Mining Incomplete Data with Many Lost and Attribute-Concept Values

摘要

This paper presents experimental results on twelve data sets with many missing attribute values, interpreted as lost values and attribute-concept values. Data mining was accomplished using three kinds of probabilistic approximations: singleton, subset and concept. We compared the best results, using all three kinds of probabilistic approximations, for six data sets with lost values and six data sets with attribute-concept values, where missing attribute values were located in the same places. For five pairs of data sets the error rate, evaluated by ten-fold cross validation, was significantly smaller for lost values than for attribute-concept values (5% significance level). For the remaining pair of data sets both interpretations of missing attribute values do not differ significantly.