Machine Learning of Credible Classifications

We present an approach to concept discovery in machine learning based on searching for maximally general credible classifications. To be credible, a classification must provide decisions for all or nearly all possible values of the condition attributes, and these decisions must be adequately supported by evidence. Our objective is to find a classification for a domain that meets predefined quality criteria. For example, a classification can be sought whose coverage of the domain exceeds a user-defined threshold and whose decisions are supported by sufficient input instances.