Research On Support Vector Machines For Imbalanced Datasets And Incremental Learning

Support vector machine (SVM) is a powerful and novel machine learning approachdeveloped in the framework of statistical learning theory. It solves the intractableproblems of traditional learning approaches, such as local minima and over-learning etc,support vector machine has good capacity of generalization. Because of havingself-contained theories and good experimental results, Support vector machines can beapplied in areas such as fingerprint, face recognition, disease detection and failuredetection in auto-controlled equipment.However, being a new theory,many aspects ofsupport vector machines are immature and incomplete currently, and more researchesand improvements should be done.In this paper, we illustrate existing algorithm and application researches on SVM,we study in detail some existing problems concerned now, such as the unbalancedproblems and bottle-neck to deal with large-scale data. To overcome the disadvantage ofSVM, some researches are carried out aimed at unbalanced problems and incrementallearning based on SVM.Study the adjustment method for unbalanced support vector machines. This paperproposes an improvement fuzzy support vector machine (FSVM) for the separatinghyper plane of two classes of unbalanced data in SVM. This method which based on thetheory of Fuzzy Support Vector,considering the capacity and dispersion degree ofsample,designs a new Membership funciton of FSVM. Simulation results show thatclassification accuracy of SVM classifier is highly improved.Study the sampling cutting technique for unbalanced support vector machines. Forthe classification information loss problem occurred in traditional random undersampling method, the dissertation proposes a new under sampling method.This newmethod using the distance of samples classified the samples,and then cuts the samplesof different category with different sample rate.Experimental results show that afterpreprocessing datasets by the above two methods, classification accuracy of SVM forimbalanced datasets will be highly improved.Study the incremental learning based on SVM. This paper combining theincremental learning method based on KKT condition and C means, propose a kind offast incremental support vector machines. This method can deal with the history andincremental redundant information efficiently, it improves the classification accuracyand rate of SVM classifier.Of cause, there are plenty of researches need to be done, support vector machinesshould be plumper in model and application based on these research.