Sample Weights Estimation And Its Application To SVW

Sample weights estimation and SVM are two tools for getting knowledge from thesample, and play a very important role in complex learning tasks. However, it is difficultfor SVM to learn large sample, incremental sample and noised sample. How to getinformation by sample weights estimation to improve SVM is a challenging problem. Forsolving these problems, we investigate sample weights estimation using the0-margin-hyperplane, based on which, we then explore SVM incremental learning andfuzzy support vector machine (FSVM). The main contributions of this thesis are outlinedas follows:1. The sample weights estimation method using the0-margin-hyperplane is proposedfor mining the classification knowledge from the training datasets, where the knowledgewill be used for SVM to finish complex learning tasks. The method constructs anappropriate hyperplane, i.e., the0-margin-hyperplane, for each sample to estimate itsweight with the classification accuracy or the wrong separating rate of the hyperplane.Thus, the weight, as a probability, predicates the importance of the sample for the futureSVM as well as the classification accuracy that SVM can achieve. That means that thenew weight is computed according to the effect of the sample for the learning machine,which can have a positive guide for itself. However, effect of the conventional weight thatuses distance for the learning machine is unforeseen. Experimental results on bothsynthetic and real datasets have shown that the performance of the proposed method isbetter than the conventional methods.2. An SVM incremental learning method based on quasi-support vectors andincremental sample is proposed. The method can receive incremental sample and discarduseless sample, and keeps only valuable sample so as to reduce the computational andmemory complexity. The method of sample weights estimation using the0-margin-hyperplane is used for choosing the quasi-support vectors, which has lower costand better results. Experimental results on both synthetic and real datasets have shown theadvantages of the proposed method.3. A new fuzzy weight function developed from the sample weights estimationmethod is proposed to construct an FSVM for reducing the effect of noises/outliers to theSVM classifier. Compared with the existing FSVM, the new fuzzy weight function is ableto suppress the potential bad training points effectively. This can help to avoid the effectof noises/outliers. Experimental results on both synthetic and real datasets have shownthat the performance of the proposed method is better than the existing methods. 4. When training a classifier with an unknown dataset, it is difficult to choose whichclassifier from SVM and FSVM. Aiming to this issue, we propose an improved FSVMwhose weights can float according to the separability of the datasets. The new FSVManalyzes the effect of the sample on the classifier according to the size of the sampleweights and the separability of the training dataset. If the sample does not affect theclassification accuracy, the FSVM will not suppress the sample. The new FSVM is able torecognize the noises adaptively, and switch between the standard SVM and the traditionalFSVM.