Analysis And Method Research On Identification Of Fraudulent Transactions

With the development of Internet technology,credit card fraud is rampant day by day.The lag of risk prevention and control is becoming one of the main obstacles to the development and profit-making of Internet financial business.Establishing a credit card risk prevention and control system to detect credit card fraudulent transactions has gradually become a research hotspot.In the field of machine learning,credit card fraud detection is a supervised binary classification problem,each transaction will be judged as legitimate or fraudulent transaction.Existing fraud detection models mainly rely on advanced machine learning algorithms,which uses historical transaction data to train a machine learning model,such as random forest,logistic regression,etc.The legitimate and fraudulent modes of transaction behavior can be obtained to distinguish between normal and fraudulent transactions.However,in the construction of credit card fraud detection model,because the original characteristics of transaction data are usually relatively simple and the information provided is very limited,how to extract the correct features from the transaction data is very important.In traditional methods,this is usually done by aggregating the historical transactions of users to analyze the amount and transaction objects,etc.,which purpose is to observe the consumer behavior patterns of users.However,this feature engineering method is very limited for information enrichment and extraction.In order to solve this problem,this paper proposes to construct a behavior rule space,which combines individual behavior analysis with group behavior analysis and extracts the relationship between individuals.Then we expands it to a new feature of transaction to describe group behavior,and combines the new behavior feature with the original feature for training model.In credit card fraud detection,there are still many challenges to build an effective fraud detection model,such as concept drift(users' habits will change over time,fraudsters will also change strategies),class imbalance(the number of legitimate transactions far exceeds fraudulent transactions),verification delay(only a small percentage of transactions are checked in time),etc.Among these problems,class imbalance and concept drift are very prominent.Firstly,the number of legitimate transactions is much more than fraudulent transactions,and traditional machine learning methods tend to classify most types,i.e.,legitimate transactions,which will greatly affect the identification of fraudulent transactions.In addition,the user's behavior isconstantly changing,which will lead to the concept drift in the transaction data,affecting the model's identification of future transactions.To solve these two problems,this paper presents an integrated learning strategy,which enables the model to retain old behavior information while learning new information.This paper compares several commonly used classification algorithms through experiments,and chooses random forest as the basic model of learning strategy.In this paper,the proposed integrated learning strategy and two popular learning strategies,forgetting strategy and balancing strategy,are experimented on several datasets.The experimental results show that the proposed strategy is superior to the existing methods in all indicators.