Database Reference
In-Depth Information
Table 3.11
(
continued
)
Option
Setting
Functionality/reasoning for selection
The model training can also stop after a specified
time has elapsed (''Stop on time (mins)'' option).
This option should only be selected if there is a
specific time constraint
Width/length
(of the out-
put grid
map)
Requires
trial and
error
Users can specify the size of the output grid map.
This size determines the maximum number of
probable clusters that will be formed by the
model. Analysts can use prior knowledge from
other clustering models to initially set the grid size
In general it is recommended to start by asking for
a quite rich solution, assuming that there are a
significant number of member records in each
cluster. The requested geometry of the solution,
for instance 3
3,haslesseffectonthe
results than the requested grid size itself
×
4or4
×
Neighborhood/
initial eta/
cycles
IBM SPSS
Modeler
default
values
Users can also intervene in the model training
process by altering the following technical settings:
neighborhood radius, initial eta value, and number
of cycles. In general IBM SPSS Modeler default
values are adequate for most situations and should
not be modified. However, users could also try the
following:
• Neighborhood: defines the number of units sym-
metrically around a ''winning'' output neuron that
will be adapted upon record assignment. Gen-
erally it should be an integer greater then 1. The
neighborhood radius should be proportional to the
requested grid size. So if we double the requested
grid size we should also double the neighborhood.
In small requested grids, the phase 1 neighbor-
hood value could be set to 1 so that the entire grid
is not affected
• Cycles: determines the number of iterations or
data passes through the network. Typically, both
phase 1 and phase 2 default values (20 and 150
respectively) should not be modified since they are
quite large and appropriate for most tasks
(
continued overleaf
)
Search WWH ::
Custom Search