提出了一种有区别方差嵌入(DVE,distinguishing variance embeddings)算法。算法建立在最大方差展开(MVU,maximumvariance unfolding)和拉普拉斯特征映射(Laplacianeigenmaps)的基础上,通过最大化非近邻点在低维空间中的距离在全局上展开嵌入流形,利用近邻点间的距离和约束有效保留了数据集的局部邻近关系。DVE算法也可以看作是对MVU算法的约束松弛改进。在易于可视化的模拟数据和具有嵌入流形结构的图像数据集上进行了实验,实验结果证明了算法的有效性。 A DVE (distinguishing variance embeddings) algorithm is proposed. The algorithm is based on maximum unwrapping (MVU) and laplacian mapping (Laplacianeigenmaps), and the embedding manifold is developed globally by maximizing the distance between non-neighbors in low- The distance and constraint between the two data sets effectively preserve the local proximity of the data set. The DVE algorithm can also be viewed as a constraint relaxation improvement on the MVU algorithm. Experiments were carried out on the visual data and the image dataset with embedded manifold structure. Experimental results proved the effectiveness of the algorithm.