Research On Nearest Neighbor Credit Scoring Model Based On Metric Learning

Consumer lending is an important business for financing institution.The Financing from banks and the consumption with credit card are generally accepted by the small enterprises and personal consumers.When facing the credit application of a new client,the financing institution need to rapidly make a decision whether or not lending to him by using credit scoring.With the rapid development of consumer credit service,many credit scoring models are developed and widely used for the credit admission decision,such as the statistic methods including discriminant analysis,Logistic regression and decision tree,etc.and the artificial intelligence methods including neural network,support vector machine,etc.while there are different advantages for every credit scoring model in real world application,there are also some limitations.The credit scoring model based on the nearest neighbor method is a nonparametric method,in which there are not assumptions about the distribution of data and the introduction of other parameters.However,it is usually relied to the selection of an appropriate distance measure.Thus,in this thesis,we will study the nearest neighbor credit scoring model based on metric learning.This thesis mainly studies the nearest neighbor credit scoring model.The appropri-ate distance measure is provided by using the proposed metric learning method.The op-timization problem of metric learning method is solved by the proposed intrinsic steepest descent algorithm,which not only increases the accuracy of learned distance metric,but also improves the classification performance of nearest neighbor credit scoring model.The main researches and innovations in this thesis are summarized as follows:1.Studying the nearest neighbor credit scoring model based on supervised metric learning.By using the proposed nonlinear supervised metric learning method,an ap-propriate distance measure is provided,which can express the distribution of data more better,and can overcome the disadvantages of Euclidean metric which are influenced by the distribution and characteristic variables of data.Then it efficiently improves the prediction accuracy of the nearest neighbor credit scoring model.There is only one con-straint condition of nonlinear supervised metric learning model,so it can be transformed to an unconstraint optimization problem,which avoids the problem that many constraint conditions will lead to the complex computations in existing metric learning models.2.By analyzing the geometrical characteristic of constraint condition,we transfer the optimization problem of proposed nonlinear supervised metric learning method into an unconstraint optimization problem,then we design an intrinsic steepest descent algo-rithm on manifolds to solve it.This algorithm can ensure the iteration will be performed along the manifolds strictly,which addresses the projection problem in traditional al-gorithms,and improves the accuracy of learned metric and the performance of nearest neighbor credit scoring model.3.Studying the nearest neighbor credit scoring model based on semi-supervised metric learning.An appropriate distance measure is provided by using the proposed nonlinear supervised metric learning method.While the metric is learned from a few la-beled client data,it better expresses the distribution of all data.Then,when facing a new credit application,we only need to compute the distances between the new client and the labeled clients,not the distances between the new client and all the clients in training set,which highly reduces the computational costs.Otherwise,The proposed nonlinear metric learning method is based on semi-supervised learning and multiple kernel learn-ing,which has two advantages.One is that it just uses a small number of labeled client data,which reduce the cost to label all clients for financing institution.The other is that it can provide a new kernel function,which addresses the choosing of kernel function in traditional nonlinear method.