Intelligent auditing and financial fraud identification method based on dynamic heterogeneous graph neural network

Abstract

The application discloses an intelligent auditing and financial fraud identification method based on a dynamic heterogeneous graph neural network. The method comprises the following steps: constructing a dynamic enterprise risk heterogeneous graph; generating a first feature vector that fuses node static features and node dynamic features, and a second feature vector that fuses edge static features, edge dynamic features and the interaction results of connected node features; obtaining a standard embedding representation of a node; obtaining a preliminary aggregated representation of the node based on an attention mechanism and the standard embedding representation of the node; modeling the global dependence of the preliminary aggregated representation of the node by using a Transformer structure to obtain an enhanced aggregated representation of the node; fusing the standard embedding representation and the enhanced aggregated representation of the node through a residual connection to obtain a final embedding representation of the node; and finally obtaining a financial fraud identification result. The application effectively solves the problems of insufficient fusion of heterogeneous features, weak global relationship learning ability and insufficient time sequence dynamic capture in the prior art.

Description

Technical Field

[0001] This invention relates to the field of intelligent auditing technology, specifically to an intelligent auditing and financial fraud detection method based on dynamic heterogeneous graph neural networks. Background Technology

[0002] Detecting corporate financial fraud is a core element in ensuring the healthy development of the capital market. Traditional methods mainly rely on statistical models such as linear regression and logistic regression. These models are based on linear assumptions and perform isolated analysis of financial data. They are unable to effectively capture the complex and non-linear relationships between various types of entities, such as enterprises, accounting items, transaction records, suppliers, and customers, and are even less able to model the dynamic evolution of financial risks over time.

[0003] With the development of deep learning technology, graph neural networks have been introduced to handle the complex relationships in financial data. Early solutions used homogeneous graph neural networks, but their single node type made it difficult to handle the naturally occurring multi-type node (heterogeneous) graph structures in the financial domain. Subsequent researchers proposed dynamic heterogeneous graph neural networks, which improve the model through type-specific embeddings and time-step features. However, existing dynamic heterogeneous graph solutions still have the following shortcomings: they fail to fully integrate the static and dynamic features of nodes and edges, resulting in imprecise extraction of heterogeneous features; when aggregating cross-type node information, they often use a single attention mechanism, lacking the ability to model global dependencies, leading to insufficient sensitivity to hidden key risk paths; and they do not optimize time-step embeddings and information propagation paths in deep networks, lacking mechanisms such as residual connections, resulting in limited ability of the model to capture the temporal evolution of the "gradual accumulation" of financial risks.

[0004] In summary, there is an urgent need in this field for a financial fraud identification method that can deeply integrate static and dynamic features, enhance global relationship learning, and optimize the propagation of temporal information. Summary of the Invention

[0005] To address the aforementioned shortcomings in existing technologies, this invention provides an intelligent auditing and financial fraud identification method based on dynamic heterogeneous graph neural networks.

[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0007] Construct a dynamic enterprise risk heterogeneous graph, which includes various types of nodes and edges, and the characteristics of the nodes and edges change dynamically with multiple time steps;

[0008] For each node in the dynamic enterprise risk heterogeneous graph, a first feature vector is generated that integrates the static and dynamic features of the node. For each edge in the dynamic enterprise risk heterogeneous graph, a second feature vector is generated that integrates the static features of the edge, the dynamic features of the edge, and the interaction results of the features of the connected nodes.

[0009] Using a type-specific parameter matrix, the first feature vector and the second feature vector are mapped to the initial embedding representation of the node and the embedding representation of the edge, respectively. The standard embedding representation of the node is obtained based on the initial embedding representation of the node and the embedding representation of the edge.

[0010] Based on the attention mechanism and the standard embedding representation of nodes, the information of the neighbor nodes of a node under different edge types is aggregated to obtain the preliminary aggregated representation of the node;

[0011] We use the Transformer structure to perform global dependency modeling on the initial aggregated representation of nodes, calculate the normalized importance weights of nodes for different edge types based on the self-attention mechanism, and obtain the enhanced aggregated representation of nodes based on the normalized importance weights of nodes for different edge types.

[0012] The standard embedding representation and the enhanced aggregate representation of a node are fused through residual connections to obtain the final embedding representation of the node;

[0013] The final embedding representation based on nodes is used to obtain financial fraud identification results through a classifier.

[0014] Furthermore, node types include enterprises, accounting subjects, transactions, suppliers / customers, long-term asset impairment, goodwill impairment, inventory impairment, and fixed asset impairment; edge types include transactions occurring within an enterprise, transactions belonging to accounting subjects, accounting subjects reconciling with accounting subjects, and cooperative suppliers / customers of an enterprise.

[0015] Furthermore, the expression for the first eigenvector is:

[0016]

[0017] in: For nodes At time step The first eigenvector, For the edge The first node, For nodes static characteristics For feature splicing operations, For nodes At time step Dynamic characteristics;

[0018] The expression for the second eigenvector is:

[0019]

[0020] in: For the edge At time step The second eigenvector, For edge symbols, For the edge static characteristics For the edge At time step The dynamic characteristics, For about and The node feature interaction function, For nodes At time step The first eigenvector, For the edge The second node.

[0021] Furthermore, using a type-specific parameter matrix, the first and second eigenvectors are mapped to the initial embedding representation of the node and the embedding representation of the edge, respectively, as expressed in the following expressions:

[0022] ,

[0023]

[0024] in: For nodes At time step The initial embedding, For the edge The first node, It is the ReLU activation function. For node type Dedicated weight matrix, For nodes type For nodes At time step The first eigenvector, For node type A dedicated bias matrix. For the edge At time step Embedded, For the edge The second node, edge type Dedicated weight matrix, For the edge type For the edge At time step The second eigenvector, edge type A dedicated bias matrix.

[0025] Furthermore, based on the attention mechanism and the standard embedding representation of nodes, the information of the neighbor nodes of a node under different edge types is aggregated to obtain the preliminary aggregated representation of the node. The specific process is as follows:

[0026] Based on the attention mechanism and the standard embedding representation of nodes, the learning importance of neighbor nodes under different edge types is determined, and its expression is as follows:

[0027] ,

[0028]

[0029] in: For nodes neighboring nodes In edge type The importance of learning below For the edge type For the edge The first node, For the edge The second node, For activation function, edge type The first training matrix below, For nodes Standard embedded representation, For neighboring nodes Standard embedded representation;

[0030] The Softmax function is used to normalize the learning importance of neighbor nodes under different edge types, so as to obtain the normalized learning importance of neighbor nodes under different edge types.

[0031] Based on the normalized learning importance of neighboring nodes under different edge types, a preliminary aggregated representation of the nodes is obtained.

[0032] Furthermore, global dependency modeling is performed on the preliminary aggregated representation of nodes using the Transformer structure. The normalized importance weights of nodes for different edge types are calculated based on the self-attention mechanism, and their expression is as follows:

[0033] ,

[0034]

[0035] in: For nodes opposite edge type Importance weights edge type Next node The key indicates, This is the matrix transpose operator. The first trainable parameter, For the dimension of the potential space, For nodes opposite edge type Normalized importance weights This is the softmax activation function.

[0036] Furthermore, an enhanced aggregate representation of nodes is obtained based on the normalized importance weights of nodes for different edge types, and its expression is as follows:

[0037]

[0038] in: For nodes Enhanced aggregation representation, For the edge type For nodes The set of edge types formed by it and its neighboring nodes. For nodes opposite edge type Normalized importance weights For the second trainable parameter, For nodes In weighted edge type The initial aggregation is represented below, This is the third trainable parameter.

[0039] Furthermore, the standard embedding representation and the enhanced aggregate representation of the node are fused through residual connections to obtain the final embedding representation of the node, whose expression is:

[0040]

[0041] in: For nodes The final embedding representation, To balance the learning weights of aggregated information and initial node information, For GELU activation function, For nodes Standard embedded representation, For nodes Enhanced aggregation representation.

[0042] The beneficial effects of this invention are as follows:

[0043] (1) The present invention constructs a dynamic enterprise risk heterogeneous graph, which includes multiple types of nodes and edges, and the features of nodes and edges change dynamically with multiple time steps. It accurately extracts the differentiated features of multiple types of nodes and edges, solves the problem of insufficient modeling accuracy of heterogeneous relationships in existing methods, and meets the modeling requirements of complex structures with multiple node types and edge types.

[0044] (2) This invention aggregates the information of neighboring nodes of a node under different edge types according to the attention mechanism and the standard embedding representation of the node to obtain the preliminary aggregated representation of the node. It then uses the Transformer structure to perform global dependency modeling on the preliminary aggregated representation of the node, calculates the normalized importance weight of the node to different edge types based on the self-attention mechanism, and obtains the enhanced aggregated representation of the node based on the normalized importance weight of the node to different edge types. This process calculates the importance weight of the relationship between nodes through the attention mechanism and combines the Transformer structure to realize the global dependency modeling of multi-relationship features, which strengthens the learning of cross-type node relationships, significantly improves the sensitivity to key risk paths, and accurately captures the hidden fraud patterns.

[0045] (3) This invention updates nodes and edges in a dynamic enterprise risk heterogeneous graph by embedding time steps, and combines the standard embedded representation and enhanced aggregation representation of residual connection nodes to effectively capture the temporal evolution of financial fraud risk, thus solving the problem of insufficient temporal evolution capture in existing methods. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the intelligent auditing and financial fraud detection method based on dynamic heterogeneous graph neural networks. Detailed Implementation

[0047] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0048] like Figure 1 As shown, the intelligent auditing and financial fraud detection method based on dynamic heterogeneous graph neural networks includes steps S1-S7, as detailed below:

[0049] S1. Construct a dynamic enterprise risk heterogeneous graph. The dynamic enterprise risk heterogeneous graph includes various types of nodes and edges, and the characteristics of the nodes and edges change dynamically with multiple time steps.

[0050] In an optional embodiment of the present invention, the node types include enterprises, accounting subjects, transactions, suppliers / customers, long-term asset impairment, goodwill impairment, inventory impairment, and fixed asset impairment; the edge types include transactions occurring in enterprises, transaction-related accounting subjects, accounting subject reconciliation with accounting subjects, and enterprise-cooperative suppliers / customers.

[0051] When a node is an enterprise, its static and dynamic characteristics are as follows:

[0052] ,

[0053]

[0054] in: For nodes static characteristics For the edge The first node, For nodes At time step The dynamic characteristics.

[0055] When the node is an accounting subject, its static and dynamic characteristics are as follows:

[0056] ,

[0057] .

[0058] When a node is a transaction, its static and dynamic characteristics are as follows:

[0059] ,

[0060] .

[0061] When a node is a supplier / customer, its static and dynamic characteristics are as follows:

[0062] ,

[0063] .

[0064] When the node represents a long-term asset impairment, its static and dynamic characteristics are as follows:

[0065] ,

[0066] .

[0067] When the node represents goodwill impairment, its static and dynamic characteristics are as follows:

[0068] ,

[0069] .

[0070] When a node represents inventory impairment, its static and dynamic characteristics are as follows:

[0071] ,

[0072] .

[0073] When the node represents fixed asset impairment, its static and dynamic characteristics are as follows:

[0074] ,

[0075] .

[0076] When an edge represents a transaction between enterprises, its static and dynamic characteristics are as follows:

[0077] ,

[0078]

[0079] in: For edge symbols, For the edge static characteristics For the edge At time step The dynamic characteristics.

[0080] When the edge represents the accounting category to which the transaction belongs, its static and dynamic characteristics are as follows:

[0081] ,

[0082] .

[0083] When the side is an accounting subject reconciling with the accounting subject, its static and dynamic characteristics are as follows:

[0084] ,

[0085] .

[0086] When the edge represents a supplier / customer of a business, its static and dynamic characteristics are as follows:

[0087] ,

[0088] .

[0089] S2. Generate a first feature vector that integrates the static and dynamic features of each node in the dynamic enterprise risk heterogeneous graph, and generate a second feature vector that integrates the static features of the edge, the dynamic features of the edge, and the interaction results of the features of the connected nodes for each edge in the dynamic enterprise risk heterogeneous graph.

[0090] In an optional embodiment of the present invention, the expression for the first feature vector is:

[0091]

[0092] in: For nodes At time step The first eigenvector, For the edge The first node, For nodes static characteristics For feature splicing operations, For nodes At time step The dynamic characteristics.

[0093] The expression for the second eigenvector is:

[0094] ,

[0095]

[0096] in: For the edge At time step The second eigenvector, For edge symbols, For the edge static characteristics For the edge At time step The dynamic characteristics, For about and The node feature interaction function, For nodes At time step The first eigenvector, For the edge The second node, This is a vector concatenation operation. For interaction weights, For interactive bias.

[0097] S3. Using type-specific parameter matrices, map the first feature vector and the second feature vector to the initial embedding representation of the node and the embedding representation of the edge, respectively. Obtain the standard embedding representation of the node based on the initial embedding representation of the node and the embedding representation of the edge.

[0098] In an optional embodiment of the present invention, the present invention utilizes a type-specific parameter matrix to map the first feature vector and the second feature vector to the initial embedding representation of the node and the embedding representation of the edge, respectively, as expressed in the following expression:

[0099] ,

[0100]

[0101] in: For nodes At time step The initial embedding, For the edge The first node, It is the ReLU activation function. For node type The dedicated weight matrix is ​​obtained by iteratively adjusting the weights on the training data using an optimization algorithm to minimize the loss function, which is the mean squared error. For nodes type For nodes At time step The first eigenvector, For node type A dedicated bias matrix is ​​obtained by iteratively adjusting the training data using an optimization algorithm to minimize the loss function, which is the mean squared error. For the edge At time step Embedded, For the edge The second node, edge type The dedicated weight matrix is ​​obtained by iteratively adjusting the weights on the training data using an optimization algorithm to minimize the loss function, which is the mean squared error. For the edge type For the edge At time step The second eigenvector, edge type The dedicated bias matrix is ​​obtained by iteratively adjusting the training data using an optimization algorithm to minimize the loss function, which is the mean squared error.

[0102] This invention obtains the standard embedding representation of a node based on the initial embedding representation of the node and the embedding representation of the edge. The specific process is as follows:

[0103] Map the initial embedding representation of a node to the same latent space based on its type to obtain the node's position at time step. The standard embedding representation of is expressed as:

[0104]

[0105] in: For nodes At time step Standard embedded representation, For batch standardization operations, For nodes The trainable weight matrix for the corresponding standard mapping is obtained by iteratively adjusting the weights on the training data using an optimization algorithm to minimize the loss function, where the loss function is the mean squared error. For nodes The bias term of the standard mapping corresponding to the type is obtained by iteratively adjusting the training data through an optimization algorithm to minimize the loss function, which uses mean squared error.

[0106] Map the edge embedding representation to the same latent space according to the edge type to obtain the edge at time step. The standard embedding representation of is expressed as:

[0107]

[0108] in: For the edge At time step Standard embedding, For the edge The trainable weight matrix for the corresponding standard mapping is obtained by iteratively adjusting the weights on the training data using an optimization algorithm to minimize the loss function, where the loss function is the mean squared error. For the edge The bias term of the standard mapping corresponding to the type is obtained by iteratively adjusting the training data through an optimization algorithm to minimize the loss function, which uses mean squared error.

[0109] Based on the node at time step Standard embedding representation and edge in time step The standard embedding representation, obtains the time step The message passed by the neighboring nodes is represented by the expression:

[0110]

[0111] in: In time step Neighbor nodes Passed to the node The message indicates that The weight matrix, representing the information passed between neighboring nodes, is obtained by iteratively adjusting the weights on the training data using an optimization algorithm to minimize the loss function, which is the mean squared error. The bias term, representing the information passed from neighboring nodes, is obtained by iteratively adjusting the algorithm on the training data to minimize the loss function, which uses mean squared error. This is for element-wise multiplication.

[0112] According to the time step The information transmitted by neighboring nodes indicates that the information is obtained at time step. The aggregated message passed by the neighboring nodes is represented by the following expression:

[0113]

[0114] in: For nodes The set of neighboring nodes, In time step Neighbor nodes pass to node The aggregated message indicates.

[0115] Based on the node at time step Standard embedding representation and in time step The aggregated messages passed by the neighboring nodes indicate that the node is at time step The standard embedding representation of is expressed as:

[0116]

[0117] in: For nodes At time step Standard embedded representation, This is a vector concatenation operation. To update the node weight matrix, an optimization algorithm iteratively adjusts it on the training data to minimize the loss function, which uses mean squared error. The bias term for updating nodes is obtained by iteratively adjusting the algorithm on the training data to minimize the loss function, which uses mean squared error. This is the ReLU activation function.

[0118] S4. Based on the attention mechanism and the standard embedding representation of nodes, the information of the neighbor nodes of a node under different edge types is aggregated to obtain the preliminary aggregated representation of the node.

[0119] In an optional embodiment of the present invention, the present invention aggregates the information of neighboring nodes of a node under different edge types based on the attention mechanism and the standard embedding representation of nodes to obtain a preliminary aggregated representation of the node. The specific process is as follows:

[0120] Based on the attention mechanism and the standard embedding representation of nodes, the learning importance of neighbor nodes under different edge types is determined, and its expression is as follows:

[0121] ,

[0122]

[0123] in: For nodes neighboring nodes In edge type The importance of learning below For the edge type For the edge The first node, For the edge The second node, As an activation function, it allows negative gradients to pass through at a certain rate, thus maintaining sensitivity to negative values ​​during training and helping to better learn the features of the data. edge type The first training matrix is ​​obtained by iteratively adjusting the training data using an optimization algorithm to minimize the loss function, which is the mean squared error. For nodes Standard embedded representation, For neighboring nodes The standard embedded representation.

[0124] The Softmax function is used to normalize the learning importance of neighbor nodes under different edge types to obtain the normalized learning importance of neighbor nodes under different edge types. Specifically, this invention considers two edge types: weighted edge type and unweighted edge type. Unweighted edge type only focuses on whether two adjacent nodes are connected, while weighted edge type also focuses on the weight of the connection between two adjacent nodes, that is, the strength attribute of the connection between two adjacent nodes.

[0125] When the edge type is unweighted edge type, its expression is:

[0126]

[0127] in: For neighboring nodes Relative to node In edge type The next The importance of normalization learning in each dimension It is a natural exponential function. For nodes neighboring nodes In edge type The next The importance of learning in each dimension For nodes In edge type The set of neighboring nodes;

[0128] When the edge type is a weighted edge type, its expression is:

[0129]

[0130] in: For neighboring nodes Relative to node In edge type The importance of normalized learning For nodes with neighboring nodes The original edge weights in this invention are obtained by weighting and quantizing the static and dynamic features of the edge.

[0131] Based on the normalized learning importance of neighboring nodes under different edge types, a preliminary aggregated representation of the nodes is obtained;

[0132] When the edge type is unweighted edge type, its expression is:

[0133]

[0134] in: For nodes In edge type The next A preliminary aggregated representation of each dimension. For nodes In edge type The set of neighboring nodes below, For dropout operations, For neighboring nodes In the standard embedding representation, the first Each feature element.

[0135] When the edge type is a weighted edge type, its expression is:

[0136]

[0137] in: For nodes In edge type The initial aggregation is represented below, edge type The second training matrix is ​​obtained by iteratively adjusting the training data using an optimization algorithm to minimize the loss function, which is the mean squared error. For neighboring nodes The standard embedded representation.

[0138] S5. Use the Transformer structure to perform global dependency modeling on the preliminary aggregated representation of nodes, calculate the normalized importance weight of nodes for different edge types based on the self-attention mechanism, and obtain the enhanced aggregated representation of nodes based on the normalized importance weight of nodes for different edge types.

[0139] In an optional embodiment of the present invention, the present invention utilizes the Transformer structure to perform global dependency modeling on the preliminary aggregated representation of nodes, and calculates the normalized importance weights of nodes for different edge types based on the self-attention mechanism, the expression of which is:

[0140] ,

[0141]

[0142] in: For nodes opposite edge type Importance weights edge type Next node The key indicates that if the selected weighted edge type is selected, then here... If the selected edge type is unweighted, here , This is the matrix transpose operator. The first trainable parameter is used to adjust the scale of learning importance. For the dimension of the potential space, For nodes opposite edge type Normalized importance weights This is the softmax activation function.

[0143] Specifically, when the edge type is unweighted edge type, the node The expressions for the key representation and the query representation are:

[0144] ,

[0145]

[0146] in: edge type Unweighted nodes The key indicates, For nodes In edge type The initial aggregation is represented below, edge type Next node The query indicates that, edge type The query below represents the weight matrix of the linear transformation, which is obtained by iteratively adjusting the weight matrix on the training data using an optimization algorithm to minimize the loss function, where the loss function is the mean squared error. For nodes Standard embedded representation, edge type The query below represents the bias term of the linear transformation, which is obtained by iteratively adjusting the training data using an optimization algorithm to minimize the loss function, which is the mean squared error.

[0147] When the edge type is weighted edge type, the node The expressions for the key representation and the query representation are:

[0148] ,

[0149] ,

[0150] in: edge type In weighted nodes The key indicates, edge type The weighted key represents the weight matrix of the linear transformation. For nodes In edge type The initial aggregation representation below, edge type The weighted key represents the bias term of the linear transformation. edge type Next node The query indicates that, edge type The query below represents the weight matrix of the linear transformation, which is obtained by iteratively adjusting the weight matrix on the training data using an optimization algorithm to minimize the loss function, where the loss function is the mean squared error. For nodes Standard embedded representation, edge type The query below represents the bias term of the linear transformation, which is obtained by iteratively adjusting the training data using an optimization algorithm to minimize the loss function, which is the mean squared error.

[0151] This invention obtains an enhanced aggregate representation of nodes based on the normalized importance weights of nodes for different edge types, and its expression is as follows:

[0152]

[0153] in: For nodes Enhanced aggregation representation, For the edge type For nodes The set of edge types formed by it and its neighboring nodes. For nodes opposite edge type Normalized importance weights The second trainable parameter is obtained by iteratively adjusting it on the training data using an optimization algorithm to minimize the loss function, where the loss function is the mean squared error. For nodes In weighted edge type The initial aggregation is represented below, The third trainable parameter is obtained by iteratively adjusting it on the training data using an optimization algorithm to minimize the loss function, which is the mean squared error.

[0154] S6. The standard embedded representation and the enhanced aggregate representation of the node are fused through residual connections to obtain the final embedded representation of the node.

[0155] In an optional embodiment of the present invention, the standard embedding representation and the enhanced aggregate representation of a node are fused through residual connections to obtain the final embedding representation of the node, the expression of which is:

[0156]

[0157] in: For nodes The final embedding representation, To balance the learning weights of aggregated information and initial node information, For GELU activation function, For nodes Standard embedded representation, For nodes Enhanced aggregation representation.

[0158] S7. The final embedding representation based on nodes is used to obtain the financial fraud identification results through a classifier.

[0159] In an alternative embodiment of the invention, the final embedded representation of the node is input into a classifier to obtain financial fraud identification results.

[0160] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for intelligent auditing and financial fraud detection based on dynamic heterogeneous graph neural networks, characterized in that, Includes the following steps: Construct a dynamic enterprise risk heterogeneous graph, which includes various types of nodes and edges, and the characteristics of the nodes and edges change dynamically with multiple time steps; For each node in the dynamic enterprise risk heterogeneous graph, a first feature vector is generated that integrates the static and dynamic features of the node. For each edge in the dynamic enterprise risk heterogeneous graph, a second feature vector is generated that integrates the static features of the edge, the dynamic features of the edge, and the interaction results of the features of the connected nodes. Using a type-specific parameter matrix, the first feature vector and the second feature vector are mapped to the initial embedding representation of the node and the embedding representation of the edge, respectively. The standard embedding representation of the node is obtained based on the initial embedding representation of the node and the embedding representation of the edge. Based on the attention mechanism and the standard embedding representation of nodes, the information of the neighbor nodes of a node under different edge types is aggregated to obtain the preliminary aggregated representation of the node; We use the Transformer structure to perform global dependency modeling on the initial aggregated representation of nodes, calculate the normalized importance weights of nodes for different edge types based on the self-attention mechanism, and obtain the enhanced aggregated representation of nodes based on the normalized importance weights of nodes for different edge types. The standard embedding representation and the enhanced aggregate representation of a node are fused through residual connections to obtain the final embedding representation of the node; The final embedding representation based on nodes is used to obtain financial fraud identification results through a classifier.

2. The intelligent auditing and financial fraud identification method based on dynamic heterogeneous graph neural networks according to claim 1, characterized in that, Node types include enterprises, accounting subjects, transactions, suppliers / customers, long-term asset impairment, goodwill impairment, inventory impairment, and fixed asset impairment; edge types include transactions occurring within an enterprise, transactions belonging to accounting subjects, accounting subjects reconciling with accounting subjects, and cooperative suppliers / customers of an enterprise.

3. The intelligent auditing and financial fraud identification method based on dynamic heterogeneous graph neural networks according to claim 1, characterized in that, The expression for the first eigenvector is: ； in: For nodes At time step The first eigenvector, For the edge The first node, For nodes static characteristics, For feature splicing operations, For nodes At time step Dynamic characteristics; The expression for the second eigenvector is: ； in: For the edge At time step The second eigenvector, For edge symbols, For the edge static characteristics, For the edge At time step The dynamic characteristics, For about and Node feature interaction function, For nodes At time step The first eigenvector, For the edge The second node.

4. The intelligent auditing and financial fraud identification method based on dynamic heterogeneous graph neural networks according to claim 1, characterized in that, Using a type-specific parameter matrix, the first and second eigenvectors are mapped to the initial embedding representation of the node and the embedding representation of the edge, respectively, as expressed in the following expressions: ， ； in: For nodes At time step The initial embedding, For the edge The first node, It is the ReLU activation function. For node type Dedicated weight matrix, For nodes type For nodes At time step The first eigenvector, For node type A dedicated bias matrix. For the edge At time step Embedded, For the edge The second node, edge type Dedicated weight matrix, For the edge type For the edge At time step The second eigenvector, edge type A dedicated bias matrix.

5. The intelligent auditing and financial fraud identification method based on dynamic heterogeneous graph neural networks according to claim 1, characterized in that, Based on the attention mechanism and the standard embedding representation of nodes, the information of the neighbor nodes of a node under different edge types is aggregated to obtain the preliminary aggregated representation of the node. The specific process is as follows: Based on the attention mechanism and the standard embedding representation of nodes, the learning importance of neighbor nodes under different edge types is determined, and its expression is as follows: ， ； in: For nodes neighboring nodes In edge type The importance of learning below For the edge type For the edge The first node, For the edge The second node, For activation function, edge type The first training matrix below, For nodes Standard embedded representation, For neighboring nodes Standard embedded representation; The Softmax function is used to normalize the learning importance of neighbor nodes under different edge types, so as to obtain the normalized learning importance of neighbor nodes under different edge types. Based on the normalized learning importance of neighboring nodes under different edge types, a preliminary aggregated representation of the nodes is obtained.

6. The intelligent auditing and financial fraud identification method based on dynamic heterogeneous graph neural networks according to claim 1, characterized in that, We use the Transformer structure to perform global dependency modeling on the initial aggregated representation of nodes, and calculate the normalized importance weights of nodes for different edge types based on the self-attention mechanism. The expression is as follows: ， ； in: For nodes opposite edge type Importance weights edge type Next node The key indicates, This is the matrix transpose operator. The first trainable parameter, For the dimension of the potential space, For nodes opposite edge type Normalized importance weights This is the softmax activation function.

7. The intelligent auditing and financial fraud identification method based on dynamic heterogeneous graph neural networks according to claim 1, characterized in that, The enhanced aggregate representation of nodes is obtained based on the normalized importance weights of nodes for different edge types, and its expression is as follows: ； in: For nodes Enhanced aggregation representation, For the edge type For nodes The set of edge types formed by it and its neighboring nodes. For nodes opposite edge type Normalized importance weights For the second trainable parameter, For nodes In weighted edge type The initial aggregation representation below, This is the third trainable parameter.

8. The intelligent auditing and financial fraud identification method based on dynamic heterogeneous graph neural networks according to claim 1, characterized in that, The standard embedding representation and the enhanced aggregate representation of a node are fused through residual connections to obtain the final embedding representation of the node, which is expressed as: ； in: For nodes The final embedding representation, To balance the learning weights of aggregated information and initial node information, For GELU activation function, For nodes Standard embedded representation, For nodes Enhanced aggregation representation.

Images