Recently, dimensionality reduction have become hot topics in the field of machine.Dimensionality reduction is the s the process of transforming data from a high dimensional space to a low dimensional space to reveal the intrinsic structure of the distribution of data. It plays a crucial role as a way of dealing with the “curse of dimensionality”.In the approximate number of multidimensional reduction method, graph based dimensionality reduction is widely studied for its simplicity, understandability and good performance. These graph based approaches take each sample as a node of the graph,and the objective of these methods is to find a subspace, in which the underlining structure of samples embedding in the high-dimensional space are preserved in the low-dimensional space. How to structure graph in the high dimensional samples is the key of dimension reduction algorithm base on graph. A good figure is combined with proper dimension reduction algorithm can have good performance. In contrast, different from traditional unsupervised dimensionality reduction approaches, semi-supervised dimensionality reduction performs dimensionality reduction under the guidance of valuable and limited labeled samples or pairwise constraints, it often gets better embedding than the unsupervised ones.In past decades, a large number of dimensionality reduction algorithms. These algorithms including the conventional algorithms e.g., PCA and LDA; the recent proposed manifold learning algorithms; algorithms based on figure and so on. However,most of existing dimensionality reduction algorithms still suffer from various open problems e.g., small sample size problem, nonlinearity of the distribution of samples,and classification problem. In order to solve some problems, further improve the efficiency and accuracy of dimension reduction algorithm, we for dimension reductionmethods about the research of the system. Specifically, the paper main work includes the following two aspects:(1)We have proposed Semi-Supervised Dimensionality Reduction based on Composite Graph, SSDRCG in short. According to the ideas of the stochastic subspace,SSDRCG put each subspace structure by means of clustering subgraph. Then each figure together constitute a composite figure. Finally we construct the new pairing with based on constraints of a semi-supervised dimensionality reduction algorithm combining to form a new dimension reduction algorithm In order to verify the algorithm’s performance, we respectively with a variety of dimension reduction methods about SSDRCG. Through the different face data sets of experiments show that the dimension reduction algorithm using the classification accuracy is higher than other methods, time complexity is lower than the ELPP.(2) An improved method based on hybrid graph of a semi-supervised dimensionality reduction algorithm, SSDRHG is short. SSDRHG using mixed graph structure strategy to further build the new figure. Semi-supervised dimensionality reduction base on pairwise constraints about using constraints, both at the same time also use the untagged data. By experiment on two face data sets and a variety of methods contrast analysis shows that under the same parameters, this method is more superior, not only the classification accuracy is higher than other related methods after dimension reduction, and the selection of neighbor parameters and noise data more robust. This method of thought can be applied to the other in learning based on graph. |