<s>
Pruning	B-Algorithm
is	O
a	O
data	B-General_Concept
compression	I-General_Concept
technique	O
in	O
machine	O
learning	O
and	O
search	B-Application
algorithms	I-Application
that	O
reduces	O
the	O
size	O
of	O
decision	O
trees	O
by	O
removing	O
sections	O
of	O
the	O
tree	O
that	O
are	O
non-critical	O
and	O
redundant	O
to	O
classify	O
instances	O
.	O
</s>
<s>
Pruning	B-Algorithm
reduces	O
the	O
complexity	O
of	O
the	O
final	O
classifier	B-General_Concept
,	O
and	O
hence	O
improves	O
predictive	O
accuracy	O
by	O
the	O
reduction	O
of	O
overfitting	B-Error_Name
.	O
</s>
<s>
One	O
of	O
the	O
questions	O
that	O
arises	O
in	O
a	O
decision	B-Algorithm
tree	I-Algorithm
algorithm	O
is	O
the	O
optimal	O
size	O
of	O
the	O
final	O
tree	O
.	O
</s>
<s>
A	O
tree	O
that	O
is	O
too	O
large	O
risks	O
overfitting	B-Error_Name
the	O
training	O
data	O
and	O
poorly	O
generalizing	O
to	O
new	O
samples	O
.	O
</s>
<s>
This	O
problem	O
is	O
known	O
as	O
the	O
horizon	B-General_Concept
effect	I-General_Concept
.	O
</s>
<s>
A	O
common	O
strategy	O
is	O
to	O
grow	O
the	O
tree	O
until	O
each	O
node	O
contains	O
a	O
small	O
number	O
of	O
instances	O
then	O
use	O
pruning	B-Algorithm
to	O
remove	O
nodes	O
that	O
do	O
not	O
provide	O
additional	O
information	O
.	O
</s>
<s>
Pruning	B-Algorithm
should	O
reduce	O
the	O
size	O
of	O
a	O
learning	O
tree	O
without	O
reducing	O
predictive	O
accuracy	O
as	O
measured	O
by	O
a	O
cross-validation	B-Application
set	O
.	O
</s>
<s>
There	O
are	O
many	O
techniques	O
for	O
tree	O
pruning	B-Algorithm
that	O
differ	O
in	O
the	O
measurement	O
that	O
is	O
used	O
to	O
optimize	O
performance	O
.	O
</s>
<s>
Pruning	B-Algorithm
processes	O
can	O
be	O
divided	O
into	O
two	O
types	O
(	O
pre	O
-	O
and	O
post-pruning	O
)	O
.	O
</s>
<s>
Pre-pruning	O
procedures	O
prevent	O
a	O
complete	O
induction	O
of	O
the	O
training	O
set	O
by	O
replacing	O
a	O
stop	O
(	O
)	O
criterion	O
in	O
the	O
induction	O
algorithm	O
(	O
e.g.	O
</s>
<s>
Pre-pruning	O
methods	O
are	O
considered	O
to	O
be	O
more	O
efficient	O
because	O
they	O
do	O
not	O
induce	O
an	O
entire	O
set	O
,	O
but	O
rather	O
trees	O
remain	O
small	O
from	O
the	O
start	O
.	O
</s>
<s>
Prepruning	O
methods	O
share	O
a	O
common	O
problem	O
,	O
the	O
horizon	B-General_Concept
effect	I-General_Concept
.	O
</s>
<s>
Post-pruning	O
(	O
or	O
just	O
pruning	B-Algorithm
)	O
is	O
the	O
most	O
common	O
way	O
of	O
simplifying	O
trees	O
.	O
</s>
<s>
Pruning	B-Algorithm
can	O
not	O
only	O
significantly	O
reduce	O
the	O
size	O
but	O
also	O
improve	O
the	O
classification	O
accuracy	O
of	O
unseen	O
objects	O
.	O
</s>
<s>
These	O
methods	O
include	O
Reduced	O
Error	O
Pruning	B-Algorithm
(	O
REP	O
)	O
,	O
Minimum	O
Cost	O
Complexity	O
Pruning	B-Algorithm
(	O
MCCP	O
)	O
,	O
or	O
Minimum	O
Error	O
Pruning	B-Algorithm
(	O
MEP	O
)	O
.	O
</s>
<s>
By	O
pruning	B-Algorithm
the	O
tree	O
at	O
an	O
inner	O
node	O
,	O
it	O
can	O
happen	O
that	O
an	O
entire	O
sub-tree	O
(	O
regardless	O
of	O
its	O
relevance	O
)	O
is	O
dropped	O
.	O
</s>
<s>
One	O
of	O
these	O
representatives	O
is	O
pessimistic	O
error	O
pruning	B-Algorithm
(	O
PEP	O
)	O
,	O
which	O
brings	O
quite	O
good	O
results	O
with	O
unseen	O
items	O
.	O
</s>
<s>
One	O
of	O
the	O
simplest	O
forms	O
of	O
pruning	B-Algorithm
is	O
reduced	O
error	O
pruning	B-Algorithm
.	O
</s>
<s>
While	O
somewhat	O
naive	O
,	O
reduced	O
error	O
pruning	B-Algorithm
has	O
the	O
advantage	O
of	O
simplicity	O
and	O
speed	O
.	O
</s>
<s>
Cost	O
complexity	O
pruning	B-Algorithm
generates	O
a	O
series	O
of	O
trees	O
where	O
is	O
the	O
initial	O
tree	O
and	O
is	O
the	O
root	O
alone	O
.	O
</s>
<s>
The	O
function	O
defines	O
the	O
tree	O
obtained	O
by	O
pruning	B-Algorithm
the	O
subtrees	O
from	O
the	O
tree	O
.	O
</s>
<s>
Once	O
the	O
series	O
of	O
trees	O
has	O
been	O
created	O
,	O
the	O
best	O
tree	O
is	O
chosen	O
by	O
generalized	O
accuracy	O
as	O
measured	O
by	O
a	O
training	O
set	O
or	O
cross-validation	B-Application
.	O
</s>
