Decision trees from uncertain learning sets

Decision trees are structures which form a partition of some given measurement space for the purpose of classifying new measurements. Typically, the construction of the decision tree is performed using a learning set (or training set) where the class of each observation is certain. An uncertain learning set is a collection of measurements and probability masses over the set of classes for each measurement. The uncertainty in the learning set affects several of the results on decision trees previously presented in the literature. This paper suggests several new splitting criteria for constructing trees using uncertain learning sets. Additional extensions are provided for developing new misclassification estimates and test sample estimates for the decision tree. Finally, an application of these techniques in sensor fusion is outlined