Page 54 -

40 2 Pattern Discrimination

variance, certainly a negligible fraction. The first 3 eigenvalues, however, are
responsible for more than 95% of the total variance, which suggests that it would
probably be adequate to use the corresponding first 3 eigenvectors (computed as
linear transformations of the original features) instead of the 10 original features.

Figure 2.16. Sorted list of the eigenvalues for the cork stoppers data (two classes).

Number of Eigefmlues
Figure 2.17. Plot of the eigenvalues for the cork stoppers data (two classes).

When using principal component analysis for dimensionality reduction, the
decision one must make is how many eigenvectors (and corresponding
eigenvalues) to retain. The Kaiser criterion discards eigenvalues below 1, which
nearly corresponds to retaining the eigenvalues responsible for an amount of