A financial credit anti-fraud method, device and medium

Abstract

The present invention relates to the field of financial credit anti-fraud, and discloses a financial credit anti-fraud method. The method uses data in credit to construct a multi-relationship time series network, and transforms node features in the network so that the representation vector retains the original data features while also retaining the role information of the original data; a neighbor aggregation method is used to update the network node features at different times in different relationships, and corresponding asynchronous timestamp information is added; finally, a gated loop late fusion model is used to fuse the features of each time step of different relationships, and the graph node joint features of the fused credit applicant are obtained, and whether the applicant is cashing out for consumption or cashing out for credit cards is accurately judged, thereby achieving the effect of efficient credit anti-fraud.

Description

Technical Field

[0001] The present invention relates to the field of financial credit anti-fraud, and in particular to a financial credit anti-fraud method, device and medium. Background Art

[0002] Initially, the method of screening applicants by manually defining rules was used to determine whether the applicants would violate the law or use false means to exchange for cash benefits. This method has high requirements for rules and requires a lot of prior knowledge. In addition, there are many different fraud methods, which are difficult to screen with simple rules. Then, with the development of machine learning and deep learning, the industry gradually used characteristics such as buyer attributes and historical behavior information, and used different machine learning classifiers, such as tree-based models and neural networks, to conduct supervised classification of applicants. Although these methods have been successful in some scenarios, they are not effective when multiple people with different roles (such as sellers and applicants) conspire to apply for fraudulent loans.

[0003] Recently, graph neural networks (GNNs) have achieved remarkable success in node classification. GNNs use the feature information of neighboring nodes to improve the representation learning of target nodes. However, they map and propagate all features together without distinguishing them. As a result, when applied in financial credit anti-fraud scenarios, applicants with neighbors of different roles have no difference in the representation space after feature mapping (for example, they are mapped to the same representation position), making it impossible for the model to accurately identify fraudsters. In addition, existing models for processing multi-relational time series graphs process time information after aggregating multi-relational information, and process relational information and time information separately, which will lose some information to a certain extent, and the model cannot capture the joint dependency information of relation and time. Summary of the invention

[0004] In order to solve the above technical problems existing in the prior art, the present invention proposes a financial credit anti-fraud method, device and medium, and its specific technical solution is as follows:

[0005] A financial credit anti-fraud method comprises the following steps:

[0006] Step 1: Construct a multi-relational time series network in financial credit: collect credit data, treat applicants, sellers / merchants, lending institutions and other users as nodes in the network, and use their respective attributes as the characteristics of each node. The connections between nodes, such as social relationships, capital transactions and device dependencies, are used as edges, and the start and end times of each edge are recorded.

[0007] Step 2: Construct a multi-relation time slice network: In a network formed by different relations, divide the graph according to the time when the connection is formed to form a network under different time slices of each relation;

[0008] Step 3: Transform the attributes of all nodes in the network and map all nodes to the same feature space, so that the distance between node pairs in the original feature space is close to the distance between them in the transformed feature space;

[0009] Step 4: Using the converted features, for each relationship, in each time slice, learn the embedding expression of the timestamp corresponding to the time slice, use the neighbor aggregation method, add the encoding vector of each time slice obtained, and update the node information features;

[0010] Step 5: Use the gated recurrent late fusion model to fuse the features of different relations at different times to obtain the total features that fuse all relations and times;

[0011] Step 6: Using the feature vectors before and after the node attribute conversion and the node feature vector obtained by the gated loop late fusion model, perform joint learning of unsupervised embedding representation and supervised classification on the applicant's user node to determine whether there is a risk of fraud.

[0012] Furthermore, the step 2 is specifically as follows: according to the time when the connection in each relationship network in the multi-relationship time series network is formed, it is divided into every X days or Y months. If two nodes have a connection edge in this time period, it is considered that the two nodes have a connection edge in this time slice, thereby obtaining the network under different time slices of each network relationship.

[0013] Furthermore, the step three is specifically as follows:

[0014] The attribute characteristics of different nodes including applicants, merchants, lending institutions and other users are converted into a space of the same dimension q. The conversion formula is as follows:

[0015] X′=XW+B

[0016] H=σ(X′W′+B′)

[0017] where X∈R n×k represents the attribute feature matrix of all nodes, X′ represents the variables between the transformation processes of all nodes, n represents the total number of all nodes, k represents the initial dimension of all nodes, W∈R k×q , W′∈R q×q , B∈R n×q , B′∈R n×q Represent the transformation matrix and bias vector respectively, which are the parameters to be learned, σ(*) is the nonlinear activation function, H∈R n×qrepresents the transformed features of the node, where q represents the unified feature dimension of all types of nodes after feature transformation. It is necessary to satisfy that the distance between the node pairs in the original feature space is close to the distance in the transformed feature space, that is, the goal of the feature mapping for the node pair (i, j) is to minimize: Considering all nodes, the goal of this part of the feature transformation is to minimize the function:

[0018]

[0019] Where n represents the number of all nodes, h i 、h j Respectively represent the transformed features of node i and node j, x i 、x j Represent the original attribute features of node i and node j respectively.

[0020] Furthermore, the step 4 is specifically as follows:

[0021] Using the attention mechanism, by calculating the weight of each neighbor node, using the weighted average method to aggregate the neighbor node information, and embedding the absolute time, the representation encoding of each time slice is obtained.

[0022]

[0023] in is the M-dimensional vector encoding of the one-dimensional timestamp t, and α and β are the parameters to be learned;

[0024] Aggregate and update the node information features. The specific aggregation update formula is as follows:

[0025]

[0026]

[0027]

[0028]

[0029] in represents the feature of node j at time slice t under the rth relationship. The symbol [*,*] represents the vector concatenation operation. Indicates that node i is a neighbor node of node j at time slice t under the rth relationship, a r (t)∈R 2q are the parameters to be learned, is the weight of node i to node j at time slice t under the rth relationship, l is the number of stacked layers of the graph neural network, L is an integer, It is the last aggregated update information feature of node j at time slice t under the rth relationship.

[0030] Furthermore, the step five is specifically as follows:

[0031] First, an early gating mechanism is used to capture the contribution of each relationship to the final fusion state. For R relationships, the fusion gate of each relationship is calculated using the attention mechanism based on the feature matrix of all relationships. Use this fusion gate to control the input value of the LSTM unit under each relationship The specific formula is as follows:

[0032]

[0033]

[0034] Among them, W rr′ are the parameters to be trained, represents the encoding vector of the rth relation at time slice t after taking into account the fusion information of other relations, ⊙ represents the Cartesian product; S = {S 1 ,S 2 ,...,S R} is the node feature of the multi-relation time series graph obtained after neighbor information aggregation, where S r ={S r (1),S r (2),....,S r (T)} represents the data of T time slices under the rth relationship, is the feature matrix of the rth relationship of the nodes including applicants, sellers, lending institutions and other users in time slice t after feature transformation, []' represents the transpose of the matrix, n represents the number of all nodes, and R represents the number of types of relational temporal networks;

[0035] Then, the late gated recurrent unit is used. The late gated recurrent unit integrates the temporal dependency information within and between relations. The late gated recurrent unit has R LSTM units. Each LSTM unit independently controls the memory under each relation and calculates the cell state and hidden state of each relation under the current time slice.

[0036] Finally, add the final cell and latent variable outputs obtained under each relationship:

[0037]

[0038]

[0039]

[0040]

[0041]

[0042]

[0043]

[0044] Where f, i, and o represent the forget gate, input gate, and output gate in the LSTM unit, respectively; w, U, and b are the weight and offset parameters of the model, respectively; c and h represent the cell state and hidden state value of the model, respectively; the initial value at t = 0 is set to 0; g is the intermediate variable of the LSTM unit, and cell is one of the units of the LSTM;

[0045] Obtain the cell and hidden variable values ​​of all previous time and all relationship information in the current time slice, propagate the fused state to the next time step; finally, use the hidden variables obtained under time slice T as the final feature output of each node.

[0046] Furthermore, the step six is ​​specifically as follows:

[0047] According to the characteristics of each node finally obtained, the fully connected layer is used to classify the applicant user nodes, so as to achieve efficient anti-fraud effect. The loss function is as follows:

[0048]

[0049]

[0050] in is the sigmoid activation function, W and b are the transformation matrix and bias vector to be learned, respectively, y i is the real label of whether the i-th user is a fraudulent user, 0 or 1, Indicates the calculated probability of whether the i-th user is a fraudulent user: between 0 and 1, L cross_entropy Represents the cross entropy loss function of the model, then the final loss function of the model is:

[0051] L=L cross_entropy +λL distance

[0052] Where λ is a custom hyperparameter, balancing the supervised loss L cross_entropy With the unsupervised loss L distance Impact on model training.

[0053] A financial credit anti-fraud device includes one or more processors for implementing the financial credit anti-fraud method.

[0054] A computer-readable storage medium stores a program, which, when executed by a processor, implements the financial credit anti-fraud method.

[0055] The beneficial effects of the present invention are:

[0056] (1) The present invention proposes a feature conversion process, which constructs financial credit related data as a multi-relational time series graph and maps all types of nodes into a representation space. The distances between node pairs in the feature space remain consistent before mapping. This not only ensures the type information of the original features of the nodes input into the model, but also provides unsupervised learning loss information, so that a better detection model can be trained even when there is relatively little labeled data.

[0057] (2) The present invention uses a gated recurrent fusion model to synchronously fuse the features of different relationships at different times, obtain the overall features that fuse each relationship and time, and capture the time dependency information within and between relationships at the same time, so as to perform anti-fraud detection more efficiently and accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1a A schematic diagram of a multi-relationship time series network in an actual scenario of an embodiment of the present invention;

[0059] Figure 1b A schematic diagram of a multi-relation time-slicing network in an actual scenario of an embodiment of the present invention;

[0060] Figure 2 It is the overall flow chart of the entire algorithm model of the present invention;

[0061] Figure 3 This is a schematic diagram of a standard long short-term memory neural network unit model;

[0062] Figure 4 The present invention is a structural diagram of a financial credit anti-fraud device. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical scheme and technical effect of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings.

[0064] like Figure 2 As shown, a financial credit anti-fraud method comprises the following steps:

[0065] Step 1: If Figure 1aAs shown in the figure, a multi-relationship time series network in financial credit is constructed: first, all available data in the credit are collected. Based on the collected data, the applicants, sellers / merchants, banks and other lending institutions and other users in the credit are used as nodes in the network. There are social relationships such as friends and communications between all users including applicants. Banks and other lending institutions need to contact customers due to business needs. There are social relationships between users and lending institutions. There are transaction relationships between all users and sellers, etc. These relationships are used as connecting edges between nodes, and the connecting edges change with time. The start time and end time of each edge are recorded. The other users are all users except the applicant.

[0066] Step 2: If Figure 1b As shown, a multi-relation time slice network is constructed: in a network formed by different relations, the graph is divided according to the time when the connection is formed to form a network under different time slices of each relation.

[0067] According to the time when the connection in each relationship network in the multi-relationship time series network is formed, it is divided into every 20 days, 1 month or 2 months. If two nodes have had an edge in this time period, it is considered that the two nodes have an edge in this time slice, thereby obtaining the network under different time slices of each network relationship.

[0068] Step 3: Convert the attribute features of all nodes in the network: Map the attribute features of all nodes to the same feature space, so that the distance between node pairs in the original feature space is close to the distance in the converted feature space.

[0069] Specifically, the attribute characteristics of different nodes including applicants, merchants, banks and other lending institutions and other users are converted into the space of the same dimension q. The conversion formula is as follows:

[0070] X′=XW+B

[0071] H=σ(X′W′+B′)

[0072] where X∈R n×k represents the attribute feature matrix of all nodes, X′ represents the variables between the transformation processes of all nodes, n represents the total number of all nodes, k represents the initial dimension of all nodes, W∈R k×q , W′∈R q×q , B∈R n×q , B′∈R n×q Represent the transformation matrix and bias vector respectively, which are the parameters to be learned, σ(*) is the nonlinear activation function, H∈R n×qrepresents the transformed features of the node, where q represents the unified feature dimension of all types of nodes after feature transformation, which must satisfy the requirement that the distance between node pairs in the original feature space is close to the distance in the transformed feature space, that is, the goal of feature mapping for node pair (i, j) is to minimize: Considering all nodes, the goal of this part of the feature transformation is to minimize the function:

[0073]

[0074] Where n represents the number of all nodes, h i 、h j Respectively represent the features of nodes i and j after transformation, x i 、x j Represent the original attribute features of nodes i and j respectively.

[0075] Step 4: Using the converted features, for each relationship, in each time slice, learn the embedding expression of the timestamp corresponding to the time slice, use the neighbor aggregation method, add the encoding vector of each time slice obtained, and update the node information features;

[0076] Specifically, the attention mechanism is used to calculate the weight of each neighbor node, aggregate the neighbor node information using the weighted average method, and embed the absolute time to obtain the representation encoding of each time slice.

[0077]

[0078] in is the M-dimensional vector encoding of the one-dimensional timestamp t, and α, β are the parameters to be learned;

[0079] Update the node information features. The specific aggregation update formula is as follows:

[0080]

[0081]

[0082]

[0083]

[0084] in represents the feature of node j at time slice t under the rth relationship. The symbol [*,*] represents the vector concatenation operation. Indicates that node i is a neighbor node of node j at time slice t under the rth relationship, a r (t)∈R 2q are the parameters to be learned, is the weight of node i to node j at time slice t under the rth relationship, l is the number of stacked layers of the graph neural network, L is an integer, It is the last aggregated update information feature of node j at time slice t under the rth relationship.

[0085] Step 5: Use the gated recurrent late fusion model to fuse the features of different relations at different times to obtain the total features that fuse all relations and times;

[0086] Specifically, when fusing the features of different relations at different times, S = {S 1 ,S 2 ,...,S R} is the node feature of the multi-relation time series graph obtained after neighbor information aggregation, where S r ={S r (1),S r (2),....,S r (T)} represents the data of all T time slices under the rth relationship, is the feature matrix of the rth relationship of all nodes (applicants, sellers, banks and other lending institutions and other users) in time slice t after feature transformation, []' represents the transpose of the matrix, n represents the number of all nodes, and R represents the number of types of relational temporal networks.

[0087] First, an early gating mechanism is used to capture the contribution of each relationship to the final fusion state. For R relationships, the fusion gate of each relationship is calculated using the attention mechanism based on the feature matrix of all relationships. Use this fusion gate to control the input value of the LSTM unit under each relationship The specific formula is as follows:

[0088]

[0089]

[0090] Among them, W rr′ are the parameters to be trained, represents the encoding vector of the rth relation at time slice t after taking into account the fusion information of other relations, ⊙ represents the Cartesian product;

[0091] Then, the late gated recurrent unit is used. The late gated recurrent unit integrates the temporal dependency information within and between relations. The late gated recurrent unit has R LSTM units, such as Figure 3 As shown in the figure, each LSTM unit independently controls the memory of each relationship and calculates the cell state and hidden state of each relationship at the current time slice;

[0092] Finally, add the final cell and latent variable outputs obtained under each relationship:

[0093]

[0094]

[0095]

[0096]

[0097]

[0098]

[0099]

[0100] Where f, i, and o represent the forget gate, input gate, and output gate in the LSTM unit, respectively; w, U, and b are the weight and offset parameters of the model, respectively; c and h represent the cell state and hidden state value of the model, respectively; the initial value at t = 0 is set to 0; g is the intermediate variable of the LSTM unit, and cell is one of the units of the LSTM;

[0101] Obtain the cell and hidden variable values ​​of the current time slice that aggregate all previous time and all relationship information, and propagate the fused state to the next time step; finally, use the hidden variables obtained at the last time slice T as the final feature output of each node.

[0102] Step 6: Using the feature vectors before and after the node attribute conversion and the node feature vector obtained by the gated loop late fusion model, perform joint learning of unsupervised embedding representation and supervised classification on the applicant's user node to determine whether there is a risk of fraud.

[0103] Specifically, according to the characteristics of each node finally obtained, the fully connected layer is used to classify the applicant user nodes, so as to achieve an efficient anti-fraud effect. The loss function is as follows:

[0104]

[0105]

[0106] in is the sigmoid activation function, W and b are the transformation matrix and bias vector to be learned, respectively, y i is the real label of whether the i-th user is a fraudulent user, 0 or 1, Indicates the calculated probability of whether the i-th user is a fraudulent user: between 0 and 1, L cross_entropyRepresents the cross entropy loss function of the model, then the final loss function of the model is:

[0107] L=L cross_entropy +λL distance

[0108] Where λ is a custom hyperparameter, balancing the supervised loss L cross_entropy With the unsupervised loss L distance Impact on model training. For example, when there is a lack of sufficient prior knowledge, the cost of manually labeling data is high, and there is relatively little labeled data in the early stage, a larger λ value can be set, and the model is trained more using unsupervised loss functions.

[0109] Corresponding to the aforementioned embodiment of a financial credit anti-fraud method, the present invention also provides an embodiment of a financial credit anti-fraud device.

[0110] See also Figure 4 A financial credit anti-fraud device provided in an embodiment of the present invention includes one or more processors for implementing a financial credit anti-fraud method in the above embodiment.

[0111] An embodiment of a financial credit anti-fraud device of the present invention can be applied to any device with data processing capabilities, and the device with data processing capabilities can be a device or apparatus such as a computer. The device embodiment can be implemented through software, or through hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, the processor of any device with data processing capabilities reads the corresponding computer program instructions in the non-volatile memory into the memory and runs them. From the hardware level, if Figure 4 As shown, it is a hardware structure diagram of a financial credit anti-fraud device of the present invention, in which any device with data processing capability is located, except Figure 4 In addition to the processor, memory, network interface, and non-volatile memory shown, any device with data processing capabilities in which the apparatus in the embodiments is located may also include other hardware, generally based on the actual functions of the device with data processing capabilities, which will not be described in detail.

[0112] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0113] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of the present invention. Ordinary technicians in this field can understand and implement it without paying creative work.

[0114] An embodiment of the present invention further provides a computer-readable storage medium on which a program is stored. When the program is executed by a processor, a financial credit anti-fraud method in the above embodiment is implemented.

[0115] The computer-readable storage medium may be an internal storage unit of any device with data processing capability described in any of the aforementioned embodiments, such as a hard disk or a memory. The computer-readable storage medium may also be an external storage device of a wind turbine, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. equipped on the device. Furthermore, the computer-readable storage medium may also include both an internal storage unit and an external storage device of any device with data processing capability. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capability, and may also be used to temporarily store data that has been output or is to be output.

[0116] The above is only a preferred implementation case of the present invention and does not limit the present invention in any form. Although the implementation process of the present invention is described in detail above, for those familiar with the art, they can still modify the technical solutions recorded in the above examples, or replace some of the technical features therein with equivalents. All modifications, equivalent replacements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A financial credit anti-fraud method, characterized in that: The following steps are involved: Step 1: Construct a multi-relational time series network in financial credit: collect credit data, treat applicants, sellers / merchants, lending institutions and other users as nodes in the network, and use their respective attributes as the characteristics of each node. The connections between nodes, such as social relationships, capital transactions and device dependencies, are used as edges, and the start and end times of each edge are recorded. Step 2: Construct a multi-relation time slice network: In a network formed by different relations, divide the graph according to the time when the connection is formed to form a network under different time slices of each relation; Step 3: Transform the attributes of all nodes in the network and map all nodes to the same feature space, so that the distance between node pairs in the original feature space is close to the distance between them in the transformed feature space; Step 4: Using the converted features, for each relationship, in each time slice, learn the embedding expression of the timestamp corresponding to the time slice, use the neighbor aggregation method to add the encoding vector of each time slice, and update the node information features; Step 5: Use the gated recurrent late fusion model to fuse the features of different relations at different times to obtain the total features that fuse all relations and times; Step 6: Using the feature vectors before and after the node attribute conversion and the node feature vectors obtained by the gated loop late fusion model, perform joint learning of unsupervised embedding representation and supervised classification on the applicant's user node to determine whether it has fraud risk; The step three is specifically as follows: The attribute characteristics of different nodes including applicants, merchants, lending institutions and other users are converted into a space of the same dimension q. The conversion formula is as follows: X′=XW+B H=σ(X′W′+B′) where X∈R n×k represents the attribute feature matrix of all nodes, X′ represents the variables between the transformation processes of all nodes, n represents the total number of all nodes, k represents the initial dimension of all nodes, W∈R k×q , W′∈R q×q , B∈R n×q , B′∈R n×q Represent the transformation matrix and bias vector respectively, which are the parameters to be learned, σ(*) is the nonlinear activation function, H∈R n×q represents the features after node transformation, where q represents the unified feature dimension of all types of nodes after feature transformation. It is necessary to satisfy that the distance between node pairs in the original feature space is close to the distance in the transformed feature space, that is, the goal of feature mapping for node pair (i, j) is to minimize: Considering all nodes, the goal of this part of the feature transformation is to minimize the function: Where n represents the number of all nodes, h i 、h j Respectively represent the transformed features of node i and node j, x i 、x j Represent the original attribute features of node i and node j respectively; The step 4 is specifically as follows: Using the attention mechanism, by calculating the weight of each neighbor node, using the weighted average method to aggregate the neighbor node information, and embedding the absolute time, the representation encoding of each time slice is obtained. in is the M-dimensional vector encoding of the one-dimensional timestamp t, and α and β are the parameters to be learned; Aggregate and update the node information features. The specific aggregation update formula is as follows: in represents the feature of node j at time slice t under the rth relationship. The symbol [*,*] represents the vector concatenation operation. Indicates that node i is a neighbor node of node j at time slice t under the rth relationship, a r (t)∈R 2q are the parameters to be learned, is the weight of node i to node j at time slice t under the rth relationship, l is the number of stacked layers of the graph neural network, L is an integer, is the last aggregated update information feature of node j at time slice t under the rth relationship; The step five is specifically as follows: First, an early gating mechanism is used to capture the contribution of each relationship to the final fusion state. For R relationships, the fusion gate of each relationship is calculated using the attention mechanism based on the feature matrix of all relationships. Use this fusion gate to control the input value of the LSTM unit under each relationship The specific formula is as follows: Among them, W rr′ are the parameters to be trained, represents the encoding vector of the rth relation at time slice t after taking into account the fusion information of other relations, ⊙ represents the Cartesian product; S = {S 1 ,S 2 ,...,S R } is the node feature of the multi-relation time series graph obtained after aggregating neighbor information, where S r ={S r (1),S r (2),....,S r (T)} represents the data of T time slices under the rth relationship, ∈R n×q is the feature matrix of the rth relationship of the nodes including applicants, sellers, lending institutions and other users in time slice t after feature transformation, [ ]' represents the transpose of the matrix, n represents the number of all nodes, and R represents the number of types of relational temporal networks; Then, the late gated recurrent unit is used. The late gated recurrent unit integrates the temporal dependency information within and between relations. The late gated recurrent unit has R LSTM units. Each LSTM unit independently controls the memory under each relation and calculates the cell state and hidden state of each relation under the current time slice. Finally, add the final cell and latent variable outputs obtained under each relationship: Where f, i, and o represent the forget gate, input gate, and output gate in the LSTM unit, respectively; w, U, and b are the weight and offset parameters of the model, respectively; c and h represent the cell state and hidden state value of the model, respectively; the initial value at t = 0 is set to 0; g is the intermediate variable of the LSTM unit, and cell is one of the units of the LSTM; Obtain the cell and hidden variable values ​​of all previous time and all relationship information in the current time slice, propagate the fused state to the next time step; finally, use the hidden variables obtained under time slice T as the final feature output of each node.

2. A financial credit anti-fraud method as claimed in claim 1, characterized in that: The step 2 is specifically as follows: according to the time when the connection in each relationship network in the multi-relationship time series network is formed, the network is divided into every X days or Y months. If two nodes have a connection edge in the time period, it is considered that the two nodes have a connection edge in the time slice, thereby obtaining the network under different time slices of each network relationship.

3. A financial credit anti-fraud method as claimed in claim 1, characterized in that: The step six is ​​specifically as follows: According to the characteristics of each node finally obtained, the fully connected layer is used to classify the applicant user nodes, so as to achieve efficient anti-fraud effect. The loss function is as follows: in is the sigmoid activation function, W and b are the transformation matrix and bias vector to be learned, respectively, y i is the real label of whether the i-th user is a fraudulent user, 0 or 1, Indicates the calculated probability of whether the i-th user is a fraudulent user: between 0 and 1, L cross_entropy Represents the cross entropy loss function of the model, then the final loss function of the model is: L=L cross_entropy +λL distance Where λ is a custom hyperparameter, balancing the supervised loss L cross_entropy With the unsupervised loss L distance Impact on model training.

4. A financial credit anti-fraud device, characterized in that: It comprises one or more processors, and is used to implement a financial credit anti-fraud method as claimed in any one of claims 1 to 3.

5. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, a financial credit anti-fraud method as claimed in any one of claims 1 to 3 is implemented.

Images