The Study On Supervised Manifold Learning Algorithms In Pattern Recognition

With the arrival of big data, artificial intelligence and pattern recognition are widely applied in more and more filed. The data scale of Face recognition, text classification and data mining is increasing in the speed of geometric rate. Those data have the common characteristic of high dimension and nonlinear distributions. With the dimension rising of the data, the method of traditional dimensionality reduction on linear data has been unable to meet the demand of complex data processing. So the emergence of manifold learning for nonlinear dimensionality reduction brings the infinite possibility. When embedding the data into low dimensional space, manifold learning can explore and preserve the inherent structure of nonlinear distributed data in high dimensional space.In this paper, the research mainly put focus on metric learning and neighbor selection during manifold learning. The main work of this thesis can be summarized as follows:(1) A new method which fuses Euclidean distance and label information is presented. For the traditional algorithm of locally linear embedding, the internal structure of data can’t be reflected effectively with the Euclidean distance during the neighbor selection. Also it often uses unsupervised methods for nonlinear dimensionality reduction, taking no consideration on the class information of data. As a result, the embedding data in the low dimensional space coordinate shows large overlap and can’t be classified properly. The paper proposes an improved Supervised LLE which fuses label information and Euclidean distance (ISO-SPLLE). The experiment result of UCI datasets shows its good performance on reduction and recognition for high dimensional data.(2) A new method which fuses Mahalanobis distance and label information is presented. With the dataset which has large similarity between samples, the algorithm firstly ascertain a Mahalaobis metric from the existing samples. Then, the Mahalanobis and label information are used to reduce the dimensionality of the new samples. Thus, the accuracy of sample recognition is improved largely while improving the similarity between samples of the same class. The validities can be confirmed through experiments on UCI datasets and remote sensing image.(3) For the new unlabeled sample, ELM is using to perform the map function which embeds the unlabeled data to the low dimensional space. When the new data need dimensionality reduction, it can be directly mapped through the trained ELM model.