| Anomaly detection is the identification of observation results that are significantly different from most data and cause concern.It is very common in software defects,natural disasters,gene expression,fraud detection and so on.The resulting data class imbalance is a common problem in real-life data sets,widely distributed in various research fields,and is also one of the hot research contents in the current frontier.Therefore,this paper selects the fraud claims data set and uses sampling technology to conduct in-depth research on the class imbalance problem,which has important application value.This paper firstly summarizes the adopted model and method theory,emphasizes the misunderstanding that sampling technology cannot be used before cross-validation,analyzes the evaluation indicators,according to the characteristics of the fraud claims dataset,selects the most suitable two evaluation indicators of recall rate and precision rate,and introduced the principle of generative adversarial network and the construction of WGANRF model.Then the datasets are processed,and the datasets have different imbalance(5.8% and 24.7% imbalance respectively),which can detect whether the imbalance of the data sets will affect the proportion of sampling method settings and the performance of different models.According to the characteristics of many classification variables in the two datasets,the methods of one hot coding and target coding are adopted for conversion,the Synthetic Minority Over-Sampling Technique method(SMOTE)and hybrid sampling method are selected,and adjust the ratio of SMOTE and random undersampling.Finally,the fraud identification models based on Naive Bayes,logistic regression,random forest,and gradient boosting methods is constructed respectively,and the prediction results are analyzed by using the evaluation indicators of the classification model,the research shows that compared with the original model,the use of sampling technology,the performance of the model will increase,and the SMOTE ratio increases,the better the effect of the model,the model using hybrid sampling is the best.To solve the problem that the precision of the dataset with high imbalance(5.8%)is not high,the new sampling method WGAN is integrated into the random forest model,which further improves the precision.In summary,the paper uses sampling technology to explore datasets with different degrees of imbalance,which provides ideas for future research. |
