Learning Bayesian network parameters from small data sets Application of noisy-or gates.pdfVIP

In Working Notes of the Workshop on ‘Bayesian and Causal Networks: From Inference to Data Mining,’ 12th European Conference on Artificial Intelligence (ECAI-2000), Berlin, Germany, 22 August 2000 Learning Bayesian Network Parameters from Small Data Sets: Application of Noisy-OR Gates Agnieszka Onis?ko 1 and Marek J. Druzdzel 2 and Hanna Wasyluk 3 Abstract. Existing data sets of cases can significantly reduce the knowledge engineering effort required to parameterize Bayesian net- works. Unfortunately, when a data set is small, many conditioning cases are represented by too few or no data records and they do not offer sufficient basis for learning conditional probability distribu- tions. We propose a method that uses Noisy-OR gates to reduce the data requirements in learning conditional probabilities. We test our method on HEPAR II, a model for diagnosis of liver disorders, whose parameters are extracted from a real, small set of patient records. Di- agnostic accuracy of the multiple-disorder model enhanced with the Noisy-OR parameters was around 6% better than the accuracy of the plain multiple-disorder model and 10% better than the single- disorder diagnosis model. 1 INTRODUCTION Bayesian networks [13] (also called belief networks) are acyclic di- rected graphs modeling probabilistic dependencies among variables. The graphical part of a Bayesian network reflects the structure of a problem, while local interactions among neighboring variables are quantified by conditional probability distributions. Bayesian net- works have earned the reputation of being powerful tools for mod- eling complex problems involving uncertain knowledge. They have been employed in practice in a variety of fields, including engineer- ing, science, and medicine with some models reaching the size of hundreds of variables. A major difficulty in applying Bayesian network models to practi- cal problems is the effort that goes in model building, i.e., obtaining the model structure and the numerical