A Comparison of Classification Systems for Rule Sets Induced from Incomplete Data by Probabilistic Approximations

摘要

In this paper, we compare four strategies used in classification systems. A classification system applies a rule set, induced from the training data set in order to classify each testing case as a member of one of the concepts. We assume that both training and testing data sets are incomplete, i.e., some attribute values are missing. In this paper, we discuss two interpretations of missing attribute values: lost values and "do not care" conditions. In our experiments rule sets were induced using probabilistic approximations. Our main results are that for lost value data sets the strength only strategy is better than conditional probability without support and that for "do not care" data sets the conditional probability with support strategy is better than strength only.