Dissimilarity coefficients which are independent of a special set of data (Q1813339)

The author studies formally the question of when functions measuring similarity or dissimilarity of elements defined on any of a class \(\{S_ i:\;i\in I\}\) of finite sets \(S_ i\) with some additional structure (such as a pre-order or affine structure) may be said to yield values which are ``independent'' of the particular sets \(S_ i\) to which they are applied. The point of view is that each set \(S_ i\) is a set of data whose elements are to be grouped or ``clustered'' by a formal algorithm, and the distinct sets \(S_ i\) and \(S_ j\) may well contain common elements. The functions \(f\) to be used for clustering are defined for example on \(S_ i\times S_ i\) (measuring distance or ``dissimilarity'' between pairs of elements) or on the power-set of \(S_ i\) (`` clustering functions'' whose values provide some index of similarity of the elements of each subset of \(S_ i\)) and are assumed to leave invariant the ordering or linear structure, if any, of the dataset. The author argues that `` category theory'' is an appropriate formal structure within which ``independence'' of function-values from particular datasets can be explored, and he provides a definition. The (unsurprising) results of the paper are essentially negative, saying for example that only trivial dissimilarity-functions satisfy his definition of independence.