Information Technology Reference
In-Depth Information
U-matrix and SOMs of three-dimensional data set
(reprinted and adapted from Rantanen et al., 2001;
with permission from Elsevier)
Figure 5.15
In this way, the SOM is partially organized from the beginning, and
narrower neighborhood functions and smaller learning-rate factors can be
selected (Kohonen, 1998). The SOM algorithm is therefore sometimes
referred to as a nonlinear adaptation of PCA (Ritter, 1995). The time it
takes to build the SOM, that is, the time when convergence occurs,
depends on the data being analyzed and parameters selected for SOM
construction. Most software programs that are available for application
of the SOMs give the possibility of automatic selection of training
parameters.
SOMs are less susceptible to noise in the data when compared to
other clustering techniques, such as hierarchical clustering (Mangiameli
et al., 1996) or k-means clustering (Chen et al., 2002). Some authors
emphasize the advantage of SOMs (as an unsupervised learning technique)
Search WWH ::
Custom Search