Medical Insurance Fraud Detection Algorithm and System Based on Multi-Layer Attention Mechanism Graph Neural Network

Abstract

The present invention relates to a medical insurance fraud detection algorithm based on a multi-layer attention mechanism graph neural network, comprising the following steps: S1 establishing a medical insurance fraud detection AHIN model; S2 selecting semantic paths and finding neighbor nodes; S3 constructing a detection MHAMFD model based on a graph neural network; S4 obtaining data for the year to be measured or data in a test set, and inputting the data into the MHAMFD model to predict medical insurance fraudsters. By constructing a detection MHAMFD model based on a graph neural network through multi-layer attention extremes of nodes, paths, subgraphs, and different times, the present invention excavates rich interaction relationships among various entity nodes in the AHIN, solves the problem that some medical insurance fraud detection methods ignore the abnormal behavior characteristics of multiple visits, and at the same time reduces the influence of noise node and path data on the final prediction task, providing a theoretical basis for solving the problem of medical insurance fraud detection.

Description

Technical Field

[0001] The present invention relates to a medical insurance fraud detection algorithm, an algorithm and system for detecting medical insurance fraud using a neural network, and particularly to an algorithm and system for detecting medical insurance fraud based on a multi-layer attention mechanism graph neural network, belonging to the field of artificial intelligence detection. Background Art

[0002] With the popularization of medical insurance, while bringing great convenience to people's medical treatment, it also provides medical security for the social masses. However, we have also witnessed an increasing number of incidents of defrauding medical insurance funds, and a large amount of medical insurance funds are lost globally every year due to medical insurance fraud. Traditional medical insurance fraud detection methods are as follows: 1. Rule-based detection methods. 2. Outlier-based detection methods. 3. User statistical feature-based detection methods. For the first solution, it not only relies on certain prior knowledge, but also has a huge workload and low efficiency, and cannot guarantee the correct discovery of fraud behaviors. In addition, with the explosive growth of medical insurance data, the number of domain experts cannot meet the needs of screening existing fraud cases, and now the means of medical insurance fraud emerge in an endless stream, and it is difficult to handle complex and changeable patterns based on rules. For the second solution, the outlier-based detection method is mainly carried out in a fixed mode. With the development of the medical insurance system, medical insurance-related services have become more and more detailed, and at the same time, fraudsters have become more professional. Medical insurance fraud behaviors are complex and changeable and have concealment. More worthy of attention is that new fraud patterns will also emerge continuously. The anomaly detection algorithm for fixed patterns lacks immunity to new fraud patterns, and the method of discovering fraud behaviors from fixed patterns is difficult to meet the current needs. For the third solution, a large amount of labeled data is required, but in actual scenarios, the data is not marked, and there is only a small amount of fraud data (patient privacy protection).

[0003] In fact, fraudulent users may not only have abnormal characteristics, but also their behaviors in the interaction relationship are abnormal. For example, a medical insurance fraud user may have many drug transactions in multiple hospitals at the same time, which is difficult to utilize by traditional feature extraction methods. Traditional medical insurance fraud detection methods cannot make full use of the interaction relationship between users, and we attempt to use technologies in other fields to solve the above problems.

[0004] Heterogeneous graph representation learning is one of the effective methods for modeling the interaction relationship between such entities, and has currently been widely used in fields such as e-commerce recommendation systems, academic network analysis, and natural language processing. By learning graph-based representations, the sequence, topological structure, geometric and other relationship features of structured data can be captured. Existing graph neural network models are not designed for specific problems and are not particularly suitable for solving medical insurance fraud problems. Therefore, it is necessary to design an effective model for medical insurance fraud detection. Summary of the Invention

[0005] The present invention mainly aims at the problem of medical insurance fund fraud detection, and proposes a medical insurance fraud detection algorithm based on a multi-layer attention mechanism graph neural network. This method not only focuses on the feature attributes of users, but also considers the behavioral attributes of multiple medical visits during the medical process. Medical insurance fraudsters not only have unusual features, but also have unusual behaviors in these interactions.

[0006] Specifically, on the one hand, the medical insurance fraud detection algorithm based on a multi-layer attention mechanism graph neural network described in the present invention includes the following steps:

[0007] S1 Establish a medical insurance fraud detection AHIN (i.e., Attribute Heterogeneous Information Network) model;

[0008] S2 Select semantic paths and find neighbor nodes;

[0009] S3 Construct a detection MHAMFD model based on a graph neural network;

[0010] S4 Obtain the data of the year to be measured or the data in the test set, and input it into the MHAMFD model to predict medical insurance fraudsters.

[0011] In the medical insurance dataset, there are millions of transaction records from a large number of users. To better understand the behavior of patients, we construct it into a medical insurance AHIN. Therefore, step S1 specifically includes:

[0012] S1-1 Extract all the medical records of patients, and construct four entities: patients, hospital departments, dates, and drugs from them; among them, the entities with a drug unit price less than 20 yuan in the medical records are excluded. This is to avoid the graph being too dense. The date unit is days.

[0013] S1-2 Model different types of objects and their interactions in the real medical scenario into an AHIN, which is represented as a heterogeneous graph , where V is the set of different types of objects, that is, the node set (including the nodes of the four entities of patients, hospital departments, dates, and drugs), is the relationship set, X is the information matrix, and let the patient node set , each patient in the dataset has a label , indicating whether the patient belongs to a medical insurance fraudster. When not belonging, it is 0, and when belonging, it is 1. The dataset is divided into a training set , a validation set , and a test set finally used to predict the probability that a patient belongs to a medical insurance fraudster , and the ratio of the training set, validation set, and test set is 1 - 3:1:3 - 1.

[0014] Among them, in order to refine the geographical area in space, we regard hospitals and departments as a whole, which means that even departments with the same name in different hospitals will be regarded as different entities. Finally, we abstract the patient's medical record as the patient prescribing a certain drug from a certain department of a certain hospital on a certain day on the heterogeneous graph.

[0015] After processing the structured medical insurance data, a heterogeneous graph G is established. It is necessary to convert the data into a low-dimensional vector representation through a graph representation learning algorithm. We adopt a method based on semantic paths to extract the structural information and rich semantics in the network to process the data. Therefore, step S2 specifically includes:

[0016] S2-1 Define meta-paths and multi-semantic paths. Specifically, a meta-path is represented as (abbreviation ) form of a path, where describes the composite relationship between objects ; a multi-path is represented as (abbreviation ) form of a path, where describes the composite relationship between objects , represents the composition operator;

[0017] S2-2 Select appropriate neighbors based on meta-paths and multi-semantic paths, specifically including:

[0018] S2-2-1 Neighbor set based on meta-paths. Given a user u in the attribute heterogeneous information network, the neighbor based on the meta-path is defined as the aggregated neighbor set of user u in the AHIN under the given meta-path;

[0019] S2-2-2 Neighbor set based on multi-path sampling. Given a user u in the attribute heterogeneous information network, the neighbor based on multi-path sampling is defined as the aggregated neighbor set of user u in the AHIN under the given multi-path;

[0020] S2-2-3 Construct heterogeneous subgraphs, which are decomposed into subgraph structures of different degrees through meta-paths and multi-paths, denoted as . Let the number of meta-paths be N. After the meta-paths are merged to form multi-paths, the heterogeneous graph will be decomposed into N meta-path subgraphs at different levels. The same applies to the multi-path subgraphs.

[0021] To further construct the MHAMFD model, we first observed the real medical scenarios and medical data, analyzed the influence of neighbors based on meta-paths and multi-paths on the detection of medical insurance fraudsters based on real data, and then proposed to introduce a model based on a multi-attention mechanism for medical insurance fraud detection. Therefore, the specific step S3 includes:

[0022] S3-1 Node-level Aggregation;

[0023] We obtain the neighbor sets based on these paths through the methods of meta-path and multi-path sampling, and hope to more accurately represent the user's embedding by integrating the feature information of the neighbors. Here we note that the neighbors of each node based on the meta-path or multi-path play different roles and exhibit different importance levels when learning the node embeddings for specific tasks. For example, in the neighbor set of the path based on P-K-P, i.e., the patient-department-patient pattern, the similarity between the features of each neighbor node and the features of the target node varies, which has different impacts on the feature representation of the target node learned through the neighbor nodes. Therefore, we introduce a node-level attention mechanism here, which can learn the different importance levels of each node and aggregate the embedding representations of these meaningful neighbors together to form the embedding representation of the node for a specific path.

[0024] Specifically, we use the self-attention method to learn the importance of each neighbor node for the target node, that is, the corresponding weights.

[0025] Specifically, it includes: S3-1-1 Given the node pair connected by the path , the importance is defined as This indicates how important the neighbor node j is for the target node i. Based on the path of the node pair , the importance formula can be expressed as follows: where represent the embeddings of node i and node j respectively, represents the path connecting the two nodes. Here refers to the deep neural network that executes the node-level attention mechanism.

[0026] For the given sampled path (meta-path and multi-path), all node pairs based on the sampled path share . This is because under the same sampled path, all neighbors of the target node are sampled through the same connection pattern. After obtaining the importance of a neighbor node j in the neighborhood for the target node i, in order to facilitate the comparison of the importance coefficients of different neighbor nodes. Therefore, S3-1-2 normalizes the importance through the softmax function , where represents the node-level attention vector of the path , Denotes a splicing operation. Here, we use LeakyReLU as the activation function, is the number of all nodes connected to node i. As can be seen from the formula in, the weight coefficients depend on their features. And the weight coefficients are also asymmetric, which means that their degrees of contribution to each other are also different. This is because when the target node changes, its neighborhood will also change accordingly. They have different neighbors, so the normalization terms (denominators) will be very different. After obtaining the normalized importance of all neighbor nodes for target node i, step S3-1-3 aggregates the meta-path or multi-path based embedding of node i through the projected features of the neighbors and the corresponding coefficients as follows: , where represents the embedding representation learned by node i on the path , is the activation function, preferably is LeakyReLU.

[0027] The embeddings of the target nodes are all aggregated by the neighbors sampled from the path . Each attention weight coefficient is generated for a single specific meta-path or multi-path, so we consider it to be specific semantics, and in this way, a kind of semantic information can be captured.

[0028] Since the heterogeneous graph exhibits a scale-free property, the variance of the graph data is relatively large. To solve the high-variance problem, we extend the node-level attention to multi-head attention to make the training process more stable. Therefore, preferably, we repeat the node-level attention mechanism K times and splice the embeddings learned each time, In a specific subgraph structure, given the path set , after the features of the neighbor nodes are learned by the node-level attention mechanism, we can obtain m sets of path-specific node embedding representations, denoted as .

[0029] S3-2 Path-level aggregation. Generally, each node in the subgraph contains multiple types of semantic information. We have learned the node embedding representations on each path in the subgraph before. The node embedding of a specific single path can only reflect node information from one aspect. To learn a more comprehensive node embedding representation, it is necessary to combine the semantic information revealed by each path into each different subgraph and fuse this semantic information. In a specific task, each path has a different impact on the node embedding. In the medical insurance heterogeneous graph we constructed, the significance of the path sampling connecting the target node is different. The embeddings learned from the set of patients who visited the same department as the target patient and the set of patients who had visit records on the same day obtained through path sampling have different degrees of importance for us to judge whether the target patient has fraudulent behavior. Therefore, in this part, we need to learn the influence of different paths on our final task. And the attention coefficient for the path can reflect the influence of different paths on the node embedding. We propose a path-level attention mechanism to automatically learn the importance of different paths. Therefore, S3-2-1 aggregates the node embeddings of all paths in the subgraph to generate a subgraph-specific embedding representation, which is expressed as follows: where represents the embedding of a specific subgraph, represents the set of paths in a specific subgraph, is the path for a specific task 's contribution degree, is the path for a specific task 's embedding representation. To understand the influence degree of different paths in the subgraph on the final task, we use the attention mechanism to assign different weights to different paths. We first measure the importance of the path-specific node embedding by the similarity between the transformed path embedding and the path-level attention vector , as follows: where represents the importance of the path, represents the path-level attention vector, is the weight matrix, is the bias vector. To make a meaningful comparison of the importance of paths, all the above parameters are shared for all paths and the embeddings of specific paths in the subgraph. After obtaining the importance of the embedding of each path in the subgraph, we perform a normalization process through the Softmax function, and then can be obtained through the following formula: S3-2-2 uses the learned weight of each path in the subgraph as a coefficient to perform a weighted sum of all path-specific embeddings to obtain a specific subgraph embedding . For a given set of node embeddings of specific paths in the subgraph as input, the contribution degree of each path can be expressed as: S3-2-3 Aggregate the embeddings of the specified paths within each subgraph individually and generate the vector embeddings specified for n subgraphs, denoted as . Note that in the aggregation of the attention mechanism at the path level, the processes of each subgraph are relatively independent. In S3-3, subgraph-level aggregation is performed to obtain the final node embedding Z. The node embeddings of a specific single path on the subgraph can only reflect node information from one aspect. As we can see in the previous section, the contribution degrees of the embeddings of the same node on each path to the final task are different. Generally speaking, when the same node is located in different network structures, it may also play different roles for the target node. Existing research indicates that different subgraphs also have different degrees of influence on the final task. Here, in order to learn the importance of different subgraphs, a subgraph-level attention mechanism is proposed. Similar to the aggregation process of the path-level attention mechanism, given a specific embedding set of the subgraph and adding the importance degree of different subgraphs for the final task, it is expressed as follows: (11), thus obtaining the final node embedding . In S3-4, we input the final node embedding Z into a multi-layer perceptron (MLP) for fraud detection and apply it to different downstream tasks. We use cross-entropy as the loss function and optimize the model weights by minimizing the function through backpropagation. The cross-entropy is expressed as:

[0030] where is the set of node indices with labels, is the embedding of the label node and the corresponding label, are the parameters of the classifier. In this graph neural network model, we use these frameworks to learn the weights and aggregate information layer by layer to obtain meaningful node embeddings. Finally, the node embeddings are used to predict whether the user is a medical insurance fraudster.

[0031] In step S4, the fraudulent medical transaction information includes all actual transaction records of the medical transactions of the detected fraudsters. On the other hand, the present invention provides a system for implementing the above detection algorithm, including at least one client, a server, and a medical institution transaction processing device, characterized in that the server is composed of multiple sub-servers that independently serve different medical institutions, and the medical institution transaction processing device includes at least one sub-transaction processing device disposed in different departments of different medical institutions. There is communication between the at least one client and the server, and between the sub-server and at least one sub-transaction processing device. Among them, the sub-server is used to receive and save the historical records of the transaction requests of the clients, and complete the medical insurance fraud detection based on the graph neural network with a multi-layer attention mechanism through the detection instructions of the sub-transaction processing device, and return the detection results to the sub-transaction processing device.

[0032] The third aspect of the present invention is to provide a non - transitory storage medium, which stores a computer - readable program that can be run by the server to implement the above - mentioned detection algorithm.

[0033] Beneficial effects brought by the technical solution of the present invention

[0034] (1) Step 1 models the medical insurance fraud detection problem as a classification problem in AHIN, providing a theoretical basis for solving the medical insurance fraud detection problem.

[0035] (2) Step 2 defines a semantic path to explore the structure of the medical insurance AHIN, mines the rich interaction relationships of each entity node in the AHIN, and solves the problem that some medical insurance fraud detection methods ignore the abnormal behavior characteristics of multiple visits.

[0036] (3) Step 3 uses a multi - layer attention mechanism to aggregate the information of neighbor nodes and the structural information of the network, reducing the influence of noise nodes and paths on the final prediction task. The finally aggregated embedding is used to predict whether a user is a medical insurance fraudster. Brief description of the drawings

[0037] Schematic diagram of the definition of meta - path (a) and multi - path (b) in FIG. 1,

[0038] Figure 2 Schematic diagram of the MHAMFD model structure,

[0039] Figure 3 Comparison chart of the percentage increase in the fraud rate of users with different numbers of medical insurance fraudster neighbors on different meta - paths PHP (left) and PtP (right),

[0040] Figure 4 The system for implementing the detection algorithm of Embodiment 2 in Embodiment 3 of the present invention. Detailed implementation manners

[0041] Embodiment 1

[0042] This embodiment explains the principle of the present invention. A medical insurance fraud detection algorithm based on a multi - layer attention mechanism graph neural network has a basic principle of using AHIN to model the real medical insurance treatment scenario, modeling the medical insurance fraud detection problem as a classification problem in AHIN, and capturing the interaction information in AHIN through Meta - path and Multi - path. Secondly, these interaction information are aggregated and learned through a multi - layer attention mechanism. Finally, the learned embedding representation information is input into a multi - layer perceptron (MLP) for fraud detection, which is the complete content of the MHAMFD model proposed by the present invention.

[0043] Among them, in step S1, after the structured medical insurance data is processed, a heterogeneous graph G is established. It is necessary to convert the data into a low-dimensional vector representation through a graph representation learning algorithm. We adopt a semantic path-based method to extract the structural information and rich semantics in the network to process the data.

[0044] In S2-1, according to the definitions of meta-paths and multi-semantic paths, this embodiment describes the meanings of meta-paths and multi-semantic paths in the medical insurance AHIN. As shown in Figure 1(a), we construct an AHIN to model the medical treatment scenarios of medical insurance fraud. The medical insurance attribute heterogeneous network contains various types of objects (i.e., patients (P), hospital departments (K), drugs (M), dates (T)), with rich attributes and relationships. In the AHIN, two users can be connected by multiple meta-paths. For example, patient - department - patient (P-K-P), patient - drug - patient (P-M-P), patient - date - patient (P-T-P). Different meta-paths always express different semantics. For example, the path P-K-P indicates that the two patients connected by this meta-path have seen a doctor in the same department. And the path P-T-P indicates that the two patients connected by this meta-path have a medical record on the same day. Two users can be connected by multiple multi-semantic paths. For example, in Figure 1(b), the path P-(KT)-P indicates that the two patients connected by this path have seen a doctor in the same department on the same day. The representation connected to the target patient through the path P-(KTM)-P has taken the same drug in the same department as the target patient on the same day.

[0045] The key to learning the representation of the target node in the heterogeneous graph G lies in how to accurately propagate and aggregate the information of the neighbors. We select appropriate neighbors based on the semantic path according to the above definitions of meta-paths and multi-semantic paths.

[0046] In S2-2, we illustrate the selected neighbor nodes in the medical insurance AHIN based on the neighbor set based on meta-paths and the neighbor set based on multi-path sampling. For example, given the meta-path P-K-P, the neighbors of patient "Zhang San" are "Li Si" and "Wang Er". Similarly, the neighbor of "Zhang San" based on the multi-path P-(KTM)-P is only "Li Si". Obviously, both the neighbors based on meta-paths and the neighbors based on multi-paths can utilize the structural information in different aspects of the AHIN. After selecting the appropriate nodes, we need to aggregate and propagate the feature information of these neighbor nodes to the target node to learn the final embedding representation of the target node.

[0047] In S3, we first observe the real medical treatment scenarios and medical data, analyze the impact of neighbors based on meta-paths and multi-paths on the detection of medical insurance fraudsters based on real data, and then propose to introduce a model based on a multi-attention mechanism for medical insurance fraud detection. We Figure 2The overall structure of the model is shown. First, we aggregate the neighbors of each user based on different meta-paths and multi-paths to integrate the structural information of multiple aspects in the AHIN for better representation learning. On this basis, we distinguish the differences between meta-paths and multi-paths through semantic attention to obtain the optimal weighted combination of specific-task and specific-semantic node embeddings. Finally, considering the importance of different heterogeneous subgraphs for the ultimate goal, we also adopt an attention mechanism at this layer of heterogeneous subgraphs to learn the preferences of different subgraphs.

[0048] From the observation of real medical insurance fraud events and datasets, it can be seen that medical insurance fraudsters usually tend to commit crimes in teams, and these people are closely aggregated together through different types of interactions. Medical insurance fraudsters are more likely to seek medical treatment in the same hospital or leave medical records during the same time period in a team-like manner. To verify the aggregation of medical insurance fraudsters under different interaction relationships, we conduct experiments on real datasets. First, we collect the meta-path-based neighbors of each patient based on two meta-paths, P-K-P and P-T-P (the neighbors based on the path P-K-P represent patients who have seen a doctor in the same department, and the neighbors based on the path P-KT-P represent patients who have seen a doctor in the same department on the same day). We count the number of fraudsters. And we divide them into multiple groups according to different paths and calculate the proportion of fraudsters in each group. As Figure 3 , the observation results show that different meta-path-based neighbors have different degrees of influence on patients. It can be seen that different semantic paths have different importance for patients, so we adopt an attention mechanism to capture these different importances.

[0049] Example 2

[0050] This example shows the test results without considering the period effect. We used the real dataset of a certain city's medical insurance bureau in 2018. The fraud samples in the Medical-1 dataset are abnormal patients discovered through methods such as kidney disease abnormalities, repeated prescribing, prescribing Alzheimer's drugs for people born in the 1980s, and being hospitalized while seeing an outpatient. Different from this, Medical-1 is a balanced sample, and the ratio of positive and negative samples is 1:2. The specific information is shown in Table 1. The classification effect of the balanced sample nodes is shown in Table 2. The anomaly detection effect of the balanced sample nodes is shown in Table 3.

[0051] The experimental results show that on the real dataset, whether it is F1 or ACC, the MHAMFD model we used shows the best performance under different training set allocations. This indicates that the MHAMFD model can better learn the semantic information between real medical insurance data and use it for fraud detection. First, heterogeneous graph embedding methods based on graph structure, such as Metapath2vec, contain network structure information to a certain extent but ignore the feature information of nodes, so the performance is relatively average. In addition, GNN-based models combine structure information and feature content and perform well in this experiment. GAT, HAN, and MHAMFD also introduce an attention mechanism to evaluate the importance of objects in the graph to improve the performance of the embedded representation. In addition, the HAN model ignores the impact of the composite semantic information of paths and subgraphs in heterogeneous graphs on the embedded representation. Therefore, MHAMFD uses metapaths and multi-paths to capture more complex semantic information in heterogeneous graphs, aggregate more complex neighborhood information, and consider the impact of different levels of subgraph structures on node embeddings. MHAMFD uses metapath and multi-path methods to decompose heterogeneous graphs into multiple subgraphs at different levels, considers the composite semantic relationships brought by the intertwining of metapaths, and improves the quality of the neighborhood of target nodes.

[0052] Table 1: Datasets used in the experiment

[0053]

[0054] Table 2: Classification effect of balanced sample nodes

[0055]

[0056] Table 3: Anomaly detection effect of balanced sample nodes

[0057]

[0058] Example 4

[0059] This example provides a system for implementing the detection algorithm of Example 2, as Figure 4 shown, including at least one client, a server, and hospital transaction processing devices. It is characterized in that the server is composed of multiple sub-servers that independently serve different hospitals (a total of mun), and the hospital transaction processing devices include at least one sub-transaction processing device in different departments (a total of mun') in different hospitals. There is communication between the at least one client and the server, and between the sub-servers and at least one sub-transaction processing device. Among them, the sub-server is used to save the historical records of transaction requests received from clients and complete the medical insurance fraud detection based on the multi-layer attention mechanism graph neural network through the detection instructions of the sub-transaction processing device, and return the detection results to the sub-transaction processing device.

Claims

1. A medical insurance fraud detection method based on a multi-layer attention mechanism graph neural network, comprising the following steps: S1 Establish a medical insurance fraud detection AHIN model; S2 Select semantic paths and find neighbor nodes; S3 Construct a detection MHAMFD model based on a graph neural network; S4 Obtain data for the year to be measured or data in the test set, and input it into the MHAMFD model to predict medical insurance fraudsters; Step S1 specifically includes: S1-1 Extract all medical records of patients, and construct four entities: patients, hospital departments, dates, and drugs from them; among them, entities with a drug unit price less than 20 yuan in the medical records are excluded, and the date unit is days; S1-2 models different types of objects and their interactions in a real medical scenario into an AHIN, represented as a heterogeneous graph , where V is a set of different types of objects, that is, a node set of nodes including four entities: patients, hospital departments, dates, and drugs is a set of relationships, and X is an information matrix Set the patient node set , for each patient in the dataset has a label , indicating whether the patient belongs to a medical insurance fraudster. When not, it is 0; when belonging, it is 1. Divide the dataset into a training set , a validation set , and a test set finally used to predict the probability of whether a patient belongs to a medical insurance fraudster , the ratio of the training set, validation set, and test set is 1 - 3:1:3 - 1; Step S2 specifically includes: S2-1 Define meta-paths and multiple semantic paths; S2-2 Select appropriate neighbors based on meta-paths and multiple semantic paths, where S2-1 includes meta-paths represented as paths of the form, where describes the composite relationship between objects ; multiple paths are represented as paths of the form, where describes the composite relationship between objects ; represents the composition operator. S2-2 includes: S2-2-1 Neighbor set based on meta-paths. Given a user u in an attribute heterogeneous information network, the neighbor based on the meta-path is defined as the aggregated neighbor set of user u in the AHIN under the given meta-path; S2-2-2 Neighbor set based on multi-path sampling. Given a user u in an attribute heterogeneous information network, the neighbor based on multi-path sampling is defined as the aggregated neighbor set of user u in the AHIN under the given multi-path; S2-2-3 Construct heterogeneous subgraphs, which are decomposed into multiple subgraph structures of different degrees through meta-paths and multiple paths, denoted as ; Step S3 includes: S3-1 Node-level aggregation; S3-2 Path-level aggregation; S3-3 Subgraph-level aggregation to obtain the final node embedding Z; S3-4 Input the final embedding Z into a multi-layer perceptron (MLP) for fraud detection, apply it to different downstream tasks, use cross-entropy as the loss function (loss), and optimize the model weights through the backpropagation minimization function, where S3-1 includes: S3-1-1 Given the node pairs connected by a path , define the importance as , and based on the node pairs of the path , the importance formula is expressed as follows: of the path ​ ， Among them , represent the embeddings of node i and node j respectively, represents the path connecting the two nodes. Here refers to the deep neural network that executes the node-level attention mechanism. For a given sampling path including meta-paths and multiple paths , all node pairs based on the sampling path are shared ; S3-1-2 normalizes the importance through the softmax function: ​ （2）， Among them represents the node-level attention vector of the path , denotes the concatenation operation, and LeakyReLU is the activation function is the number of all nodes connected to node i; S3-1-3 Aggregate the meta-path or multi-path-based embeddings of node i through the projection features of neighbors and corresponding coefficients, as follows: ， Among them represents the embedding representation learned by node i on the path and is the activation function S3-2 includes: S3-2-1 Aggregate the node embeddings of all paths in the subgraph to generate a subgraph-specific embedding representation, expressed as follows: ， Among them represents the embedding of a specific sub - graph, represents the set of paths in the specific sub - graph, is the contribution of the path for a specific task ; By measuring the similarity between the transformed path embedding and the path - level attention vector the path of the specific task ; the importance of the path - specific node embedding is measured as follows: ​ ， where represents the importance of the path, represents the attention vector at the path level, is the weight matrix, is the bias vector; after normalization by the softmax function, is obtained by the following formula: ； S3-2-2 uses the weights learned for each path in the subgraph as coefficients to perform a weighted sum of all path-specific embeddings to obtain a specific subgraph embedding , for a given set of node embeddings of specific paths in the subgraph as input, the contribution of each path is expressed as: ； S3-2-3 Aggregate the embeddings of the specified paths within each subgraph separately and generate the vector embeddings specified for the n subgraphs, denoted as ; S3-3 includes a given set of subgraph-specific embeddings plus the importance degree of different subgraphs for the final task, expressed as follows: （9） （10） （11）， where represents the importance degree of the sub - graph, represents the attention vector at the path level, is the weight matrix, is the bias vector, P is the total number of all sub - graphs in the sub - graph set. Finally, fuse the information contained in each specific sub - graph and aggregate the node embedding representations of the specific sub - graphs to obtain the final node embedding (12); The cross-entropy in S3-4 is expressed as: , where is the set of node indices with labels, and are the embeddings of the label nodes and the corresponding labels, are the parameters of the classifier.

2. The method according to claim 1, wherein In S3-1-3, the node-level attention mechanism is repeated K times, and the embeddings learned each time are concatenated to obtain ， In a specific subgraph structure, a given path set , after the features of neighbor nodes are learned through the node-level attention mechanism, m sets of path-specific node embedding representations are obtained, denoted as , LeakyReLU.

3. A system for implementing the medical insurance fraud detection method based on the multi-layer attention mechanism graph neural network according to any one of claims 1-2, characterized in that, It includes at least one client, a server, and a medical institution transaction processing device. It is characterized in that the server is composed of multiple sub-servers that independently serve different medical institutions. The medical institution transaction processing device includes at least one sub-transaction processing device set in different departments of different medical institutions. Communication occurs between the at least one client and the server, and between the sub-server and at least one sub-transaction processing device. Among them, the sub-server is used to save the historical records of transaction requests received from clients, and complete the medical insurance fraud detection based on the multi-layer attention mechanism graph neural network through the detection instructions of the sub-transaction processing device, and return the detection results to the sub-transaction processing device.

4. A non-transitory storage medium, characterized in that, It stores a computer-readable program that can be run by the server to implement the medical insurance fraud detection method based on the multi-layer attention mechanism graph neural network as described in any one of claims 1-2.

Images