K-Nearest Neighbour Classification for Interval-Valued Data
Résumé
This paper studies the problem of providing predictions with a K-nn approach when data have partial features given in the form of intervals. To do so, we adopt an optimistic approach to replace the ill-known values, that requires to compute sets of possible and necessary neighbours of an instance. We provide an easy way to compute such sets, as well as the decision rule that follows from them. Our approach is then compared to a simple imputation method in different scenarios, in order to identify those ones where it is advantageous.
Origine : Fichiers produits par l'(les) auteur(s)