Summary
International Symposium on Nonlinear Theory and its Applications
2008
Session Number:B4L-B
Session:
Number:B4L-B2
Training and Scoring of Probabilistic Classifiers
Jochen Brocker,
pp.-
Publication Date:2008/9/7
Online ISSN:2188-5079
Summary:
This contribution discusses aspects of a learning theory for probabilistic classifiers. Classical statistical learning theory focusses mainly on classifiers which give unequivocal response to an input. This is, the output of the classifier is always one of the class labels (after appropriate conversion). Such classifiers are often inadequate though, for example if the classifier is to be used by a community of users with heterogeneous cost?loss profiles. Recently there has been increasing interest in classifiers which provide a probabilistic rather than a deterministic answer, since probability assignments allow for more informed decision making in the face of uncertain risks. The present contribution discusses how to evaluate or “score” probability assignments, leading to the concept of scoring rules. As will be demonstrated, scoring rules need to have certain properties in order to guarantee this evaluation to be logically consistent. Furthermore, scoring rules allow to formulate the training of a probabilistic classifier as empirical risk minimisation, rendering large parts of the theory of statistical learning applicable to the present problem.