Mining Incomplete Data Using Global and Saturated Probabilistic Approximations Based on Characteristic Sets and Maximal Consistent Blocks

Abstract

In this paper we discuss incomplete data sets with missing attribute values interpreted as “do not care” conditions. For data mining, we use two types of probabilistic approximations, global and saturated. Such approximations are constructed from two types of granules, characteristic sets and maximal consistent blocks. We present results of experiments on mining incomplete data sets using four approaches, combining two types of probabilistic approximations, global and saturated, with two types of granules, characteristic sets and maximal consistent blocks. We compare these four approaches, using an error rate computed as the result of ten-fold cross validation. We show that there are significant differences (5% level of significance) between these four approaches to data mining. However, there is no universally best approach. Hence, for an incomplete data set, the best approach to data mining should be chosen by trying all four approaches.