is
the
g
r
a
dient
of
the
c
os
t
f
unc
ti
on
a
t
.
Unlike
the
OL
S
method,
L
-
B
F
GS
doe
s
not
r
e
quir
e
matr
ix
inver
s
io
n,
making
it
mo
r
e
s
c
a
lable
a
nd
e
f
f
icie
nt
f
or
ha
ndli
ng
lar
ge
da
tas
e
ts
.
I
t
it
e
r
a
ti
ve
ly
a
djus
ts
the
c
oe
f
f
icie
nts
b
y
f
oll
owing
the
gr
a
dient
of
the
c
os
t
f
unc
ti
on
,
gr
a
dua
ll
y
c
onve
r
ging
to
the
opti
mal
s
olut
ion.
T
his
make
s
L
-
B
F
GS
pa
r
ti
c
ular
ly
a
dva
ntage
ous
f
or
s
c
e
na
r
ios
whe
r
e
the
da
tas
e
t
s
ize
or
the
numbe
r
o
f
f
e
a
tur
e
s
is
lar
ge
[
23]
.
De
s
pit
e
the
s
im
pli
c
it
y
a
nd
int
e
r
pr
e
tabili
ty
o
f
li
ne
a
r
r
e
gr
e
s
s
ion,
it
is
e
s
s
e
nti
a
l
to
e
va
luate
the
un
de
r
lyi
ng
a
s
s
umpt
ions
—
s
uc
h
a
s
li
ne
a
r
it
y,
indep
e
nde
nc
e
,
homos
c
e
da
s
ti
c
it
y
(
c
ons
tant
va
r
ianc
e
of
e
r
r
or
s
)
,
a
nd
nor
malit
y
o
f
e
r
r
or
ter
ms
—
to
e
ns
ur
e
the
va
li
dit
y
a
nd
r
e
li
a
bil
it
y
of
the
model
’
s
pr
e
dictions
.
B
y
c
a
r
e
f
u
ll
y
s
e
lec
ti
ng
the
a
ppr
opr
iate
method
a
nd
va
li
da
ti
ng
the
a
s
s
umpt
ions
,
li
ne
a
r
r
e
gr
e
s
s
ion
r
e
mains
a
powe
r
f
ul
tool
f
or
unde
r
s
tanding
a
nd
pr
e
dicting
the
r
e
lations
hips
withi
n
the
da
ta
a
c
r
os
s
va
r
ious
domains
[
24]
.
2.
1.
3.
S
c
or
in
g
m
e
t
r
ics
T
o
f
it
the
li
ne
a
r
r
e
gr
e
s
s
ion
model
us
ing
thes
e
methods
,
we
f
ir
s
t
pr
e
pa
r
e
the
da
ta
by
c
ons
oli
da
ti
ng
the
s
e
lec
ted
f
e
a
tur
e
s
int
o
a
s
ingl
e
ve
c
tor
us
ing
a
Ve
c
t
or
As
s
e
mbl
e
r
.
T
he
da
tas
e
t
is
then
divi
de
d
int
o
t
r
a
i
ning
a
nd
tes
t
s
e
ts
,
whic
h
a
ll
ows
us
to
e
va
luate
the
model's
pe
r
f
or
manc
e
.
E
va
luation
metr
ics
s
uc
h
a
s
R
M
S
E
a
nd
M
S
E
a
r
e
us
e
d
to
a
s
s
e
s
s
how
we
ll
the
model
ge
ne
r
a
li
z
e
s
to
uns
e
e
n
da
ta.
T
he
s
e
metr
ics
a
r
e
e
s
s
e
nti
a
l
f
or
de
ter
mi
ning
the
a
c
c
ur
a
c
y
of
our
p
r
e
dictions
,
of
f
e
r
ing
ins
ight
s
i
nto
the
model’
s
e
f
f
e
c
ti
ve
ne
s
s
a
nd
it
s
a
bil
it
y
to
ha
n
dle
ne
w
da
ta
[
25]
.
M
S
E
is
a
wide
ly
us
e
d
metr
ic
f
or
e
va
luating
the
a
c
c
ur
a
c
y
of
pr
e
dictive
models
.
I
t
qua
nti
f
ies
the
mea
n
of
the
s
qua
r
e
d
dif
f
e
r
e
nc
e
s
be
twe
e
n
pr
e
dicte
d
a
nd
obs
e
r
ve
d
va
lues
.
M
S
E
e
s
s
e
nti
a
ll
y
mea
s
ur
e
s
the
a
ve
r
a
ge
magnitude
of
the
s
qua
r
e
d
de
viations
a
c
r
os
s
a
ll
da
ta
point
s
,
pr
ovidi
ng
a
de
tailed
a
s
s
e
s
s
ment
of
model
pe
r
f
or
manc
e
.
T
his
metr
ic
is
va
luable
f
o
r
unde
r
s
tanding
the
ove
r
a
ll
qua
li
ty
of
the
model's
p
r
e
dictio
ns
,
a
s
it
c
a
ptur
e
s
the
e
xtent
of
pr
e
diction
e
r
r
or
s
in
a
c
onti
nu
ous
manne
r
[
26]
.
R
M
S
E
is
a
nother
ke
y
met
r
ic
that
o
f
f
e
r
s
a
s
tr
a
ight
f
or
wa
r
d
mea
s
ur
e
of
pr
e
diction
e
r
r
or
.
B
y
taking
the
s
qua
r
e
r
oot
o
f
the
M
S
E
,
R
M
S
E
pr
e
s
e
nts
a
n
e
r
r
or
metr
ic
that
maintains
the
s
a
me
unit
s
a
s
the
tar
ge
t
va
r
iable
,
making
it
mor
e
int
uit
ive
.
R
M
S
E
plac
e
s
a
higher
e
mphas
is
on
lar
ge
r
e
r
r
or
s
due
to
the
s
qua
r
ing
of
dif
f
e
r
e
nc
e
s
,
whic
h
mea
ns
it
pe
na
li
z
e
s
s
igni
f
ica
nt
de
viations
mo
r
e
.
T
his
c
ha
r
a
c
ter
is
ti
c
make
s
R
M
S
E
pa
r
ti
c
ular
ly
u
s
e
f
ul
f
or
unde
r
s
tanding
the
model's
pe
r
f
o
r
manc
e
with
r
e
s
pe
c
t
to
outl
ie
r
pr
e
dictions
[
27]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
15
,
No.
1
,
F
e
br
ua
r
y
20
25
:
711
-
718
714
2.
2
.
App
li
c
a
t
ion
m
e
t
h
od
I
n
thi
s
a
na
lys
is
,
a
c
ombi
na
ti
on
of
mac
hine
lea
r
ning
tec
hniques
,
including
li
ne
a
r
r
e
gr
e
s
s
ion
a
nd
S
pa
r
k's
dis
tr
ibut
e
d
c
omput
ing
c
a
pa
bil
it
ies
,
we
r
e
e
mpl
oye
d
to
pr
e
dict
taxi
tr
ip
r
e
ve
nue
s
in
Ne
w
Yor
k
C
it
y
f
or
the
ye
a
r
2023
.
L
e
ve
r
a
ging
S
pa
r
k's
powe
r
f
ul
da
ta
p
r
oc
e
s
s
ing
platf
or
m,
the
a
na
lys
is
a
im
e
d
to
pr
ovide
a
c
c
ur
a
te
r
e
ve
nue
pr
e
dictions
by
incor
po
r
a
ti
ng
ke
y
f
e
a
tur
e
s
s
uc
h
a
s
pa
s
s
e
nge
r
c
ount,
t
r
ip
dis
tanc
e
,
f
a
r
e
a
mount
,
a
nd
t
ip
a
mount
.
T
he
uti
li
z
a
ti
on
of
li
ne
a
r
r
e
gr
e
s
s
ion,
a
we
ll
-
e
s
tabli
s
he
d
a
nd
int
e
r
pr
e
table
modeling
tec
hnique,
e
ns
ur
e
d
a
c
ompr
e
he
ns
ive
a
nd
e
f
f
e
c
ti
ve
a
ppr
oa
c
h
to
r
e
ve
n
ue
pr
e
diction.
F
u
r
ther
mo
r
e
,
S
pa
r
k's
dis
tr
ibut
e
d
c
o
mput
ing
c
a
pa
bil
it
ies
e
na
bled
the
e
f
f
icie
nt
ha
ndli
ng
of
lar
ge
-
s
c
a
l
e
da
tas
e
t
s
,
a
ll
owing
f
or
ti
mely
a
nd
a
c
c
ur
a
te
pr
e
dictions
e
ve
n
with
mas
s
ive
a
mount
s
of
da
ta
.
2.
2.
1.
Dat
a
u
s
e
d
T
he
da
tas
e
t
uti
li
z
e
d
in
thi
s
a
na
lys
is
c
on
s
is
ted
of
Ne
w
Yor
k
C
it
y
taxi
tr
ip
da
ta
f
or
the
ye
a
r
2023
,
s
our
c
e
d
f
r
om
P
a
r
que
t
f
il
e
s
.
T
he
s
e
f
il
e
s
c
ontain
de
tailed
inf
or
mation
a
bout
taxi
tr
ips
,
including
a
tt
r
ibu
tes
s
uc
h
a
s
pickup
da
tetim
e
,
pa
s
s
e
nge
r
c
ount,
tr
ip
d
is
tanc
e
,
f
a
r
e
a
mount
,
ti
p
a
mount
,
a
nd
tot
a
l
a
mount
.
T
h
e
da
tas
e
t
wa
s
meticulous
ly
c
lea
ne
d
a
nd
p
r
e
pr
oc
e
s
s
e
d
to
e
ns
ur
e
da
ta
qua
li
ty
a
nd
r
e
li
a
bil
it
y
f
o
r
s
ubs
e
que
nt
a
na
lys
is
.
I
nva
li
d
r
e
c
or
ds
a
nd
mi
s
s
ing
va
lues
we
r
e
f
il
te
r
e
d
o
ut,
a
nd
the
p
ickup
da
tetim
e
c
olum
n
wa
s
c
a
s
t
to
a
da
te
type
f
or
tempor
a
l
a
na
lys
is
.
T
his
r
e
f
ined
da
tas
e
t
s
e
r
ve
d
a
s
the
f
ounda
ti
on
f
or
buil
ding
a
nd
tr
a
ini
ng
th
e
li
ne
a
r
r
e
gr
e
s
s
ion
model
f
or
r
e
ve
nue
pr
e
diction
[
28]
.
T
hr
ough
e
xplor
a
to
r
y
da
ta
a
na
lys
is
a
nd
f
e
a
tur
e
e
nginee
r
ing,
ins
ight
s
we
r
e
e
xtr
a
c
ted
f
r
om
the
da
tas
e
t
to
e
nha
nc
e
the
pr
e
dictive
model's
pe
r
f
or
manc
e
.
Ke
y
f
e
a
tur
e
s
s
uc
h
a
s
pa
s
s
e
nge
r
c
ount,
tr
ip
dis
ta
nc
e
,
f
a
r
e
a
mount
,
a
nd
ti
p
a
mount
we
r
e
identif
ied
ba
s
e
d
on
t
he
ir
potential
im
pa
c
t
on
t
r
ip
r
e
ve
nue
s
.
T
he
s
e
f
e
a
tur
e
s
we
r
e
then
us
e
d
to
tr
a
in
the
li
ne
a
r
r
e
gr
e
s
s
ion
model,
whic
h
s
e
r
ve
d
a
s
the
pr
e
dictive
e
ngine
f
or
e
s
ti
mating
taxi
tr
ip
r
e
ve
nue
s
.
B
y
leve
r
a
ging
S
pa
r
k's
dis
tr
ibut
e
d
c
omp
uti
ng
c
a
pa
bil
it
ies
,
the
model
wa
s
a
ble
to
e
f
f
icie
ntl
y
pr
oc
e
s
s
a
nd
a
na
lyze
lar
ge
-
s
c
a
le
da
ta
s
e
ts
,
pr
ovidi
ng
s
take
holder
s
with
a
c
c
ur
a
te
a
nd
ti
mely
r
e
ve
nue
p
r
e
dictions
.
2.
2.
2.
P
r
oc
e
s
s
T
his
pr
oc
e
s
s
outl
ines
a
da
ta
-
dr
iven
a
ppr
oa
c
h
f
or
pr
e
dicting
taxi
tr
ip
r
e
ve
nue
s
in
Ne
w
Yor
k
C
it
y
f
or
2023.
I
t
be
gins
with
da
ta
loading
a
nd
ini
ti
a
l
pr
oc
e
s
s
ing
,
whe
r
e
S
pa
r
k
is
c
onf
igur
e
d
f
or
e
f
f
icie
nt
ha
ndli
ng
of
lar
ge
da
tas
e
ts
.
T
he
da
ta,
s
tor
e
d
in
pa
r
que
t
f
or
ma
t
on
Ha
doop
dis
tr
ibut
e
d
f
il
e
s
ys
tem
(
HD
F
S
)
,
is
ve
r
if
ied,
loade
d,
a
nd
c
ombi
ne
d
int
o
a
s
ingl
e
da
ta
f
r
a
me
f
or
the
e
nti
r
e
ye
a
r
.
F
oll
owing
thi
s
,
da
ta
c
lea
ning
a
nd
pr
e
pr
oc
e
s
s
ing
e
ns
ur
e
the
da
ta
s
e
t's
int
e
gr
it
y
by
r
e
movi
ng
r
ows
with
null
or
invalid
va
lues
,
r
e
duc
ing
th
e
da
ta
to
37,
000,
870
r
ows
.
F
e
a
tur
e
s
e
lec
ti
on
a
nd
e
nginee
r
ing
identif
y
ke
y
ins
ight
s
,
including
pe
a
k
ope
r
a
ti
ona
l
pe
r
iods
a
nd
f
e
a
tur
e
c
or
r
e
lations
,
s
e
tt
ing
the
s
tage
f
or
model
tr
a
ini
ng
.
M
ode
l
tr
a
ini
ng
a
nd
e
va
luation
invol
ve
s
s
pli
tt
ing
t
he
da
ta
int
o
t
r
a
ini
ng
a
nd
tes
ti
ng
s
e
ts
a
nd
a
pply
ing
two
li
ne
a
r
r
e
gr
e
s
s
ion
methods
—
nor
mal
e
qua
ti
ons
a
nd
L
-
B
F
GS.
T
he
models
a
r
e
e
va
luate
d
us
ing
R
M
S
E
a
nd
M
S
E
metr
ics
f
or
a
c
c
ur
a
c
y.
P
e
r
f
or
manc
e
e
va
luati
on
a
nd
vis
ua
li
z
a
ti
on
e
xa
mi
ne
f
e
a
tur
e
im
pa
c
ts
a
n
d
ve
ndor
metr
ics
,
s
uc
h
a
s
tr
ip
c
ounts
,
a
ve
r
a
ge
f
a
r
e
s
,
a
nd
ti
ps
,
while
ins
ight
s
a
nd
de
c
is
ion
-
making
leve
r
a
ge
thes
e
r
e
s
ult
s
to
opti
mi
z
e
taxi
ope
r
a
ti
ons
a
nd
e
nha
nc
e
c
us
tom
e
r
s
a
ti
s
f
a
c
ti
on
.
T
his
s
tr
uc
tur
e
d
a
na
lys
is
of
f
e
r
s
a
c
ti
ona
ble
ins
ight
s
to
im
pr
ove
s
e
r
vice
e
f
f
icie
nc
y
a
nd
p
r
of
it
a
bi
li
ty.
3.
RE
S
UL
T
S
AN
D
DI
S
CU
S
S
I
ON
T
he
a
na
lys
is
f
oc
us
e
s
on
e
va
luating
the
pe
r
f
o
r
man
c
e
of
two
major
taxi
ve
ndo
r
s
in
Ne
w
Yor
k
C
it
y.
Us
ing
c
ompr
e
he
ns
ive
tr
ip
da
ta,
ke
y
metr
ics
s
uc
h
a
s
the
tot
a
l
number
of
tr
ips
,
a
ve
r
a
ge
tr
ip
d
is
tanc
e
,
a
ve
r
a
ge
f
a
r
e
a
mount
,
a
nd
a
ve
r
a
ge
t
ip
a
mount
a
r
e
a
na
lyze
d
to
a
s
s
e
s
s
e
a
c
h
ve
ndor
's
ope
r
a
ti
ona
l
e
f
f
icie
nc
y
a
n
d
mar
ke
t
pos
it
ioni
ng.
T
he
f
oll
owing
s
e
c
ti
ons
pr
ov
ide
a
de
tailed
e
xa
mi
na
ti
on
of
thes
e
metr
ics
,
h
ighl
igh
ti
ng
the
s
tr
e
ngths
a
nd
we
a
kne
s
s
e
s
of
ve
ndor
1
a
nd
ve
ndor
2.
3
.
1
.
P
e
r
f
or
m
an
c
e
an
alys
is
f
or
e
ac
h
ve
n
d
or
I
n
thi
s
a
na
lys
is
,
we
e
xa
mi
ne
the
pe
r
f
or
manc
e
of
tw
o
major
taxi
ve
ndor
s
in
Ne
w
Yor
k
C
it
y
us
ing
ke
y
metr
ics
de
r
ived
f
r
om
c
omp
r
e
he
ns
ive
tr
ip
da
ta.
B
y
e
va
luating
the
tot
a
l
number
o
f
tr
ips
,
a
ve
r
a
ge
tr
ip
dis
tanc
e
,
a
ve
r
a
ge
f
a
r
e
a
mount
,
a
nd
a
ve
r
a
ge
ti
p
a
mount
,
we
a
im
to
unde
r
s
tand
the
ope
r
a
ti
ona
l
e
f
f
icie
nc
y
a
nd
mar
ke
t
pos
it
ioni
ng
of
e
a
c
h
ve
ndor
.
T
he
da
ta
s
pa
ns
a
s
igni
f
ica
nt
pe
r
iod
a
nd
pr
ovides
a
r
obus
t
f
ound
a
ti
on
f
or
c
ompar
ing
thes
e
ve
ndor
s
'
e
f
f
e
c
ti
ve
ne
s
s
in
mee
ti
ng
pa
s
s
e
nge
r
de
mand
a
nd
ge
ne
r
a
ti
ng
r
e
ve
nue
.
T
he
f
oll
owing
pa
r
a
gr
a
phs
de
lve
int
o
e
a
c
h
metr
ic,
o
f
f
e
r
ing
ins
ight
s
int
o
the
s
tr
e
ngths
a
nd
we
a
kne
s
s
e
s
of
v
e
ndor
1
a
nd
v
e
ndor
2.
As
s
hown
in
F
igu
r
e
1
,
ve
ndor
2
de
mons
tr
a
tes
a
s
i
gnif
ica
ntl
y
higher
volum
e
of
tot
a
l
tr
ips
c
ompar
e
d
to
ve
ndor
1.
S
pe
c
if
ica
ll
y,
ve
ndor
2
r
e
c
or
de
d
27
,
4
71,
887
tr
ips
,
whe
r
e
a
s
ve
ndor
1
r
e
c
or
de
d
9,
528
,
9
83
tr
ips
.
T
his
dis
pa
r
it
y
indi
c
a
tes
that
ve
ndor
2
ha
s
a
lar
ge
r
s
ha
r
e
of
the
mar
ke
t,
whic
h
c
ould
be
due
to
a
va
r
iety
of
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
A
nalys
is
of
big
data
fr
om
N
e
w
Y
or
k
taxi
tr
ip
2023
:
r
e
v
e
nue
pr
e
diction
us
ing
…
(
Sar
a
R
houas
)
715
f
a
c
tor
s
s
uc
h
a
s
a
mor
e
e
xtens
ive
f
lee
t,
mor
e
e
f
f
icie
nt
dis
pa
tch
a
nd
r
outi
ng
s
ys
tems
,
or
s
tr
onge
r
br
a
nd
r
e
c
ognit
ion.
T
he
higher
tr
ip
volum
e
a
ls
o
s
ugge
s
ts
that
ve
ndor
2
is
be
tt
e
r
a
t
mee
ti
ng
pa
s
s
e
nge
r
de
m
a
nd
a
nd
potentially
ha
s
wide
r
ope
r
a
ti
ona
l
c
ove
r
a
ge
a
c
r
os
s
Ne
w
Yor
k
C
it
y
.
T
h
is
lar
ge
volum
e
of
t
r
ips
p
r
ovides
ve
ndor
2
with
a
r
obus
t
r
e
ve
nue
ba
s
e
a
nd
e
nha
nc
e
s
it
s
a
bil
it
y
to
ge
ne
r
a
te
s
igni
f
ica
nt
income
f
r
om
a
high
nu
mber
o
f
s
e
r
vice
tr
a
ns
a
c
ti
ons
.
I
n
F
igu
r
e
2
we
c
a
n
s
e
e
that
the
a
ve
r
a
ge
tr
ip
dis
tanc
e
f
or
ve
ndor
2
is
s
li
ghtl
y
longer
than
that
f
o
r
ve
ndor
1,
with
ve
ndor
2
a
ve
r
a
ging
3.
64
mi
les
pe
r
tr
ip
a
nd
ve
ndor
1
a
ve
r
a
ging
3
.
42
m
il
e
s
.
W
hil
e
the
dif
f
e
r
e
nc
e
may
s
e
e
m
mi
nim
a
l,
it
ha
s
im
po
r
tant
im
pli
c
a
ti
ons
f
or
r
e
ve
nue
.
L
onge
r
tr
ips
typi
c
a
ll
y
r
e
s
ult
in
higher
f
a
r
e
s
,
c
ontr
ibut
ing
mo
r
e
s
igni
f
ica
ntl
y
to
tot
a
l
r
e
ve
nue
.
Ve
ndor
2
’
s
s
li
ghtl
y
longer
a
ve
r
a
ge
tr
ip
dis
tanc
e
c
ould
indi
c
a
te
that
they
s
e
r
ve
a
r
e
a
s
with
gr
e
a
ter
dis
tanc
e
s
be
twe
e
n
c
omm
on
pick
-
up
a
nd
dr
op
-
of
f
point
s
or
that
they
a
tt
r
a
c
t
tr
ips
that
tend
to
c
ove
r
mo
r
e
dis
tanc
e
.
T
his
c
ould
be
a
r
e
s
ult
of
s
tr
a
tegic
ope
r
a
ti
ona
l
d
e
c
is
ions
or
a
f
oc
us
on
a
r
e
a
s
with
higher
f
a
r
e
potential.
T
he
longer
tr
ip
dis
tanc
e
s
mi
ght
a
ls
o
s
ugge
s
t
that
ve
ndo
r
2
ha
s
a
higher
pr
opor
ti
on
of
t
r
ips
to
a
nd
f
r
om
majo
r
hubs
li
ke
a
ir
por
ts
or
bus
ines
s
dis
tr
icts
,
whic
h
typi
c
a
ll
y
invol
ve
gr
e
a
ter
dis
tanc
e
s
.
Ve
ndor
2
a
ls
o
outper
f
or
ms
ve
ndor
1
in
ter
ms
of
a
v
e
r
a
ge
f
a
r
e
a
mount
,
with
a
n
a
ve
r
a
ge
f
a
r
e
of
$19
.
67
c
ompar
e
d
to
ve
ndor
1’
s
$18
.
71.
T
his
di
f
f
e
r
e
nc
e
i
n
f
a
r
e
a
mount
s
is
li
ke
ly
li
nke
d
to
the
longer
a
ve
r
a
ge
tr
ip
dis
tanc
e
s
mentioned
e
a
r
li
e
r
.
Highe
r
a
ve
r
a
ge
f
a
r
e
s
not
only
boos
t
pe
r
-
tr
ip
r
e
ve
nue
but
a
ls
o
s
ugge
s
t
th
a
t
ve
ndor
2
may
be
ope
r
a
ti
ng
mor
e
in
pr
e
mi
um
s
e
gments
of
the
mar
ke
t
whe
r
e
pa
s
s
e
nge
r
s
a
r
e
will
ing
to
pa
y
mor
e
f
or
be
tt
e
r
s
e
r
vice
or
c
onve
nienc
e
.
Additi
ona
ll
y
,
the
h
igher
f
a
r
e
s
c
ould
be
a
r
e
s
ult
of
e
f
f
e
c
ti
ve
dyna
mi
c
pr
icing
s
tr
a
tegie
s
,
whe
r
e
ve
ndor
2
a
djus
ts
pr
ice
s
ba
s
e
d
on
de
mand
a
nd
s
upply
c
ondit
ions
to
maximi
z
e
r
e
ve
nue
.
T
his
a
bil
it
y
to
c
omm
a
nd
higher
f
a
r
e
s
s
tr
e
ngthens
ve
ndor
2’
s
ove
r
a
ll
f
inanc
ial
pe
r
f
or
manc
e
a
nd
c
ompetit
ive
a
dva
ntage
in
the
mar
ke
t
.
T
he
a
ve
r
a
ge
ti
p
a
mount
is
a
nother
a
r
e
a
whe
r
e
ve
n
dor
2
lea
ds
,
with
a
n
a
ve
r
a
ge
ti
p
of
$3.
65
c
ompar
e
d
to
ve
ndor
1’
s
$3.
26
.
T
ips
a
r
e
of
ten
indi
c
a
ti
ve
of
c
us
tom
e
r
s
a
ti
s
f
a
c
ti
on
a
nd
s
e
r
vice
qua
li
ty.
T
he
higher
a
ve
r
a
ge
ti
ps
f
or
ve
ndor
2
s
ugge
s
t
that
pa
s
s
e
nge
r
s
pe
r
c
e
ive
the
s
e
r
vice
qua
li
ty
to
be
be
tt
e
r
o
r
f
e
e
l
mo
r
e
s
a
ti
s
f
ied
with
their
r
ides
.
T
his
c
ould
be
due
to
va
r
ious
f
a
c
tor
s
s
uc
h
a
s
c
lea
ne
r
ve
hicle
s
,
mo
r
e
c
our
teous
dr
iver
s
,
b
e
tt
e
r
r
ide
e
xpe
r
ienc
e
s
,
or
mor
e
r
e
li
a
ble
s
e
r
vice
.
Highe
r
ti
p
s
c
ontr
ibut
e
dir
e
c
tl
y
to
the
dr
ive
r
s
'
e
a
r
nings
a
nd
c
a
n
a
ls
o
boos
t
ove
r
a
ll
dr
iver
mo
r
a
le
a
nd
r
e
tention.
F
r
o
m
a
bus
ines
s
pe
r
s
pe
c
ti
ve
,
higher
ti
ps
indi
c
a
te
a
pos
it
ive
c
us
tom
e
r
e
xpe
r
ienc
e
,
whic
h
is
c
r
uc
ial
f
or
c
us
tom
e
r
loyalty
a
nd
r
e
pe
a
t
bus
ines
s
.
Ve
ndor
2’
s
higher
tr
ip
vo
lum
e
,
longer
a
ve
r
a
ge
tr
ip
dis
tanc
e
,
higher
a
ve
r
a
ge
f
a
r
e
a
mount
,
a
nd
gr
e
a
ter
a
ve
r
a
ge
ti
p
a
mount
c
oll
e
c
ti
ve
ly
pa
int
a
pictur
e
o
f
a
mor
e
domi
na
nt
a
nd
f
inanc
ially
s
uc
c
e
s
s
f
ul
ope
r
a
tor
.
T
he
higher
tr
ip
volum
e
indi
c
a
tes
a
lar
ge
r
ope
r
a
ti
ona
l
s
c
a
le
a
nd
be
tt
e
r
mar
ke
t
pe
ne
tr
a
ti
on
,
while
the
lo
nge
r
tr
ip
dis
tanc
e
s
a
nd
higher
f
a
r
e
a
mount
s
s
ugge
s
t
a
f
oc
us
on
higher
-
va
lue
s
e
gment
s
of
the
mar
ke
t.
T
he
gr
e
a
ter
a
ve
r
a
ge
ti
ps
r
e
f
lec
t
s
upe
r
ior
s
e
r
vice
qua
li
ty,
lea
ding
to
higher
c
us
tom
e
r
s
a
ti
s
f
a
c
ti
on
a
nd
loyalty.
T
he
s
e
f
a
c
tor
s
c
ombi
ne
d
pos
it
ion
Ve
ndor
2
a
s
a
mor
e
r
obus
t
a
nd
c
ompetit
ive
playe
r
in
Ne
w
Yor
k
C
it
y's
taxi
indus
tr
y,
with
a
s
tr
onge
r
a
bil
it
y
to
ge
ne
r
a
te
r
e
ve
nue
a
nd
s
us
tain
long
-
ter
m
gr
owth
c
ompar
e
d
to
v
e
ndor
1.
F
igur
e
1
.
T
o
tal
tr
ips
pe
r
ve
ndor
I
D
F
igur
e
2
.
Ave
r
a
ge
t
r
ip
dis
tanc
e
pe
r
ve
ndor
I
D
3
.
2.
Anal
ys
is
of
t
h
e
r
e
gr
e
s
s
ion
p
e
r
f
or
m
an
c
e
s
I
n
a
na
lyzing
ye
ll
ow
taxi
t
r
ip
f
a
r
e
pr
e
diction,
li
ne
a
r
r
e
gr
e
s
s
ion
models
we
r
e
e
mpl
oye
d
to
unde
r
s
tand
the
im
pa
c
t
of
va
r
ious
f
a
c
tor
s
on
the
tot
a
l
f
a
r
e
.
T
wo
methods
,
OL
S
a
nd
L
-
B
F
GS,
we
r
e
us
e
d
to
bu
il
d
thes
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
15
,
No.
1
,
F
e
br
ua
r
y
20
25
:
711
-
718
716
models
.
B
oth
methods
of
f
e
r
dis
ti
nc
t
a
dva
ntage
s
in
ter
ms
of
c
omput
a
ti
ona
l
e
f
f
icie
nc
y
a
nd
s
c
a
labili
ty,
making
them
s
uit
a
ble
f
or
dif
f
e
r
e
nt
c
ontexts
de
pe
nding
on
the
s
ize
a
nd
c
ompl
e
xit
y
of
the
da
tas
e
t.
T
his
s
e
c
ti
on
de
lves
int
o
the
r
e
s
ult
s
obtaine
d
f
r
o
m
both
r
e
gr
e
s
s
ion
met
hods
,
pr
ovidi
ng
a
de
tailed
c
ompa
r
is
on
of
their
pe
r
f
or
manc
e
metr
ics
,
c
omput
a
ti
ona
l
r
e
quir
e
ments
,
a
nd
the
s
i
gnif
ica
nc
e
of
the
de
r
ived
c
oe
f
f
icie
nts
.
B
y
e
xa
m
ini
ng
the
c
oe
f
f
icie
nts
a
nd
their
im
pli
c
a
ti
ons
,
we
ga
in
ins
ight
s
int
o
the
pr
im
a
r
y
dr
ive
r
s
of
taxi
f
a
r
e
s
,
e
nha
n
c
ing
our
unde
r
s
tanding
of
f
a
r
e
s
tr
uc
tur
e
s
a
nd
c
us
tom
e
r
be
ha
vior
s
.
As
s
hown
in
T
a
ble
1
,
bo
th
the
OL
S
a
nd
L
-
B
F
GS
li
ne
a
r
r
e
gr
e
s
s
ion
models
yielde
d
ne
a
r
ly
identica
l
c
oe
f
f
icie
nts
,
de
mons
tr
a
ti
ng
the
r
obus
tnes
s
of
the
f
indi
ngs
.
T
he
c
oe
f
f
icie
nt
f
or
pa
s
s
e
nge
r
c
ount
is
a
ppr
oxim
a
tely
0.
0702,
ind
ica
ti
ng
that
e
a
c
h
a
ddit
ional
pa
s
s
e
ng
e
r
ha
s
a
s
mall
but
pos
it
ive
im
pa
c
t
on
the
tot
a
l
f
a
r
e
.
T
h
is
s
ugge
s
ts
that
while
ha
ving
mor
e
pa
s
s
e
n
ge
r
s
s
li
ghtl
y
incr
e
a
s
e
s
the
f
a
r
e
,
their
inf
luenc
e
is
r
e
latively
mi
nim
a
l
c
ompar
e
d
to
other
f
a
c
tor
s
.
T
he
tr
ip
dis
tanc
e
c
oe
f
f
icie
nt,
a
r
ound
0.
0010
,
a
ls
o
s
hows
a
ve
r
y
s
mall
im
pa
c
t
on
the
tot
a
l
f
a
r
e
,
indi
c
a
ti
ng
that
t
r
ip
d
is
tanc
e
c
ontr
ibut
e
s
mar
ginally
to
f
a
r
e
c
a
lcula
ti
ons
.
T
his
s
mall
im
pa
c
t
mi
ght
r
e
f
lec
t
a
f
a
r
e
s
tr
uc
tur
e
whe
r
e
f
ixe
d
c
os
ts
or
ti
me
-
ba
s
e
d
c
h
a
r
ge
s
a
r
e
mor
e
s
igni
f
ica
nt
than
dis
tanc
e
,
potentially
due
to
mi
nim
um
f
a
r
e
poli
c
ies
or
the
inclus
ion
of
in
it
ial
s
e
r
vice
f
e
e
s
that
ove
r
s
ha
dow
the
dis
tanc
e
-
b
a
s
e
d
c
omponent.
T
a
ble
1.
T
he
r
e
s
ult
s
of
e
a
c
h
method
M
e
tr
ic
O
L
S
L
-
B
F
G
S
T
r
a
in
in
g
t
im
e
(
s
e
c
onds
)
37.96
122.69
R
M
S
E
4.691598018
4.6915980186
M
S
E
22.01109196
22.01109197
P
a
s
s
e
nge
r
c
ount
c
o
e
f
f
ic
ie
nt
0.070184284
0.070184285
T
r
ip
d
is
ta
nc
e
c
o
e
f
f
ic
ie
nt
0.0009725934876
0.00097259340671
F
a
r
e
a
mount
c
oe
f
f
ic
ie
nt
1.0036740054
1.0036740051
T
ip
a
mount
c
oe
f
f
ic
ie
nt
1.35752071718
1.357520719
I
nt
e
r
c
e
pt
4.008011703
4.008011700
T
he
f
a
r
e
a
mount
,
with
a
c
oe
f
f
icie
nt
of
a
bout
1.
00
37,
s
hows
a
ne
a
r
one
-
to
-
one
r
e
lations
hip
with
the
tot
a
l
f
a
r
e
,
c
onf
ir
mi
ng
that
ba
s
e
f
a
r
e
c
a
lcula
ti
ons
a
r
e
the
p
r
im
a
r
y
de
ter
mi
na
nt
of
the
tot
a
l
f
a
r
e
.
I
n
c
ont
r
a
s
t,
the
ti
p
a
mount
,
with
a
c
oe
f
f
icie
nt
of
a
ppr
oxim
a
tely
1
.
3575,
indi
c
a
tes
that
ti
ps
s
igni
f
ica
ntl
y
boos
t
the
t
otal
f
a
r
e
.
T
his
higher
c
oe
f
f
icie
nt
s
ugge
s
ts
that
ti
pping
not
only
a
dds
dir
e
c
tl
y
to
the
f
a
r
e
but
a
ls
o
c
or
r
e
lat
e
s
with
s
c
e
na
r
ios
invol
ving
higher
s
e
r
vice
qua
li
ty
or
mor
e
e
xpe
ns
ive
r
ides
.
T
he
int
e
r
c
e
pt
,
a
r
ound
4.
0080
,
r
e
pr
e
s
e
nts
the
ba
s
e
li
ne
tot
a
l
f
a
r
e
,
e
ns
ur
ing
a
mi
nim
um
c
ha
r
g
e
r
e
ga
r
dles
s
of
other
f
a
c
tor
s
.
T
h
is
ba
s
e
li
ne
unde
r
s
c
or
e
s
the
im
por
tanc
e
of
ini
ti
a
l
f
e
e
s
in
the
f
a
r
e
s
tr
uc
tur
e
.
C
oll
e
c
ti
ve
ly,
thes
e
c
oe
f
f
icie
nts
r
e
ve
a
l
that
while
p
a
s
s
e
nge
r
c
ount
a
nd
tr
ip
dis
tanc
e
play
s
e
c
onda
r
y
r
oles
,
the
f
a
r
e
a
mount
a
nd
ti
ps
a
r
e
c
r
uc
ial
d
r
iver
s
o
f
the
to
tal
f
a
r
e
,
r
e
f
lec
ti
ng
a
f
a
r
e
s
tr
uc
tur
e
he
a
vil
y
inf
luenc
e
d
by
ba
s
e
c
ha
r
ge
s
a
nd
c
us
tom
e
r
ti
pping
be
ha
vior
.
T
he
f
a
r
e
a
mount
,
with
a
c
oe
f
f
icie
nt
of
1
.
0037,
s
h
ows
a
ne
a
r
one
-
to
-
one
r
e
lations
hip
with
the
tot
a
l
f
a
r
e
,
c
onf
i
r
mi
ng
that
ba
s
e
f
a
r
e
c
a
lcula
ti
ons
a
r
e
the
pr
im
a
r
y
f
a
c
tor
.
M
e
a
nwhile,
the
ti
p
a
mount
,
with
a
c
oe
f
f
icie
nt
of
1
.
3575,
ha
s
a
mo
r
e
s
igni
f
ica
nt
inf
lu
e
nc
e
,
indi
c
a
ti
ng
that
ti
ps
not
only
incr
e
a
s
e
the
f
a
r
e
dir
e
c
tl
y
but
a
ls
o
c
or
r
e
late
with
s
c
e
na
r
ios
invol
ving
h
igher
s
e
r
vice
qua
li
ty
or
mo
r
e
e
xpe
ns
ive
r
ides
.
T
he
i
nter
c
e
pt,
a
r
ound
4.
0080
,
e
ns
ur
e
s
a
mi
n
im
um
f
a
r
e
,
e
mphas
izing
the
im
por
tanc
e
o
f
ba
s
e
c
ha
r
ge
s
.
Ove
r
a
ll
,
f
a
r
e
a
mount
a
nd
ti
ps
a
r
e
the
main
dr
ive
r
s
of
the
tot
a
l
f
a
r
e
,
with
pa
s
s
e
nge
r
c
ount
a
nd
tr
ip
dis
tanc
e
playing
s
maller
r
oles
.
T
he
OL
S
a
nd
L
-
B
F
GS
li
ne
a
r
r
e
gr
e
s
s
ion
models
we
r
e
us
e
d
to
pr
e
dict
taxi
f
a
r
e
s
,
with
both
s
howin
g
ne
a
r
ly
identica
l
pe
r
f
or
manc
e
met
r
ics
.
T
he
OL
S
m
ode
l,
us
ing
the
nor
mal
e
qua
ti
ons
method,
ha
d
a
n
R
M
S
E
of
4.
6916
a
nd
a
n
M
S
E
of
22
.
0111,
a
nd
c
ompl
e
ted
in
37.
96
s
e
c
onds
,
making
it
e
f
f
icie
nt
f
or
da
tas
e
ts
that
f
it
withi
n
memor
y
li
mi
ts
.
T
his
e
f
f
icie
nc
y
c
omes
f
r
o
m
the
c
los
e
d
-
f
or
m
s
olut
ion
of
the
Nor
mal
E
qua
ti
o
ns
,
whic
h
a
ll
ows
f
or
quick
c
a
lcula
ti
ons
whe
n
da
ta
s
ize
is
ma
na
ge
a
ble.
T
he
L
-
B
F
GS
model,
a
n
it
e
r
a
ti
ve
opti
mi
z
a
ti
on
met
hod
f
or
lar
ge
r
da
tas
e
ts
,
a
c
hieve
d
the
s
a
me
R
M
S
E
a
nd
M
S
E
a
s
the
OL
S
model
.
How
e
ve
r
,
it
s
c
omp
utational
ti
me
wa
s
s
igni
f
ica
ntl
y
longer
,
a
t
122
.
69
s
e
c
onds
,
r
e
f
lec
ti
ng
it
s
it
e
r
a
ti
ve
na
tur
e
.
De
s
pit
e
thi
s
,
the
L
-
B
F
GS
method
is
mor
e
f
lexible
a
nd
s
c
a
lable
,
m
a
king
it
s
uit
a
ble
f
or
lar
ge
da
tas
e
ts
that
e
xc
e
e
d
memor
y
li
mi
ts
.
I
ts
pe
r
f
or
manc
e
a
nd
c
oe
f
f
icie
nt
a
li
gnment
with
the
OL
S
model
c
onf
ir
m
it
s
e
f
f
e
c
ti
ve
ne
s
s
in
c
a
ptur
ing
t
he
da
tas
e
t’
s
li
ne
a
r
r
e
lations
hips
.
C
ompar
ing
the
two
models
,
bo
th
s
howe
d
s
im
il
a
r
pr
e
dictive
a
c
c
ur
a
c
y,
bu
t
the
OL
S
method
wa
s
f
a
s
ter
a
nd
mor
e
e
f
f
icie
nt
f
or
s
maller
da
tas
e
ts
,
wh
il
e
the
L
-
B
F
GS
method
e
xc
e
ll
e
d
in
ha
ndli
ng
lar
g
e
r
,
mo
r
e
c
ompl
e
x
da
tas
e
ts
.
T
he
c
hoice
be
twe
e
n
the
two
d
e
pe
nds
on
the
da
tas
e
t
s
ize
a
nd
c
omput
a
ti
ona
l
ne
e
ds
,
with
OL
S
f
a
vor
e
d
f
o
r
s
pe
e
d
a
nd
L
-
B
F
GS
f
or
s
c
a
labili
ty.
Unde
r
s
tanding
thes
e
t
r
a
de
-
of
f
s
e
ns
ur
e
s
the
a
ppr
opr
iate
model
is
us
e
d
f
or
e
f
f
icie
nt
a
nd
a
c
c
ur
a
te
a
na
lys
is
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
A
nalys
is
of
big
data
fr
om
N
e
w
Y
or
k
taxi
tr
ip
2023
:
r
e
v
e
nue
pr
e
diction
us
ing
…
(
Sar
a
R
houas
)
717
4.
CONC
L
USI
ON
T
he
gr
owing
int
e
r
e
s
t
in
big
da
ta
a
nd
mac
hine
l
e
a
r
ning
ha
s
r
e
volut
ioni
z
e
d
numer
ous
indus
tr
ies
,
including
ur
ba
n
tr
a
ns
por
tation
.
L
e
ve
r
a
ging
th
e
s
e
a
dva
nc
e
d
tec
hnologi
e
s
a
ll
ows
f
or
mor
e
inf
or
med
de
c
is
ion
-
making,
ope
r
a
ti
ona
l
e
f
f
icie
nc
y,
a
nd
e
n
ha
nc
e
d
c
us
tom
e
r
e
xpe
r
ienc
e
s
.
I
n
thi
s
c
ontext,
a
na
lyzing
e
xtens
ive
da
tas
e
t
s
,
s
uc
h
a
s
thos
e
ge
ne
r
a
ted
by
Ne
w
Yor
k
C
it
y's
ye
ll
ow
taxi
s
e
r
vice
s
,
pr
ov
ides
va
luable
ins
ight
s
int
o
the
pe
r
f
or
manc
e
a
nd
mar
ke
t
dyna
mi
c
s
of
c
ompeting
ve
ndor
s
.
T
his
s
tudy
ha
r
ne
s
s
e
s
the
powe
r
of
big
da
ta
a
nd
mac
hine
lea
r
n
ing
to
e
va
luate
the
o
pe
r
a
ti
ona
l
metr
ics
of
two
majo
r
taxi
ve
ndor
s
,
of
f
e
r
ing
a
de
tailed
c
ompar
is
on
of
thei
r
e
f
f
e
c
ti
ve
ne
s
s
in
mee
ti
ng
pa
s
s
e
nge
r
de
mand
a
nd
ge
ne
r
a
ti
ng
r
e
ve
nue
.
I
n
c
onc
lus
ion,
the
int
e
gr
a
ti
on
of
b
ig
da
ta
a
nd
mac
hine
lea
r
ning
in
a
na
lyzing
Ne
w
Yor
k
C
it
y's
ye
ll
ow
taxi
indus
tr
y
r
e
ve
a
ls
ve
ndor
2
a
s
the
mor
e
domi
na
nt
a
nd
f
inanc
ially
s
uc
c
e
s
s
f
ul
ope
r
a
tor
.
Highe
r
tr
ip
volum
e
s
,
longer
a
ve
r
a
ge
tr
ip
dis
tanc
e
s
,
h
igher
f
a
r
e
a
mount
s
,
a
nd
gr
e
a
ter
ti
ps
pos
it
ion
ve
ndo
r
2
a
s
a
s
tr
onge
r
c
ompetit
or
with
a
be
tt
e
r
a
bil
it
y
to
mee
t
pa
s
s
e
nge
r
de
mands
a
nd
ge
ne
r
a
te
r
e
ve
nue
.
T
he
s
e
ins
ight
s
a
r
e
ins
tr
um
e
ntal
f
or
both
ve
ndor
s
in
opti
mi
z
ing
their
ope
r
a
ti
ons
,
im
pr
o
ving
s
e
r
vice
qua
li
ty,
a
nd
making
da
ta
-
dr
iven
de
c
is
ions
that
e
nha
nc
e
c
us
tom
e
r
s
a
ti
s
f
a
c
ti
on
a
nd
ope
r
a
ti
ona
l
e
f
f
i
c
ienc
y.
T
h
is
s
tudy
e
xe
mpl
if
ies
the
t
r
a
ns
f
or
mative
potential
of
big
da
ta
a
nd
mac
hine
lea
r
ning
in
ur
ba
n
tr
a
ns
por
tation,
pa
ving
the
wa
y
f
or
mor
e
e
f
f
e
c
ti
ve
a
nd
c
o
mpetit
ive
s
e
r
vice
de
li
ve
r
y.
RE
F
E
RE
NC
E
S
[1
]
B.
It
ri
,
Y
.
M
o
h
ame
d
,
B.
O
mar,
E
.
M.
L
at
i
fa,
M.
L
ah
c
en
,
an
d
O
.
A
d
i
l
,
“H
y
b
r
i
d
mac
h
i
n
e
l
e
arn
i
n
g
fo
r
s
t
o
ck
p
ri
ce
p
red
i
ct
i
o
n
i
n
t
h
e
Mo
ro
cca
n
b
a
n
k
i
n
g
s
ect
o
r,
”
In
t
e
r
n
a
t
i
o
n
a
l
Jo
u
r
n
a
l
o
f
E
l
ect
r
i
c
a
l
a
n
d
C
o
m
p
u
t
er
E
n
g
i
n
ee
r
i
n
g
,
v
o
l
.
1
4
,
n
o
.
3
,
p
p
.
3
1
9
7
–
3
2
0
7
,
J
u
n
.
2
0
2
4
,
d
o
i
:
1
0
.
1
1
5
9
1
/
i
j
ece.
v
1
4
i
3
.
p
p
3
1
9
7
-
3
2
0
7
.
[2
]
Q
.
H
u
,
L
.
Z
h
u
,
C.
Ch
an
g
,
an
d
W
.
Z
h
an
g
,
“A
t
ru
n
cat
ed
t
h
ree
-
t
erm
co
n
j
u
g
at
e
g
ra
d
i
e
n
t
met
h
o
d
w
i
t
h
co
m
p
l
e
x
i
t
y
g
u
ara
n
t
ee
s
w
i
t
h
ap
p
l
i
cat
i
o
n
s
t
o
n
o
n
co
n
v
e
x
reg
res
s
i
o
n
p
ro
b
l
em,
”
A
p
p
l
i
ed
Nu
m
er
i
ca
l
M
a
t
h
e
m
a
t
i
c
s
,
v
o
l
.
1
9
4
,
p
p
.
8
2
–
9
6
,
D
ec.
2
0
2
3
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
ap
n
u
m.
2
0
2
3
.
0
8
.
0
0
6
.
[3
]
A
.
L
.
Bu
rt
o
n
,
“O
rd
i
n
ar
y
l
eas
t
s
q
u
ares
(
l
i
n
ear)
reg
r
es
s
i
o
n
,
”
i
n
Th
e
E
n
cycl
o
p
e
d
i
a
o
f
R
e
s
ea
r
ch
M
et
h
o
d
s
i
n
Cr
i
m
i
n
o
l
o
g
y
a
n
d
Cr
i
m
i
n
a
l
Ju
s
t
i
ce
,
W
i
l
ey
,
2
0
2
1
,
p
p
.
5
0
9
–
5
1
4
.
[4
]
S.
Rh
o
u
a
s
,
A
.
E
l
A
t
t
ao
u
i
,
an
d
N
.
E
l
H
am
i
,
“O
p
t
i
m
i
za
t
i
o
n
o
f
t
h
e
p
red
i
ct
i
o
n
p
erf
o
rman
ce
i
n
t
h
e
fu
t
u
re
e
x
ch
a
n
g
e
rat
e,
”
i
n
2
0
2
3
9
t
h
In
t
e
r
n
a
t
i
o
n
a
l
Co
n
f
er
e
n
ce
o
n
O
p
t
i
m
i
z
a
t
i
o
n
a
n
d
A
p
p
l
i
ca
t
i
o
n
s
(ICO
A
)
,
O
ct
.
2
0
2
3
,
p
p
.
1
–
6
,
d
o
i
:
1
0
.
1
1
0
9
/
ICO
A
5
8
2
7
9
.
2
0
2
3
.
1
0
3
0
8
8
5
8
.
[5
]
A
.
E
l
A
t
t
a
o
u
i
,
S
.
R
h
o
u
a
s
,
a
n
d
N
.
E
l
H
a
m
i
,
“
E
T
L
a
p
p
l
i
e
d
t
o
k
l
a
r
n
a
e
-
c
o
mm
er
c
e
d
a
t
a
s
e
t
,
”
i
n
2
0
2
3
9
th
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
O
p
t
i
m
i
z
a
t
i
o
n
a
n
d
A
p
p
l
i
c
a
t
i
o
n
s
(I
C
O
A
)
,
O
ct
.
2
0
2
3
,
p
p
.
1
–
4
,
d
o
i
:
1
0
.
1
1
0
9
/
I
C
O
A
5
8
2
7
9
.
2
0
2
3
.
1
0
3
0
8
8
0
8
.
[6
]
M.
B.
U
l
ak
,
A
.
Y
az
i
ci
,
a
n
d
M.
A
l
j
arra
h
,
“V
a
l
u
e
o
f
c
o
n
v
en
i
en
ce
f
o
r
t
ax
i
t
r
i
p
s
i
n
N
ew
Y
o
rk
C
i
t
y
,
”
Tr
a
n
s
p
o
r
t
a
t
i
o
n
R
es
e
a
r
c
h
P
a
r
t
A
:
P
o
l
i
cy
a
n
d
P
r
a
ct
i
ce
,
v
o
l
.
1
4
2
,
p
p
.
8
5
–
1
0
0
,
D
ec.
2
0
2
0
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
t
ra.
2
0
2
0
.
1
0
.
0
1
6
.
[7
]
M.
S.
A
n
s
ar,
Y
.
Ma,
S
.
Ch
en
,
K
.
T
an
g
,
an
d
Z
.
Z
h
an
g
,
“In
v
e
s
t
i
g
a
t
i
n
g
t
h
e
t
ri
p
co
n
f
i
g
u
red
ca
u
s
a
l
effect
o
f
d
i
s
t
ra
c
t
ed
d
ri
v
i
n
g
o
n
ag
g
res
s
i
v
e
d
r
i
v
i
n
g
b
e
h
av
i
o
r
f
o
r
e
-
h
ai
l
i
n
g
t
ax
i
d
r
i
v
er
s
,
”
Jo
u
r
n
a
l
o
f
Tr
a
f
f
i
c
a
n
d
Tr
a
n
s
p
o
r
t
a
t
i
o
n
E
n
g
i
n
eer
i
n
g
(E
n
g
l
i
s
h
E
d
i
t
i
o
n
)
,
v
o
l
.
8
,
n
o
.
5
,
p
p
.
7
2
5
–
7
3
4
,
O
ct
.
2
0
2
1
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
j
t
t
e.
2
0
2
0
.
1
2
.
0
0
1
.
[8
]
X
.
D
o
n
g
,
E
.
G
u
erra,
an
d
M.
S.
Ry
ers
o
n
,
“I
n
v
e
s
t
i
g
a
t
i
n
g
t
h
e
rec
o
v
er
y
o
f
fo
r
-
h
i
re
-
v
e
h
i
c
l
e,
t
ax
i
,
an
d
a
i
rt
ra
i
n
a
t
t
w
o
N
ew
Y
o
rk
C
i
t
y
ai
r
p
o
r
t
s
d
u
r
i
n
g
t
h
e
CO
V
ID
-
1
9
p
an
d
e
mi
c,
”
Tr
a
ve
l
B
e
h
a
v
i
o
u
r
a
n
d
S
o
ci
e
t
y
,
v
o
l
.
3
3
,
O
ct
.
2
0
2
3
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
t
b
s
.
2
0
2
3
.
1
0
0
6
4
6
.
[9
]
D
.
K
at
i
ć,
H
.
K
rs
t
i
ć,
I.
Iš
t
o
k
a
O
t
k
o
v
i
ć,
an
d
H
.
Beg
i
ć
J
u
ri
či
ć,
“Co
mp
ar
i
n
g
mu
l
t
i
p
l
e
l
i
n
ear
reg
res
s
i
o
n
an
d
n
e
u
ral
n
et
w
o
r
k
mo
d
e
l
s
fo
r
p
re
d
i
c
t
i
n
g
h
eat
i
n
g
en
erg
y
co
n
s
u
m
p
t
i
o
n
i
n
s
c
h
o
o
l
b
u
i
l
d
i
n
g
s
i
n
t
h
e
Fed
erat
i
o
n
o
f
Bo
s
n
i
a
an
d
H
erzeg
o
v
i
n
a,
”
Jo
u
r
n
a
l
o
f
B
u
i
l
d
i
n
g
E
n
g
i
n
ee
r
i
n
g
,
v
o
l
.
9
7
,
N
o
v
.
2
0
2
4
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
j
o
b
e.
2
0
2
4
.
1
1
0
7
2
8
.
[1
0
]
B.
V
.
Su
ry
a
V
ar
d
h
a
n
,
M.
K
h
e
d
k
a
r,
I.
Sri
v
a
s
t
a
v
a,
an
d
S.
K
.
Pat
ro
,
“Im
p
act
o
f
i
n
t
eg
ra
t
ed
cl
a
s
s
i
fi
er
—
re
g
res
s
i
o
n
map
p
e
d
s
h
o
r
t
t
erm
l
o
a
d
fo
reca
s
t
i
n
g
o
n
p
o
w
er
s
y
s
t
em
man
ag
eme
n
t
i
n
a
g
ri
d
co
n
n
ec
t
ed
m
u
l
t
i
e
n
erg
y
s
y
s
t
e
ms
,
”
E
l
ec
t
r
i
c
P
o
we
r
S
y
s
t
e
m
s
R
e
s
ea
r
ch
,
v
o
l
.
2
3
0
,
May
2
0
2
4
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
e
p
s
r.
2
0
2
4
.
1
1
0
2
2
2
.
[1
1
]
M.
A
rman
u
r
Rah
ma
n
,
A
.
H
o
s
s
e
n
,
J
.
H
o
s
s
e
n
,
V
.
C,
T
.
Bh
u
v
a
n
es
w
ari
,
an
d
A
.
Su
l
t
a
n
a,
“T
o
w
ard
s
mach
i
n
e
l
earn
i
n
g
-
b
a
s
ed
s
e
l
f
-
t
u
n
i
n
g
o
f
h
ad
o
o
p
-
s
p
ark
s
y
s
t
em,
”
In
d
o
n
es
i
a
n
Jo
u
r
n
a
l
o
f
E
l
ec
t
r
i
ca
l
E
n
g
i
n
eer
i
n
g
a
n
d
Co
m
p
u
t
e
r
S
ci
e
n
ce
,
v
o
l
.
1
5
,
n
o
.
2
,
p
p
.
1
0
7
6
–
1
0
8
5
,
A
u
g
.
2
0
1
9
,
d
o
i
:
1
0
.
1
1
5
9
1
/
i
j
eecs
.
v
1
5
.
i
2
.
p
p
1
0
7
6
-
1
0
8
5
.
[1
2
]
A
.
Man
co
n
i
,
M.
G
n
o
cch
i
,
L
.
Mi
l
an
e
s
i
,
O
.
Maru
l
l
o
,
a
n
d
G
.
A
rma
n
o
,
“Frami
n
g
A
p
ach
e
s
p
ar
k
i
n
l
i
fe
s
ci
e
n
c
es
,
”
H
el
i
yo
n
,
v
o
l
.
9
,
n
o
.
2
,
Feb
.
2
0
2
3
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
h
el
i
y
o
n
.
2
0
2
3
.
e1
3
3
6
8
.
[1
3
]
P.
J
h
a,
A
.
T
i
w
ari
,
N
.
Bh
ari
l
l
,
M.
Rat
n
ap
ar
k
h
e,
M.
Mo
u
n
i
k
a,
an
d
N
.
N
a
g
en
d
ra,
“A
p
ach
e
s
p
ar
k
b
as
e
d
k
er
n
el
i
zed
fu
zzy
c
l
u
s
t
er
i
n
g
framew
o
rk
f
o
r
s
i
n
g
l
e
n
u
c
l
eo
t
i
d
e
p
o
l
y
mo
rp
h
i
s
m
s
eq
u
e
n
ce
an
a
l
y
s
i
s
,
”
Co
m
p
u
t
a
t
i
o
n
a
l
B
i
o
l
o
g
y
a
n
d
Ch
em
i
s
t
r
y
,
v
o
l
.
9
2
,
J
u
n
.
2
0
2
1
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
co
m
p
b
i
o
l
c
h
em.
2
0
2
1
.
1
0
7
4
5
4
.
[1
4
]
M.
B.
A
l
-
Mas
a
d
eh
,
M.
S.
A
zmi
,
a
n
d
S.
S.
Sy
e
d
A
h
ma
d
,
“T
i
n
y
d
at
a
b
l
o
ck
i
n
s
av
i
n
g
h
a
d
o
o
p
d
i
s
t
r
i
b
u
t
e
d
fi
l
e
s
y
s
t
em
w
as
t
ed
memo
r
y
,
”
In
t
er
n
a
t
i
o
n
a
l
Jo
u
r
n
a
l
o
f
E
l
ect
r
i
c
a
l
a
n
d
Co
m
p
u
t
er
E
n
g
i
n
e
er
i
n
g
,
v
o
l
.
1
3
,
n
o
.
2
,
p
p
.
1
7
5
7
–
1
7
7
2
,
A
p
r.
2
0
2
3
,
d
o
i
:
1
0
.
1
1
5
9
1
/
i
j
ece.
v
1
3
i
2
.
p
p
1
7
5
7
-
1
7
7
2
.
[1
5
]
F.
A
s
h
k
o
u
t
i
,
K
.
K
h
amfo
r
o
o
s
h
,
an
d
A
.
Sh
e
i
k
h
ah
m
ad
i
,
“D
I
-
M
o
n
d
ri
a
n
:
d
i
s
t
r
i
b
u
t
e
d
i
m
p
ro
v
ed
M
o
n
d
ri
a
n
fo
r
s
at
i
s
fac
t
i
o
n
o
f
t
h
e
L
-
d
i
v
er
s
i
t
y
p
ri
v
acy
m
o
d
e
l
u
s
i
n
g
A
p
ach
e
s
p
ark
,
”
I
n
f
o
r
m
a
t
i
o
n
S
c
i
e
n
ces
,
v
o
l
.
5
4
6
,
p
p
.
1
–
2
4
,
Feb
.
2
0
2
1
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
i
n
s
.
2
0
2
0
.
0
7
.
0
6
6
.
[1
6
]
Y
.
L
i
u
an
d
S.
Cao
,
“T
h
e
an
al
y
s
i
s
o
f
aero
b
i
c
s
i
n
t
e
l
l
i
g
e
n
t
fi
t
n
e
s
s
s
y
s
t
em
f
o
r
n
e
u
ro
r
o
b
o
t
i
cs
b
as
e
d
o
n
b
i
g
d
at
a
an
d
mach
i
n
e
l
earn
i
n
g
,
”
H
el
i
yo
n
,
v
o
l
.
1
0
,
n
o
.
1
2
,
J
u
n
.
2
0
2
4
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
h
e
l
i
y
o
n
.
2
0
2
4
.
e3
3
1
9
1
.
[1
7
]
S.
G
Pu
ro
h
i
t
an
d
V
.
Sw
amy
,
“
E
n
h
an
c
i
n
g
d
a
t
a
p
u
b
l
i
s
h
i
n
g
p
ri
v
acy
:
s
p
l
i
t
-
a
n
d
-
m
o
u
l
d
,
a
n
al
g
o
r
i
t
h
m
fo
r
eq
u
i
v
al
en
t
s
p
ec
i
fi
ca
t
i
o
n
,
”
In
d
o
n
es
i
a
n
J
o
u
r
n
a
l
o
f
E
l
ec
t
r
i
ca
l
E
n
g
i
n
eer
i
n
g
a
n
d
C
o
m
p
u
t
er
S
ci
e
n
ce
,
v
o
l
.
3
3
,
n
o
.
2
,
Feb
.
2
0
2
4
,
d
o
i
:
1
0
.
1
1
5
9
1
/
i
j
eec
s
.
v
3
3
.
i
2
.
p
p
1
2
7
3
-
1
2
8
2
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
15
,
No.
1
,
F
e
br
ua
r
y
20
25
:
711
-
718
718
[1
8
]
Y
.
H
u
a
n
g
,
W
.
X
u
,
P.
Su
k
j
a
i
ru
n
g
w
at
t
an
a,
a
n
d
Z
.
Y
u
,
“L
earn
ers
’
co
n
t
i
n
u
an
ce
i
n
t
en
t
i
o
n
i
n
m
u
l
t
i
m
o
d
a
l
l
a
n
g
u
ag
e
l
earn
i
n
g
ed
u
cat
i
o
n
:
a
n
i
n
n
o
v
at
i
v
e
mu
l
t
i
p
l
e
l
i
n
ear
reg
r
es
s
i
o
n
m
o
d
e
l
,
”
H
e
l
i
y
o
n
,
v
o
l
.
1
0
,
n
o
.
6
,
Mar.
2
0
2
4
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
h
e
l
i
y
o
n
.
2
0
2
4
.
e
2
8
1
0
4
.
[1
9
]
C.
K
l
ei
b
er,
“Fi
n
i
t
e
s
am
p
l
e
eff
i
ci
e
n
cy
o
f
O
L
S
i
n
l
i
n
ear
reg
res
s
i
o
n
mo
d
el
s
w
i
t
h
l
o
n
g
-
mem
o
ry
d
i
s
t
u
rb
a
n
ce
s
,
”
E
co
n
o
m
i
c
s
Let
t
e
r
s
,
v
o
l
.
7
2
,
n
o
.
2
,
p
p
.
1
3
1
–
1
3
6
,
A
u
g
.
2
0
0
1
,
d
o
i
:
1
0
.
1
0
1
6
/
S
0
1
6
5
-
1
7
6
5
(
0
1
)0
0
4
2
3
-
2.
[2
0
]
I.
A
h
mad
et
a
l
.
,
“Sp
at
i
al
co
n
fi
g
u
ra
t
i
o
n
o
f
g
ro
u
n
d
w
a
t
er
p
o
t
en
t
i
a
l
zo
n
es
u
s
i
n
g
O
L
S
reg
res
s
i
o
n
met
h
o
d
,
”
Jo
u
r
n
a
l
o
f
A
f
r
i
c
a
n
E
a
r
t
h
S
ci
e
n
ces
,
v
o
l
.
1
7
7
,
May
2
0
2
1
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
j
afrears
c
i
.
2
0
2
1
.
1
0
4
1
4
7
.
[2
1
]
E
.
G
h
y
s
el
s
an
d
H
.
Q
i
an
,
“E
s
t
i
mat
i
n
g
MID
A
S
re
g
r
es
s
i
o
n
s
v
i
a
O
L
S
w
i
t
h
p
o
l
y
n
o
m
i
al
p
a
ramet
er
p
r
o
fi
l
i
n
g
,
”
E
co
n
o
m
et
r
i
c
s
a
n
d
S
t
a
t
i
s
t
i
cs
,
v
o
l
.
9
,
p
p
.
1
–
1
6
,
J
a
n
.
2
0
1
9
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
ec
o
s
t
a.
2
0
1
8
.
0
2
.
0
0
1
.
[2
2
]
A
.
Bemp
o
rad
,
“A
n
L
-
BFG
S
-
B
ap
p
ro
ach
fo
r
l
i
n
e
ar
an
d
n
o
n
l
i
n
ear
s
y
s
t
em
i
d
en
t
i
f
i
ca
t
i
o
n
u
n
d
er
l
1
an
d
g
ro
u
p
-
L
as
s
o
reg
u
l
ari
za
t
i
o
n
,
”
a
r
X
i
v:2
4
0
3
.
0
3
8
2
7
,
Mar.
2
0
2
4
.
[2
3
]
F.
A
l
p
a
k
et
a
l
.
,
“A
mach
i
n
e
-
l
ear
n
i
n
g
-
acce
l
erat
e
d
d
i
s
t
ri
b
u
t
ed
L
BFG
S
met
h
o
d
fo
r
fi
e
l
d
d
e
v
el
o
p
me
n
t
o
p
t
i
m
i
zat
i
o
n
:
al
g
o
ri
t
h
m,
v
al
i
d
a
t
i
o
n
,
an
d
ap
p
l
i
ca
t
i
o
n
s
,
”
Co
m
p
u
t
a
t
i
o
n
a
l
G
e
o
s
c
i
en
ce
s
,
v
o
l
.
2
7
,
n
o
.
3
,
p
p
.
4
2
5
–
4
5
0
,
J
u
n
.
2
0
2
3
,
d
o
i
:
1
0
.
1
0
0
7
/
s
1
0
5
9
6
-
0
2
3
-
1
0
1
9
7
-
3.
[2
4
]
D
.
Ch
an
g
,
S.
Su
n
,
an
d
C.
Z
h
an
g
,
“A
n
accel
erat
e
d
l
i
n
ear
l
y
co
n
v
erg
e
n
t
s
t
o
c
h
as
t
i
c
L
-
BFG
S
al
g
o
ri
t
h
m,
”
IE
E
E
Tr
a
n
s
a
c
t
i
o
n
s
o
n
Ne
u
r
a
l
Ne
t
wo
r
ks
a
n
d
Le
a
r
n
i
n
g
S
ys
t
e
m
s
,
v
o
l
.
3
0
,
n
o
.
1
1
,
p
p
.
3
3
3
8
–
3
3
4
6
,
N
o
v
.
2
0
1
9
,
d
o
i
:
1
0
.
1
1
0
9
/
T
N
N
L
S.
2
0
1
9
.
2
8
9
1
0
8
8
.
[2
5
]
M.
W
.
L
i
emo
h
n
,
A
.
D
.
Sh
an
e,
A
.
R.
A
zari
,
A
.
K
.
Pet
ers
en
,
B.
M.
Sw
i
g
er,
an
d
A
.
Mu
k
h
o
p
ad
h
y
a
y
,
“RMSE
i
s
n
o
t
en
o
u
g
h
:
g
u
i
d
el
i
n
e
s
t
o
ro
b
u
s
t
d
a
t
a
-
mo
d
el
co
m
p
ari
s
o
n
s
fo
r
mag
n
e
t
o
s
p
h
eri
c
p
h
y
s
i
c
s
,
”
Jo
u
r
n
a
l
o
f
A
t
m
o
s
p
h
er
i
c
a
n
d
S
o
l
a
r
-
Te
r
r
es
t
r
i
a
l
P
h
y
s
i
c
s
,
v
o
l
.
2
1
8
,
J
u
l
.
2
0
2
1
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
j
a
s
t
p
.
2
0
2
1
.
1
0
5
6
2
4
.
[2
6
]
S.
Rh
o
u
as
,
A
.
E
l
A
t
t
ao
u
i
,
an
d
N
.
E
l
H
ami
,
“E
n
h
an
c
i
n
g
cu
rren
c
y
p
red
i
ct
i
o
n
i
n
i
n
t
ern
at
i
o
n
al
e
-
co
mmerce
:
Bay
es
i
an
-
o
p
t
i
mi
zed
ran
d
o
m
f
o
res
t
ap
p
r
o
ach
u
s
i
n
g
t
h
e
K
l
ar
n
a
d
at
as
e
t
,
”
In
t
e
r
n
a
t
i
o
n
a
l
Jo
u
r
n
a
l
o
f
E
l
ect
r
i
c
a
l
a
n
d
Co
m
p
u
t
e
r
E
n
g
i
n
eer
i
n
g
,
v
o
l
.
1
4
,
n
o
.
3
,
J
u
n
.
2
0
2
4
,
d
o
i
:
1
0
.
1
1
5
9
1
/
i
j
e
ce.
v
1
4
i
3
.
p
p
3
1
7
7
-
3
1
8
6
.
[2
7
]
S.
H
ad
i
y
o
s
o
,
H
.
N
u
g
ro
h
o
,
T
.
L
.
E
raw
at
i
Raj
a
b
,
an
d
K
.
Su
ren
d
ro
,
“D
a
t
a
p
re
d
i
c
t
i
o
n
fo
r
cas
e
s
o
f
i
n
co
rrec
t
d
at
a
i
n
mu
l
t
i
-
n
o
d
e
el
ect
r
o
card
i
o
g
ram
mo
n
i
t
o
ri
n
g
,
”
In
t
e
r
n
a
t
i
o
n
a
l
Jo
u
r
n
a
l
o
f
E
l
ect
r
i
c
a
l
a
n
d
Co
m
p
u
t
e
r
E
n
g
i
n
eer
i
n
g
,
v
o
l
.
1
2
,
n
o
.
2
,
p
p
.
1
5
4
0
–
1
5
4
7
,
A
p
r.
2
0
2
2
,
d
o
i
:
1
0
.
1
1
5
9
1
/
i
j
ece.
v
1
2
i
2
.
p
p
1
5
4
0
-
1
5
4
7
.
[2
8
]
“T
L
C
t
ri
p
reco
r
d
d
at
a,
”
Ta
xi
a
n
d
Li
m
o
u
s
i
n
e
C
o
m
m
i
s
s
i
o
n
.
h
t
t
p
s
:
/
/
w
w
w
.
n
y
c.
g
o
v
/
s
i
t
e/
t
l
c
/
ab
o
u
t
/
t
l
c
-
t
ri
p
-
rec
o
rd
-
d
at
a.
p
a
g
e
(acces
s
ed
J
u
n
.
1
3
,
2
0
2
4
).
B
I
OG
RA
P
HI
E
S
OF
AU
T
HO
RS
Sa
ra
R
h
o
ua
s
i
s
cu
rren
t
l
y
p
u
r
s
u
i
n
g
h
er
Ph
.
D
.
i
n
co
m
p
u
t
er
s
ci
e
n
ce
at
t
h
e
N
a
t
i
o
n
a
l
Sch
o
o
l
o
f
A
p
p
l
i
e
d
Sci
e
n
ces
,
Ib
n
T
o
fai
l
U
n
i
v
ers
i
t
y
i
n
Mo
ro
cc
o
.
Sh
e
earn
e
d
h
er
e
n
g
i
n
eeri
n
g
d
eg
ree
i
n
In
d
u
s
t
ri
a
l
E
n
g
i
n
eeri
n
g
i
n
2
0
1
9
fro
m
t
h
e
s
am
e
i
n
s
t
i
t
u
t
i
o
n
.
H
er
acad
em
i
c
ex
p
er
i
en
ce
i
n
c
l
u
d
es
a
s
t
r
o
n
g
fo
c
u
s
o
n
au
t
o
m
o
b
i
l
e
t
ech
n
o
l
o
g
i
e
s
,
w
i
t
h
a
p
art
i
cu
l
ar
i
n
t
eres
t
i
n
b
ra
k
i
n
g
s
y
s
t
ems
.
Sh
e
h
a
s
carri
e
d
o
u
t
res
earch
i
n
t
h
e
fi
e
l
d
o
f
o
p
t
i
mi
zat
i
o
n
al
g
o
ri
t
h
m
s
an
d
h
er
res
earc
h
i
n
t
eres
t
s
e
x
t
e
n
d
t
o
area
s
s
u
ch
a
s
b
i
g
d
a
t
a,
i
n
t
er
o
p
era
b
i
l
i
t
y
,
art
i
fi
c
i
al
i
n
t
el
l
i
g
en
ce,
mac
h
i
n
e
l
earn
i
n
g
,
an
d
d
eep
l
ear
n
i
n
g
.
Sh
e
h
as
au
t
h
o
r
ed
an
d
c
o
-
au
t
h
o
re
d
s
ev
era
l
p
u
b
l
i
ca
t
i
o
n
s
i
n
b
o
t
h
co
n
fere
n
ces
an
d
s
c
i
en
t
i
f
i
c
j
o
u
r
n
al
s
.
Sh
e
can
b
e
co
n
t
ac
t
e
d
at
emai
l
:
rh
o
u
a
s
.
s
ara@
g
ma
i
l
.
c
o
m
.
N
o
re
l
i
s
l
a
m
El
Ha
m
i
i
s
a
p
ro
fes
s
o
r
o
f
co
mp
u
t
er
s
c
i
e
n
ce
at
t
h
e
N
at
i
o
n
al
Sch
o
o
l
o
f
A
p
p
l
i
ed
Sc
i
en
ce
s
,
Ib
n
T
o
fai
l
U
n
i
v
er
s
i
t
y
,
i
n
K
en
i
t
ra,
Mo
ro
cc
o
.
H
e
ear
n
ed
a
d
i
p
l
o
ma
o
f
s
t
a
t
e
en
g
i
n
eer
i
n
2
0
0
0
,
s
p
ec
i
al
i
z
i
n
g
i
n
co
m
p
u
t
er
an
d
t
e
l
ec
o
mmu
n
i
ca
t
i
o
n
s
fro
m
t
h
e
Po
l
y
t
ech
n
i
c
Facu
l
t
y
o
f
Mo
n
s
(FPMS)
i
n
Bel
g
i
u
m.
H
e
h
o
l
d
s
a
Ph
.
D
.
i
n
co
mp
u
t
er
s
c
i
en
ce
fro
m
t
h
e
N
at
i
o
n
al
In
s
t
i
t
u
t
e
o
f
A
p
p
l
i
e
d
Sci
en
ce
s
(IN
S
A
)
o
f
Ro
u
en
,
Fr
an
ce,
as
w
el
l
as
a
Ph
.
D
.
i
n
ap
p
l
i
ed
mat
h
ema
t
i
c
s
an
d
co
m
p
u
t
er
s
c
i
en
ce
fr
o
m
M
o
h
amme
d
V
U
n
i
v
ers
i
t
y
i
n
Rab
a
t
,
Mo
r
o
cco
.
H
i
s
w
o
r
k
i
n
cl
u
d
e
s
n
u
mero
u
s
s
ch
o
l
ar
l
y
p
u
b
l
i
ca
t
i
o
n
s
i
n
c
o
n
fere
n
ces
a
n
d
j
o
u
rn
a
l
s
.
H
e
can
b
e
co
n
t
act
e
d
at
emai
l
:
n
o
rel
i
s
l
am@
o
u
t
l
o
o
k
.
co
m.
Evaluation Warning : The document was created with Spire.PDF for Python.