Using data as observers: A new paradigm for prototypes selection

Abstract

The prototype selection is a bottleneck for lot of data analysis procedures. This paper proposes a new deterministic selection of prototypes based on a pairwise comparison between data. Data is ranked relative to each data.We use the paradigm of the observer situated on the data. The ranks relative to this data gives the viewpoint of the observer to the dataset. Two observers provide a link between them if they have no data between them from their respective viewpoints. The links are directed to obtain a directed graph where data is the set of vertices of the graph. The observers move using the directed graph. They reach a prototype when they arrive at a viewpoint with no outgoing connexion of the directed graph. This method proposes both the prototype selection and the structuration of the dataset through the directed graph. The paper also presents an assessment with three kinds of datasets. The method seems particularly useful when the classes are hardly distinguishable with classical clustering methods.