Insurance fraud detection method based on heterogeneous graph neural network

Abstract

The invention discloses an insurance fraud detection method based on a heterogeneous graph neural network, and the method comprises the steps: carrying out the preprocessing of medical insurance reimbursement original data, and constructing a heterogeneous graph structure model in the original data; carrying out attribute mapping according to attribute information of nodes in the heterogeneous graph structure model, and converting the attribute information into feature vector spaces of the nodes; learning the heterogeneous graph structure model by adopting a cross-relation-based multi-head attention system to obtain a global embedded vector; and detecting and sensing the nodes of the medical insurance reimbursement item by using the global embedded vector to obtain a detection result. Node and relation weights are learned through a heterogeneous graph neural network model, and information is gradually aggregated among layers of the model, so that not only is deep fusion of node embedding and relation embedding promoted, but also understanding of the model on complex interaction between nodes and relations is enhanced to more accurately capture key features of reimbursement items, and the reimbursement efficiency is improved. According to the invention, which reimbursement items may have illegal behaviors can be identified, so that the accuracy and efficiency of illegal detection are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an insurance fraud detection method and computing device based on a heterogeneous graph neural network. Background Art

[0003] Traditional medical insurance fraud detection methods include: 1. Rule-based methods; 2. Statistics-based methods; 3. Clustering-based methods.

[0004] Among them, the rule-based approach relies on the knowledge of medical experts and historical data to develop a series of rules. If a claim or medical record violates these rules, it may be marked as fraudulent. This method is simple and direct, but it relies too much on expert experience and is difficult to adapt to new fraud patterns.

[0005] Statistical methods use statistical models to analyze data and identify fraud by detecting anomalies in the data. However, for large-scale medical insurance reimbursement data, statistical methods require a lot of computing resources and are very inefficient.

[0006] Clustering-based anomaly detection methods assume that normal data points tend to cluster together, while anomalies do not belong to these clusters or belong to smaller clusters. However, medical insurance fraud data is often class-unbalanced, and this method may produce high false positive or false negative rates when processing such data.

[0007] In reality, medical insurance fraud is generally divided into two types: the first is that the hospital's charging standards are inconsistent with relevant government documents, resulting in illegal charges such as over-standard charges and repeated charges. The second is that patients use their own or other people's medical insurance cards to falsely hospitalize in the hospital and fabricate medical service items in order to siphon medical insurance funds.

[0008] These two types of medical insurance fraud are essentially interactions between patients and hospitals. Traditional medical insurance fraud detection methods cannot fully utilize this interactive relationship to discover and solve the above problems. Summary of the invention

[0009] One of the purposes of the embodiments of the present invention is to provide an insurance fraud detection method and device based on a heterogeneous graph neural network. In response to the problem of medical insurance fraud, a heterogeneous graph neural network fraud detection method that performs relationship perception through a multi-layer attention mechanism is proposed.

[0010] In order to solve the above technical problems, in a first aspect, an embodiment of the present invention provides an insurance fraud detection method based on a heterogeneous graph neural network, the method comprising:

[0011] Preprocessing the original medical insurance reimbursement data to construct a heterogeneous graph structure model in the original data;

[0012] Performing attribute mapping according to the attribute information of the nodes in the heterogeneous graph structure model and converting it into a feature vector space of the nodes;

[0013] The heterogeneous graph structure model is learned by adopting a multi-head attention system based on cross-relationships to obtain a global embedding vector;

[0014] The global embedding vector is used to detect and perceive the nodes of the medical insurance reimbursement items to obtain detection results.

[0015] Preferably, the information set of the heterogeneous graph structure model is:

[0016] G = {V, ε, X};

[0017] Wherein, V is a set of nodes of different types in a heterogeneous graph, and the node includes one of the four entities of reimbursement record, department, patient and time;

[0018] ε is a set of relationship types, including reimbursement item-patient, reimbursement item-time, and reimbursement item-department;

[0019] X is a feature matrix, which is the attribute features of the nodes in the heterogeneous graph;

[0020] Reimbursement Item Collection For each reimbursement item u∈U, there is a label y∈{0,1}, which indicates whether the reimbursement item is a violation item. 0 indicates that it is a violation item, and 1 indicates that it is not a violation item.

[0021] Preferably, attribute mapping is performed according to the attribute information of the nodes in the heterogeneous graph structure model, and the conversion into the feature vector space of the nodes specifically includes:

[0022] The attribute information of the reimbursement project node is mapped into a vector with the same dimension through word2vec, and the output is:

[0023] f embedding =CBOW(x);

[0024] Among them, CBOW is the continuous bag of words model, which is one of the methods used by word2vec;

[0025] The time data of the node is mapped to the feature vector space through time2Vec, and the mathematical expression of time2Vec is:

[0026] f time (t) = [w 0 t+b 0,sin(w 1 t+b 1 ),…,sin(w k t+b k )];

[0027] Among them, t is the original time series, w 0 t+b 0 is the time linear part, sin(w i t+b i ) is the periodic part of time, w and b are a set of learning parameters.

[0028] Preferably, the adopting of a cross-relation-based multi-head attention system to learn the heterogeneous graph structure model to obtain a global embedding vector specifically includes:

[0029] According to the association mode between different nodes, the heterogeneous graph is divided into subgraphs according to the meta-path, and the meta-path includes: φ = {reimbursement item-time-reimbursement item, reimbursement item-patient-reimbursement item, reimbursement item-department-reimbursement item, patient-reimbursement item-patient};

[0030] Through the cross-relation multi-head attention mechanism, multi-head attention calculation is performed on the interactions of nodes in the heterogeneous graph subgraph to obtain the global embedding vector.

[0031] Preferably, the multi-head attention calculation is performed on the interaction of nodes in the heterogeneous graph subgraph through the cross-relation multi-head attention mechanism to obtain the global embedding vector specifically including:

[0032] Automatically adjust the weight coefficient of each node in each different meta-path to obtain the embedded representation of each node in the heterogeneous graph subgraph;

[0033] The embedding representations of nodes on different heterogeneous graph subgraphs and the relationship embeddings between different meta-paths are fused to obtain a global embedding vector.

[0034] Preferably, the automatic adjustment of the weight coefficient of each node in each different meta-path to obtain the embedded representation of each node in the heterogeneous graph subgraph specifically includes:

[0035] A self-attention mechanism is introduced when each meta-path is passed to calculate the importance of a node v to its neighbor node u. The importance is expressed as:

[0036]

[0037] Among them, f is the feature vector of the node, ‖ represents the concatenation of two sets of vectors, σ is the activation function, represents the attention vector between nodes;

[0038] For the metapath φi The message passing in the heterogeneous graph is calculated by the following method to obtain the embedding of each node in the heterogeneous graph subgraph:

[0039]

[0040] Preferably, the step of fusing the embedding representations of nodes on different heterogeneous graph subgraphs and the embedding of relations between different meta-paths to obtain a global embedding vector specifically includes:

[0041] For the metapath φ i , its embedding representation is calculated as follows:

[0042]

[0043] in, is the relational semantic embedding at level l-1, and at level 0 is f is the embedding representation of the reimbursement item, and is a learnable parameter of linear propagation;

[0044] The obtained node embedding representation of the meta-path With relational embedding Through weighted aggregation of multiple independent attention heads, where N is the number of nodes on the meta-path, i = 1, 2, 3, ..., M, and M is the number of relations, a global embedding vector is obtained.

[0045] Preferably, the node embedding representation of the obtained meta-path is With relational embedding The weighted aggregation through multiple independent attention heads specifically includes:

[0046] For each independent attention head, first embed the meta-path Perform linear transformation to obtain three matrices Q, K and V:

[0047]

[0048] in, are three trainable parameter matrices, d k is the projection dimension;

[0049] Relation embedding for different meta-paths Perform a linear transformation to project the relationship matrix into a different subspace:

[0050]

[0051] The attention weights are calculated based on the relationship matrix, that is, the embedding of different nodes is weighted based on the interaction information of each relationship type. Specifically, for the lth attention head, the attention weights of each node under different relationship types are: The calculation formula is:

[0052]

[0053] in, is the learnable parameter vector of the lth attention head to calculate the attention weight, is the embedding matrix R of the i-th relationship i A row in represents the interaction representation of the node under this relationship, and the attention weight of each node under different relationship types is obtained.

[0054] According to the calculated attention weights, the relation embeddings are weighted and summed to obtain the weighted relation representation of each node i in the relation matrix under the lth attention head:

[0055]

[0056] Concatenate the outputs of all attention heads: We get a concatenated long vector:

[0057]

[0058] Linearly transform this long vector to represent the final fusion and obtain the final global embedding vector:

[0059]

[0060] Preferably, the use of the global embedding vector to detect and sense the nodes of the medical insurance reimbursement items to obtain the detection results specifically includes:

[0061] A multi-layer perceptron is used to classify nodes of reimbursement items. The input of the multi-layer perceptron is the global embedding vector z final , the output is the classification probability distribution y of each node, the size of y is N*C, C is the number of categories, indicating the probability of each node belonging to a different category;

[0062] The cross entropy loss function is used for back propagation to update the parameters, and its calculation formula is:

[0063]

[0064] Y i represents the true label of the node in category i, y i Represents the probability of the model output on category i.

[0065] In a third aspect, an embodiment of the present invention further provides a computing device, the computing device comprising a processor and a memory, the memory being used to store a computer program, the computer program comprising program instructions, the processor being configured to call the program instructions to execute the method as described above.

[0066] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor or a computer, the processor or the computer executes the method as described above.

[0067] Compared with the prior art, the insurance fraud detection method based on heterogeneous graph neural network provided by the embodiment of the present invention has at least the following beneficial effects:

[0068] The embodiment of the present invention learns node and relationship weights through a heterogeneous graph neural network model framework and gradually aggregates information between the layers of the model, which not only promotes the deep fusion of node embedding and relationship embedding, but also enhances the model's understanding of the complex interactions between nodes and relationships. Through carefully designed embedding representations, the embodiment of the present invention can more accurately capture the key features of reimbursement items. Ultimately, these embedding representations are applied to prediction tasks to identify which reimbursement items may have violations, thereby improving the accuracy and efficiency of violation detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] The preferred implementation modes will be described below in a clear and understandable manner with reference to the accompanying drawings to further illustrate the above-mentioned characteristics, technical features, advantages and implementation methods of the present invention.

[0070] Figure 1 This is a flow chart of an insurance fraud detection method based on a heterogeneous graph neural network according to an embodiment of the present invention;

[0071] Figure 2 A schematic diagram of a heterogeneous graph model of an insurance fraud detection method based on a heterogeneous graph neural network according to an embodiment of the present invention;

[0072] Figure 3 A schematic diagram of a neural network structure of a multi-head attention mechanism learning of an insurance fraud detection method based on a heterogeneous graph neural network according to an embodiment of the present invention;

[0073] Figure 4 This is a schematic diagram of the structure of a computing device for implementing an insurance fraud detection method based on a heterogeneous graph neural network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0074] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the specific implementation methods of the present invention will be described below with reference to the accompanying drawings. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings and other implementation methods can be obtained based on these drawings without creative work.

[0075] In order to simplify the drawings, only the parts related to the invention are schematically shown in each figure, and they do not represent the actual structure of the product. In addition, in order to simplify the drawings and facilitate understanding, in some figures, only one of the parts with the same structure or function is schematically drawn or marked. In this article, "one" not only means "only one", but also means "more than one".

[0076] The following mainly takes some specific embodiments as examples to describe in detail the implementation of the technical solution of the present invention.

[0077] like Figure 1 As shown, in order to achieve the purpose of the invention, the embodiment of the present invention provides an insurance fraud detection method based on a heterogeneous graph neural network provided by the embodiment of the present invention, and the method includes:

[0078] S1. Preprocessing the original data of medical insurance reimbursement to construct a heterogeneous graph structure model in the original data;

[0079] S2, performing attribute mapping according to the attribute information of the nodes in the heterogeneous graph structure model, and converting it into a feature vector space of the nodes;

[0080] S3, using a multi-head attention system based on cross-relationships to learn the heterogeneous graph structure model to obtain a global embedding vector;

[0081] S4. Using the global embedding vector, detect and perceive the nodes of the medical insurance reimbursement items to obtain detection results.

[0082] The insurance fraud detection method based on heterogeneous graph neural network in the embodiment of the present invention adopts a cross-relation multi-head attention mechanism, which specifically includes:

[0083] First, the original data of the hospital's medical insurance reimbursement is preprocessed to abstract the graph structure relationship in the data, as follows:

[0084] Based on the real original reimbursement record data, the field information related to the violation detection is extracted, such as personal number, medical insurance settlement number, consultation number, department name, medical insurance project code, etc.

[0085] Based on historical experience data, for example, in actual medical insurance examinations, the medical reimbursement items of patients in a department are generally violated because the department's charging standards do not meet the charging standards, and these violations are mainly reflected in the time when the project costs occurred. Therefore, the embodiment of the present invention constructs four entities, namely, reimbursement record, department, patient, and time, from these fields, so as to realize the message transmission of all reimbursement items in the calculation process.

[0086] By preprocessing the real original reimbursement record data set and constructing the graph structure model, a heterogeneous graph such as Figure 2 As shown, the information set of the heterogeneous graph is:

[0087] G = {V, ε, X};

[0088] Among them, V is a set of nodes of different types in the heterogeneous graph, including four entities: reimbursement record, department, patient, and time;

[0089] ε is a set of relationship types, including reimbursement item-patient; reimbursement item-time; reimbursement item-department;

[0090] X is the feature matrix, which is the attribute of the nodes in the heterogeneous graph. For example, the characteristics of the reimbursement items are the medical insurance item code, unit price, quantity, etc. The characteristics of the patients are personal number, gender, disease type, etc., and time is a single date attribute.

[0091] Reimbursement Item Collection In the dataset, each reimbursement item u∈U has a label y∈{0,1} indicating whether the reimbursement item is an illegal item. 0 indicates that the item is an illegal item, and 1 indicates that the item is not an illegal item.

[0092] The embodiment of the present invention divides the data set into a training set Train = {U train ,Y train}, validation set Val = {U val ,Y val} and the test set Test for model performance testing = {U test ,Y test}.

[0093] That is, at the beginning of model training, the data set is divided into training set, validation set and test set with a ratio of 4:1:1, and finally the F1 score and accuracy are used to judge the effect of the model. Finally, the reimbursement project nodes of three types of violations, namely, over-standard charges, repeated charges and restricted medication, can be identified in the test set.

[0094] The attribute mapping is performed according to the attribute information of the nodes in the heterogeneous graph structure model, and the conversion into the feature vector space of the nodes specifically includes:

[0095] The attribute information of the reimbursement project node is mapped into a vector with the same dimension through word2vec, and the output is:

[0096] f embedding =CBOW(x);

[0097] Among them, CBOW is the continuous bag of words model, which is one of the methods used by word2vec;

[0098] The time data of the node is mapped to the feature vector space through time2Vec, and the mathematical expression of time2Vec is:

[0099] f time (t) = [w 0 t+b 0 ,sin(w 1 t+b 1 ),…,sin(w k t+b k )];

[0100] Among them, t is the original time series, w 0 t+b 0 is the time linear part, sin(w i t+b i ) is the periodic part of time, w and b are a set of learning parameters.

[0101] In the embodiment of the present invention, since the node attributes of entities such as reimbursement items, patients, departments, and patients usually exist in the form of text or strings, but character-type data is not easy to be directly used for calculation, the embodiment of the present invention maps the attribute information of the reimbursement item node into a vector with the same dimension through word2vec, and its output is:

[0102] f embedding =CBOW(x);

[0103] Among them, CBOW is the continuous bag of words model, which is one of the methods used by word2vec;

[0104] The embodiment of the present invention retains important information in the output vector while ensuring that the amount of calculation is not too large, thereby not only making the data suitable for subsequent computing tasks but also retaining key information in the original data.

[0105] In particular, when dealing with time entities, especially in the medical insurance reimbursement process, the periodic and continuous characteristics of time data are crucial for the analysis of cost incurred and patient hospital stay.

[0106] In order to effectively map the time information to the vector space, the embodiment of the present invention may use the time2Vec method, which can map the time data to the vector space and capture its linear and periodic characteristics. The mathematical expression of time2Vec can be expressed as: time (t) = [w 0 t+b 0 ,sin(w 1 t+b 1 ),…,sin(w k t+b k )].

[0107] Among them, t is the original time series, w 0 t+b 0 is the time linear part, sin(w i t+b i ) is the periodic part of time (which can capture the periodic pattern of temporal data), w and b are a set of learnable parameters that represent the concept of weight coefficients in the representation layer of time2Vec.

[0108] By setting k to 31, the output of time2Vec will be a 32-dimensional vector containing 1 linear part and 31 periodic parts.

[0109] Through the above technical features, the embodiment of the present invention uses a fully connected neural network to encode the character data in the data and uses the time2Vec algorithm to encode the time data and map them into high-dimensional vectors.

[0110] In the heterogeneous graph constructed above, two nodes can pass messages through a semantic path, which is called a meta-path. The meta-path expresses a kind of association between nodes of different types, that is, a message passing method. The heterogeneous graph is divided into subgraphs according to the meta-path, and then the subgraphs are subjected to meta-path learning (message passing) to obtain the embedded representation of the meta-path, that is, a set of vectors.

[0111] Then, a relation embedding representation is constructed and a heterogeneous graph neural network model is constructed using cross-relational attention, and the model is trained and predicted.

[0112] For example, in actual implementation, the present invention uses a real data set of medical insurance reimbursement in a certain year. In this data set, there may be violations such as over-standard charges, repeated charges, and restricted medication.

[0113] First, the embodiment of the present invention extracts fields related to violation determination from the medical insurance reimbursement data, including personal number, medical insurance settlement number, consultation number, department name, medical insurance item code, cost, time and other information.

[0114] Based on the extracted data, a heterogeneous graph containing four types of entities (patients, reimbursement items, departments, and time) was constructed.

[0115] The relationships between nodes include but are not limited to:

[0116] The relationship between the reimbursement item and the patient (i.e. which patient was reimbursed for the medical item);

[0117] The relationship between the reimbursement item and the department (i.e. which department performed the reimbursed medical item);

[0118] The relationship between reimbursement items and time (i.e. the time when the reimbursement items occurred), etc.

[0119] For example, Claim-Time-Claim can represent two claims submitted by the same patient at different times, which can be used to analyze claim patterns within a specific time period or to compare claim activities across different time periods.

[0120] Encode the text attributes of the patient, department, and reimbursement item nodes, and use word2vec to map the attribute information of the reimbursement item nodes into vectors with the same dimension.

[0121] In addition, the time2Vec algorithm is used for time nodes to map time data (such as the date when reimbursement occurs, the time of patient hospitalization and discharge) into a 32-dimensional vector with periodic features.

[0122] The method of learning the heterogeneous graph structure model by adopting a multi-head attention system based on cross-relationships to obtain a global embedding vector specifically includes:

[0123] According to the association mode between different nodes, the heterogeneous graph is divided into subgraphs according to the meta-path, and the meta-path includes: φ = {reimbursement item-time-reimbursement item, reimbursement item-patient-reimbursement item, reimbursement item-department-reimbursement item, patient-reimbursement item-patient};

[0124] Through the cross-relation multi-head attention mechanism, multi-head attention calculation is performed on the interactions of nodes in the heterogeneous graph subgraph to obtain the global embedding vector.

[0125] The embodiment of the present invention realizes message transmission between different types of entities by constructing meta-paths, and according to the association method between different entities, the embodiment of the present invention designs four meta-paths: φ = {reimbursement item-time-reimbursement item, reimbursement item-patient-reimbursement item, reimbursement item-department-reimbursement item, patient-reimbursement item-patient} is used to describe possible application scenarios.

[0126] The embodiment of the present invention designs four meta-paths for the heterogeneous graph to describe the interaction relationship between different types of nodes. Through the cross-relationship multi-head attention mechanism, the model performs multi-head attention calculations on the interactions of nodes in different meta-paths and automatically adjusts the importance of each node in different relationships.

[0127] Preferably, the multi-head attention calculation is performed on the interaction of nodes in the heterogeneous graph subgraph through the cross-relation multi-head attention mechanism to obtain the global embedding vector specifically including:

[0128] Automatically adjust the weight coefficient of each node in each different meta-path to obtain the embedded representation of each node in the heterogeneous graph subgraph;

[0129] The embedding representations of nodes on different heterogeneous graph subgraphs and the relationship embeddings between different meta-paths are fused to obtain a global embedding vector.

[0130] The embodiment of the present invention first extracts the semantic information of the relationship through a concise linear propagation mechanism and adding the feature vector (ie, node embedding representation) of the original reimbursement item.

[0131] Preferably, the automatic adjustment of the weight coefficient of each node in each different meta-path to obtain the embedded representation of each node in the heterogeneous graph subgraph specifically includes:

[0132] A self-attention mechanism is introduced when each meta-path is passed to calculate the importance of a node v to its neighbor node u. The importance is expressed as:

[0133]

[0134] Among them, f is the feature vector of the node, ‖ represents the concatenation of two sets of vectors, σ is the activation function, represents the attention vector between nodes;

[0135] For the metapath φ i The message passing in the heterogeneous graph is calculated by the following method to obtain the embedding of each node in the heterogeneous graph subgraph:

[0136]

[0137] For the same metapath φ i The importance of nodes in the meta-path is different. Therefore, the embodiment of the present invention introduces a self-attention mechanism when each meta-path message is transmitted to calculate the importance of a node v to its neighbor node u (for example, the neighbor nodes of a project node can be time, patient and department), which can be expressed as:

[0138]

[0139] Among them, f is the feature vector of the node, ‖ represents the concatenation of two sets of vectors, σ is the activation function, Represents the attention vector between nodes. The attention weight matrix obtained here is asymmetric, because the degree of influence between nodes is different. Finally, for the meta-path φ i The message passing in , that is, the node embedding can be calculated as follows:

[0140]

[0141] In the subgraph of a heterogeneous graph, the node embedding of each node carries multi-dimensional semantic information.

[0142] Although the previous step has described how to obtain the feature vector of the node, namely the node embedding representation, through each meta-path, these embedding representations can often only reflect the information of the node in a relationship, which reflects the connection pattern of the node and carries rich semantic information.

[0143] In order to build a more comprehensive node embedding view, it is necessary to integrate the semantic information contained in different meta-paths and the semantic information in each relationship into their respective subgraph embeddings and effectively fuse this information to perform relationship embedding.

[0144] The relation embedding representation of the embodiment of the present invention adopts a cross-relation multi-head attention mechanism. Through the cross-relation multi-head attention mechanism, the correlation between different meta-paths can be learned, and weighted aggregation is performed according to the importance of each meta-path.

[0145] Preferably, the step of fusing the embedding representations of nodes on different heterogeneous graph subgraphs and the embedding of relations between different meta-paths to obtain a global embedding vector specifically includes:

[0146] For the metapath φ i , its embedding representation is calculated as follows:

[0147]

[0148] in, is the relational semantic embedding at level l-1, and at level 0 is f is the embedding representation of the reimbursement item, and is a learnable parameter of linear propagation;

[0149] The obtained node embedding representation of the meta-path With relational embedding Through weighted aggregation of multiple independent attention heads, where N is the number of nodes on the meta-path, i = 1, 2, 3, ..., M, and M is the number of relations, a global embedding vector is obtained.

[0150] The embedding representation of the reimbursement item is added here to enhance the semantic details of the reimbursement item in the process of relation embedding learning and ensure that key reimbursement item information is not lost during message transmission. Therefore, in the process of relation embedding learning, special attention is paid to the connections related to the entity of the reimbursement item, so as to predict violations in the reimbursement item.

[0151] In order to achieve the effective fusion of relation embedding and meta-path embedding, the embodiment of the present invention adopts multi-head attention across relations to first embed the nodes of the obtained meta-path (N is the number of nodes on the meta-path, 32 is the embedding dimension set at the beginning) and relational embedding i=1,2,3,...,M, where M is the number of relations, and weighted aggregation is performed through multiple independent attention heads.

[0152] Preferably, the node embedding representation of the obtained meta-path is With relational embedding The weighted aggregation through multiple independent attention heads specifically includes:

[0153] For each independent attention head, first embed the meta-path Perform linear transformation to obtain three matrices Q, K and V:

[0154]

[0155] in, are three trainable parameter matrices, d k is the projection dimension;

[0156] Relation embedding for different meta-paths Perform a linear transformation to project the relationship matrix into a different subspace:

[0157]

[0158] The attention weights are calculated based on the relationship matrix, that is, the embedding of different nodes is weighted based on the interaction information of each relationship type. Specifically, for the lth attention head, the attention weights of each node under different relationship types are: The calculation formula is:

[0159]

[0160] in, is the learnable parameter vector of the lth attention head to calculate the attention weight, is the embedding matrix R of the i-th relationship i A row in represents the interaction representation of the node under this relationship, and the attention weight of each node under different relationship types is obtained.

[0161] According to the calculated attention weights, the relation embeddings are weighted and summed to obtain the weighted relation representation of each node i in the relation matrix under the lth attention head:

[0162]

[0163] Concatenate the outputs of all attention heads: We get a concatenated long vector:

[0164]

[0165] Linearly transform this long vector to represent the final fusion and obtain the final global embedding vector:

[0166]

[0167] Since traditional heterogeneous graph neural network models process each meta-path independently to capture the relationship between different node types, each meta-path represents a specific entity interaction pattern and does not fully utilize the potential associations between meta-paths.

[0168] Through the above technical solution, in the embodiment of the present invention, different meta-paths may share certain key nodes (such as a doctor or a department), and the behavior patterns of these nodes may show abnormalities in different contexts. Through cross-meta-path perception, the interaction patterns of these nodes in different meta-paths can be effectively captured, and potential fraud behaviors can be discovered.

[0169] Preferably, the use of the global embedding vector to detect and sense the nodes of the medical insurance reimbursement items to obtain the detection results specifically includes:

[0170] A multi-layer perceptron is used to classify nodes of reimbursement items. The input of the multi-layer perceptron is the global embedding vector z final , the output is the classification probability distribution y of each node, the size of y is N*C, C is the number of categories, indicating the probability of each node belonging to a different category;

[0171] The cross entropy loss function is used for back propagation to update the parameters, and its calculation formula is:

[0172]

[0173] Y i represents the true label of the node in category i, y i Represents the probability of the model output on category i.

[0174] That is, in the embodiment of the present invention, after obtaining the global embedding vector, the embodiment of the present invention uses a multi-layer perceptron (MLP) to perform node classification of reimbursement items for fraud detection.

[0175] MLP is a simple feedforward neural network, which usually includes several fully connected layers, activation functions and the Softmax function in the last layer to obtain the probability of each node belonging to which category.

[0176] The input to the MLP is the embedding vector z of the node final , the output is the classification probability distribution y of each node, the size of y is N*C, C is the number of categories, indicating the probability of each node belonging to a different category.

[0177] Then use the cross entropy loss function for back propagation to update the model parameters. The calculation formula is:

[0178]

[0179] Y i represents the true label of the node in category i, y i Represents the probability of the model output on category i.

[0180] The embodiment of the present invention learns node and relationship weights through a heterogeneous graph neural network model framework and gradually aggregates information between the layers of the model, which not only promotes the deep fusion of node embedding and relationship embedding, but also enhances the model's understanding of the complex interactions between nodes and relationships. Through carefully designed embedding representations, the embodiment of the present invention can more accurately capture the key features of reimbursement items. Ultimately, we apply these embedding representations to prediction tasks to identify which reimbursement items may have violations, thereby improving the accuracy and efficiency of violation detection.

[0181] In a third aspect, an embodiment of the present invention further provides a computing device, which includes a processor or a calculator and a memory, wherein the memory is used to store a computer program, and the computer program includes program instructions, and the processor or calculator is configured to call the program instructions to execute the method as described above.

[0182] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor or a computer, the processor or the computer executes the method as described above.

[0183] like Figure 4 As shown, an embodiment of the present application provides an in-vehicle computing device, the in-vehicle computing device 1000 includes a processor or a calculator (not shown in the figure) 1001 and a memory 1002, and the processor or the calculator 1001 and the memory 1002 can be interconnected through a communication bus 1003. The communication bus 1003 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus 1003 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the memory 1002 is used to store a computer program, the computer program includes program instructions, the processor 1001 is configured to call the program instructions, and the above program includes a method for executing some or all of the steps in the aforementioned method.

[0184] The processor 1001 may be a general purpose central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits for controlling the execution of the above program.

[0185] The memory 1002 may be a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM) or other types of dynamic storage devices that can store information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a compact disc (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical disc, laser disc, optical disc, digital versatile disc, Blu-ray disc, etc.), a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory may exist independently and be connected to the processor via a bus. The memory may also be integrated with the processor.

[0186] The electronic device 1000 may further include a communication module 1004 and a display 1005. The communication module 1004 may be connected to the optical tracking device for communication. The communication module 1004 may be a wireless communication module (eg, a WiFi module, a Bluetooth module, etc.) or a wired communication module.

[0187] In addition, the electronic device 1000 may also include common components such as a communication interface (eg, a USB interface, a microphone interface, etc.), an antenna, etc., which will not be described in detail here.

[0188] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the described order of actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0189] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0190] In the several embodiments provided in the present application, it should be understood that the disclosed device can be implemented in other ways. For example, the device embodiments described above are only schematic, such as the division of the units, which is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of devices or units can be electrical or other forms.

[0191] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0192] In addition, the functional units in the various embodiments of the application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated units may be implemented in the form of hardware or in the form of software program modules.

[0193] If the integrated unit is implemented in the form of a software program module and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a memory, including a number of instructions to enable a computer device (which can be a personal computer, server or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned memory includes: U disk, read-only memory (ROM), random access memory (RAM), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.

[0194] A person of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable memory, which can include: a flash drive, a read-only memory, a random access memory, a magnetic disk or an optical disk, etc.

[0195] The embodiments of the present application are introduced in detail above. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method and core idea of ​​the present application. At the same time, for general technical personnel in this field, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

[0196] It should be noted that the above embodiments can be freely combined as needed. The above are only preferred embodiments of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. An insurance fraud detection method based on heterogeneous graph neural network, characterized in that: The method comprises: Preprocessing the original medical insurance reimbursement data to construct a heterogeneous graph structure model in the original data; Performing attribute mapping according to the attribute information of the nodes in the heterogeneous graph structure model and converting it into a feature vector space of the nodes; The heterogeneous graph structure model is learned by adopting a multi-head attention system based on cross-relationships to obtain a global embedding vector; The global embedding vector is used to detect and perceive the nodes of the medical insurance reimbursement items to obtain detection results.

2. The insurance fraud detection method based on heterogeneous graph neural network according to claim 1, characterized in that: The information set of the heterogeneous graph structure model is: G = {V, ε, X}; Wherein, V is a set of nodes of different types in a heterogeneous graph, and the node includes one of the four entities of reimbursement record, department, patient and time; ε is a set of relationship types, including reimbursement item-patient, reimbursement item-time, and reimbursement item-department; X is a feature matrix, which is the attribute features of the nodes in the heterogeneous graph; Reimbursement Item Collection For each reimbursement item u∈U, there is a label y∈{0,1}, which indicates whether the reimbursement item is a violation item. 0 indicates that it is a violation item, and 1 indicates that it is not a violation item.

3. The insurance fraud detection method based on heterogeneous graph neural network according to claim 1, characterized in that: Performing attribute mapping according to the attribute information of the nodes in the heterogeneous graph structure model and converting into the feature vector space of the nodes specifically includes: The attribute information of the reimbursement project node is mapped into a vector with the same dimension through word2vec, and the output is: f embedding =CBOW(x); Among them, CBOW is the continuous bag of words model, which is one of the methods used by word2vec; The time data of the node is mapped to the feature vector space through time2Vec, and the mathematical expression of time2Vec is: f time (t)=[w0t+b0,sin(w1t+b1),…,sin(w k t+b k )]; Among them, t is the original time series, w0t+b0 is the time linear part, sin(w i t+b i ) is the periodic part of time, w and b are a set of learning parameters.

4. The insurance fraud detection method based on heterogeneous graph neural network according to claim 1, characterized in that: The adopting of a cross-relation-based multi-head attention system to learn the heterogeneous graph structure model to obtain a global embedding vector specifically includes: According to the association mode between different nodes, the heterogeneous graph is divided into subgraphs according to the meta-path, and the meta-path includes: φ = {reimbursement item-time-reimbursement item, reimbursement item-patient-reimbursement item, reimbursement item-department-reimbursement item, patient-reimbursement item-patient}; Through the cross-relation multi-head attention mechanism, multi-head attention calculation is performed on the interactions of nodes in the heterogeneous graph subgraph to obtain the global embedding vector.

5. The insurance fraud detection method based on heterogeneous graph neural network according to claim 4, characterized in that: The cross-relation multi-head attention mechanism is used to perform multi-head attention calculation on the interaction of nodes in the heterogeneous graph subgraph to obtain the global embedding vector, which specifically includes: Automatically adjust the weight coefficient of each node in each different meta-path to obtain the embedded representation of each node in the heterogeneous graph subgraph; The embedding representations of nodes on different heterogeneous graph subgraphs and the relationship embeddings between different meta-paths are fused to obtain a global embedding vector.

6. The insurance fraud detection method based on heterogeneous graph neural network according to claim 5, characterized in that: The automatic adjustment of the weight coefficient of each node in each different meta-path to obtain the embedded representation of each node in the heterogeneous graph subgraph specifically includes: A self-attention mechanism is introduced when each meta-path is passed to calculate the importance of a node v to its neighbor node u. The importance is expressed as: Among them, f is the feature vector of the node, ‖ represents the concatenation of two sets of vectors, σ is the activation function, represents the attention vector between nodes; For the metapath φ i The message passing in the heterogeneous graph is calculated by the following method to obtain the embedding of each node in the heterogeneous graph subgraph:

7. The insurance fraud detection method based on heterogeneous graph neural network according to claim 6, characterized in that: The method of fusing the embedding representations of nodes on different heterogeneous graph subgraphs and the embedding of relations between different meta-paths to obtain a global embedding vector specifically includes: For the metapath φ i , its embedding representation is calculated as follows: in, is the relational semantic embedding at level l-1, and at level 0 is f is the embedding representation of the reimbursement item, and is a learnable parameter of linear propagation; The obtained node embedding representation of the meta-path With relational embedding Through weighted aggregation of multiple independent attention heads, where N is the number of nodes on the meta-path, i = 1, 2, 3, ..., M, and M is the number of relations, a global embedding vector is obtained.

8. The insurance fraud detection method based on heterogeneous graph neural network according to claim 7, characterized in that: The node embedding representation of the obtained meta-path is With relational embedding The weighted aggregation through multiple independent attention heads specifically includes: For each independent attention head, first embed the meta-path Perform linear transformation to obtain three matrices Q, K and V: in, are three trainable parameter matrices, d k is the projection dimension; Relation embedding for different meta-paths Perform a linear transformation to project the relationship matrix into a different subspace: The attention weights are calculated based on the relationship matrix, that is, the embedding of different nodes is weighted based on the interaction information of each relationship type. Specifically, for the lth attention head, the attention weights of each node under different relationship types are: The calculation formula is: in, is the learnable parameter vector of the lth attention head to calculate the attention weight, is the embedding matrix R of the i-th relationship i A row in represents the interaction representation of the node under this relationship, and the attention weight of each node under different relationship types is obtained. According to the calculated attention weights, the relation embeddings are weighted and summed to obtain the weighted relation representation of each node i in the relation matrix under the lth attention head: Concatenate the outputs of all attention heads: We get a concatenated long vector: Linearly transform this long vector to represent the final fusion and obtain the final global embedding vector:

9. The insurance fraud detection method based on heterogeneous graph neural network according to claim 1, characterized in that: The use of the global embedding vector to detect and perceive the nodes of the medical insurance reimbursement items to obtain the detection results specifically includes: A multi-layer perceptron is used to classify nodes of reimbursement items. The input of the multi-layer perceptron is the global embedding vector z final , the output is the classification probability distribution y of each node, the size of y is N*C, C is the number of categories, indicating the probability of each node belonging to a different category; The cross entropy loss function is used for back propagation to update the parameters, and its calculation formula is: Y i represents the true label of the node in category i, y i Represents the probability of the model output on category i.

10. A computing device, characterized in that: The computing device comprises a processor and a memory, wherein the memory is used to store a computer program, the computer program comprises program instructions, and the processor is configured to call the program instructions to execute the method according to any one of claims 1 to 9.

Images