NEGATIVE INFORMATION INTEGRATION IN PROBABILISTIC CLASSIFIERS

Abstract

This study presents a procedure for integrating negative information into robust Bayesian inference. In the proposed procedure, negative information is codified as linear restrictions of the conditional probability intervals that quantify the uncertainty in the relationship between the classifying variables. During the inference, the robust Bayesian classifier is converted into a credal classifier. The classifier topology is the same as the Naive Bayesian classifier, and the optimization problems related to the inferences are solved by multilinear optimization. Since the objective of an inference is to compute the posterior probability interval of each class, by integrating the negative information, the inference procedure might obtain more precise intervals than those obtained by a robust Bayesian classifier. This might favor the use of the decision criterion called interval dominance when selecting plausible labels and defining a course of action for a given object of interest. The effectiveness of the procedure is illustrated with an example.