Improving recall of k-nearest neighbor algorithm for classes of uneven size

The k-nearest neighbor algorithm is one of the most suitable method of classification for its simplicity, adaptability and performance. The real problem arises when classes do overlap and when samples size is unevenly distributed between categories. Many studies present optimization techniques on discriminant metrics, on weighting the features, on using probabilistic measures or adjusting the prototypes position. Classes that are represented by a small sample size are overwhelmed by the large number of prototypes of dominated groups. In this paper we describe a method of weighting the prototypes for each class of the k nearest neighbors to cope with the uneven distribution of data. The proposed method increases the classification rate in terms of recall measure.