Weighted nearest centroid neighbourhood

Abstract

A novel binary classifier based on nearest centroid neighbours is presented. The proposed method uses the well known idea behind the classic k-Nearest Neighbours (k-NN) algorithm: one point is similar to others that are close to it. The new proposal relies on an alternative way of computing neighbourhoods that is better suited to the distribution of data by considering that a more distant neighbour must have less influence than a closer one. The relative importance of any neighbour in a neighbourhood is estimated using the SoftMax function on the implicit distance. Experiments are carried out on both simulated and real data sets. The proposed method outperforms alternatives, providing a promising new research line.