Zementis © - Confidential PMML Overview
Ze
me
nti
s ©
2
Mo
dels
PM
ML
de
fines a
sta
ndard
no
t only
to
repre
sen
t data
-min
ing m
odels
, but
als
o d
ata
han
dli
ng
and d
ata
tran
sfo
rma
tio
ns
(pre
-and
post-
pro
cessin
g)
PM
ML
Pre
dic
tive M
od
el M
ark
up
Lan
gu
ag
e
Tra
nsfo
rmati
on
s
•P
MM
L is a
n X
ML
-based
la
ng
ua
ge
used to
define s
tatistical and d
ata
min
ing m
odels
and to
share
these b
etw
een c
om
plia
nt applic
ations.
•It
is a
ma
ture
sta
nd
ard
develo
ped b
y the D
MG
(Data
Min
ing G
roup)
to a
void
pro
prieta
ry issues
and incom
patibili
ties a
nd to d
eplo
y m
odels
.
•P
MM
L a
llow
s fo
r th
e c
lear
sepa
ration o
f ta
sks:
Model develo
pm
ent vs.
model deplo
ym
en
t. A
s a
consequence, scie
ntists
can focus o
n b
uild
ing the
best m
odel.
•P
MM
L e
limin
ate
s n
eed fo
r custo
m m
odel
deplo
ym
en
t and e
nsure
s s
cala
bili
ty a
nd
relia
bili
ty.
Ze
me
nti
s ©
-C
on
fid
en
tia
l3
Matu
red
an
d S
up
po
rted
by I
nd
ustr
yM
atu
red
an
d S
up
po
rted
by I
nd
ustr
y
PM
ML
PM
ML
In
du
str
y S
up
po
rt
�D
ata
Min
ing G
roup htt
p:/
/ww
w.d
mg.o
rg
�M
atu
re s
tanda
rd
�C
urr
en
t ve
rsio
n 4
.0 (
just re
leased)
�A
ctive g
roup a
nd
consta
nt enhancem
ents
�V
endor
independent consort
ium
�In
dustr
y s
uppo
rters
�M
ajo
r P
laye
rs:
IBM
, O
racle
, S
AP
, M
icro
soft
�A
naly
tics: K
XE
N, S
AS
, S
alford
, S
PS
S, Z
em
entis
�B
I: M
icro
str
ate
gy, T
era
data
, T
ibco
�O
pen
Sourc
e: K
NIM
E, R
, R
apid
-I
�O
thers
: E
quifax,
FIC
O,
Open D
ata
Gro
up
, V
isa,
Perv
asiv
e
�D
ata
Min
ing G
roup htt
p:/
/ww
w.d
mg.o
rg
�M
atu
re s
tanda
rd
�C
urr
en
t ve
rsio
n 4
.0 (
just re
leased)
�A
ctive g
roup a
nd
consta
nt enhancem
ents
�V
endor
independent consort
ium
�In
dustr
y s
uppo
rters
�M
ajo
r P
laye
rs:
IBM
, O
racle
, S
AP
, M
icro
soft
�A
naly
tics: K
XE
N, S
AS
, S
alford
, S
PS
S, Z
em
entis
�B
I: M
icro
str
ate
gy, T
era
data
, T
ibco
�O
pen
Sourc
e: K
NIM
E, R
, R
apid
-I
�O
thers
: E
quifax,
FIC
O,
Open D
ata
Gro
up
, V
isa,
Perv
asiv
e
Ze
me
nti
s ©
4
PM
ML
Co
mp
on
en
ts
�A
Data
Dic
tio
na
ryde
fines a
ll th
e r
aw
data
fie
lds (
inclu
din
g m
issin
g v
alu
e
str
ate
gy a
nd
outlie
r tr
eatm
en
t).
�S
evera
l D
ata
Tra
nsfo
rma
tio
ns
str
ate
gie
s a
llow
for
inte
lligent
extr
action o
f fe
atu
re d
ete
cto
rs f
rom
ra
w d
ata
(“d
ata
massagin
g”)
.
�A
com
pre
hensiv
e lis
t of D
ata
-Min
ing
M
od
els
off
ers
po
we
r and f
lexib
ility
.
�P
ost-
pro
cessin
g o
f re
sults a
llow
for
tailo
red d
ecis
ions.
�M
odel E
xpla
nation a
llow
s for
perf
orm
ance e
va
luation.
Ze
me
nti
s ©
5
PM
ML
Fil
es
3)
Po
st-
Pro
ce
ssin
g
Scalin
g o
f m
odel outp
uts
can b
e p
erf
orm
ed w
ith
PM
ML e
lem
ent
Targ
ets
1)
Pre
-Pro
ce
ssin
g
PM
ML e
lem
ents
Tra
nsfo
rma
tions,
Min
ing
Schem
aand F
un
ctions
allo
w fo
r e
ffective p
re-
pro
cessin
g
2)
Mo
dels
PM
ML a
llow
s for
severa
l
pre
dic
tive m
odelin
g
techniq
ues to b
e f
ully
expre
ssed
PM
ML
Ze
me
nti
s ©
6
PM
ML
: D
ata
Pre
-Pro
ce
ssin
gP
MM
L:
Data
Pre
-Pro
ce
ssin
g
�D
ata
Dic
tio
nary
: A
llow
s fo
r th
e e
xplic
it s
pecific
ation o
f valid
, in
valid
and m
issin
g v
alu
es.
�M
inin
g S
ch
em
a:
Used to d
efine t
he a
ppro
priate
tre
atm
ent
to
be a
pplie
d to m
issin
g a
nd invalid
valu
es.
�T
ran
sfo
rma
tio
ns: A
llow
for
variable
dis
cre
tization,
norm
aliz
ation, and m
appin
g w
ith h
andlin
g o
f m
issin
g a
nd
defa
ult v
alu
es.
�B
uil
t-in
Fu
ncti
on
s: A
rith
metic e
xpre
ssio
ns, handlin
g o
f da
te
and tim
e a
s w
ell
as s
trin
gs. A
lso u
sed for
imple
menting IF
-
TH
EN
-ELS
E logic
and B
oole
an o
pera
tions.
�D
ata
Dic
tio
nary
: A
llow
s fo
r th
e e
xplic
it s
pecific
ation o
f valid
, in
valid
and m
issin
g v
alu
es.
�M
inin
g S
ch
em
a:
Used to d
efine t
he a
ppro
priate
tre
atm
ent
to
be a
pplie
d to m
issin
g a
nd invalid
valu
es.
�T
ran
sfo
rma
tio
ns: A
llow
for
variable
dis
cre
tization,
norm
aliz
ation, and m
appin
g w
ith h
andlin
g o
f m
issin
g a
nd
defa
ult v
alu
es.
�B
uil
t-in
Fu
ncti
on
s: A
rith
metic e
xpre
ssio
ns, handlin
g o
f da
te
and tim
e a
s w
ell
as s
trin
gs. A
lso u
sed for
imple
menting IF
-
TH
EN
-ELS
E logic
and B
oole
an o
pera
tions.
1
Da
ta P
re-P
roc
es
sin
g
Ze
me
nti
s ©
7
Da
ta P
re-P
roc
es
sin
g:
PM
ML
Exam
ple
Arb
itra
ry P
iece
wis
e L
inear
Fu
ncti
on
Th
is P
MM
L c
od
e im
ple
men
ts:
Var_
b:=
inte
rpo
late
(Var_
a,(
(100,0
),(2
00,1
),(8
00,3
),(9
00,4
)))
See h
ttp
://w
ww
.dm
g.o
rg/v
3-2
/Tra
nsfo
rmati
on
s.h
tml -
loo
k f
or
ele
men
t N
orm
Co
nti
nu
ou
s.
Arb
itra
ry P
iece
wis
e L
inear
Fu
ncti
on
Th
is P
MM
L c
od
e im
ple
men
ts:
Var_
b:=
inte
rpo
late
(Var_
a,(
(100,0
),(2
00,1
),(8
00,3
),(9
00,4
)))
See h
ttp
://w
ww
.dm
g.o
rg/v
3-2
/Tra
nsfo
rmati
on
s.h
tml -
loo
k f
or
ele
men
t N
orm
Co
nti
nu
ou
s.
Ze
me
nti
s ©
8
Mo
deli
ng
Ele
men
tsM
od
eli
ng
Ele
men
ts
�P
MM
L a
llow
s for
severa
l p
red
icti
ve m
od
elin
gte
chniq
ues to b
e
expre
ssed d
irectly. S
upport
ed techniq
ues w
hic
h h
ave their o
wn
ele
ments
are
:
�R
egre
ssio
n a
nd
Ge
ne
ral R
egre
ssio
n
�N
eu
ral N
etw
ork
s
�S
uppo
rt V
ecto
r M
ach
ine
s
�D
ecis
ion
Tre
es
�N
aïv
e B
aye
s
�C
luste
rin
g
�S
eque
nce
s
�R
ule
Se
ts
�A
sso
cia
tion
Rule
s
�T
ime
-Se
rie
s (
as o
f P
MM
L 4
.0)
�T
ext
Mo
de
ls
�S
uppo
rt f
or
Mu
ltip
le M
od
els
�P
MM
L a
llow
s for
severa
l p
red
icti
ve m
od
elin
gte
chniq
ues to b
e
expre
ssed d
irectly. S
upport
ed techniq
ues w
hic
h h
ave their o
wn
ele
ments
are
:
�R
egre
ssio
n a
nd
Ge
ne
ral R
egre
ssio
n
�N
eu
ral N
etw
ork
s
�S
uppo
rt V
ecto
r M
ach
ine
s
�D
ecis
ion
Tre
es
�N
aïv
e B
aye
s
�C
luste
rin
g
�S
eque
nce
s
�R
ule
Se
ts
�A
sso
cia
tion
Rule
s
�T
ime
-Se
rie
s (
as o
f P
MM
L 4
.0)
�T
ext
Mo
de
ls
�S
uppo
rt f
or
Mu
ltip
le M
od
els
2
Ea
sy E
xp
res
sio
n o
f P
red
icti
ve
Mo
de
ls
Ze
me
nti
s ©
10
The P
MM
L c
ode b
elo
w im
ple
ments
score
post-
pro
cessin
g.
It u
ses t
he P
MM
L e
lem
ent T
arg
ets
for
checkin
g
boundaries (
min
and m
ax
) and
to
rescale
(re
scale
Co
nsta
nt
and r
escale
Fac
tor)
the
origin
al score
genera
ted b
y m
odel
See h
ttp:/
/ww
w.d
mg.o
rg/v
3-2
/Ta
rgets
.htm
l
The P
MM
L c
ode b
elo
w im
ple
ments
score
post-
pro
cessin
g.
It u
ses t
he P
MM
L e
lem
ent T
arg
ets
for
checkin
g
boundaries (
min
and m
ax
) and
to
rescale
(re
scale
Co
nsta
nt
and r
escale
Fac
tor)
the
origin
al score
genera
ted b
y m
odel
See h
ttp:/
/ww
w.d
mg.o
rg/v
3-2
/Ta
rgets
.htm
l
Da
ta P
os
t-P
roc
es
sin
g:
PM
ML
Exam
ple
3
Ze
me
nti
s ©
11
Ap
pli
cati
on
s
Se
rvic
e P
rovid
ers
Exte
rnal
Ven
do
rs D
ivis
ion
s
On
e S
tan
da
rd,
On
e P
roc
es
s
Ze
me
nti
s ©
12
Model Deployment
Model Building
Model Building
PM
ML
= E
as
y M
od
el
De
plo
ym
en
t
PMML
Ze
me
nti
s ©
13
PM
ML
-Z
em
en
tis
Co
ntr
ibu
tio
ns
•A
DA
PA
: A
decis
ion e
ngin
e that deplo
ys m
odels
expre
ssed in P
MM
L a
nd e
xecute
s
them
in r
eal-tim
e. N
ow
availa
ble
as a
serv
ice o
n th
e A
mazon C
loud.
•P
MM
L C
on
vert
er:
Valid
ate
s, convert
s,
and c
orr
ects
old
and n
ew
PM
ML c
ode.
Availa
ble
at th
e D
MG
website a
nd a
t h
ttp
://w
ww
.zem
entis.c
om
/pm
ml.htm
.
•C
on
trib
uti
ng
Mem
ber
of
the
DM
G:
Subm
itte
d s
evera
l pro
posals
for
PM
ML
4.0
and
already w
ork
ing w
ith o
the
r m
em
bers
on P
MM
L 4
.1.
•C
ode c
ontr
ibuto
r fo
r th
e R
PM
ML
packag
e(a
vaila
ble
on C
RA
N).
•P
MM
L A
rtic
les
: R
Journ
al and S
IGK
DD
Explo
rations N
ew
sle
tter.
Availa
ble
for
dow
nlo
adin
g a
t http:/
/ww
w.z
em
entis.c
om
/manual.htm
•P
MM
L B
log
s: S
evera
l blo
gs o
n P
MM
L topic
s (
htt
p://a
dapasuppo
rt.z
em
entis.c
om
and
http:/
/ww
w.p
redic
tive-a
naly
tics.info
).
Ze
me
nti
s ©
14
Th
an
k Y
ou
!
U.S.A Headquarters
Asia Office
E-m
ail
:in
fo@
zem
entis.c
om
19/F., Unit A
Ho Lee Commercial Building
38-44 D’AguilarStreet
Central, Hong Kong (S.A.R.)
Tel: +852 2868-0878
Fax: +852 2845-6027
6125 Cornerstone Court East
Suite 250
San Diego, CA, 92121
Tel: +1 619 330-0780
Fax: +1 858 535-0227