A Credit Card Customer Segmentation Model Based On Improved Support Vector Machine

More and more world-class foreign-funded banks have flooded into Chinese money market after China entered into WTO and reformed of the monetary system, which intensifies the competition among banks. If banks don't want to be defeated in this competition, they should segment customers by their value. Then it can use limited resources to server for the high-valued customers, realizing maximization of resource value.However, for the ever-increasing numbers of customers and business volume, the data volume of bank is very massive. The traditional methods based on experience or simple statistics can't meet requirement. Luckily, the emergence of data mining technology provides new solution for customer segmentation under massive data volume.SVM (Support Vector Machine) is a new and effective machine learning method, which is advanced on the base of statistics theory. SVM can resolve a series problems, such as non-linear, over-learn, high-dimension, local minimum value. Now, more and more learners pay attention to SVM, which now becomes the hot in the research field of Artificial Intelligent and Machine Learning.On the base of research in and out of China, the author analyzed a series of indexes which will influence customer value, and advanced a all-round credit card customer segmentation index system. In this paper, SVM was applied in credit customer segmentation after improved for concerning practical application problems, such as unbalanced distributed data, different misclassification cost, fuzzy subordination, low learning speed intrigued by large number of support vector and multi-class classification.The main research job done by author is listed as follows:Firstly, construct an all-round credit card customer segmentation indicator system for personal characteristic will influence consuming behavior, which embody certain customer value. Enhance, we suggest classify customers from the view of personal characterize, consuming behavior and customer contribution.Secondly, introduce a support vector machine classification method based on cost-sensitive learning. Most traditional classification methods based on data mining is universal. However, for many practical problems, different misclassification has different cost. Taking this into consideration, author introduces cost-sensitive learning into segmentation. It is worth to notice that, a function is designed to computer misclassification cost, while most of research set misclassification cost according experience.Thirdly, advance SVM based on Fuzzy C-Mean (FCM) Cluster. One customer may be subordinated to many classes at the same time. Take this problem into consider, author adopts Fuzzy C-Mean to resolve those. Standard SVM will produce many support vectors, which intrigued low learning speed. For solving this problem, the solution is not adding a variation as most paper done, but using FCM to cluster Sample. Then we will get c pair sample distributed on the boundary. Take this c pair sample as support vector. Use this support vector and sample to train SVM. This method can reduce the number of support vector, and then improve learning speed.Fourthly, advance a multi-class SVM based on distance binary tree. Standard SVM only can classify 2 classes. However, most practical applications are multi-classes problems. Enhance, we should expand Standard SVM to multi-class SVM to fit practical application requirement. In this paper, we adopt binary tree based on distance to resolve the structure problem of traditional binary tree, and improve the generalized ability.Fifthly, design a customer segmentation system which is based on data mining. Under the basis of aforementioned chapters, design a customer segmentation system based on improved SVM to segment customers.