Transductive Learning And Model Transformation Based On SVM And Its Application Problems In Enterprises Credit Assessment

Credit assessment is the key step in loan business of commercial banks.The assessments can have significant impact on bank lending decisions and profitability. The rationality and reliability of credit assessment will greatly affect the achievements of a bank.The commercial banks in china need some much better assessment methods to improve their competition ability. Nowadays, the traditional ratio-analysis method is popularly adopted. Many Chinese scholars have done related research in this field and acquired certain achievements. However, because of the short term of credit assessment developed in China and the incompleteness of historical data, neither the traditional statistic methods nor the intelligent training algorithms, such as Neural Network (NN), could do a good generalization ability in it. Considering the developing status of credit assessment in China, we introduce Support Vector Machine (SVM) technique, which is a universal learning algorithm and based on the small sample-size learning theory, to research this problem.In this paper, we research two parts of contents:First, transductive learning method is an extension of standard SVM in semi-supervised learning problem. However, the existing TSVM algorithm has some drawbacks, such as slow training speed, too much back learning steps, and unstable learning performance, etc. Thispaper presents an improved TSVM algorithm-----PPTSVM, which join the thinking of postprobability. Experiments results show the superiority of PPTSVM over TSVM in computational efficiency and classification accuracy.Second, since the credit assessment is a multi-classification question, but the support vector machine (SVM) is a binary-classification method. We also research the SVM multi-class classification SVM.The work is consisted of two parts. The first part transforms the multi-class problem to the two-class problem through varying the description of the problem. This thesis proposes a SVM multi-class classification method based on boosting in subspace. The second part is about binary tree SVM. The algorithm takes the separability between classes as metric and generates the node of binary tree by the clustering method. Experimental results show that the above-mentioned two algorithms both achieve the well result. And the method based on clustering is the best.