基于随机子空间的线性嵌入推斥图及其在半监督学习中的应用

Graph construction is a key step for graph based semi-supervised classification. Inspiring the success of the Local Linear Embedding method, this paper present a novel method called Random Subspace Method based Linear Neighborhoods Embedding Repulsion graph (RSMLNER). In the proposed framework, the geodesic distance is first used to define neighborhood system, and only the neighbors with same class label are then used for data reconstruction in the neighborhood system, resulting that one can construct a good graph which can describe the data manifold more effectively. Moreover, the random sub-space method is used for noise suppression. Experiments on the artificial dataset, Binary Alphadigits dataset, Newsgroups dataset and UCI dataset show the feasibility and effectiveness of the proposed method.