Page 85 -

72 3 Data Clustering

squared deviations. per variable. Notice that ~f)is just the variance of feature j
multiplied by (n-1).

P A0 A
AAPN

RDES
- T102 NAZO 0 A~203
@4
z 0.2 .
4 0 FEZ03
u MGO RMCS MNU
*-
-0.2 - c;O RCSG RCHO
RMFX

Factor 1
Figure 3.18. Factor loadings graph for the Rocks dataset. Factor 1 is highly
correlated with chemical features and Factor 2 with physical features.

Figure 3.19. Scatter plot of the Rocks data in the factor space with three identified
clusters.

Figure 3.20 shows the cluster merit indexes for the several values of c,
computed with formula (3-7a). Factor 1 has the most important contribution in
terms of the decrease of the within-cluster error. Inspection of Figure 3.20 suggests