Analysis of Distance Matrices for Studying Data Structures and Separating, Classes

This paper demonstrates that the computation of Euclidian distances between objects generates an intrinsic geometry, which, besides containing all the information explicitly present in the original space, also provides new information about the object interrelationships. In other words, by computing distances the data are mapped onto a new, and more flexible, reference frame. Such properties of the distance function also permit the derivation of non linear descriptors of the data space, and can be usefully exploited for pattern recognition purposes. In particular, it is demonstrated that this approach is able to separate embedded classes, which have been repeatedly reported in QSAR research.