Data Mining, also called Knowledge Discovery in Databases, is a rather young research area, which has emerged in response to the flood of data we are faced with nowadays. It has taken up the challenge to develop techniques that can help humans to discover useful patterns in their data. One such technique---which certainly is among the most important, since it can be used for frequent data mining tasks like classifier construction and dependence analysis---are graphical models (also called inference networks) and especially learning such models from a dataset of sample cases. In this thesis I review the basic ideas of graphical modeling, with a special focus on the recently introduced possibilistic networks, for which I try to provide a clearer semantical foundation. Furthermore, I study the principles of learning graphical models from data and discuss several algorithms that have been suggested for this task. The main achievements of this thesis are enhancements of such learning algorithms: A projection method for database induced possibility distributions, a naive Bayes style possibilistic classifier, and several new evaluation measures and search methods.
Data Mining, Learning from Data, Graphical Models, Possibility Theory