Abstract: This paper proposes a new algorithm named Kernel-kNN. To begin with, an approach for information energy metric is proposed, which is used to learn the nearest neighbor. This method overcomes the inconvenience for distance metric expression with high dimensional data set, and improves the consistency between the class similarity and the distance. Meanwhile, the traditional kNN is extended to an nonlinear form, and semidefinite programming is usd to learn the globally optimal metric matrix. The main characteristic of the proposed algorithm is that it is suitable for high dimensional data set, and can improve the classification performance efficiently. Experiments and analysis on many data sets have shown that Kernel-kNN can get the common performance in low dimensional data, and have a significant improvement on large scale data in high dimensions.