Semi-supervised Dimensional Reduction For Discriminative Clustering Analysis

In real applications, collection of a large number of unlabeled data for data analysis is easy.However, collecting data labels is difficult for the limited resource. By making use of the smallamount of valuable label information, semi-supervised learning improves the performance ofmachine learing. But the label information is hard to obtain, and we can often find inconsistentlabel in it. In addition, massive high-dimensional data brought us good returns, but also greatlyincreased the difficulty of machine learning.In this paper, we discuss the semi-supervised learning methods on data dimensionalityreduction and analyze the shortcoming of existing algorithms. In the first research part of thisthesis, we found the problem of contradictory constraints in prior information and proposed amethod of adjusting the inconsistent constraints. We use prior information to detect and revisethe contradictory constraints which avoid using unsupervised method to adjust supervisedinformation. It improves the accuracy of semi-supervised learning by revising contradictoryconstraints. The second research part of this thesis proposed a new objective function, whichmake full use of the information underlying the data. This function maintains the global andlocal structure of the whole data manifold in the low dimensional embedding subspace. For itreduce the invalid data in low dimensional data, the method has more area of application thanprevious algorithm. The algorithms were used in discriminative clustering analysis, and theexperimental results show better effectiveness of these two algorithms.