Except that eigenvectors play different role in facial recognition as from pagerank.

In the former, eigenvectors are an orthogonal basis for representing the set of human faces: first, the vary in this respect, then next most significantly in this respect, and so on.

In the latter, the primary eigenvector tells you which nodes are the most significant. The eigenvector not looking like a good site (or typical face) has little to do with whether it's informative about the goodness of a site.

Also, the point of using morality eigenvectors is to quantify relative morality of positions, not to find a compromise between currently-popular positions, so the bimodality issue is not a problem in this respect.