Une approche probabiliste pour le classement d'objets incomplètement connus dans un arbre de décision. (A Probabilistic Approach to Classify Incomplete Objects in a Decision Tree)

We describe in this thesis an approach to fill missing values in decision trees during theclassification phase. This approach is derived from the ordered attribute trees (OAT) method,proposed by Lobo and Numao in 2000, which builds a decision tree for each attribute and usesthese trees to fill the missing attribute values. It is based on the Mutual Information between theattributes and the class. Our approach extends this method by taking the dependence betweenthe attributes into account when constructing the attributes trees, and provides a probabilitydistribution as a result when classifying an incomplete object (instead of the most probableclass). We present our approach and we test it on some real databases. We also compare ourresults with those given by the C4.5 method and OAT.We also propose a k-nearest neighbours algorithm which calculates for each object from thetest data its frequency in the learning data. We compare these frequencies with the classificationresults given by our approach, C4.5 and OAT. Finally, we calculate the complexity of construct-ing the attribute trees and the complexity of classifying a new instance with missing values usingour classification algorithm, C4.5 and OAT.