Research On Credit Evaluation Method Based On Ensemble Learning And Federated Learning

With the reform and opening up and the continuous promotion of China’s economic quality development process,the consumer finance industry of China is booming.However,financial institutions often suffer from serious information asymmetry when approving customers’ loan requests.For financial institutions,the information asymmetry means that managers have no immediate access to complete information about the customers before transactions,which makes it difficult for financial institutions to accurately evaluate the customers’ creditworthiness.This situation results in financial institutions being exposed to more defaults.Therefore,there is an urgent need for financial institutions to transform their profitability from a crude credit model based on user scale to an intensive credit model based on user quality.In this transformation process,the credit risk management capability of financial institutions plays an important role.The superior risk management capability of a financial institution is the core competitiveness in the fierce market competition.Therefore,how to build a more scientific,reasonable and effective credit evaluation model has become a key factor affecting the development of financial institutions.The credit evaluation model is essentially a classification model that discriminates between good and bad credit of loan applicants.In previous studies,many researchers have conducted extensive researches on credit evaluation models and proposed numerous integrated models for credit evaluation.However,there are still some shortcomings at the current stage.For example,the existing integrated credit evaluation models only focus on a certain aspect of the model for improvement,and rarely consider multiple aspects of the credit evaluation model in a comprehensive manner.In addition,when uniting data from multiple data sources for collaborative modeling,the existing credit evaluation models are unable to guarantee the data privacy and security.Hence,based on the existing research results,this paper systematically reviews the research status of credit evaluation models from two aspects: the research of credit evaluation models based on ensemble learning and the research of credit evaluation models based on federated learning.Subsequently,this paper provides a detailed introduction of relevant knowledge and techniques,such as information asymmetry theory,feature processing methods,noise detection methods,classifiers and classifier ensemble methods,and federated learning.On this basis,this paper presents two models applicable to different credit evaluation scenarios.(1)For the credit evaluation scenario in which models are constructed with a single data source,an integrated credit evaluation model based on multi-stage hybrid is proposed.First,a new feature selection method,FSCM-GA,is proposed to address the problem of high-dimensional features of data,which effectively reduces the negative impact of redundant features on the model;in addition,a stacking mechanism-based class noise detection method is proposed to address the class noise of data,which effectively enhances the adaptability of the model to class noise and further improves the model performance;finally,a novel classifier assignment method MBCG is proposed to combine with the weighted voting ensemble method to construct the resulting credit evaluation ensemble model.The experiments prove the effectiveness and superiority of the proposed model in the credit evaluation scenario where models are constructed with a single data source.(2)For the credit evaluation scenario in which model are constructed with multiple data sources,a credit evaluation model based on the federated knowledge transfer(Fed KT)algorithm is proposed.First of all,a federated learning framework is adopted to propose a modeling manner that can joint multiple data sources while protecting data privacy and security;second,a new federated algorithm Fed KT is proposed to cope with the Non-IID(Non-independent identical distribution)data distribution formed by class imbalance in credit data,which effectively avoids the learning insufficiency problem of local models;finally,the experiments prove the effectiveness and superiority of the proposed model in the credit evaluation scenario where models are constructed with multiple data sources.This paper uses multi-stage ensemble learning and federated learning for credit evaluation model construction,which can better adapt to the changing credit evaluation scenarios.This paper enriches the research and application of artificial intelligence technology in the field of credit evaluation with good theoretical value.In addition,the research results of this paper are beneficial to prevent systemic financial risks arising in the field of consumer finance in China,and are of great practical value to financial institutions in achieving efficient credit risk management,conducting large-scale collaborative modeling and protecting the privacy and security of customer data.