Research On Medical Insurance Anomaly Detection Based On Deep Learning

With the vigorous development of China's social medical insurance industry,medical insurance fraud has become increasingly rampant.As the forms of medical insurance fraud are complex and diverse,it is more difficult to identify medical insurance fraud.The work of anomaly detection of medical insurance is facing enormous challenges,which has a great impact on the national economy.Therefore,the use of modern information technology to effectively detect medical insurance fraud and ensure the safe use of medical insurance funds is the urgent task of current medical insurance risk prevention and control.This thesis mainly uses deep learning technology to apply the generative adversarial network and transfer learning algorithm to the medical insurance anomaly detection,and optimize its effect in anomaly detection.The main work of this thesis is as follows:(1)This thesis first preprocesses the medical insurance dataset,and improves the boundary ambiguity of the Smote method when generating minority-class samples.The improved Smote method measures the boundaries of minority-class samples and adopts different construction methods for samples with different boundaries,finally proposed the NF-Smote method.And verified in the experiment that the NF-Smote method has a positive impact on the final classification effect.(2)To solve the problem of imbalanced data in the classification of medical insurance dataset,this thesis improves the generator network of SSGAN to generate only minority-class samples in the training process,and modifies the output layer of discriminator network to a Softmax classifier which has there neurons,UbSSGAN algorithm is proposed.In order to reduce the number difference between positive and negative samples and alleviate the imbalance,the generated minority-class samples and the original samples are input into the discriminator network for training.By comparing with other three algorithms,experiments show that UbSSGAN has the best classification effect.(3)In order to alleviate the contradiction in TrAdaboost between the large demand of samples for auxiliary domain and the difficulty of obtaining real samples,the generator network in UbSSGAN is used to participate in the construction of auxiliary domain.At the same time,in order to further enhance the importance of minority-class samples in the classification process,this thesis optimizes the method of TrAdaboost on the adjustment of samples in the auxiliary domain and original domain: In auxiliary domain,improve the weights of incorrect-classified minority-class samples which are true samples and reduce the weights of incorrect-classified minority-class which are produced by generator.And increase the weights of incorrect-classified samples in original domain by different proportions.The adjustment of weighs and integration methods on the weak classifier in TrAdaboost are improved too,and the G-means value is used as the weight adjustment coefficient of the weak classifier.By transferring the discriminator network in UbSSGAN,the basic classification network is constructed which is needed in the boosting,and the Ub-TrAdaboost algorithm is proposed.Experiments show that the classification performance of Ub-TrAdaboost method on the medical insurance dataset has been significantly improved.