Data Classification Algorithm Based On Kernel Fisher Discriminant

With the rapid development of Internet and other technologies, large amounts of datagenerate every day, so how to extract data from a mass of useful information becomes the newchallenge. Data mining technology came out in such demand. The classification has importantapplications in speech recognition and image classification as an important branch of datamining. Recently, the algorithm research and application on how to use the known sample toimprove the classification performance has attracted more and more attention on the academicand industry. However, with the increasing requirements of robustness, adaptability andclassification accuracy, the current classification algorithm may not meet the application needs,which needs further theoretical analysis.In this thesis, the work is mainly based on Kernel Fisher Discriminant (KFD) algorithm,specific studies are as follows:(1) To the traditional linear classification, a new iterative Fisher discrimination algorithm isgiven in this paper: first initialization is done by K-means algorithm, then the iteration isimplemented by using Fisher discrimination, this algorithm achieves good results on linearclassification problem.(2) On the kernel function selection problem of the non-linear classification algorithm (KFD),this paper compares nine commonly used kernel function of the classification performance underthe method of correct resample t-test using the rule of the information gain rate. We point outthat in the absence of priori condition, RBF function has the best classification ability.(3) Considering the parameter optimization problem of RBF kernel in KFD algorithm, wefirst point out the variation while changes from0to, then we verify the results byexperimental results.(4) On the problem of seeking the RBF kernel parameter, we propose a new "trichotomy" tofind the optimal RBF kernel parameter. Compared to commonly used optimization algorithms(cross-validation method and gradient descent), this new method makes up the lack of theparameter optimization.(5) To solve the shortcomings of traditional KFD algorithm, this article systematically studies of the weighted kernel fisher discriminant algorithm (WKFD), then we compare thecommonly used distance such as the centroid distance, the group average distance, Ward distance,and get a better weight function.