Data Mining Techniques for Segmentation - Data Mining Techniques in CRM: Inside Customer Segmentation - page 92

called input neurons or units. The output layer is a two-dimensional grid map, con-
sisting of the output neurons which will form the derived clusters. Each input neu-
ron is connected to each of the output neurons with ''strengths'' or weights. These
weights (which are analogous to the cluster centers referred to in the K-means
procedure) are initially set at random and are refined as the model is trained.
Input records are presented to the output layer and the output neurons
''compete'' to ''win'' them. Each record ''gravitates'' toward the output neuron with
the most similar pattern characteristics and is assigned to it. Record assignment is
based on the Euclidean distance: each record's input values are compared to the
centers of the output neurons and the ''closest'' output neuron ''wins'' the record.
This assignment also results in the adaptation or ''learning'' of the output
neuron's characteristics and the adjustment of its corresponding weights (cluster
center) so that they better match the pattern of the assigned record. Thus, the
output neuron's center is updated and moved closer to the characteristics of
the assigned record, in a way similar to that of K-means. Therefore, if another
record with similar characteristics is presented to the output neuron, it will have
a greater chance of winning it. The outline of a self-organizing map is graphically
represented in Figure 3.9.
Data are passed through the network a specified number of times. The
training process includes two phases, also referred to as cycles: a first cycle of rough
estimations and large-scale changes in weights and a second refinement cycle with
smaller changes.
After ''winning'' a record, the cluster center of the ''winning'' neuron is
adapted by taking into account a portion of the difference between the current
Figure 3.9 Self-organizing map graphical representation.