Scalable SVM Classifer With Application In Credit Rating

As a novel machine learning approach based on statistical learning, support vector machine (SVM) solve the practical problems such as nonlinear, high-dimension and small number of samples. It has obtained considerable researching attentions in the areas of pattern classification, target tracking and classification, image processing etc. However, SVM still has shortage such as imperfect robustness, low efficiency and high communicational burden for distributed learning. In order to attach these problems, several improved SVM classifiers are presented and applied in credit rating in this dissertation. Namely, introducing posterior probability of samples into binary tree of SVM, extending the posterior probability SVM (PPSVM) to multiclass case based on Fisher ratio separability, distributed scalable SVM for peer-to-peer sensor networks, etc. Specifically, the main contributions can be formulated as follows:(1) Posterior probability based SVM is robust against outliers and noises, even in the case of fuzzy or error class labels, while having less support vectors (SVs) and in turn decreased computational complexity. Hence, posterior probability is introduced for recently proposed fast classifier c-BTS, namely, a novel posterior probability based binary tree of SVMs (P2BTS) is given. Experimental results illustrate that P2BTS can obtain higher classification accuracy compared with c-BTS, while obviously less binary classifiers are necessary. This of course increases the decision burden and comparing times for P'BTS.(2) Considering the Fisher ratio seperability, PPSVM are extended to multiclass problem and two decision tree structure of PPSVM are proposed. One is the some-against-some PPSVM binary tree and the other is the one-against-rest PPSVM binary tree. The performance analysis for both approaches is also included. Both theoretical analysis and experimental results reveals that both approaches need (n-1) binary PPSVM classifier and converge respectively,O(log2n) and O((n!-1)/n). Moreover, both approaches improve the classifying accuracy while need less SVs and binary classifier.(3) Based on average consensus algorithm, a totally distributed scalable SVM (DS2VM) is proposed for peer-to-peer ad hoc sensor networks. One of the main advantages of the novel approach is that it only requires local samples during training, and then only information exchanges between neighbors make each agent in the network reaches network-wide agreement. This makes the novel approach scalable. Furthermore, a new consensus filter is introduced with its convergence and stability analysis. Experiments for UCI machine learning datasets shown the DS2VM achieve comparable accuracy with centralized learning, while the communicational burden is decreased.(4) Finally, the improved SVM algorithms are applied to the credit rating systems. Besal New Capital Accord is encouraging the banks develop their inner credit rating systems. The problem of credit rating is essentially nonlinear with small samples. What's more, the samples are always noise corrupted and containing outliers, it is unavoidable to have fuzzy or even error class labels. Therefore, the improved SVM classifiers introduced in the earlier chapters are applied to the problem of credit rating. A hierarchal standard rating decision support system is formulated based on the improved SVM classifiers. Practical analysis and experiments illustrate the effectiveness of the system.