Characteristic Relations for Incomplete Data: A Generalization of the Indiscernibility Relation

This paper shows that attribute-value pair blocks, used for many years in rule induction, may be used as well for computing indiscernibility relations for completely specified decision tables. Much more importantly, for incompletely specified decision tables, i.e., for data with missing attribute values, the same idea of attribute-value pair blocks is a convenient tool to compute characteristic sets, a generalization of equivalence classes of the indiscernibility relation, and also characteristic relations, a generalization of the indiscernibility relation. For incompletely specified decision tables there are three different ways lower and upper approximations may be defined: singleton, subset and concept. Finally, it is shown that, for a given incomplete data set, the set of all characteristic relations for the set of all congruent decision tables is a lattice.