Experiments on rule induction from incomplete data using three probabilistic approximations

We present results of experiments on rule induction using three probabilistic approximations: lower, middle, and upper. Our results were conducted on four typical series of incomplete data sets with 5% increments of missing attribute values. Two interpretations of missing attribute values were used: lost and “do not care” conditions. We conclude that the best approach (choice of the interpretation of missing attribute values and selection of the best type of approximation) depends on a data set. Probabilistic approximations are constructed from characteristic sets. The number of distinct probabilities associated with characteristic sets is much larger for data sets with “do not care” conditions than with data sets with lost values. Therefore, for data sets with “do not care” conditions the number of probabilistic approximations is also larger.