Selecting Adaptive Number of Nearest Neighbors in k-Nearest Neighbor Classifier Apply Diabetes Data

Natalia Labuda¹, Julia Seeliger¹, Tomasz Gedrande¹, Karol Kozak^1,2

1. Medical Faculty, Dresden University of Technology, Carl Gustav Carus University Hospital Dresden.
2. Faculty of Management, Finances and Informatics, Wroclaw University of Economy.

Abstract: The k-nearest neighbours (knn) is a simple but effective method of classification. K is the most important parameter in medical data classification based on k-nearest neighbor algorithm (knn). The major drawback with respect to knn is dependency on the selection of a “good value” for k. The value of k is usually determined by the cross-validation method but if k is too large, big classes will overwhelm small ones. On the other hand, if k is too small, the advantage of the knn algorithm will not be exhibited. Therefore, it is very likely that a fixed k value will result in a bias on large classes. In this paper we propose a modified k-nearest neighbor method, which uses different k values for different regions in an entire data set, rather than a fixed k value for a complete data set. The number of nearest neighbors is selected locally based on P-value Rate criteria. We apply the modified knn method to diagnose type II diabetes dataset which includes 768 samples from diabetic patients taken from Pima Indians Dataset.

Pages: 1 – 13 | Full PDF Paper