Research And Application Of Medical Insurance Fraud Detection System Based On Relational Network And Active Learning

In recent years,with the aging of society and the pressure of economic downturn,the income of medical insurance funds has gradually slowed down.At the same time,the national medical insurance system is constantly improving,the scope of medical insurance participation is gradually expanding,and medical demand has been released in a large amount,making the growth of medical insurance fund expenditures continuously higher than income growth.In many areas,there have been cases of unsatisfactory or even severe deficits.On the other hand,due to the imperfect medical insurance fund system and supervision system,it is easier for criminals to defraud medical insurance funds for cash,and various cases of medical insurance fraud emerge in an endless stream.How to effectively protect the safe and effective use of medical insurance funds has gradually become a hot topic in all sectors of society.Due to the sensitivity of medical data and the particularity of data samples,traditional medical insurance fraud research methods have been uncomfortable with the current situation.The main difficulties in research are as follows:(1)The methods of medical insurance fraud are endless,and in recent days,gangs have committed crimes,and the operation is extremely concealed.It is difficult to see the clues from the conventional data dimension;(2)Among the existing medical data,the patients who have been judged to be fraud are scarce and the data of unknown fraud or not is massive,naturally,there is a shortage of samples labels that are common in machine learning,the existing solutions are limited by the data itself,and it is difficult to improve the generalization ability of the model.(3)In the existing medical insurance fraud literature,most of the medical data used for medical treatment is used,and the important information about the social relationship generated in the medical treatment behavior is ignored.To solve the above problems,this paper proposes a medical insurance fraud detection solution based on relational network and active learning,which fully considers the potential information in the patient doctor network,and proposes the OCGVAE medical insurancefraud detection framework based on the graph convolutional neural network(GCN)algorithm.OCGVAE is a one-class classification detection algorithm,which is based on the patient doctor network.By using the information of the patient doctor network to make up for the imbalance of the sample,the medical insurance fraud detection under the small data training sample have achieved.In order to solve the problem of high cost of manual labelling,this paper proposes an active learning strategy to annotate the medical insurance data.The ideal performance can be achieved with fewer data samples.The main innovations and research work are summarized in the following sections:(1)A patient-doctor relationship network model was proposed.This paper analyzes existing fraud cases and finds that fraudulent patients can be directly or indirectly linked through the doctor,and the patient-doctor mathematical relationship is established to improve the classification effect of the model..Experiments have shown that higher classification accuracy can be achieved in the algorithm using the patient doctor network,which is on average 19% higher.(2)The GCN algorithm is proposed to deal with the problem of medical insurance fraud detection.This paper effectively and rationally utilizes the social relationship network generated by patients during the medical treatment process.The GCN algorithm uses this network information to learn the topology information between network nodes,even under the small data label.,also achieved the ideal accuracy.(3)This paper improved the decoding layer of the variational auto-encoding(VAE)algorithm,combined with the real data set of medical insurance fraud,and proposed the OCGVAE medical insurance fraud detection framework.The input of the OCGVAE algorithm is the fraud sample label and the entire patient doctor network(including all node information and weight information between the nodes).The algorithm is a one-class classification algorithm that learns the prior distribution of fraud samples.To predict the label of unlabelled samples,solving the problem of extreme imbalance of samples.At the same time,the model has link prediction ability,and its accuracy rate reaches 80%.(4)It is proposed to use active learning strategies to help medical insurance fraud data labeling and reduce labor costs.This paper designed three data selection labeling strategies,namely maximum entropy(MEs),maximum probability(MPs)and random(Rs)selection active learning strategies.In this paper,four sets of contrast experiments with a learning rate of 0.1,0.01 and a classifier threshold of 0.5 and 0.8 are designed.The MEs data labeling strategy proposed in this paper can achieve the best results in each group of parameters,which proves that the maximum entropy strategy can effectively reduce the cost of manual labeling.Accuracy rate reaches 97%.