Abstract: Classification methods based on k-nearest neighbor (kNN) are the basis of realizing various high-performance pattern recognition techniques. However, these methods are easily affected by the selection of neighborhood parameter k. It is difficult to get the neighborhood of various data sets without knowing the characteristics of data sets. This paper introduces a new supervised classification method, named extended natural neighbor (ENaN) method, and proves that this method provides a better classification result without artificial selection of neighborhood parameters. Different from the previous kNN based method, ENaN method predicts different k in different stages. Therefore, whether in the training stage or in the test stage, ENaN method can learn more information from the dynamic neighborhood information, and provide better classification results. Simulations on real data of different types and scales all prove the effectiveness of our proposed method. 

Key words: classification analysis, nearest neighbor, parameter-free, extended natural neighbor, dynamic neighborhood