by

The main goal of a relatively new scientific discipline, known as Knowledge Discovery in Databases or Data Mining, is to provide methods capable of finding patterns, regularities or knowledge implicitly contained in the data so that we can gain a deeper and better understanding of the phenomenon under study. Because of the very fast growing nature of information, it is necessary to propose novel approaches in order to process this information in a quick, efficient and reliable way. In this dissertation, I use a graphical modelling data mining technique, called a Bayesian network, because of its simplicity, robustness and consistency in representing and handling relevant probabilistic interactions among variables of interest. Firstly, I present an existing algorithmic procedure, which belongs to a class of algorithms known as constraint-based algorithms, that builds Bayesian networks from data based on mutual information and conditional mutual information measures and its performance using simulated and real databases. Secondly, because of the limitations shown by this algorithm, I propose a first extension of such a procedure testing its performance using these same datasets. Thirdly, since this improved algorithm does no