<s>
In	O
machine	O
learning	O
,	O
Platt	B-General_Concept
scaling	I-General_Concept
or	O
Platt	B-General_Concept
calibration	I-General_Concept
is	O
a	O
way	O
of	O
transforming	O
the	O
outputs	O
of	O
a	O
classification	B-General_Concept
model	I-General_Concept
into	O
a	O
probability	B-General_Concept
distribution	I-General_Concept
over	I-General_Concept
classes	I-General_Concept
.	O
</s>
<s>
The	O
method	O
was	O
invented	O
by	O
John	O
Platt	O
in	O
the	O
context	O
of	O
support	B-Algorithm
vector	I-Algorithm
machines	I-Algorithm
,	O
</s>
<s>
Platt	B-General_Concept
scaling	I-General_Concept
works	O
by	O
fitting	O
a	O
logistic	O
regression	O
model	O
to	O
a	O
classifier	B-General_Concept
's	O
scores	O
.	O
</s>
<s>
Consider	O
the	O
problem	O
of	O
binary	B-General_Concept
classification	I-General_Concept
:	O
for	O
inputs	O
,	O
we	O
want	O
to	O
determine	O
whether	O
they	O
belong	O
to	O
one	O
of	O
two	O
classes	O
,	O
arbitrarily	O
labeled	O
and	O
.	O
</s>
<s>
Platt	B-General_Concept
scaling	I-General_Concept
is	O
an	O
algorithm	O
to	O
solve	O
the	O
aforementioned	O
problem	O
.	O
</s>
<s>
i.e.	O
,	O
a	O
logistic	O
transformation	O
of	O
the	O
classifier	B-General_Concept
scores	O
,	O
where	O
and	O
are	O
two	O
scalar	O
parameters	O
that	O
are	O
learned	O
by	O
the	O
algorithm	O
.	O
</s>
<s>
The	O
parameters	O
and	O
are	O
estimated	O
using	O
a	O
maximum	O
likelihood	O
method	O
that	O
optimizes	O
on	O
the	O
same	O
training	O
set	O
as	O
that	O
for	O
the	O
original	O
classifier	B-General_Concept
.	O
</s>
<s>
Platt	O
himself	O
suggested	O
using	O
the	O
Levenberg	B-Algorithm
–	I-Algorithm
Marquardt	I-Algorithm
algorithm	I-Algorithm
to	O
optimize	O
the	O
parameters	O
,	O
but	O
a	O
Newton	B-Algorithm
algorithm	I-Algorithm
was	O
later	O
proposed	O
that	O
should	O
be	O
more	O
numerically	B-Algorithm
stable	I-Algorithm
.	O
</s>
<s>
Platt	B-General_Concept
scaling	I-General_Concept
has	O
been	O
shown	O
to	O
be	O
effective	O
for	O
SVMs	B-Algorithm
as	O
well	O
as	O
other	O
types	O
of	O
classification	O
models	O
,	O
including	O
boosted	B-Algorithm
models	O
and	O
even	O
naive	B-General_Concept
Bayes	I-General_Concept
classifiers	I-General_Concept
,	O
which	O
produce	O
distorted	O
probability	O
distributions	O
.	O
</s>
<s>
It	O
is	O
particularly	O
effective	O
for	O
max-margin	O
methods	O
such	O
as	O
SVMs	B-Algorithm
and	O
boosted	B-Algorithm
trees	O
,	O
which	O
show	O
sigmoidal	O
distortions	O
in	O
their	O
predicted	O
probabilities	O
,	O
but	O
has	O
less	O
of	O
an	O
effect	O
with	O
well-calibrated	O
models	O
such	O
as	O
logistic	O
regression	O
,	O
multilayer	B-Algorithm
perceptrons	I-Algorithm
,	O
and	O
random	B-Algorithm
forests	I-Algorithm
.	O
</s>
<s>
An	O
alternative	O
approach	O
to	O
probability	O
calibration	B-General_Concept
is	O
to	O
fit	O
an	O
isotonic	B-General_Concept
regression	I-General_Concept
model	O
to	O
an	O
ill-calibrated	O
probability	O
model	O
.	O
</s>
<s>
This	O
has	O
been	O
shown	O
to	O
work	O
better	O
than	O
Platt	B-General_Concept
scaling	I-General_Concept
,	O
in	O
particular	O
when	O
enough	O
training	O
data	O
is	O
available	O
.	O
</s>
