Instance-Based Prediction in the Framework of Possibility Theory

A possibilistic framework for instance-based prediction is presented which formalizes the generalization beyond experience by means of fuzzy rules. In comparison with related instance-based approaches such as the well-known Nearest Neighbor classifier, this method distinguishes itself by the following: First, by suggesting (guaranteed) degrees of possibility for competing outcomes rather than making precise predictions, it takes the uncertain character of similarity-based inference into account. Second, the possibilistic framework can easily be extended so as to cope with incompletely specified cases. Thirdly, the close connection between possibility theory and fuzzy sets suggests the extension of the basic model by means of fuzzy set-based (linguistic) modeling techniques. This paper especially highlights two of these aspects, namely the modeling of uncertainty and the handling of incomplete infor-