Research On Key Issues Of Heterogeneous Information Network Based Healthcare Insurance Anti-fraud Detection

With the development of the Internet and big data,more and more enterprises and government agencies use electronic information systems to conduct trading activities or provide services such as credit card business,healthcare insurance,and automobile insurance.Informatizat,ion also enables fraudsters to have new ways of fraud while facilitating legitimate users.Hundreds of billions of dollars in economic losses per year due to fraud worldwide.Fraud detection in areas such as credit,card and property insurance has been extensively studied.However,due to its unique characteristics,fraud detection in some areas such as healthcare insurance fraud faces more challenges and difficulties.Healthcare insurance data refers to data obtained during the healthcare in-surance business process,including healthcare insurance institutions,individuals,admissions and funds.It has the following characteristics:1)The domain knowledge is coarser in granularity.2)Discreteness.3)Redundancy.In addition to the particularity of healthcare insurance data,the types of healthcare insurance fraud are more abundant,and they are also more compli-cated than fraud in other application areas.1)There are many types of healthcare insurance fraud entities.2)Some smart fraudsters know enough domain knowledge and they can clev-erly evade anti-fraud rules by camouflage.3)When a fraudster organizes joint fraud,analyzing the individual’s behavior can not reveal the fraud record.4)Healthcare insurance covers different types of subjects,such as hospitals,patients,diagnosis and treatment,programs,etc.There are complex correlation relationships among different types of entities.Most of the data-driven research in healthcare fraud is focused on statisti-cal analysis and the use of machine learning algorithms like clustering,k-nearest neighbor,decision trees,neural networks,etc.However,these methods always have high false positive rate because normal patients may have some behaviors that violate behavior patterns while fraudsters may try their best to add normal behaviors so that they look "normal".An effective fraud detection method for healthcare insurance requires inter-pretability and high accuracy.In response to the above challenges faced by anti-fraud,this paper adopts heterogeneous information network to model healthcare insurance data,then discusses and studies from the perspectives of community division/maximum clique mining/frequent subgraph mining of heterogeneous in-formation network.This paper proposes anti-fraud algorithms for different types of fraud in healthcare insurance.The main research contents and contributions include the following aspects:1)This study proposes patient,cluster divergence based healthcare insurance fraudster detection method to address the issue of camouflage of drug trafficking fraudsters.The method combines the patient’s time behavior with the hetero-geneous network community detection algorithm to counter the camouflage of the fraudster behavior.This study defines the patient admission records simi-larity calculation and clusters the records.Then performs semantic extraction on each cluster category which can help to understand the meaning behind each cluster category.When there is a large conflict between the similarity of the patient and the similarity of the patient’s healthcare insurance treatmentbehav-ior,the probability of the patient being a suspected fraudster is higher.Patient Cluster Divergence based Healthcare Insurance Fraudster Detection-PCDHIFD considers the admission records of each patient throughout the period.Since the camouflage behavior of the fraudsters usually only lasts for a short period of time,the method can detect the healthcare insurance fraudsters from the interference of the fraudsters’ camouflage.The experimental results show that PCDHIFD can significantly improve the detection accuracy of fraudsters to 87%in the presence of camouflage,and the performance in f-measure is better than existing algorithm by more than 15%.2)Aiming at the problem of collusive fraud,this paper proposes abnormal group based joint medical fraud detection method,which reduces the compu-tational complexity through two-stage H-map based maximal sub-graph mining algorithm,which can help detect joint fraud.This study classifies the mined abnormal groups and distinguish whether the ab normal groups are caused by conspiracy fraud or periodicity.Eventually,the abnormal groups resulting from joint fraud are handed over to the manual inspection as suspected fraudsters.In the context of healthcare insurance,in order to solve the problems of joint fraud for illegal cash,this study construct person similarity adjacency graph and mine abnormal groups through the maximal clique mining algorithm.Further.the abnormal groups are classified(joint fraud/periodically generated).Patients included in abnormal groups due to joint fraud are considered as suspected fraud-sters.Abnormal Group based Joint Fraud Detection(AGJFD)method can dis-tinguish suspicious joint fraudsters from those who happen to be highly similar by periodicity,thus ensuring high precision of anti-fraud results.In addition,this study reduces the computational complexity by proposing a two-stage H-map-based maximal clique mining algorithm.A large number of experiments on the healthcare insurance data set show that the method of this study can obtain 94%in terms of accuracy.3)To solve the problem of false reporting of chronic diseases,this paper pro-poses frequent graph mining and community detection based chronic disease fraud detection method,which redefines the candidate set selection scheme and can help to understand the rare diseases.The progression of chronic diseases is very useful for discovering chronic disease fraud and reducing medical costs.Frequent graph mining and community detection based Chronic Disease Progression Mining-HNCDPM,considers different treatment phases of the same disease and obtains two modes-the temporal patterns of different chronic diseases(indicating the temporal relationship between different kinds of chronic diseases)and different treatment stages of the same chronic diseases(indicating different treatment op-tions for different stages of chronic disease).These two modes can be used to help detect chronic disease fraud.HNCDPM is able to exploit the clinical pathways of rare diseases such as leukemia,which is impossible for traditional subgraph min-ing methods because rare disease nodes will be removed in the process of finding frequent candidate sets.At the same time.HNCDPM considers different phases of treatment patterns for the same disease,which makes more sense for under-standing the progression of chronic disease.Existing disease progression mining method considers the same diagnosis as the same disease and ignores the different treatment phases of the same disease.Experiments conducted on real healthcare insurance dataset show that the method of this studycan exceed existing method by about 20%in F-measure.4)To solve the problem of suspicious patient fraudster detection,this study proposes an admission graph based healthcare insurance patient fraudster detec-tion method.The method proposes the concept of admission graph and captured the relationship between patients,admission and hospitals.Because of the com-plexity and continuity of healthcare.it is difficult to identify fraud by a single admission record.Existing fraud detection methods often consider the behavior of entities and treat the fraudster as an exception,ignoring complex relationships between different entities.This study uses the correlation between the credibil-ity of patients,the reality of admission and the authority of hospitals to detect patient fraudsters.The method shows how information in the graph indicates the cause of the fraudster and reveals important clues about the different types of fraudsters.Experimental results show that this method can identify potential patient fraudsters more accurately than existing methods by about 10%.