Abstract: Previous work on metric learning for k Nearest Neighbor（kNN） has focused on purely discriminative approach.A approach is proposed to learn a metric by analyzing the probability distribution on nearest neighbors provided that the underlying generative model is known.Experiments show that this learned local metric can improve the performance of the discriminative kNN approach using simple class conditional generative model.

Key words: k nearest neighbor, metric learning, generative model, discriminative model

摘要： 已有的关于k近邻测度学习算法的工作主要集中于纯区分模型。在假定隐含的生成模型已知的情况下，提出了一种通过分析样本的k个近邻点的概率密度学习测度的方法。实验表明，这种基于类的生成模型假设学习到的局部测度可以有效改善kNN区分模型的性能。

关键词: k近邻, 测度学习, 生成模型, 区分模型