基于LPP和Lasso的kNN回归算法

kNN Regression Algorithm Based on LPP and Lasso

针对k NN回归算法中k值固定且未考虑样本相关性的影响,提出一种基于LPP和Lasso的最近邻算法.该算法通过局部保持投影与稀疏编码相结合,使训练样本对每一个测试样本都进行重构,重构过程中,LPP用于保持原始数据的局部结构,l1-范式确保每个测试样本被k个不同数目的最近邻样本预测,以此解决k NN算法中k值固定问题.在UCI数据集上得到的实验结果表明,改进算法在线性回归中的预测能力优于传统k NN算法.

This paper proposed a newnearest neighbor algorithm based LPP and Lasso,for solving the fixed k value problem and correlation among the samples wasn’ t considered of k NN algorithms. The proposed algorithm combined Locality Preserving Projections（ LPP） with sparse coding（ e.g.,Lasso） to restructure test samples with the training data.During the reconstruction process,LPP was used to preserve the local structures of the data and the l1-norm was used to learn different k value for various samples.Experiments results on UCI datasets showed that the proposed methods were superior to traditional k NN algorithm in terms of regression performance.