Semi-supervised Support Vector Machine For Credit Prediction Using Simulated Annealing And Coupled Simulated Annealing

With the development of China's economy,enterprise credit and personal credit are increasingly valued by society.However,due to the lack of Chinese enterprises credit database,many enterprises,especially small enterprises,have no credit information.In recent years,the development of the Internet has promoted the digitization of information,and a large amount of credit data can be obtained through the network.At the same time,traditional supervised learning tends to have lower prediction accuracy when facing the problem of small amount of labeled data and unbalanced data.Semi-supervised support vector machine(SVM)can study with a small amount of labeled data and a large amount of unlabeled data,and it can overcome the problems of unbalanced credit data and insufficient sample information.However,the parameters of the semi-supervised support vector machine have a great influence on the algorithm.The selection of the parameters are often based on human experience.To solve the problem,this thesis proposes an algorithm by using simulated annealing to find the optimal hyperparameters of the semi-supervised support vector machine.And then the coupled simulated annealing semi-supervised support vector machine was proposed,with the coupled local optimal technique.The main work of this thesis is as follows:(1)The analysis and experimental comparisons on semi-supervised support vector machine.This thesis mainly analyzes the support vector machine and the semi-supervised support vector machine,and conducts comparative experiments on the enterprise credit information data and personal credit data.The experimental results show that the semi-supervised SVM has good robustness in credit prediction.(2)For the problem that the traditional semi-supervised SVM hyperparameters needs artificial experience selection,this thesis proposes a credit prediction algorithm based on simulated annealing semi-supervised SVM.It uses the simulated annealing method to find the optimal hyperparameters of the semi-supervised SVM.Compared with the traditional semi-supervised algorithm,the experimental results show that the precision of the simulated annealing semi-supervised SVM is increased by 13% at most,the false positive rate is reduced by 10% at most,the accuracy rate is increased by 11% at most,and the F-1 measure is increased by 13% at most.(3)For the single annealing problem of the simulated annealing semi-supervised SVM a credit prediction method based on coupled simulated annealing semi-supervised support vector machine is proposed.The coupled simulated annealing technique is used to share the information between multiple simulated annealing processes.It can reduce the number of iterations and accelerate the iterative convergence on fixed temperature.In addition,the coupled simulated annealing semi-supervised support vector machine with variance control temperature is proposed for the initial acceptance temperature of the coupled simulated annealing method.The results on the credit dataset show that the coupled simulated annealing semi-supervised SVM improves the precision of the credit data by 16% at most,the false positive rate is reduced by 30%,and the F-1 measure is increased by 3% at most.The initial acceptance temperature comparison test shows that the coupled simulated annealing semi-supervised support vector machine with variance control acceptance temperature can reduce the dependence on the initial acceptance temperature.