Prototype Selection Using Boosted Nearest-Neighbors

Abstract

We present a new approach to Prototype Selection (PS), that is, the search for relevant subsets of instances. It is inspired by a recent classification technique known as Boosting, whose ideas were previously unused in that field. Three interesting properties emerge from our adaptation. First, the accuracy, which was the standard in PS since Hart and Gates, is no longer the reliability criterion. Second, PS interacts with a prototype weighting scheme, i. e., each prototype receives periodically a real confidence, its significance, with respect to the currently selected set. Finally, Boosting as used in PS allows to obtain an algorithm whose time complexity compares favorably with classical PS algorithms.

Three types of experiments lead to the following conclusions. First, the optimally reachable prototype subset is almost always more accurate than the whole set. Second, more practically, the output of the algorithm on two types of experiments with fourteen and twenty benchmarks compares favorably to those of five recent or state-of-the-art PS algorithms. Third, visual investigations on a particular simulated dataset gives evidence in that case of the relevance of the selected prototypes.