Research On Abnormal Transaction Detection Algorithm Based On Frequency Domain Features

With the rapid growth of Internet commerce,identifying abnormal transactions in massive transaction information and conducting risk control are the primary chal-lenges currently faced by financial inspection departments.As the amount of data has increased significantly,the workload of researchers has increased dramatically,which has triggered an upsurge in the application of machine learning methods to detect vul-nerabilities.Although machine learning methods have been widely used in image,lan-guage,and speech processing,there is still a long way to go in identifying abnormal transactions.In particular,the data imbalance characteristic of transaction samples is a major bottleneck hindering the optimization of classification method performance.Therefore,how to optimize the financial transaction data itself and improve the accu-racy of machine learning methods for detecting abnormal transactions,to better assist banks in making decisions,is a problem that needs to be solved urgently.This paper focuses on the problem of abnormal credit card transaction detection.On the one hand,it focuses on the use of frequency-domain features for sample ampli-fication,and on the other hand,based on the semi-supervised collaborative strategy to optimize the classifier and carry out related research.The specific content is as follows:(1)Aiming at the problem of how to effectively use the correlation characteristics of transactions for data augmentation,a method of obtaining frequency domain features based on FFT is proposed.Strengthen the application of frequency-domain features in data processing,and reduce over-fitting to improve accuracy.This method is based on the invariance of the information between the frequency domain and the canonical domain,and the characteristics of the frequency domain characteristics reflecting the sample relevance,and uses SFFT to transform between the frequency domain and the canonical domain.In frequency domain amplification,WGAN is used as an improved basis to generate counter-frequency domain samples from the noise,and finally,the data is transformed to the regular domain to complete the generation of amplified data.The experimental verification results show that this sample amplification method can gener-ate more diverse feature data and effectively reduce the degree of overfitting.Compared with the unamplified data,the accuracy is improved by 5.66%,and the degree of over-fitting is also reduced by 7.18%.Compared with the SMOTE method,the accuracy is improved by no less than 2.26%,and the degree of overfitting is reduced by 1.95%.(2)Aiming at the problem of how to reduce the cost of data labeling and improve the accuracy of abnormal data recognition,a multi-level partition model based on a semi-supervised collaboration strategy is designed and implemented.In this method,the unsupervised method of covariance estimation is selected as the first-level parti-tion model,and the experimental data is roughly divided.The improved decision tree algorithm based on migration learning is used as the secondary partition model to re-alize the partition model that reduces the cost of labeling and improves the accuracy of recognition.By adjusting the two custom model configuration parameters set:the first-level division parameter and the second-level migration parameter,the results on multiple sample sets show that the classification effect has been improved.The exper-imental verification results show that the improved method performs better than other algorithms on data sets such as RWC.Taking the RWC data set as an example,com-pared with the XGBoost algorithm,the AUC has an increase of 1.88%,the F1-score has an increase of 1.12%,and the overfitting ratio has also dropped from 0.1433 to 0.1071.In summary,this article conducts an in-depth study on the anomaly detection prob-lem of credit card transaction data,using frequency-domain features,adversarial net-works,and machine learning classification methods to improve the accuracy of anomaly detection while maintaining efficiency of the algorithm.