Research On Credit Fraud Detection Based On Machine Learning

Credit loan is an important form for the state to mobilize and allocate funds in a paid way,and it is a powerful lever for economic development.Credit business occupies an important position in banking business,and although the possibility of fraud is relatively rare,the impact of fraud can be very large,hindering the smooth operation of the country’s digital economy.The phenomenon of credit fraud has become a problem that cannot be ignored,so it is very important to study timely and accurate credit fraud detection methods.Credit fraud detection belongs to the category of anomaly detection.The key point of anomaly detection is that there are not enough abnormal samples in the available data set,that is,the data set is class imbalanced.With the wide application of complex data generation models based on deep neural networks,this paper proposes a two-stage anomaly detection scheme,which maximizes the effective information of the available data and detect hierarchically in prediction.The scheme combines a single classifier and a binary classifier.The single classifier based on the GANomaly can learn the feature distribution of non-fraudulent transactions,and iteratively screen out the high probability non-fraudulent samples in prediction to ease the difficulty of discrimination.In addition,this paper simulates the statistical distribution of fraudulent samples through the Conditional Unrolled GAN,and oversamples the minority class samples to balance the training set,so as to improve the discriminative performance of the binary classification model.This paper adopts the F1-Score to evaluate the prediction.The experiment compares the oversampling effect of the Conditional Unrolled GAN and the traditional oversampling algorithm,and compares the discriminative effect of the two-stage detection scheme proposed in this paper with other machine learning schemes.The empirical results confirm the feasibility and effectiveness of the oversampling technique based on the Conditional Unrolled GAN,determine the advantages of the detection scheme combining a single classifier and a binary classifier,and show that the GAN has the advantages of processing unbalanced data.