With this observation, the method described in [18] may be understood as a method
that seeks the binary vector y that maximises the quadratic form by choosing y so
that it has small correlation with those 3−5 eigenvectors corresponding to the smallest eigenvalues and hence correlates strongly with the eigenvectors corresponding
to large eigenvalues.
Precision and recall results for the PLSA and PCA clustering strategies are reproduced in Figure 25.10 for an average attack of size 10%. Similar results have
been obtained for random and bandwagon attacks. The PLSA and PCA clustering