In neighborhood rough sets, the neighborhood classifier is an intuitive and effective classification method in data mining. However, the neighborhood radius, as a key factor to determine the classification performance of the neighborhood classifier, has some defects in construction. The construction of the neighborhood radius is not universal due to the lack of the training stage. When the data samples are not uniformly distributed and then empty neighborhoods appear, the classifier fails. Aiming at these problems, a near neighborhood classifier with adaptive radius selection(NNC-AR) is proposed. Firstly, for training samples, a training neighborhood radius based on K-nearest neighbor is defined. For test samples, an adaptive neighborhood radius is defined to overcome the subjectivity of the artificial parameter in the traditional neighborhood radius. Finally, for the test samples with classifier failure, an approximate neighborhood radius is defined to improve the generalization ability of the classifier. The experimental results show that the F1 score and classification accuracy of NNC-AR model are significantly improved.