Research On Regularized Classifier Based On Sample Prior Information

In pattern recognition, regularization technology is widely applied to design the classifiers and achieves great success. However, there are three main problems in approving the regular-ization technology. The first is how to increase the generalization ability of the classifiers. The second is how to dig more priori information of original data. The last is how to arrange both global and local information in design progress. So the main aim of this thesis is to fusion more data priori structural information and gain better generalization ability both globally and local-ly. The thesis proposes an effective Three-fold Structural Multiple Empirical Kernel Learning (TSMEKL) model on the generally priori information of samples. Then it proposes a classifier algorithm based on a matrix weight with local priori sample information, called Double MHKS (MHKS). The contribution of this thesis is as follows:Firstly, in the classical multiple kernel learning frameworks, most regularization classi-fiers focus on the information of certain feature spaces. But they ignore the priori information between samples and the presentation under different feature spaces. The proposed TSMEKL simultaneously utilizes the space, the class and the cluster information in the way from globali-ty to locality. TSMEKL owns three-fold structural information and can guide the classification performance to an improving trend. The main advantage of the developed three-fold structural learning framework is considering different folds of priori information so as to improve the per-formance. The experimental results demonstrate the feasibility and effectiveness of TSMEKL. Furthermore, we discuss the theoretical and experimental generalization risk bound of the pro-posed algorithm.Secondly, traditional regularization classifiers are based on vectorial samples, leading to ignore the original information of matrix patterns. Trying to apply more structure of data itself, the thesis proposes a classifier algorithm based on the matrix weight, called DMHKS, which can process matrix pattern directly. DMHKS gets a satisfying performance in experiments and achieves higher training speed than other compared algorithms. Moreover, the influence of the parameters on the performance is also discussed.Last but not the least, this thesis designs two regularization classifiers based on the priori information of global samples and single sample. TSMEKL is also a effective framework with the high compatibility of other algorithms. And the motivation of DMHKS can be applied on other classifier to speed up calculation. Both this two algorithm are enlightening and powerfully expandable and contain important research meaning and theoretical value.