Supervised Neighborhood Preserving Embedding Algorithm And Its Application

At the information age, data is complex and high-dimensional in data mining,machine vision, artificial intelligence and other fields, such as face data, gene expression data, and text data. If we directly perform learning tasks using these high-dimensional data,we would encounter the computational complexity, the curse of dimensionality and other unavoidable problems. One effective solution is to reduce the dimension of high-dimensional data. Manifold learning has made remarkable achievements in nonlinear dimension reduction, and attracted substantial attention.Neighborhood preserving embedding(NPE) is a typical unsupervised manifold learning method, which has been widely used in dimension reduction, clustering and visualization. This thesis focuses on neighborhood preserving embedding algorithm and extends it to supervised versions. The main work of this thesis is described in the following.(1) By introducing discriminate neighborhood embedding(DNE) to utilize the label information of training samples, this thesis proposes a novel supervised learning method,called discriminate neighborhood embedding-based neighborhood preserving embedding(DNE-NPE). In the novel method, the embedded points in the same class should be close to each other, while the embedded points in the different classes should be far away from each other. Experiments on face databases demonstrate the effectiveness of the proposed algorithm.(2) By introducing attraction points to utilize the label information of training samples,this thesis presents an attraction point-based NPE(AP-NPE) algorithm. In the proposed method, training samples belonging to the same class share the same attraction point and each embedded sample in the subspace should be attracted to its attraction point.Experiments on face databases show the proposed algorithm has a better performance than existing supervised algorithm.(3) This thesis introduces sparse representation to determine neighborhood based onAP-NPE, and then proposes an attraction point-based sparse NPE algorithm(AP-MSPP-NPE). The new algorithm can not only preserve the local neighborhood information but also take local sparse reconstructive relationship of data into account.Simulation experiments on face databases show the effectiveness of the proposed algorithm.