Sampling Algorithms for Unsupervised Prototype Selection

The
$K$
Nearest Neighbor classifier (
$K$
-NN) is one of the most widely used classification algorithms. It is simple, non-parametric, and does not have a training phase, but it requires storing all observations (training set). The class for an unknown sample
$x_{q}$
is labeled using a similarity/distance function
$d$
to compute the
$K$
most similar elements (neighbors). The latter makes that
$K$
-NN prediction process unfeasible for large datasets. Two main approaches are studied to optimize the prediction process: a pre-computed index to speed up the neighbors' computation and by generating a subset of representative elements called prototypes. This work presents an analysis of sampling algorithms to perform unsupervised prototype selection. Each proposed algorithm's performance is measured by the reproduction capacity of the original data set class distribution and the performance in several classification tasks.