Describing the Result of a Classifier to the End-User: Geometric-based Sensitivity

摘要

This paper addresses the issue of supporting the end-user of a classifier, when it is used as a decision support system, to classify new cases. We consider several kinds of classifiers: Statistical or machine learning classifiers, which are built on data, but also direct model-based classifiers that are built to solve a particular problem (like in viability or control problems). The end-user relies mainly on global information (like error rates or global sensitivity analysis) to assess the quality of the result given by the system. Class membership probability, if available, describes certainly the local statistical viewpoint. But it doesn't take into account other contextual information: Cases with high value of class membership probability can also be close to the decision boundary. In the case of numerical state space, we propose to use the decision boundary of the classifier (which always exists, even implicitly), to describe the situation of a particular case: The distance of a case to the decision boundary measures the robustness of the decision to a change in the input data. Other geometric concepts, such as the maximal maximal ball, can present a precise picture of the situation to the end-user. We show the interest of such a geometric study on different examples.