Hamid Nasiri - Computer Engineering and Information Technology Department Amirkabir University of Technology (Tehran Polytechnic)Tehran, Iran
Saeed Shiry Ghidary - Computer Engineering and Information Technology Department Amirkabir University of Technology (Tehran Polytechnic)Tehran, Iran
Mohammad Mehdi Ebadzadeh

Classification is a broad ranging research field and different algorithms proposed in this area, one of which is K nearest neighbor (KNN) algorithm. This algorithm has a simple structure and easy implementation. Its performance depends on three main factors including similarity measure for voting, distance function and appropriate value for the parameter K among which the value of K is particularly significant, So that if it is not correctly selected, algorithm performance would remarkably reduce. We proposed a novel method for adaptive selection of parameter k in this paper. In this method, an optimal K-value for each training instance is obtained and used to classify a test instance by KNN algorithm. Evaluation tests on standard datasets and comparing obtained results with conventional methods show that the presented method has an acceptable performance compared to other methods and improves classification accuracy as well.

K Nearest Neighbor Algorithm; Adaptive KNN Algorithm; Nearest Neighbor Classification; Pattern Classification