Databases Reference
In-Depth Information
Figure 7.5
Matrices used to compute clustering correlation coefficient
that the SOM algorithm utilizes a topologically fixed centroid structure and,
based on that topology, when one centroid is updated, neighboring centroids are
also updated. Suggested topologies for constructing a SOM are a single-
dimension linear ordering of centroids (Figure 7.6a), a two-dimensional grid
(Figure 7.6b), and a two-dimensional honeycomb pattern (Figure 7.6c).
The steps of the algorithm are:
1. Select topology including number of centroids (cells).
2. Randomly initialize all centroids
3. Repeat:
a. select next observation
b. locate centroid closest to the observation (winning centroid)
c. update winning centroid and other centroids in winner's neighborhood,
nudging each centroid closer to the selected observation until a
(a) linear
(b) grid
(c) honeycomb
Figure 7.6
SOM Topologies
