Research On Credit Card Fraud Transaction Identification Based On Modified LLM Model

The use of credit cards has greatly facilitated people's living expenses,and credit card consumption payment has become one of the main ways of domestic consumer payment.Drawing on the experience of foreign mature credit card business management,while the scale of credit card business maintains high growth,domestic banks are also strengthening and paying attention to credit card risk management.Credit card fraud transactions are one of the major risks of current credit card business.Credit card transaction fraud is extremely concealed.Due to the rapid development of information technology,the fraudulent means and fraud methods of fraudsters frequently change and escalate.This means that the distribution of transaction data will change with time,that is,the concept drift problem.At the same time,the growth of credit card business is accompanied by a sharp expansion of card transaction data volume.Data mining is a process of obtaining information and knowledge that people do not understand beforehand but may have high value from massive data.It has been widely used in credit risk management systems.The data mining technology involved in the research of fraud detection model mainly includes three aspects:(1)application research of classification algorithm,such as traditional decision tree and logistic regression classifier,emerging deep learning method,etc.;(2)Data balancing techniques for unbalanced data sets;(3)Incremental learning methods for solving conceptual drift problems.Comparing existing research,it can be found that most data mining models with excellent fraudulent transaction detection capabilities do not have good interpretability,which hinders the further research of fraudulent transactions by practitioners and analyzes the source of fraud.Therefore,based on the existing research,this paper does the following work:(1)The mixed model(Logit Leaf Model,LLM)composed of decision tree and logistic regression is used to construct the credit card fraud detection model.The model itself has good performance.Interpretable and excellent classification ability.(2)In order to adapt to the extreme imbalance of the credit card transaction data set,this study improves the classification model itself,introduces the loss sensitivity metric based on information gain,and proposes LLM based on transaction loss sensitivity.On the other hand,Combining data balancing techniques in the data preprocessing phase makes the data set tend to be balanced.(3)Under the scenario that the fraud mode changes with time,the research uses the incremental learning method to continuously update the training data set,while retaining some old data sets,continuously training the new model and combining the old and new models,in order to update the model knowledge.Through the combination of various data mining techniques above,a complete transaction loss-sensitive LLM fraud detection model is constructed.Finally,based on the real data set for experimental analysis,the experiment is divided into three stages: the first stage,based on the transaction loss sensitive LLM and six benchmark models for comparative analysis,using crossvalidation and grid search to optimize the model parameters;In the second stage,the influence of three classical data sampling techniques on the classification effect of the model is compared and the corresponding analysis is given.In the third stage,the incremental learning method is introduced on the basis of the previous two steps to verify the fraud detection model.Validity and extraction of classification rules for model interpretability analysis.