Financial risk control cold start modeling method based on unsupervised field self-adaption

Abstract

The invention relates to the field of credit risk assessment, and aims to provide a financial risk control cold start modeling method based on unsupervised field self-adaption. The method comprises the steps of data input and preprocessing, variational auto-encoder mapping, domain alignment based on adversarial, sample adaptive weighting based on weighted adaptation, pseudo-tag iterative optimization, parameter optimization and result output. The risk control modeling algorithm can be applied to the cold start stage of new service online lack of label samples, is suitable for a cold start scene without label samples, and is higher in precision and better in effect in a financial risk control scene. The problem of negative migration in existing migration learning and the problem of data heterogeneity in the financial risk control field can be effectively solved. The model training and deployment framework is high in generalization, can be effectively applied to other service scenes, and has good adaptability and mobility.

Description

technical field [0001] The invention relates to the field of credit risk assessment, in particular to an adaptive financial risk control cold-start modeling method based on an unsupervised field. Background technique [0002] Credit risk forecasting is a set of decision-support technologies that assist lenders in extending consumer credit. These technologies affect who gets a loan, the amount of the loan, the interest rate on the loan, and what appropriate business strategies the lender will set to improve profit margins. Generally, the risk control of lending institutions includes three stages: pre-lending risk control, lending risk control, and post-lending risk control. [0003] Pre-loan risk control usually uses Application Scoring (Application Scoring), the main purpose of which is to identify the overdue risk of users in the customer acquisition stage. It is generally used in processes such as access, credit line granting, risk pricing, and expenditure approval. Risk...

AI Technical Summary

Problems solved by technology

In the initial stage of new product launch, we often face the problem of lacking or even no historical mark samples, such as figure 1 As shown, the user's label often needs to go through a performance period after the observation point to be determined, and the performance period is usually 3 months or even longer. At this time, it is difficult for modelers to establish data-driven rule decisions and supervised quantitative scoring. model, this dilemma is also known as the cold start problem of the wind control system

[0006] On the whole, the above inventions mainly have the following problems in the field of cold start of financial risk control: (1) Supervised methods based on XGBoost and logistic regression require a large number of label samples, which are not suitable for cold start scenarios of financial risk control; (2) The current migration learning-based methods, whether it is sample-based migration or fine-tuning-based migration methods, can alleviate the problem of lack of label samples in the initial stage of risk control modeling to a certain extent by migrating from the source domain to the target domain, but still requires a small amount of assisted modeling of target domain label samples, which cannot be directly applied to unlabeled financial risk control cold-start scenarios; (3) Unsupervised methods based on isolated forests are mainly used in fraud detection and anomaly recognition, which have very obvious category differences Detection problems, without the help of other business scenario experience, the application in the financial risk control cold start scenario has the problem of high classification contingency and low accuracy

[0009] 1) Supervised methods based on XGBoost and logistic regression require a large number of labeled samples, which are not suitable for cold start scenarios of financial risk control; 2) The current methods based on migration learning, whether it is sample-based migration or fine-tuning-based migration methods Migration from the source domain to the target domain can alleviate the problem of lack of label samples in the early stage of risk control modeling to a certain extent, but still requires a small number of target domain label samples to assist in modeling, and cannot be directly applied to financial risk control cooling without labels. In the start-up scenario; 3) Unsupervised methods based on isolated forests are mainly used in fraud detection, anomaly recognition and other anomaly detection problems with obvious category differences, and without the help of other business scenario experience, they are applied in financial risk control cold start scenarios There is a problem of large contingency and low accuracy of classification

Images