A Modified Nearest Neighbor Classification Approach Based on Class-Wise Local Information

Abstract

Nearest Neighbor (NN) is a nonparametric classification approach, which is simple yet effective. In NN, the classification decision only uses the information of distance, thus its classification performance is always undermined by the outliers. To solve such a problem, a modified nearest neighbor classification method is proposed by using class-wise local information. For a given test sample, its corresponding nearest neighbor in each class can found and the corresponding distance information can be derived. At the same time, the class distributions in the neighborhood of the test sample’s nearest neighbors in each class can be obtained, which is considered as the class-wise local information and can represent the possibility for the nearest neighbor in each class to be the outlier. The classification decision criterion is designed by jointly utilizing the distance information and the class-wise local information. Experimental results show that the proposed method is rational and effective.