In general, classification is a precise affair: presented with an object, the classifier will assign it a single class. However, in many real applications, it is more useful to trade uncertainty about the results for imprecision. Our work is based on classifiers that can yield possibilistic valuations as output, that may have been obtained from a labeled data set either directly as such, by possibilistic classifiers, or by applying transformations to the output of other valuation-based classifiers, as the probabilistic ones. In this work, we propose the use of a parameterized family of imprecise classifiers and a set of indices to help the decision maker to choose one of them: the parameter creates a series of imprecise classifiers, varying from accurate but imprecise to more precise but less accurate. In our framework, the elicitation consists in deriving the level cuts from normalized possibility distributions on the potential classification of a given element of interest. Imprecise classifiers are obtained as we vary the level cut value alpha between 0 and 1. As alpha increases, the classification tends to be more precise but less accurate. We also addressed the question of combining imprecise classifiers, in particular, the approach is extended to deal with the aggregated results.
- Acerca del IIIA
- 25 aniversario