| Many concrete problems involving uncertainty can be viewed as classification problems in which an instance described by a number of features has to be classified in one of several distinct classes. The family of Bayesian network classifiers has gained considerable popularity for solving such problems. These models include a single class variable with arcs pointing directly to multiple feature variables. While the family of Bayesian network classifiers provides for univariate classification only, not all classification problems are one-dimensional: many real problems require that an instance be assigned to a most likely combination of classes instead of to a single class. In two recent papers, we introduced the concept of multi-dimensionality in Bayesian network classifiers by allowing multiple dependent class variables. The goal of the proposed research is to further develop the framework of multi-dimensional Bayesian network classifiers. More specifically, the following research results are aimed at: - insight in properties of expressiveness and practicability of various members of the family of multi-dimensional Bayesian network classifiers, compared to other types of statistical classification model; - learning algorithms for automated construction from data for a variety of members of the family of multi-dimensional Bayesian network classifiers. The proposed research is in line with, yet sufficiently complementary to a recently funded affiliated project in Spain. |
