The Study Of Novel Dimensionality Reduction Methods And Application In Intelligent Recognition

In the field of intelligent pattern recognition, feature reduction methods of are widely used to preprocess high-dimensional raw data in order to improve the effectiveness of intelligent recognition. Feature extraction and feature selection, two primary feature reduction methods, are explored in depth and have been applied successfully to deal with certain pattern recognition problems. While dealing with some concrete recognition problems, we find that the current classical feature reduction methods and parts of improved ones show shortcomings of low robustness and weak generalization ability to certain extent. Therefore, in order to improve their ability, new feature reduction methods are proposed through integrating other intelligent processing technology into them. They are presented as follows.Part one consists of chapter 2 and 3, in which two novel feature selection methods are discussed respectively. In chapter 2, in order to solve the weak generalization ability problem of the potential support vector machine (P-SVM), within-class scatter matrix in Fisher linear discriminant criterion is used to reconstruct an objective function of P-SVM and hence a generalized potential support features sSelection method (GPSFM) is proposed, which has stronger generalization capability. GPSFM not only has the advantages of P-SVM to some extent but also has the characteristics of low redundant features selection, high selection speed and nicer adaptive abilities. So compared with the traditional P-SVM, this new method has much stronger abilities in both feature selection and classification. In chapter 3, since the traditional fuzzy clustering method, FCM, is sensitive to noisy data and features, sample points and features are weighted simultaneously to reconstruct FCM objective function and hence a new robust fuzzy clustering algorithm with the capability of feature ranking-FCA-is proposed. The FCA can rank the samples'features simultaneously according to their contribution degrees when realizing clustering efficiently. Therefore, the proposed algorithm FCA is robust and can be used to extract the sample's features. The reasonable value range of feature and sample weighting parameter is defined according to certain geometry, which is proved theoretically in this chapter.Part two consists of chapter 4 and 5, in which two novel feature extraction methods are discussed respectively. In chapter 4, it is considered that linear Laplacian discrimination (LLD) suffers from the small size sample problem (SSS) and/or the type of the sample space when it is used. So a contextual-distance metric based Laplacian maximum margin criterion (CLMMC) is proposed by using contextual-distance metric and integrating maximum margin criterion (MMC) into the LLD. The proposed criterion can obviously decrease the dependence on the sample space since the contextual-distance metric focuses more on intrinsic structure of a cluster of samples than on the type of it. And it is of higher adaptability and efficiency by using a new algorithm to compute contextual-distance metric and applying QR-discomposition when high-dimensional vector data are dealt with. The basic properties of CLMMC and its relation to LLD are also discussed. In chapter 5, since maximum scatter difference discriminant criterion (MSD) has the following drawbacks:the effectiveness of MSD depends greatly on the definition of the parameterηand it is a type of hard classification, which can't reflect the real world objectively to certain extent, a novel fuzzy maximum scatter difference discriminant criterion is presented by using fuzzy technology Based on this criterion, the novel fuzzy clustering algorithm, FMSDC, is also presented here. The proposed algorithm realizes clustering by iterative optimizing procedure and at the same time obtains the optimal discriminant vectors which reduces dimensionality. The parameterηin the fuzzy criterion can be appropriately determined in terms of a certain principle. Hence, the sensibility aroused by parameterηis decreased to certain extent.Part three consists of chapter 6 and 7, in which two within-class scatter SVM based methods are studied. In chapter 6, the following is discussed:minimum within-class scatter support vector machines (MCSVMs) suffer from the small size sample problem. By integrating the tensor theory into it, the objective function of MCSVMs is reconstructed and then a new matrix pattern based MCSVMs (MCSVMsmatrix) and a corresponding nonlinear kernel method, Ker-MCSVMsmatrix, are presented. The MCSVMsmatrix not only solves the small size sample problem of MCSVMs but also reduces the time/place complexity. And the Ker-MCSVMsmatrix makes the matrix pattern nonlinear for the first time. In chapter 7, in order to overcome the drawback of the traditional SVM which can't suffeciently reflect the internal geometric structure and discriminant information of samples, a global and local preserving based semi-supervised support vector machines:GLSSVMs, is presented in this paper by integrating the basic theories of the locality preserving projections (LPP) and the linear discriminant analysis (LDA) into the SVM. This method fully considers the global and local geometric structure between the samples, shows the global and local underlying discriminant information and meets the consistency assuming which the semi-supervised method must coincides with.