Analogical proportions are statements of the form "a is to b as c is to d", where a, b, c, d are tuples of attribute values describing items. The mechanism of analogical inference, empirically proved to be efficient in classification and reasoning tasks, started to be better understood when the characterization of the class of classification functions with which the analogical inference always agrees was established for Boolean attributes. The purpose of this paper is to study the case of finite attribute domains that are not necessarily two-valued, i.e., when attributes are nominal. In particular, we describe the more stringent class of "hard" analogy preserving (HAP) functions f : X1×· · ·×Xm → X over finite domains X1,. .. , Xm, X for binary classification purposes. This description is obtained in two steps. First we observe that such AP functions are almost affine, that is, their restriction to any S1×· · ·×Sm, where Si ⊆ Xi and |Si| ≤ 2 (1 ≤ i ≤ m), can be turned into an affine function by renaming variable and function values. We then use this result together with some universal algebraic tools to show that they are essentially unary or quasi-linear, which provides a general representation of HAP functions. As a by-product, in the case when X1 = · · · = Xm = X, it follows that this class of HAP functions constitutes a clone on X, thus generalizing several results by some of the authors in the Boolean case.