Design Of Matrix-pattern-oriented Classification Machine With The Universum

Pattern recognition mainly refers to the processing of information and phenomena, as well as describes and classifies different objects. Classifier design is the core content of pattern recognition. It relates to the knowledge of statistics, nonlinear algebraic and computer science. Regularization learning has been demonstrated to improve the generalization performance ef-fectively due to the introduced priori knowledge. Matrixized learning, as one of regularization learning, can boost the classification and meanwhile reduce the computational complexity when dealing with matrix data. So it can solve the problems of vector-oriented classifiers. The suc-cess of the matrixized learning is attributed to the potentially-incorporated structural knowledge of matrix data themselves. Compared with the traditional vector-oriented classifier, matrix clas-sifier algorithms are more suitable to access matrix data such as image data. But it only focuses on the structural information of the matrix sample itself, and pays no attention to the distri-bution information of whole data set. Thus it can be improved in classification performance. Universum learning does better in the prior domain information of distribution. It introduces Universum samples which do not belong to either class of the classification problem to help froming the classification hyperplane, so it can improve classification accuracy effectively. But at present, Universum learning only be used in vector-oriented classifier design. There are some defects in the process of selecting Universum samples, such as long computing time and bad Universum samples. In this paper, we mainly discuss and do some research to solve these problems.Firstly, we generalize the matrixized learning through taking advantage of the Universum samples. Doing so can not only introduce the structural knowledge of the individual matrix data themselves, but also get a priori domain knowledge of the whole data distribution. In imple-mentation, we incorporate our previous matrix-pattern-oriented modification of Ho-Kashyap algorithm named MatMHKS with the available Universum regularization term, and thus get-ting a regularized matrix-pattern-oriented classification machine with the Universum called UMatMHKS for short. There are several advantages of the proposed algorithm:1) the Uni-versum learning is introduced to the classifier. Hence the classifier can take full account of the relationship of the data distribution, and make full use of the priori domain information of distri-bution;2) it is the first time to compare the Universum learning and matrix learning. It does not only consider the prior domain information, but also inherit the advantages of matrix learning, paying attention to the structure information of sample itself. By analyzing the Rademacher complexity, we demonstrate UMatMHKS has tighter generalization risk boundary through the theoretical analysis and experiments. The experimental results here validate that the proposed UMatMHKS can effectively improve classification accuracy over both MatMHKS and some other state-of-the-art regularized algorithms.Then, we also discuss the method of Universum samples selection and generation. We combine the advantages of selection and generation algorithms, propose a new algorithm called Creating In-Between Universum (CIBU). The proposed algorithm has two main advantages:1) neighbor matrix is introduced to select the boundary samples;2) the idea of (?)Mean is introduced to control the computing time. The experimental results proved that CIBU can improve the classification performance as well as control the computing time effectively.