<s>
In	O
computer	B-General_Concept
science	I-General_Concept
,	O
a	O
logistic	B-Algorithm
model	I-Algorithm
tree	I-Algorithm
(	O
LMT	O
)	O
is	O
a	O
classification	B-General_Concept
model	O
with	O
an	O
associated	O
supervised	B-General_Concept
training	I-General_Concept
algorithm	O
that	O
combines	O
logistic	O
regression	O
(	O
LR	O
)	O
and	O
decision	B-Algorithm
tree	I-Algorithm
learning	I-Algorithm
.	O
</s>
<s>
Logistic	B-Algorithm
model	I-Algorithm
trees	I-Algorithm
are	O
based	O
on	O
the	O
earlier	O
idea	O
of	O
a	O
model	O
tree	O
:	O
a	O
decision	O
tree	O
that	O
has	O
linear	B-General_Concept
regression	I-General_Concept
models	I-General_Concept
at	O
its	O
leaves	O
to	O
provide	O
a	O
piecewise	O
linear	B-General_Concept
regression	I-General_Concept
model	I-General_Concept
(	O
where	O
ordinary	O
decision	O
trees	O
with	O
constants	O
at	O
their	O
leaves	O
would	O
produce	O
a	O
piecewise	O
constant	O
model	O
)	O
.	O
</s>
<s>
In	O
the	O
logistic	O
variant	O
,	O
the	O
LogitBoost	B-Algorithm
algorithm	O
is	O
used	O
to	O
produce	O
an	O
LR	O
model	O
at	O
every	O
node	O
in	O
the	O
tree	O
;	O
the	O
node	O
is	O
then	O
split	O
using	O
the	O
C4.5	B-Algorithm
criterion	O
.	O
</s>
<s>
Each	O
LogitBoost	B-Algorithm
invocation	O
is	O
warm-started	O
from	O
its	O
results	O
in	O
the	O
parent	O
node	O
.	O
</s>
<s>
The	O
basic	O
LMT	O
induction	O
algorithm	O
uses	O
cross-validation	B-Application
to	O
find	O
a	O
number	O
of	O
LogitBoost	B-Algorithm
iterations	O
that	O
does	O
not	O
overfit	B-Error_Name
the	O
training	O
data	O
.	O
</s>
<s>
A	O
faster	O
version	O
has	O
been	O
proposed	O
that	O
uses	O
the	O
Akaike	O
information	O
criterion	O
to	O
control	O
LogitBoost	B-Algorithm
stopping	O
.	O
</s>
