A Model-Free Subject Selection Method for Active Learning Classification Procedures

To construct a classification rule via an active learning method, during the learning process, users select training subjects sequentially, without knowing their labels, based on the model learned at the current stage. For a parametric-model-based classification rule, methods of statistical experimental design are popular guidelines for selecting new learning subjects. However, there is a lack of a counterpart for non-parametric-model-based classifiers, such as support vector machines. Thus, we propose a subject selection scheme via an extended influential index for the area under a receiver operating characteristic curve, which is applicable to general classifiers with continuous scores.