The large number of variables and the relatively small amount of statistical data propose the use of multivariable Bayesian approaches to the feature selection problem in many domains. In the Bayesian analysis of feature relevance, an a posteriori probability distribution is estimated over the subsets of predictor variables, instead of the selection of a single set of predictors based on some criteria. However the number of possible subsets with significant probabilities can be numerous and the estimation of the a posteriori probabilities also brings some uncertainty as it is done by Monte Carlo methods. The interpretation of such results is a very challenging, quite complex activity. To support this interpretationthe Bioinformatics workgroup at the BME MIT faculty developed many concepts and methods, for example the concept of sub- and super k-Markov Blanket Set, which allows a scalable analysis of multivariate relevance using multivariable projections of the results.
The goal of this diploma work is to support the user in computing probabilities over the space of k-Markov Blanket Sets using already developed approximation methods. Furthermore to support the visualization of probabilities over the space of k-Markov Blanket Sets, including the visualization of the probabilities of the Markov Blanket Memberships. In addition, the software tool developed in the diploma work ensures the visualization of Markov Blanket Graphs that spans the space of Markov Blanket Sets and ensures the visualization and design of standard Bayesian networks as well.