3-108
Chapter 3 Introduction to Predictive Modeling: Decision Trees
Tree Variations: Maximum Branches 1
r» • l-.vfi*J r.,*_. . .. • ' u • 1 O'ffin >•rtrriQn r,, i... L**il
—
,
U » «n >» n t h
"1
j | J M Input 0*.r» :
M a j v n u m D«pD>
! Ljlf S C * -UunMU'ol f?u*< l
S * W f
l :.••> - HM i -
IhB
1
Maximum branches in split
* ;
Exhaustive search size limit
BUM
- M*mn J i_. ,.- • • Ml* u**'. ' A s » » r n # r t Fractal - N I-E - I
h
rvrittr
En*
Oil
141 T w o fields in the Properties panel affect the number of splits in a tree. T h e M a x i m u m Branch property sets an upper limit on the number of branches emanating from a node. W h e n this number is greater than the default of 2. the number of possible splits rapidly increases. T o save computation time, a limit is set in the Exhaustive property as to h o w m a n y possible splits are explicitly examined. W h e n this n u m b e r is exceeded, a heuristic algorithm is used in place of the exhaustive search described above. T h e heuristic algorithm alternately merges branches and reassigns consolidated groups of observations to different branches. T h e process stops w h e n a binary split is reached. A m o n g all candidate splits considered, the one with the best worth is chosen. T h e heuristic algorithm initially assigns each consolidated group of observations to a different branch, even if the number of such branches is more than the limit allowed in the final split. At each merge step, the two branches that degrade the worth of the partition the least are merged. After the two branches are merged, the algorithm considers reassigning consolidated groups of observations to different branches. Each consolidated group is considered in turn, and the process stops w h e n no group is reassigned.