Feature Discovery in Non-metric Pairwise Data

by Julian Laub Klaus-Robert Muller

This is a paper about how to analysis pairwise "distance" or similarity matrices. Since no all similarity matrices can be transformed into a Gram matrix (as we do in MDS), it is interesting to take a deeper insight into the details.

Basically, we may imagine there are two metrics, one for similarity and another for dissimilarity (penalizing the similarity in human perception). By applying a spetral transformation, we may use metric methods if the spectra can be fixed (no negative).

The problem is how we may utilize the negative part of the spectra.