PREDICTING SURVIVAL IN MALIGNANT SKILL MELANOMA USING BAYESIAN NETWORKS AUTOMATICALLY INDUCED BY GENETIC ALGORITHMS - AN EMPIRICAL-COMPARISON BETWEEN DIFFERENT APPROACHES

In this work we introduce a methodology based on genetic algorithms fo r the automatic induction of Bayesian networks from a file containing cases and variables related to the problem. The structure is learned b y applying three different methods: The Cooper and Herskovits metric f or a general Bayesian network, the Markov blanket approach and the rel axed Markov blanket method. The methodologies are applied to the probl em of predicting survival of people after 1, 3 and 5 years of being di agnosed as having malignant skin melanoma. The accuracy of the obtaine d models, measured in terms of the percentage of well-classified subje cts, is compared to that obtained by the so-called Naive-Bayes. In the four approaches, the estimation of the model accuracy is obtained fro m the 10-fold cross-validation method. (C) 1998 Elsevier Science B.V. All rights reserved.