Methods, computing systems and computer program products implement embodiments of the present invention that include selecting a dataset that includes instances, each of the instances having respective features, and determining an expected distribution of the instances among multiple categories. A first classification is generated that includes, for each of the instances based on their respective features, one or more first categories and a corresponding confidence score for each of the one or more categories. One or more of the instances classified into the given category are allocated to each given category, based on their respective confidence scores, using the allocated one or more instances, a second classification is generated.