Abnormal account detection method and device based on graph network, equipment and medium
By generating transaction graph networks from multiple perspectives based on graph network, training and adjusting models, the problem of low detection accuracy of abnormal accounts in the prior art is solved, and higher detection accuracy is achieved.
Patent Information
- Application Number
- CN202411053301.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-01
- Publication Date
- 2024-11-26
AI Technical Summary
When detecting abnormal accounts, the transaction information is analyzed from a single perspective, making it difficult to obtain comprehensive information of the account, resulting in low detection accuracy, especially when abnormal accounts conduct normal transactions at the same time.
A graph network-based method is used to generate multiple transaction graph networks from different perspectives. By training the initial model and adjusting the model parameters until the training cut-off conditions are met, an abnormal account detection model is obtained, so that the set of accounts to be tested is detected.
Through the transaction network diagram from multiple perspectives, the accuracy of abnormal account detection can be improved, and the accuracy rate can be increased by more than 10%.
Smart Images

Figure CN119026047A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of data processing, and in particular to a method, device, equipment and medium for detecting abnormal accounts based on a graph network. Background Art
[0002] With the continuous development and promotion of science and technology, the means for criminals to commit fraud or other illegal operations are endless, but these illegal means all involve the flow of funds, and funds will circulate in accounts. Therefore, the account transaction information can be used to detect and identify the accounts used by criminals, that is, abnormal accounts.
[0003] However, the transactions involved in the accounts are very complex. At present, the transaction information used for account detection is transaction information from a single perspective. The mining of account transaction information is not sufficient, and it is difficult to obtain more comprehensive information about the accounts. The accuracy of abnormal account detection is also low. Summary of the invention
[0004] The embodiments of the present application provide a method, apparatus, device and medium for detecting abnormal accounts based on a graph network, which can improve the accuracy of abnormal account detection.
[0005] In a first aspect, an embodiment of the present application provides an abnormal account detection method based on a graph network, comprising: generating N transaction graph networks of an account sample set according to the transaction relationship, transaction characteristics and transaction type of the sample accounts in the account sample set, the sample accounts having labels, the labels representing normal accounts or abnormal accounts, one transaction graph network corresponding to one transaction type, each transaction graph network comprising nodes and edges, the nodes representing the sample accounts, the nodes having the transaction characteristics of the sample accounts, the edges representing the transaction relationship, and N being an integer greater than 1; inputting the N transaction graph networks into an initial model to obtain abnormal prediction results of the sample accounts, the initial model comprising a preliminary mapping sub-model, a feature update sub-model and a result determination sub-model; obtaining a loss function based on the abnormal prediction results, labels and mapping characteristics of the sample accounts as nodes in the N transaction graph networks; if the loss function does not meet the training cutoff condition, adjusting the model parameters of the initial model until the loss function meets the training cutoff condition, and obtaining an abnormal account detection model; inputting the N transaction graph networks obtained based on the account set to be tested into the abnormal account detection model to obtain abnormal accounts in the account set to be tested output by the abnormal account detection model.
[0006] In a second aspect, an embodiment of the present application provides an abnormal account detection device based on a graph network, including: a graph network generation module, which is used to generate N transaction graph networks of an account sample set according to the transaction relationship, transaction characteristics and transaction type of the sample accounts in the account sample set, the sample accounts have labels, the labels represent normal accounts or abnormal accounts, one transaction graph network corresponds to one transaction type, each transaction graph network includes nodes and edges, the nodes represent the sample accounts, the nodes have the transaction characteristics of the sample accounts, the edges represent the transaction relationship, and N is an integer greater than 1; a model processing module, which is used to input the N transaction graph networks into an initial model to obtain the sample accounts. The abnormal prediction results of this account, the initial model includes a preliminary mapping sub-model, a feature update sub-model and a result judgment sub-model; and, based on the abnormal prediction results of the sample account, the label and the mapping features of the sample account as a node in N transaction graph networks, a loss function is obtained; and, if the loss function does not meet the training cutoff condition, the model parameters of the initial model are adjusted until the loss function meets the training cutoff condition to obtain an abnormal account detection model; and, based on the N transaction graph networks obtained based on the account set to be tested, the N transaction graph networks are input into the abnormal account detection model to obtain the abnormal accounts in the account set to be tested output by the abnormal account detection model.
[0007] In a third aspect, an embodiment of the present application provides an abnormal account detection device based on a graph network, comprising: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the abnormal account detection method based on a graph network of the first aspect is implemented.
[0008] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having computer program instructions stored thereon. When the computer program instructions are executed by a processor, the graph network-based abnormal account detection method of the first aspect is implemented.
[0009] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the abnormal account detection method based on a graph network of the first aspect.
[0010] The embodiments of the present application provide a method, device, equipment and medium for detecting abnormal accounts based on a graph network, which can generate N transaction graph networks under different perspectives corresponding to N types of transactions according to the transaction information of account samples in an account sample set. The initial model is trained using the transaction graph network under multiple perspectives, and the model parameters of the model are determined to continue iterative training according to the loss function of the trained model until the loss function meets the training cutoff condition, and an abnormal account detection model is obtained, and the N transaction network graphs under different perspectives corresponding to the test account set are processed using the abnormal account detection model to obtain the test account that is an abnormal account in the test account set. The transaction network graphs under multiple perspectives can be used to more deeply explore the relationship information contained in the transaction graph network of accounts under multiple dimensions, and the abnormal account detection model can extract more comprehensive information from the transaction graph network of the test account set under multiple different perspectives, and perform more accurate detection, thereby improving the accuracy of abnormal account detection. BRIEF DESCRIPTION OF THE DRAWINGS
[0011] In order to more clearly illustrate the technical solution of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.
[0012] Figure 1 A flowchart of an abnormal account detection method based on a graph network provided in one embodiment of the present application;
[0013] Figure 2 A schematic diagram of an example of a transaction graph network of transfer types of a sample account set provided in an embodiment of the present application;
[0014] Figure 3 A schematic diagram of an example of a transaction graph network of consumption types of a sample account set provided in an embodiment of the present application;
[0015] Figure 4 A schematic diagram of an example of a model framework for detecting sample accounts provided in an embodiment of the present application;
[0016] Figure 5 A flowchart of an abnormal account detection method based on a graph network provided in another embodiment of the present application;
[0017] Figure 6 A flowchart of an abnormal account detection method based on a graph network provided in yet another embodiment of the present application;
[0018] Figure 7 A schematic diagram of the structure of an abnormal account detection device based on a graph network provided in one embodiment of the present application;
[0019] Figure 8 A schematic diagram of the structure of an abnormal account detection device based on a graph network provided in one embodiment of the present application. DETAILED DESCRIPTION
[0020] The features and exemplary embodiments of various aspects of the present application will be described in detail below. In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are intended only to explain the present application, rather than to limit the present application. For those skilled in the art, the present application can be implemented without the need for some of these specific details. The following description of the embodiments is only to provide a better understanding of the present application by showing examples of the present application. It should be noted that the acquisition, storage, use, and processing of information and data in the embodiments of the present application are authorized by users or relevant agencies, in compliance with the relevant provisions of national laws and regulations.
[0021] With the continuous development and promotion of science and technology, criminals have come up with endless means to commit fraud or other illegal operations. However, these illegal means all involve capital transactions, and the funds will circulate in the account. Therefore, the account transaction information can be used to detect and identify the accounts used by criminals, that is, abnormal accounts. However, the transactions involved in the account are very complex. At present, the transaction information used for account detection is transaction information from a single perspective. The mining of account transaction information is not sufficient, and it is difficult to obtain more comprehensive information about the account. The accuracy of abnormal account detection is also low. For example, in many cases, in addition to abnormal transactions, abnormal accounts will also make normal consumption. The transaction information of normal consumption will affect the detection results during the detection of abnormal accounts, and abnormal accounts may be misjudged as normal accounts.
[0022] The present application provides a method, device, equipment, medium and program product for detecting abnormal accounts based on a graph network. Starting from the transaction type of the account's transaction, a transaction graph network of the account under multiple different perspectives can be generated. The transaction graph network under each perspective corresponds to a transaction type, thereby mining the relationship information contained in the transaction graph network of the account under multiple dimensions to a greater extent. A transaction graph network training model of accounts under multiple different perspectives is used to obtain an abnormal account detection model that can detect whether an account is an abnormal account. The abnormal account detection model obtained by training can be used to detect the set of accounts to be tested. The abnormal account detection model can extract more comprehensive information from the transaction graph network of the set of accounts to be tested under multiple different perspectives, and perform more accurate detection, thereby improving the accuracy of abnormal account detection.
[0023] The following describes the graph network-based abnormal account detection method, device, equipment, medium and program product provided by the present application.
[0024] The first aspect of the present application provides an abnormal account detection method based on a graph network, which can be applied to the scenario of detecting abnormal accounts. Abnormal accounts may include but are not limited to accounts involved in illegal activities such as telecommunications fraud and pyramid schemes. The accounts in the embodiments of the present application may include card accounts such as bank cards, and may also include non-card accounts such as electronic wallets, which are not limited here. The abnormal account detection method based on a graph network can be executed by an abnormal account detection device, equipment, etc. based on a graph network, which is not limited here. Figure 1 A flowchart of an abnormal account detection method based on a graph network provided in an embodiment of the present application is shown in FIG. Figure 1 As shown, the abnormal account detection method based on graph network may include steps S101 to S105.
[0025] In step S101, N transaction graph networks of the account sample set are generated according to the transaction relationships, transaction features and transaction types of the sample accounts in the account sample set.
[0026] The account sample set includes a plurality of sample accounts, and the sample accounts have labels, and the labels represent normal accounts or abnormal accounts. That is, the sample accounts include accounts known to be normal accounts or abnormal accounts. In some examples, there are both normal accounts and abnormal accounts in the sample accounts.
[0027] The transaction relationship of sample accounts may include direct transaction relationship and indirect transaction relationship. For example, if there is a direct flow of resources, such as capital flow, between two sample accounts, it can be considered that there is a direct transaction relationship between the two sample accounts; if two sample accounts conduct transactions with the same merchant, it can be considered that there is an indirect transaction relationship between the two sample accounts.
[0028] The transaction characteristics of the sample accounts can be obtained based on the original transaction data of the sample accounts. The original transaction data of the sample accounts can be cleaned up first, such as removing abnormal data, supplementing missing data, etc. Then, transaction characteristics can be constructed from multiple dimensions based on the original transaction data. For example, transaction characteristics related to abnormal transactions can be constructed from dimensions such as transaction time, transaction channel, and number of transactions. In some examples, transaction characteristics may include but are not limited to the total number of transactions, the number of nighttime transactions, the amount of nighttime transaction resources, the number of transactions in a specific channel, etc. The transaction type can characterize the transaction method. For example, the transaction type may include but is not limited to the transfer type and the consumption type. The transfer type characterizes the transaction as a transfer method, and the consumption type characterizes the transaction as a transaction method between an account and a merchant; further, the transfer type can be further subdivided into multiple types, and the consumption type can be further subdivided into multiple types. Transaction features can be constructed separately under different transaction types. For example, if the transaction type is a transfer type, the transaction features may include but are not limited to the total number of transfers, the number of night transfers, the amount of night transfer resources, the number of ATM (Automated Teller Machine) withdrawals, etc.; if the transaction type is a consumption type, the transaction features may include but are not limited to the total number of consumptions, the number of large-amount consumptions, the amount of large-amount consumption resources, the number of consumptions at merchants of designated categories, etc.
[0029] Based on the above information of the sample accounts in the account sample set, N transaction graph networks are generated, where N is an integer greater than 1. One transaction graph network corresponds to one transaction type, that is, the sample account has transactions of N transaction types. Each transaction graph network of the account sample set can reflect the transaction characteristics of the sample accounts themselves in the account sample set under one transaction type condition and the relationship between the sample accounts. Each transaction graph network includes nodes and edges. The nodes represent the sample accounts, the nodes have the transaction characteristics of the sample accounts, and the edges represent the transaction relationships. The transaction network graph can be represented by G=(V,E r ,X,C) represents, where G is the transaction network graph; V is the node; E r is the edge corresponding to transaction type r. If transaction type r represents transfer, then G is the transaction network graph corresponding to the transfer type. If transaction type r represents consumption, then G is the transaction network graph corresponding to the consumption type. X is the transaction feature of the node in the transaction network graph, X = {x 1 ,…,x n}∈R d , x i is the transaction feature of the ith node in the transaction network graph; R d Represents a feature with dimension d; C is the label of the sample account, which can also be considered as the label of the node; in some examples, the dimension d of the feature is smaller than the number of account samples in the account sample set, and further, the dimension d of the feature is much smaller than the number of account samples in the account sample set.
[0030] In some examples, nodes representing sample accounts can be generated based on the transaction characteristics of the sample accounts in each transaction type; the transaction relationships of the sample accounts in transactions of each transaction type are obtained; and edges are set between the nodes of the sample accounts that have transaction relationships in transactions of the same transaction type to form N transaction graph networks corresponding to N transaction types.
[0031] The node is connected to the transaction features of the corresponding sample account in the transaction type. The corresponding relationship can be established, and the transaction features can also be stored in the node. The way of establishing the connection between the node and the transaction features is not limited here. In the case of different types of transactions, other sample accounts having a transaction relationship with the same sample account may be different. For example, in the case of the transfer type, other sample accounts having a transaction relationship with the sample account A1 may include sample accounts A2 and A3, and in the case of the consumption type, other sample accounts having a transaction relationship with the sample account A1 may include sample accounts A3, A4 and A5. For a transaction type, edges are set between the nodes representing the sample accounts having a transaction relationship under the transaction of the transaction type to form a transaction graph network of the transaction type of the account sample set. If N = 2, two transaction graph networks can be formed for an account sample set; if N = 3, three transaction graph networks can be formed for an account sample set. Each transaction graph network of the account sample set is an isomorphic graph of the sample account under a transaction type. The transaction graph networks are generated according to the transaction type, which can make the propagation of transaction anomalies in the transaction graph network more concentrated and targeted.
[0032] In some examples, when the transaction type is a transfer type, the sample account that initiates the transfer has a transaction relationship with the sample account that receives the transfer. That is, when the transaction type is a transfer type, the nodes of the sample accounts that have a transfer relationship are connected by edges. For example, Figure 2 A schematic diagram of an example of a transaction graph network of a transfer type of a sample account set provided in an embodiment of the present application, such as Figure 2 As shown in the figure, the dots in the transaction graph network are nodes representing sample accounts, and the connecting lines between nodes are edges. Figure 2There has been a transfer transaction between any two sample accounts represented by nodes connected by edges. In the case where the transaction type is a transfer type, the sample accounts that conduct transactions with the same merchant that meet the preset abnormal conditions have a transaction relationship. The preset abnormal conditions may include transaction conditions with abnormal transaction risks, which can be set according to scenarios, needs, experience, etc., and are not limited here. For example, the preset abnormal conditions may include but are not limited to one or more of the following: large-amount consumption in a specified type of merchant in a short period of time, small-amount consumption tests in a small merchant in a short period of time, consumption in a risk merchant that conducts transactions with multiple abnormal accounts in a short period of time, etc. In the case where the transaction type is a consumption type, the nodes of the sample accounts that consume at the same merchant and meet the preset abnormal conditions are connected by edges. Specifically, the nodes of the sample accounts can be connected to the nodes of the merchant where the consumption occurs first, and then the nodes of the sample accounts that consume at the same merchant and meet the preset abnormal conditions can be connected by edges, and the nodes of the merchant and the edges connected to the merchant are deleted, that is, the merchant is shielded from the transaction graph network, thereby obtaining the transaction graph network of the consumption type of the sample account set. In some examples, the edges set between nodes may be different depending on the preset abnormal conditions met. Figure 3 A schematic diagram of an example of a transaction graph network of consumption types of a sample account set provided in an embodiment of the present application, such as Figure 3 As shown, the dots in the transaction graph network are nodes representing sample accounts, and the connecting lines between nodes are edges. The transaction relationships between some nodes meet the preset abnormal condition 1, the transaction relationships between some nodes meet the preset abnormal condition 2, and the transaction relationships between some nodes meet the preset abnormal condition 3. Figure 3 The edges corresponding to different preset abnormal conditions may be different.
[0033] In step S102, N transaction graph networks are input into the initial model to obtain abnormal prediction results of sample accounts.
[0034] The initial model includes a preliminary mapping sub-model, a feature update sub-model and a result determination sub-model. The preliminary mapping sub-model can map the transaction features of nodes in N transaction graph networks to the feature space to obtain initial mapping features. The same sample account can be mapped to N nodes in N transaction graph networks to obtain N initial mapping features. The feature update sub-model can combine the initial mapping features of the node itself and the initial mapping features of neighboring nodes connected by edges to update the initial mapping features of the node itself to obtain mapping features. The same sample account can correspond to N mapping features. The result determination sub-model can output anomaly prediction results based on the N mapping features corresponding to each sample account. The anomaly prediction results of the sample account can characterize whether the model predicts that the sample account is a normal account or an abnormal account.
[0035] For example, Figure 4A schematic diagram of an example of a model framework for detecting sample accounts provided in an embodiment of the present application, such as Figure 4 As shown, the transaction network including both transfer relations and consumption relations can be split into transaction graph network 1 and transaction graph network 2 according to the transaction type. The transaction type corresponding to transaction graph network 1 is the transfer type, and the transaction type corresponding to transaction graph network 2 is the consumption type. The nodes in transaction graph network 1 and the nodes in transaction graph network 2 are feature mapped respectively to obtain mapping feature 1 and mapping feature 2. Mapping feature 1 corresponds to the nodes in transaction graph network 1, and mapping feature 2 corresponds to the nodes in transaction graph network 2. Mapping feature 1 and mapping feature 2 are fused and then input into the classifier to obtain the abnormal prediction results of each sample account.
[0036] In step S103, a loss function is obtained based on the abnormal prediction results and labels of the sample accounts and the mapping features of the sample accounts as nodes in the N transaction graph networks.
[0037] The label of the sample account represents whether the sample account is actually a normal account or an abnormal account. The abnormal prediction result of the sample account may be consistent with the label of the sample account, or it may be inconsistent. The loss function can be obtained by combining the abnormal prediction result of the sample account, the label, and the mapping characteristics of the nodes of the sample account in the N transaction graph networks. The loss function can reflect the difference between the abnormal prediction result of the model and the actual situation.
[0038] In step S104, if the loss function does not meet the training cutoff condition, the model parameters of the initial model are adjusted until the loss function meets the training cutoff condition to obtain an abnormal account detection model.
[0039] The training cut-off condition can be set according to the scenario, demand, experience, etc. For example, the training cut-off condition may include but is not limited to the condition that the loss function is within the expected range. If the loss function does not meet the training cut-off condition, it means that the detection effect of the trained model has not reached the expectation, and further training is required. The model parameters of the initial model can be adjusted, and the above training process is repeated until the loss function meets the training cut-off condition, and the trained model is determined as the abnormal account detection model. The model parameters of the initial model may include at least a part of the model parameters of the preliminary mapping sub-model, the model parameters of the feature update sub-model, and the model parameters of the result determination sub-model. The abnormal account detection model includes a preliminary mapping sub-model, a feature update sub-model, and a result determination sub-model. Of course, the preliminary mapping sub-model, the feature update sub-model, and the result determination sub-model in the abnormal account detection model are all trained sub-models. The abnormal account detection model can be used to detect the account to be tested in the set of accounts to be tested, and determine whether the account to be tested is a normal account or an abnormal account.
[0040] In step S105, N transaction graph networks obtained based on the account set to be tested are input into an abnormal account detection model to obtain abnormal accounts in the account set to be tested output by the abnormal account detection model.
[0041] The method for generating the N transaction graph networks corresponding to the account set to be tested is the same as the method for generating the N transaction graph networks corresponding to the sample account set in the above embodiment. For specific content, please refer to the relevant content in the above embodiment, which will not be repeated here.
[0042] The abnormal account detection model processes the N transaction graph networks corresponding to the set of accounts to be tested in the same manner as the initial model processes the N transaction graph networks corresponding to the account sample set in the above embodiment. The specific content of the processing of the N transaction graph networks corresponding to the account sample set by the initial model in the embodiment of the present application can be referred to, and will not be repeated here. The abnormal account detection model can output the abnormal detection results of the accounts to be tested in the set of accounts to be tested, and the abnormal detection results of the accounts to be tested indicate whether the accounts to be tested are normal accounts or abnormal accounts.
[0043] In the embodiment of the present application, N transaction graph networks under different perspectives corresponding to N types of transactions can be generated according to the transaction information of the account samples in the account sample set. The initial model is trained using the transaction graph network under multiple perspectives, and the model parameters of the model are determined to continue iterative training according to the loss function of the trained model until the loss function meets the training cutoff condition, and the abnormal account detection model is obtained, so as to use the abnormal account detection model to process the N transaction network graphs under different perspectives corresponding to the test account set, and obtain the test account that is an abnormal account in the test account set. Through the transaction network graphs under multiple perspectives, the relationship information contained in the transaction graph network of the account under multiple dimensions can be mined to a greater extent. The abnormal account detection model can extract more comprehensive information from the transaction graph network of the test account set under multiple different perspectives, and perform more accurate detection, thereby improving the accuracy of abnormal account detection. After verification, the abnormal account detection method based on the graph network provided in the embodiment of the present application can improve the accuracy by more than 10% compared with the abnormal account detection method in the related art.
[0044] In some embodiments, the preliminary mapping sub-model can perform feature mapping on the transaction graph network, the feature updating sub-model can perform feature updating, and the result determination sub-model can predict whether an account is an abnormal account. Figure 5 A flowchart of an abnormal account detection method based on a graph network provided in another embodiment of the present application is provided. Figure 5 and Figure 1 The difference is that Figure 1 Step S102 in the above example can be specifically broken down into the following steps: Figure 5 Steps S1021 to S1023 in .
[0045] In step S1021, N transaction graph networks are input into a preliminary mapping sub-model, and the transaction features of the nodes in the N transaction graph networks are mapped to the same feature space using model parameters through the preliminary mapping sub-model to obtain initial mapping features of the nodes in the N transaction graph networks.
[0046] The preliminary mapping sub-model can map the transaction features of the input nodes to the latent space through a projection layer. The latent space is a low-dimensional, abstract representation space. The initial mapping features mapped to the latent space can contain the potential features of the data. Each node in each transaction graph network can obtain the initial mapping features. For example, the initial mapping features obtained by mapping the transaction features to the feature space can be shown as follows (1):
[0047]
[0048] in, is the initial mapping feature of the i-th node in a transaction graph network; k is the order of the convolutional layer in the preset model; σ() is the ReLU activation function; is the transaction feature of the i-th node in a transaction graph network; W 1 and b 1 are all model parameters, and the model parameters are learnable model parameters.
[0049] In step S1022, the initial mapping features of the node are input into the feature updating sub-model, and the initial mapping features of the node are updated by the feature updating sub-model using the model parameters and the initial mapping features of the node's neighbor nodes to obtain the mapping features of the node.
[0050] Neighbor nodes are connected to nodes through edges, and other nodes that are directly connected to a node through edges are neighbor nodes of the node. There is a transaction relationship between neighbor nodes and nodes. Correspondingly, neighbor nodes can reflect the characteristics of nodes to a certain extent. The initial mapping features of the neighbor nodes around the node can be aggregated by the message propagation paradigm based on the graph neural network to enhance the initial mapping features of the node and obtain the mapping features of the node. The mapping features of the node can better reflect the characteristics of the node in the anomaly detection scenario, and the mapping features of the node can also improve the performance of the model for the training of the initial model.
[0051] In some examples, the model parameters of the feature update sub-model may include attention parameters, update parameters, and aggregation parameters, and the attention parameters, update parameters, and aggregation parameters are all learnable model parameters. The attention weight of the neighbor node can be calculated based on the initial mapping feature of the node, the initial mapping feature of the neighbor node, the dimension of the transaction feature, and the attention parameter; the first sum of the product of the initial mapping feature, the attention weight, and the aggregation parameter of each neighbor node is calculated; the first product of the initial mapping feature of the node and the update parameter is calculated, and the sum of the first product and the first sum is determined as the mapping feature of the node.
[0052] The attention parameters may include the attention parameters of the node and the attention parameters of the neighboring nodes. The normalized exponential function may be used to process the dimensions of the node's attention parameters, the node's initial mapping features, the neighboring node's attention parameters, the neighboring node's initial mapping features, and the transaction features to obtain a multi-classification probability, which is the attention weight of the neighboring node, that is, the degree of influence or importance of the neighboring node on the node. Different neighboring nodes of a node have different degrees of influence or importance on the node, and this difference can be reflected by the attention weight. For example, the attention weight of the neighboring node can be obtained according to the following formula (2):
[0053]
[0054] Among them, α i,j is the attention weight of the neighboring nodes of the ith node in the transaction graph network; softmax() is the normalized exponential function; is the initial mapping feature of the i-th node in the transaction graph network; W 2 is the attention parameter of the i-th node in the transaction graph network; is the initial mapping feature of the neighboring nodes of the ith node in the transaction graph network; W 3 is the attention parameter of the neighboring nodes of the i-th node in the transaction graph network; d is the dimension of the transaction feature.
[0055] The attention weight of the neighbor node is obtained, and the node-related features in the initial mapping features of the neighbor node can be extracted using the attention weight and aggregation parameter. The important features in the initial mapping features of the node itself are extracted using the update parameter, and the initial mapping features of the node are enhanced to obtain the mapping features of the node. For example, the mapping features can be obtained according to the following formula (3):
[0056]
[0057] in, is the mapping feature of the i-th node in the transaction graph network; is the initial mapping feature of the i-th node in the transaction graph network; N(i) is the set of neighbor nodes of the i-th node in the transaction graph network; W 4 is the update parameter; W 5 is the aggregation parameter; α i,j is the attention weight of the neighboring nodes of the i-th node in the transaction graph network; is the initial mapping feature of the neighboring nodes of the i-th node in the transaction graph network; is the first sum in the above embodiment; is the first product.
[0058] For each node in the N transaction graph networks, the mapping features of the nodes can be obtained in the above manner, that is, if a sample account has corresponding nodes in N transaction graph networks, then the sample account corresponds to N mapping features. For example, N = 2, the transaction type includes transfer type and consumption type, then the sample account corresponds to 2 mapping features, which are the mapping features from the transfer perspective and consumption perspective mapping features
[0059] In step S1023, the result judgment sub-model obtains the abnormal prediction result according to the mapping features and model parameters representing the same sample account in N transaction graph networks.
[0060] The result determination sub-model can fuse the mapping features of the same sample account from different perspectives, and can also perform classification and discrimination based on the fused features, that is, output anomaly prediction results. In some examples, the mapping features of the nodes representing the same sample account in N transaction graph networks can be merged to obtain the comprehensive features of the sample account; the comprehensive features of the sample account are processed using the model parameters to obtain anomaly prediction results. The merging of multiple mapping features can be specifically implemented as the concatenation of multiple mapping features, or can be implemented as other merging methods, which are not limited here. For example, the anomaly prediction result of a sample account can be obtained according to the following equations (4) and (5):
[0061]
[0062] Among them, z (k) To characterize the comprehensive characteristics of the nodes of a sample account; To characterize the mapping features of a node of a sample account from the perspective of transfer; To characterize the mapping features of a node of a sample account from the consumption perspective; Indicates that the features will be mapped and mapping features Splicing; is the abnormal prediction result; W 6 and b 2are all model parameters of the result determination sub-model.
[0063] In some embodiments, in the fusion stage, i.e., the merging stage, of mapping features at different perspectives, contrastive learning may be used to optimize mapping features at different perspectives, and an optimization loss function may be introduced into the loss function to bring mapping features of nodes with the same labels closer to each other and mapping features of nodes with different labels farther away from each other, so that mapping features at different perspectives can be more targeted for anomaly detection. Figure 6 A flowchart of an abnormal account detection method based on a graph network provided in another embodiment of the present application is provided. Figure 6 and Figure 1 The difference is that Figure 1 Step S103 in the above example can be specifically broken down into Figure 5 Steps S1031 to S1033 in .
[0064] In step S1031, an optimization loss function is obtained based on mapping features of nodes with the same label, mapping features of nodes with different labels, the number of nodes with the same label, and a preset temperature coefficient.
[0065] The mapping features of nodes with the same label in the optimization loss function can be used to make the mapping features of nodes with the same label close to each other in the feature space. The mapping features of nodes with different labels in the optimization loss function can be used to make the mapping features of nodes with different labels far away from each other in the feature space. The temperature coefficient can control the degree of differentiation of the mapping features of nodes with different labels in the model. The temperature coefficient can be set according to the scenario, requirements, experience, etc.
[0066] In some examples, the first parameter can be calculated based on the mapping characteristics of the node with a different label from the first node, the mapping characteristics of the first node, and the temperature coefficient; the second parameter can be calculated based on the mapping characteristics of the node with the same label as the first node, the temperature coefficient, and the first parameter; the optimization loss function can be obtained based on the second parameter obtained by taking the node of each sample account in the account sample set as the first node and the number of nodes with the same label as the first node. For example, the optimization loss function can be obtained according to the following formula (6):
[0067]
[0068] Among them, l fc is the optimization loss function; I is the account sample set; P(i) is the set of nodes in the account sample set that have the same label as the node of the i-th sample account; |P(i)| is the number of nodes in the account sample set that have the same label as the node of the i-th sample account; is the mapping feature of the node of the i-th sample account in the account sample set; is the node in the account sample set that has the same label as the node of the i-th sample account; A(i) is the node set in the account sample set that has different labels from the node of the i-th sample account; is the mapping feature of the nodes in the account sample set that have different labels from the node of the i-th sample account; τ is the temperature coefficient; is the first parameter; is the second parameter.
[0069] In step S1032, a cross entropy loss function is obtained according to the abnormal prediction result of the node, the label of the node and the number of nodes.
[0070] In the embodiment of the present application, abnormal account detection is a binary classification task, and the abnormal detection result is a binary classification result. The cross entropy loss function can be used as part of the loss function that needs to be optimized in the end. The cross entropy loss function can measure the difference between the predicted result and the actual result, that is, the difference between the abnormal prediction result and the label. For example, the cross entropy loss function can be obtained according to the following formula (7):
[0071]
[0072] Among them, l ce is the cross entropy loss function; M is the number of nodes; y i is the label of the i-th node; is the abnormal prediction result of the i-th node.
[0073] In step S1033, a loss function is obtained according to the optimized loss function, the cross entropy loss function and the preset balance coefficient.
[0074] The optimization loss function and the cross entropy loss function may be different in magnitude. The optimization loss function and the cross entropy loss function can be kept consistent in magnitude by using the balancing coefficient. For example, the loss function can be obtained according to the following formula (8):
[0075] l=l ce +βl fc (8)
[0076] Among them, l is the loss function; l ce is the cross entropy loss function; β is the balance coefficient; l fc To optimize the loss function.
[0077] In some embodiments, the above step S105 can be specifically refined as follows: inputting N transaction graph networks obtained based on the account set to be tested into the abnormal account detection model, using model parameters to map the attributes of the nodes in the N transaction graph networks to the same feature space, and obtaining the initial mapping features of the nodes in the N transaction graph networks; updating the initial mapping features of the nodes using the model parameters and the initial mapping features of the node's neighbor nodes to obtain the mapping features of the nodes, and the neighbor nodes are connected to the nodes through edges; obtaining abnormal prediction results based on the mapping features and model parameters that characterize the same account to be tested in the N transaction graph networks; and determining the abnormal accounts in the account set to be tested based on the abnormal prediction results.
[0078] The abnormal account detection model processes the N transaction graph networks corresponding to the account set to be tested, which is similar to the preset model's processing of the N transaction graph networks corresponding to the account sample set. For details, please refer to the relevant descriptions in the above embodiments, which will not be repeated here.
[0079] A second aspect of the present application provides an abnormal account detection device based on a graph network. Figure 7 A schematic diagram of the structure of an abnormal account detection device based on a graph network provided in an embodiment of the present application is shown in FIG. Figure 7 As shown, the abnormal account detection device 200 based on a graph network may include a graph network generation module 201 and a model processing module 202 .
[0080] The graph network generation module 201 can be used to generate N transaction graph networks of the account sample set according to the transaction relationships, transaction characteristics and transaction types of the sample accounts in the account sample set. The sample accounts have labels, and the labels represent normal accounts or abnormal accounts. One transaction graph network corresponds to one transaction type. Each transaction graph network includes nodes and edges. The nodes represent the sample accounts, and the nodes have the transaction characteristics of the sample accounts. The edges represent the transaction relationships. N is an integer greater than 1.
[0081] The model processing module 202 can be used to: input N transaction graph networks into the initial model to obtain abnormal prediction results of sample accounts, the initial model includes a preliminary mapping sub-model, a feature update sub-model and a result judgment sub-model; and, based on the abnormal prediction results of the sample accounts, labels and mapping features of the sample accounts as nodes in the N transaction graph networks, obtain a loss function; and, if the loss function does not meet the training cutoff condition, adjust the model parameters of the initial model until the loss function meets the training cutoff condition to obtain an abnormal account detection model; and, input the N transaction graph networks obtained based on the account set to be tested into the abnormal account detection model to obtain abnormal accounts in the account set to be tested output by the abnormal account detection model.
[0082] In some embodiments, the graph network generation module 201 can be specifically used to: generate nodes representing sample accounts based on the transaction characteristics of the sample accounts in each transaction type; obtain the transaction relationships of the sample accounts in transactions of each transaction type; set edges between the nodes of the sample accounts that have transaction relationships in transactions of the same transaction type, to form N transaction graph networks corresponding to N transaction types.
[0083] In some examples, the transaction type includes a consumption type and a transfer type. When the transaction type is a transfer type, the sample account initiating the transfer has a transaction relationship with the sample account receiving the transfer. When the transaction type is a transfer type, the sample accounts that conduct transactions with the same merchant that meet the preset abnormal conditions have a transaction relationship.
[0084] In some embodiments, the model processing module 202 can be used to: input N transaction graph networks into a preliminary mapping sub-model, and use the model parameters through the preliminary mapping sub-model to map the transaction features of the nodes in the N transaction graph networks to the same feature space to obtain the initial mapping features of the nodes in the N transaction graph networks; input the initial mapping features of the nodes into a feature update sub-model, and use the model parameters and the initial mapping features of the node's neighbor nodes through the feature update sub-model to update the initial mapping features of the nodes to obtain the mapping features of the nodes, and the neighbor nodes are connected to the nodes through edges; and obtain the abnormal prediction results through the result judgment sub-model based on the mapping features and model parameters that characterize the same sample account in the N transaction graph networks.
[0085] In some examples, the model parameters of the feature update sub-model include an attention parameter, an update parameter, and an aggregation parameter. The model processing module 202 can be specifically used to: calculate the attention weight of the neighbor node according to the initial mapping feature of the node, the initial mapping feature of the neighbor node, the dimension of the transaction feature, and the attention parameter; calculate the first sum of the product of the initial mapping feature, the attention weight, and the aggregation parameter of each neighbor node; calculate the first product of the initial mapping feature of the node and the update parameter, and determine the sum of the first product and the first sum as the mapping feature of the node.
[0086] In some examples, the model processing module 202 can be specifically used to: merge the mapping features of nodes representing the same sample account in N transaction graph networks to obtain comprehensive features of the sample account; use model parameters to process the comprehensive features of the sample account to obtain abnormal prediction results.
[0087] In some embodiments, the model processing module 202 can be used to: obtain an optimized loss function based on the mapping features of nodes with the same label, the mapping features of nodes with different labels, the number of nodes with the same label, and a preset temperature coefficient; obtain a cross entropy loss function based on the abnormal prediction results of the nodes, the labels of the nodes, and the number of nodes; obtain a loss function based on the optimized loss function, the cross entropy loss function, and a preset balance coefficient.
[0088] In some examples, the model processing module 202 can be specifically used to: calculate a first parameter based on a mapping feature of a node with a different label from the first node, a mapping feature of the first node, and a temperature coefficient; calculate a second parameter based on a mapping feature of a node with the same label as the first node, a temperature coefficient, and the first parameter; and obtain an optimized loss function based on the second parameter obtained by using the node of each sample account in the account sample set as the first node, and the number of nodes with the same label as the first node.
[0089] In some embodiments, the model processing module 202 is also used to: input N transaction graph networks obtained based on the set of accounts to be tested into the abnormal account detection model, use model parameters to map the attributes of the nodes in the N transaction graph networks to the same feature space, and obtain the initial mapping features of the nodes in the N transaction graph networks; use the model parameters and the initial mapping features of the node's neighbor nodes to update the initial mapping features of the node to obtain the mapping features of the node, and the neighbor nodes are connected to the node through edges; obtain abnormal prediction results based on the mapping features and model parameters that characterize the same account to be tested in the N transaction graph networks; and determine the abnormal accounts in the set of accounts to be tested based on the abnormal prediction results.
[0090] It should be noted that the graph network-based abnormal account detection device 200 is a device corresponding to the above-mentioned graph network-based abnormal account detection method. All implementation methods in the above-mentioned method embodiments are applicable to the embodiments of the device and can achieve the same technical effects.
[0091] The third aspect of the present application also provides an abnormal account detection device based on a graph network. Figure 8 A schematic diagram of a structure of an abnormal account detection device based on a graph network provided in an embodiment of the present application, such as Figure 8 As shown, the graph network-based abnormal account detection device 300 includes a memory 301 , a processor 302 , and a computer program stored in the memory 301 and executable on the processor 302 .
[0092] In some examples, the processor 302 may include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiments of the present application.
[0093] The memory 301 may include a read-only memory (ROM), a random access memory (RAM), a disk storage medium device, an optical storage medium device, a flash memory device, an electrical, optical or other physical / tangible memory storage device. Therefore, generally, the memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., a memory device) encoded with software including computer executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the abnormal account detection method based on a graph network in an embodiment of the present application.
[0094] The processor 302 runs a computer program corresponding to the executable program code by reading the executable program code stored in the memory 301, so as to implement the abnormal account detection method based on the graph network in the above embodiment.
[0095] In some examples, the graph network-based abnormal account detection device 300 may further include a communication interface 303 and a bus 304. Figure 8 As shown, the memory 301, the processor 302, and the communication interface 303 are connected via a bus 304 and communicate with each other.
[0096] The communication interface 303 is mainly used to realize the communication between the modules, devices, units and / or equipment in the embodiment of the present application. The communication interface 303 can also be used to access input devices and / or output devices.
[0097] The bus 304 includes hardware, software or both, and couples the components of the abnormal account detection device 300 based on the graph network to each other. For example, but not limitation, the bus 304 may include an accelerated graphics port (Accelerated Graphics Port, AGP) or other graphics bus, an enhanced industry standard architecture (Enhanced Industry Standard Architecture, EISA) bus, a front side bus (Front Side Bus, FSB), a hypertransport (Hyper Transport, HT) interconnection, an industry standard architecture (Industry Standard Architecture, ISA) bus, an infinite bandwidth interconnection, a low pin count (Low pin count, LPC) bus, a memory bus, a micro channel architecture (Micro Channel Architecture, MCA) bus, a peripheral component interconnect (Peripheral Component Interconnect, PCI) bus, a PCI-Express (PCI-E) bus, a serial advanced technology attachment (Serial Advanced Technology Attachment, SATA) bus, a video electronics standards association local (Video Electronics Standards Association Local Bus, VLB) bus or other suitable buses or a combination of two or more of these. Where appropriate, the bus 304 may include one or more buses. Although embodiments of the present application describe and illustrate a particular bus, the present application contemplates any suitable bus or interconnect.
[0098] The fourth aspect of the present application provides a computer-readable storage medium, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the abnormal account detection method based on the graph network in the above-mentioned embodiment can be implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here. Among them, the above-mentioned computer-readable storage medium may include a non-transitory computer-readable storage medium, such as a read-only memory (Read-Only Memory, referred to as ROM), a random access memory (Random Access Memory, referred to as RAM), a disk or an optical disk, etc., which is not limited here.
[0099] The fifth aspect of the present application provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the abnormal account detection method based on the graph network in the above-mentioned embodiment and can achieve the same technical effect. To avoid repetition, it will not be repeated here.
[0100] It should be clear that each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. For device embodiments, equipment embodiments, and computer-readable storage medium embodiments, the relevant parts can refer to the description part of the method embodiment. The present application is not limited to the specific steps and structures described above and shown in the figures. Those skilled in the art can make various changes, modifications and additions, or change the order between the steps after understanding the spirit of the present application. In addition, for the sake of brevity, a detailed description of known method technologies is omitted here.
[0101] The above reference is according to the method of the embodiment of the present application, the flow chart of the device (system) and the computer program product and / or the block diagram described various aspects of the present application.It should be understood that each square box in the flow chart and / or the block diagram and the combination of each square box in the flow chart and / or the block diagram can be realized by computer program instructions.These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer or other programmable data processing device to produce a machine so that these instructions executed by the processor of the computer or other programmable data processing device enable the realization of the function / action specified in one or more square boxes of the flow chart and / or the block diagram.Such a processor can be but is not limited to a general-purpose processor, a special-purpose processor, a special application processor or a field programmable logic circuit.It can also be understood that each square box in the block diagram and / or the flow chart and the combination of the square boxes in the block diagram and / or the flow chart can also be realized by the dedicated hardware that performs the specified function or action, or can be realized by the combination of dedicated hardware and computer instructions.
[0102] Those skilled in the art should understand that the above embodiments are exemplary rather than restrictive. Different technical features appearing in different embodiments can be combined to achieve beneficial effects. Based on the study of the drawings, the specification and the claims, those skilled in the art should be able to understand and implement other variations of the disclosed embodiments. In the claims, the term "comprising" does not exclude other devices or steps; the quantifier "one" does not exclude multiple; the terms "first" and "second" are used to indicate names rather than to indicate any specific order. Any figure marks in the claims should not be understood as limiting the scope of protection. The functions of multiple parts appearing in the claims can be implemented by a single hardware or software module. The appearance of certain technical features in different dependent claims does not mean that these technical features cannot be combined to achieve beneficial effects.
Claims
1. A method for detecting abnormal accounts based on graph networks, characterized in that: include: According to the transaction relationship, transaction characteristics and transaction type of the sample accounts in the account sample set, N transaction graph networks of the account sample set are generated, the sample accounts have labels, the labels represent normal accounts or abnormal accounts, one transaction graph network corresponds to one transaction type, each transaction graph network includes nodes and edges, the nodes represent the sample accounts, the nodes have the transaction characteristics of the sample accounts, the edges represent the transaction relationship, and N is an integer greater than 1; Inputting N transaction graph networks into an initial model to obtain abnormal prediction results of sample accounts, wherein the initial model includes a preliminary mapping sub-model, a feature update sub-model, and a result determination sub-model; Based on the abnormal prediction results and labels of the sample accounts and the mapping features of the sample accounts as nodes in the N transaction graph networks, a loss function is obtained; If the loss function does not meet the training cutoff condition, adjusting the model parameters of the initial model until the loss function meets the training cutoff condition, thereby obtaining an abnormal account detection model; N transaction graph networks obtained based on the account set to be tested are input into the abnormal account detection model to obtain abnormal accounts in the account set to be tested output by the abnormal account detection model.
2. The method according to claim 1, characterized in that The generating of N transaction graph networks of the account sample set according to the transaction relationship, transaction characteristics and transaction type of the sample accounts in the account sample set includes: Generate nodes representing sample accounts according to the transaction characteristics of the sample accounts in each transaction type; Obtain the transaction relationship of the sample accounts in each transaction type; Edges are set between the nodes of sample accounts that have transaction relationships in transactions of the same transaction type, forming N transaction graph networks corresponding to N transaction types.
3. The method according to claim 2, characterized in that Transaction types include consumption type and transfer type; When the transaction type is a transfer type, the sample account initiating the transfer and the sample account receiving the transfer have a transaction relationship; When the transaction type is a transfer type, the sample accounts that conduct transactions with the same merchant that meet the preset abnormal conditions have a transaction relationship.
4. The method according to claim 1, characterized in that: The N transaction graph networks are input into the initial model to obtain the abnormal prediction results of the sample accounts, including: Inputting N transaction graph networks into the preliminary mapping sub-model, mapping the transaction features of the nodes in the N transaction graph networks to the same feature space using the model parameters through the preliminary mapping sub-model, and obtaining the initial mapping features of the nodes in the N transaction graph networks; Inputting the initial mapping feature of the node into the feature updating sub-model, updating the initial mapping feature of the node by using the model parameters and the initial mapping features of the node's neighboring nodes through the feature updating sub-model to obtain the mapping feature of the node, wherein the neighboring nodes are connected to the node through edges; The result determination sub-model obtains an abnormal prediction result based on the mapping features and model parameters that characterize the same sample account in N transaction graph networks.
5. The method according to claim 4, characterized in that The model parameters of the feature update sub-model include attention parameters, update parameters and aggregation parameters; The updating of the initial mapping features of the node by the feature updating sub-model using the model parameters and the initial mapping features of the neighboring nodes of the node to obtain the mapping features of the node includes: Calculate the attention weight of the neighbor node according to the initial mapping feature of the node, the initial mapping feature of the neighbor node, the dimension of the transaction feature and the attention parameter; Calculate a first sum of the product of the initial mapping feature, the attention weight and the aggregation parameter of each neighbor node; A first product of the initial mapping feature of the node and the update parameter is calculated, and a sum of the first product and the first sum is determined as the mapping feature of the node.
6. The method according to claim 4, characterized in that The abnormal prediction result is obtained by the result determination sub-model according to the mapping features and model parameters representing the same sample account in N transaction graph networks, including: Merge the mapping features of the nodes representing the same sample account in N transaction graph networks to obtain the comprehensive features of the sample account; The comprehensive characteristics of the sample accounts are processed using model parameters to obtain abnormal prediction results.
7. The method according to claim 1, characterized in that The loss function is obtained based on the abnormal prediction results and labels of the sample accounts and the mapping features of the sample accounts as nodes in the N transaction graph networks, including: Based on the mapping features of nodes with the same label, the mapping features of nodes with different labels, the number of nodes with the same label, and a preset temperature coefficient, an optimized loss function is obtained; According to the abnormal prediction results of the nodes, the labels of the nodes and the number of nodes, the cross entropy loss function is obtained; The loss function is obtained according to the optimized loss function, the cross entropy loss function and a preset balance coefficient.
8. The method according to claim 7, characterized in that The optimization loss function is obtained based on the mapping features of nodes with the same label, the mapping features of nodes with different labels, the number of nodes with the same label, and a preset temperature coefficient, including: Calculate a first parameter according to a mapping feature of a node having a label different from that of the first node, the mapping feature of the first node, and a temperature coefficient; Calculate a second parameter according to a mapping feature of a node having the same label as the first node, a temperature coefficient, and the first parameter; The optimization loss function is obtained based on the second parameter obtained by using the node of each sample account in the account sample set as the first node and the number of nodes with the same label as the first node.
9. The method according to claim 1, characterized in that: The step of inputting N transaction graph networks obtained based on the account set to be tested into the abnormal account detection model to obtain abnormal accounts in the account set to be tested output by the abnormal account detection model includes: Inputting N transaction graph networks obtained based on the set of accounts to be tested into the abnormal account detection model, mapping the attributes of the nodes in the N transaction graph networks to the same feature space using model parameters, and obtaining initial mapping features of the nodes in the N transaction graph networks; The initial mapping features of the node are updated using the model parameters and the initial mapping features of the node's neighbor nodes to obtain the mapping features of the node, and the neighbor nodes are connected to the node through edges; According to the mapping features and model parameters representing the same account to be tested in N transaction graph networks, the abnormal prediction results are obtained; According to the abnormal prediction result, the abnormal accounts in the set of accounts to be tested are determined.
10. An abnormal account detection device based on graph network, characterized in that: include: A graph network generation module, used to generate N transaction graph networks of the account sample set according to the transaction relationship, transaction characteristics and transaction type of the sample accounts in the account sample set, wherein the sample accounts have labels, and the labels represent normal accounts or abnormal accounts. One transaction graph network corresponds to one transaction type, and each transaction graph network includes nodes and edges, wherein the nodes represent the sample accounts, the nodes have the transaction characteristics of the sample accounts, and the edges represent the transaction relationship, and N is an integer greater than 1; A model processing module, used to input N transaction graph networks into an initial model to obtain abnormal prediction results of sample accounts, wherein the initial model includes a preliminary mapping sub-model, a feature update sub-model and a result determination sub-model; And, used to obtain a loss function based on the abnormal prediction results, labels and mapping features of the sample accounts as nodes in N transaction graph networks; and, used to adjust the model parameters of the initial model if the loss function does not meet the training cutoff condition until the loss function meets the training cutoff condition to obtain an abnormal account detection model; and, used to input N transaction graph networks obtained based on the account set to be tested into the abnormal account detection model to obtain abnormal accounts in the account set to be tested output by the abnormal account detection model.
11. An abnormal account detection device based on a graph network, characterized in that: include: a processor and a memory storing computer program instructions; When the processor executes the computer program instructions, the abnormal account detection method based on a graph network as described in any one of claims 1 to 9 is implemented.
12. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer program instructions, and when the computer program instructions are executed by a processor, the abnormal account detection method based on a graph network as described in any one of claims 1 to 9 is implemented.
13. A computer program product, characterized in that It includes a computer program, which, when executed by a processor, implements the abnormal account detection method based on a graph network as described in any one of claims 1 to 9.
