Developing uncertainty measures for classification using information theoretic techniques in induction and validation

Learning or classifying under uncertainty and using the results of learning in subsequent deductive inference are discussed. The relationship between information theoretic techniques and validation techniques is an important component of this investigation. The specific focus is on developing the framework for a general learning paradigm and presenting techniques for uncertainty representations in clustering within this framework. A probabilistic inference network is used to reason with the results of the clustering and classification on a given data set and possibly some a priori information. The properties of appropriate uncertainty representations are developed to define specific requirements for the uncertainty measures in cluster analysis. This approach selects the distribution which uses the information provided by the cluster analysis without imposing any further bias on class assignment