In Table 6, we present a brief summary of the characteristics of these datasets: the first column shows the code of
the dataset, the second shows the number of training cases, the third includes the number of missing values, the fourth
shows the percentage of occurrence of the most frequent value of the class, the fifth and sixth show the number of
continuous and discrete attributes, respectively, and the last column shows the number of different values for the class.
For the experiment, two algorithms were implemented to build decision trees. In the first algorithm, the training
cases with missing values are distributed following the approach of C4.5; that is, the probability distribution function
of the attribute under consideration is used to distribute the cases with missing values in this attribute. In the second
algorithm (our method), the cases with missing values are distributed taking into account the probability distribution
obtained using the attribute and the class.
Our method is associated with the total percentage of missing attribute values in the training cases, and these
missing values are distributed in a random way among the attributes. The method works independently of the number
of attributes with missing values, that is, missing values can appear in any attribute. We have checked that the proposed
model does not fail when we increase the number of attributes with missing values and the total percentage of missing
values holds.