Research On Manifold Learning Methods Base On Feature Space Projection

Recently,manifold learning has aroused a great deal of interest in the field of machine learning.Manifold learning methods can automatically detect low dimensional nonlinear manifolds in high-dimensional data space,and maintain the original structure.So these approaches have achieved good performance in data visualization.However,as a kind of nonlinear feature extraction method,manifold learning methods show many shortcomings,such as sensitivity to noise and weak discriminability when faced with classification tasks.In order to overcome these shortcomings,based on the feature space projection,two feature extraction methods are proposed in this thesis.The main work of this thesis is summarized as following two aspects:(1)The original local linear embedding method is sensitive to outliers,a feature space projection based local linear embedding(FSP-based LLE)method is proposed to enhance the robustness.The projections in a neighborhood of all data points are calculated firstly in FSP-based LLE,low-dimensional embedding is constructed by the reconstruction weights,which are obtained by minimizing the reconstruction error for all projections.(2)A novel method named feature space distance metric learning(FSDML)is developed,distances between any two feature spaces are involved instead of distances between any two points.To advance the discriminability,the inter-class data separability and the intra-class data locality are all employed for graph embedding,and subspace is obtained through discriminant analysis.Experimental results on several artificial datasets indicate that FSP-based LLE is more robust than LLE.And,the proposed FSDML are evaluated by some state-of-art methods,experiments on some benchmark datasets have shown that the proposed method is more effective and efficient.