Complexity of Rule Sets Induced from Incomplete Data with Attribute-concept Values and 'Do Not Care' Conditions

摘要

In this paper we study the complexity of rule sets induced from incomplete data sets with two interpretations of missing attribute values: attribute-concept values and "do not care" conditions. Experiments are conducted on 176 data sets, using three kinds of probabilistic approximations (lower, middle and upper) and the MLEM2 rule induction system. The goal of our research is to determine the interpretation and approximation that produces the least complex rule sets. In our experiment results, the size of the rule set is smaller for attribute-concept values for 12 combinations of the type of data set and approximation, for one combination the size of the rule sets is smaller for "do not care" conditions and for the remaining 11 combinations the difference in performance is statistically insignificant (5% significance level). The total number of conditions is smaller for attribute-concept values for ten combinations, for two combinations the total number of conditions is smaller for "do not care" conditions, while for the remaining 12 combinations the difference in performance is statistically insignificant. Thus, we may claim that attribute-concept values are better than "do not care" conditions in terms of rule complexity.