Estimators based on AUC for classification and optimal subsampling for them
Maximum likelihood is based on optimizing estimators for the parameters of the model or computing good predictions of the responses. For classification, Generalized Linear Models, such as logistic regression, are frequently considered. Maximum likelihood is then appropriate for fitting the model, but not necessarily for the aim of classification. In this work we offer a criteria based in the Area Under the Curve (AUC), first for estimating the parameters, then for computing optimal designs for training the model. Some illustrative examples help to understand the main ideas.
Palabras clave: Classification Area under the curve Maximum likelihood Optimal design of experiments