Data Mining Techniques for Segmentation - Data Mining Techniques in CRM: Inside Customer Segmentation - page 87

would be to use another clustering model, save the resulting cluster centers, and
load them as initial seeds for the K-means training.
On the other hand, K-means advantages include its speed and scalability: it
is one of the fastest clustering models and it can efficiently handle long and wide
datasets with many records and many input clustering fields.

Recommended K-means Options

Figures 3.6 and 3.7 and Table 3.9 present the recommended options for the IBM
SPSS Modeler K-means modeling node. As outlined earlier, a key characteristic of

Table 3.9 IBM SPSS Modeler recommended K-means options.

Option Setting Functionality/reasoning for selection

Number of
clusters

Requires trial
and error.
Analysts
should evalu-
ate different
clustering
solutions

Analysts should experiment with different num-
bers of clusters to find the solution that best fits
their specific business goals
Worth trying alternatives: users can get an
indication of the underlying number of clusters
by trying other clustering techniques which
incorporate specific criteria and automatically
detect the number of clusters

Generate dis-
tance field

Selected It generates an additional field which denotes

the distance of each record from the center of
the assigned cluster. It can be used to assess
whether a case is a typical representative of
its cluster or lies far apart from the rest of the
other members

Show cluster
proximity

Selected It creates a proximity matrix of the distances

between the cluster centers. It can be used to
assess the separation of the revealed clusters

Maximum
iterations
and change
tolerance

Keep the
defaults,
unless

The algorithm stops if, after a data pass (iteration),
the cluster centers do not change (0 change
tolerance) or after the specified number of
iterations has been completed, regardless of
the change of the cluster centers
Users should browse the results and examine the
solution steps and the change of the cluster
centers at each iteration. In the case of non-
convergence after the specified number of
iterations, they should increase the specified
number of iterations and rerun the algorithm

...