Abstract: Most existing typical semi-supervised dimensionality reduction algorithms often ignore the manifold features of data or use them inappropriately，while focusing on how to use supervised information.Therefore those algorithms show poor performance and can not be used in many fields.This paper presents a semi-supervised complex structure data dimensionality reduction method，called CSDDR，which keeps the global and local structure of the whole data manifold in the low dimensional embedding subspace.An appropriate objective function makes the method has more areas of application，and the experimental results show the effectiveness of the method.

Key words: Semi-Supervised Dimensionality Reduction（SSDR）, manifold assumption, constraints, objective function, clustering analysis

摘要： 现有的一些典型半监督降维算法，往往在利用标记信息的同时却忽略了样本数据本身的流形特征，或者是对流形特征使用不当，导致算法性能表现不佳并且应用领域狭窄。针对上述问题提出了半监督复杂结构数据降维方法，同时保持样本数据的全局与局部的流形特征。通过设置适当的目标函数，使算法结果能有更广泛的应用场合，实验证明了算法的有效性。

关键词: 半监督降维, 流形假设, 约束对, 目标函数, 聚类分析