关键词: 两层分类, 距离学习, 基于概率, AdaBoost, 平均绝对误差

Abstract: For finite samples, KNN heavily depends on an appropriate distance while distance learning on previous studies doesn’t fully consider the distribution of the sample. In this paper, a new probability-based two-level nearest neighbor classification algorithm for adaptive metrics（PTLNN） is proposed. This proposed algorithm is divided into two levels, it uses Euclidean distance in the low-level to determine an unlabeled sample local subspace set; at the high-level, it uses AdaBoost to extract information from subspace. With principle of minimizing the mean absolute error, it defines a probability-based adaptive distance nearest neighbor classifier. The proposed algorithm PTLNN combines the advantages of KNN and AdaBoost algorithm, fully considers the sample distributions in finite sample to reduce error rate, and has good stability in noisy data to reduce AdaBoost’s overfitting phenomenon. In contrast to other algorithms, experimental results show PTLNN can achieve better results.

Key words: two-level classification, metric learning, probability-based, AdaBoost, mean-absolute error