Based On Machine Learning And Deep Learning Research On Credit Card Anti-Fraud

In recent years,due to the upgrading of consumption,more and more people have started to consume their credit.With the development of technology,the credit industry has also developed rapidly,at the same time,The risk of fraud should not be underestimated.For fraud risk,the definition of fraud samples is a challenge,and the proportion of fraud samples in the overall population is very small.Due to the nonlinear,strong temporal,high noise,and imbalanced characteristics of fraud data,traditional machine learning techniques cannot effectively predict models.In view of this,this article aims to publicly available fraud datasets.Due to the large amount of dirty data in the dataset,feature cleaning and filtering are first performed on the dataset.In financial scenarios,the stability of the model is highly valued.Therefore,the core point of this article is how the model can have the same discrimination and stability in samples outside of time.Therefore,in feature filtering,in addition to traditional filtering based on IV,PSI,etc.We also attempted a feature selection method based on the consistency of badrate trends.In terms of algorithm selection,various single submodels such as traditional logistic regression algorithm(Logistic),boosting based tree model(Xgboost),and deep learning(MLP)were attempted,and the performance of different submodels was compared.Compared with different single machine learning models,they often have inherent flaws.For example,logistic regression models have strong interpretability but low model accuracy,while deep learning algorithms have strong fitting ability but are difficult to explain.Especially in financial risk control scenarios,we still need a certain degree of interpretability.So we chose an integrated model approach to build the model,combining different sub models.The combined model showed a 3%improvement in KS,AUC,and other indicators compared to a single model.