Modification Of Matrix-Pattern Classifier And Applied To Bank Card Transaction System

Pattern recognition is one of the most important application of the automation services, it appears in all areas of our life. The purpose of pattern recognition is to identify objects by computer algorithm which simulates the human ability of recognize object. Pattern classification is one of the most typical applications of pattern recognition, for example: identification of benign tumor, personal identity authentication, E-mail spam classification, etc. The traditional classifier model uses pattern which is represented as a vector, the one-dimensional vector uses numerical value to represent features of the object. The2-D pattern will be converted to vector by a given rule before it can be used in a vector-pattern classifier.In recent years, the Matrix-oriented mode (2D mode) has drawn a lot of attentions. Many designs and researches which is based on the matrix-pattern have been published. There also have some matrix-pattern classifier which is based on the vector-based classifier that accepts the raw input of matrix-pattern. These algorithms are defined as matrix-oriented classifier algorithms, for example:2D-PCA, Mat-LDA, Mat-LSSVM and so on. In this paper, we made some modification on two existing instance of matrix-pattern classifiers:Matrix-pattern Least Square Support Vector Machine (MatLSSVM) and Matrix-pattern Ho-Kashyap Classifier with Regularization Learning (MatMHKS).This paper contains three parts.1) Based on the existing algorithm MatLSSVM and MatMHKS, we changed the process procedure of the right weighted vector into random data, it cause the sample data can be projected to different random spaces, then we use Adaboost or Bagging Algorithms to assemble multiple random projection matrix-pattern classifier into one classifier, we call the new methods Random Projection Matrix-pattern Support Vector Machine (RPMatLSSVM) and Random Projection Matrix-pattern Ho-Kashyap Classifier with Regularization Learning (RPMatMHKS).2) Furthermore, the matrix-pattern classifier has a well-known problem called matrix-pattern-dependence that when the matrix classifier process the different matrix pattern of the sample data, the result is different. To minimize the effect of the matrix-pattern-dependence due to the bad matrix-pattern causing non-ideal result, we use multiple matrix-pattern in the sample process of algorithm RPMatLSSVM and RPMatMHKS. We have designed two new classifiers based on the multiple matrix pattern method:Random Projection Matrix-pattern Support Vector Machine with Multiple Matrix-pattern (MRPLSSVM) and Random Projection Matrix-pattern Ho-Kashyap Classifier with Regularization Learning with Multiple Matrix-pattern (MRPMHKS). 3) In the end of this paper, we have coded the MRPLSSVM algorithm integrate it into an existing recognition system of banking card transactions pattern. We run the bench test and compared the result with the original classification system.