A Feature-Weighted Rule for the K-Nearest Neighbor

Abstract

The K-nearest Neighbor algorithm is a well-known non-parametric algorithm used for classifying. The algorithm is a simple, intuitive and preferable choice for many machine-learning models. Having that in mind, the negatives of the method should not be overlooked – the sensitivity of the k value, the choosing method of the neighbors and the voting mechanism.This paper reviews some state-of-art weight algorithms and motivated by their ideas proposes a solution for weight function. Unlike most weight functions, the proposed solution uses the features of the neighbors instead of just their distances. Some experiments are conducted on both real-world datasets and on well-known experimental ones. Some future improvements are targeted and the advantages and disadvantages are discussed.