Account state model construction method and apparatus, computer device, and storage medium
By constructing a user feature map and training a graph convolutional neural network model, and dynamically adjusting the importance of nodes, the problem of low prediction accuracy of traditional models is solved, and more efficient account status assessment is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHAOLIAN CONSUMER FINANCE CO LTD
- Filing Date
- 2022-09-15
- Publication Date
- 2026-05-22
AI Technical Summary
Traditional account status models have low prediction accuracy and are difficult to effectively assess the operational status of user accounts.
By constructing a user feature map, generating initial feature vectors for nodes and fusing them with an association matrix, and using a graph convolutional neural network to train an account state model, the importance weights of nodes are dynamically adjusted to improve the model's prediction accuracy.
It improves the predictive accuracy of the account status model, enabling better mining of data features and adapting to the needs of different application scenarios.
Smart Images
Figure CN115618008B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of computer technology, and in particular to a method, apparatus, computer device, and storage medium for constructing an account status model. Background Technology
[0002] With the development of computer technology, more and more users can perform online operations through mobile applications, websites and other channels. In this process, the user's submitted feature data is analyzed and mined to determine the current operational status of the user's account.
[0003] In traditional techniques, a user device relationship graph is constructed to obtain the status scores of all relevant device nodes connected to each media node, and these scores are used as confidence scores. Then, the operational status of the target user account is analyzed. However, the model prediction accuracy is low. Summary of the Invention
[0004] Therefore, it is necessary to provide an account status model construction method, apparatus, computer device, and computer-readable storage medium to address the aforementioned technical problems, which can effectively improve the accuracy of account status model predictions.
[0005] A method for constructing an account status model, characterized in that the method includes:
[0006] Obtain historical user characteristic data;
[0007] Construct a user feature map based on historical user feature data;
[0008] The initial feature vector of each node is generated based on the node attribute features of each node in the user feature map and the initial value of the model training parameters. The initial value of the model training parameters is the initial value of the parameters to be trained in the account status model.
[0009] The association matrix is obtained by fusing the initial feature vectors of the nodes. The association matrix is used to characterize the association relationship between the nodes and the importance of each node.
[0010] Feature input sample data is obtained based on user feature maps, correlation matrices, and initial feature vectors of nodes;
[0011] An account status model is trained based on feature input sample data. The account status model is used to evaluate the operational status of the target account.
[0012] In one embodiment, generating the corresponding initial feature vector for each node based on the node attribute features of each node in the user feature map and the initial values of the model training parameters includes:
[0013] Obtain the non-linear activation function;
[0014] The original node characteristics of each node are determined based on the node attributes of each node.
[0015] Initial feature vectors for each node are generated based on nonlinear activation functions, original node features, and initial values of model training parameters.
[0016] In one embodiment, generating the corresponding initial feature vector for each node based on the node attribute features of each node in the user feature map and the initial values of the model training parameters includes:
[0017] When the node is a contact information node, the initial feature vector of the adjacent node corresponding to the adjacent node of the contact information node is obtained. The contact information node is a node whose node attribute is contact information.
[0018] The initial feature vector of the node corresponding to the connection method node is determined based on the mean of the initial feature vectors of the adjacent nodes.
[0019] In one embodiment, the association matrix of each node is obtained by fusing the initial feature vectors of the nodes, including:
[0020] Retrieve the user identifier node from each node;
[0021] The initial feature vector of the user is obtained based on the account channel information, geographical information, and application status information of the user identification node;
[0022] The association matrix is obtained by fusing the user's initial feature vector with the initial feature vectors of the corresponding nodes of other nodes.
[0023] In one embodiment, feature input sample data is obtained based on the user feature map, the correlation matrix, and the initial feature vectors of the nodes, including:
[0024] Determine the sub-graphs in the user feature graph based on the user identifier nodes in each node;
[0025] The user feature map weight matrix is obtained from the sub-map.
[0026] The node degree matrix is obtained by fusing the weight matrix and the correlation matrix of the user feature graph.
[0027] The subgraph degree matrix is obtained based on the subgraph nodes and the correlation matrix in the subgraph.
[0028] Feature input sample data is obtained based on the user feature graph weight matrix, correlation matrix, node degree matrix, subgraph degree matrix, and node initial feature vectors.
[0029] In one embodiment, an account status model is trained based on feature input sample data, including:
[0030] The feature input sample data is normalized to obtain the standard input sample data;
[0031] The account status model is obtained by training a graph convolutional neural network model with standard input sample data.
[0032] In one embodiment, standard input sample data is input into a graph convolutional neural network model for training to obtain an account status model, including:
[0033] Input standard input sample data into the graph convolutional neural network model to obtain the actual output value of the model;
[0034] The actual output values of the model are classified according to a preset classification algorithm to obtain the actual estimated account status corresponding to the input sample data;
[0035] When the error between the actual predicted account status and the preset standard predicted account status in the standard input sample data is greater than the preset error threshold, return to the step of inputting the standard input sample data into the graph convolutional neural network model to obtain the actual output value of the model.
[0036] When the error between the actual predicted account status and the preset standard predicted account status in the standard input sample data is less than or equal to the preset error threshold, the training is complete and the account status model is obtained.
[0037] An account status model construction apparatus, characterized in that the apparatus comprises:
[0038] The data acquisition module is used to acquire historical user characteristic data;
[0039] The sample data generation module is used to construct a user feature map based on historical user feature data; generate corresponding initial feature vectors for each node based on the node attribute features of each node in the user feature map and the initial values of the model training parameters, where the initial values of the model training parameters are the initial values of the parameters to be trained in the account status model; obtain an association matrix by fusing the initial feature vectors of the nodes, which is used to characterize the association relationship and importance of each node; and obtain feature input sample data based on the user feature map, the association matrix, and the initial feature vectors of the nodes.
[0040] The training module is used to train an account status model based on feature input sample data. The account status model is used to evaluate the operational status of the target account.
[0041] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps:
[0042] Obtain historical user characteristic data;
[0043] Construct a user feature map based on historical user feature data;
[0044] The initial feature vector of each node is generated based on the node attribute features of each node in the user feature map and the initial value of the model training parameters. The initial value of the model training parameters is the initial value of the parameters to be trained in the account status model.
[0045] The association matrix is obtained by fusing the initial feature vectors of the nodes. The association matrix is used to characterize the association relationship between the nodes and the importance of each node.
[0046] Feature input sample data is obtained based on user feature maps, correlation matrices, and initial feature vectors of nodes;
[0047] An account status model is trained based on feature input sample data. The account status model is used to evaluate the operational status of the target account.
[0048] A computer-readable storage medium storing a computer program, which, when executed by a processor, performs the following steps:
[0049] Obtain historical user characteristic data;
[0050] Construct a user feature map based on historical user feature data;
[0051] The initial feature vector of each node is generated based on the node attribute features of each node in the user feature map and the initial value of the model training parameters. The initial value of the model training parameters is the initial value of the parameters to be trained in the account status model.
[0052] The association matrix is obtained by fusing the initial feature vectors of the nodes. The association matrix is used to characterize the association relationship between the nodes and the importance of each node.
[0053] Feature input sample data is obtained based on user feature maps, correlation matrices, and initial feature vectors of nodes;
[0054] An account status model is trained based on feature input sample data. The account status model is used to evaluate the operational status of the target account.
[0055] The aforementioned account status model construction method, apparatus, computer equipment, and storage medium acquire historical user feature data, construct a user feature map based on the historical user feature data, and then generate corresponding initial feature vectors for nodes based on the node attribute features of each node in the user feature map and the initial values of the model training parameters. The initial values of the model training parameters are the initial values of the parameters to be trained in the account status model. An association matrix is obtained by fusing the initial feature vectors of the nodes. The association matrix is used to characterize the association relationship and importance of each node. Feature input sample data is obtained based on the user feature map, the association matrix, and the initial feature vectors of the nodes. The account status model is trained based on the feature input sample data. The account status model is used to evaluate the operational status of the target account. In this way, by incorporating the initial values of the model training parameters into the process of generating the initial feature vectors of the nodes, new training parameters are continuously generated as the model iterates. The feature vectors of each node are then regenerated based on the new parameters generated in each iteration, accelerating the convergence of the model. Furthermore, the correlation matrix is obtained by fusing the initial feature vectors of the nodes. During model training, the importance weights assigned to each node are dynamically adjusted according to the importance of different nodes. This allows for better mining of the importance of different nodes based on specific application scenarios, enabling the trained model to better mine data features and effectively improve the accuracy of account status model prediction. Attached Figure Description
[0056] Figure 1 This is an application environment diagram of an account status model construction method in one embodiment;
[0057] Figure 2 This is a flowchart illustrating an account status model construction method in one embodiment;
[0058] Figure 3 This is a schematic diagram of the process for generating initial feature vectors for nodes in one embodiment;
[0059] Figure 4 This is a flowchart illustrating the process of generating the initial feature vector for the contact information node in another embodiment;
[0060] Figure 5 This is a flowchart illustrating the process of generating an association matrix in one embodiment;
[0061] Figure 6 This is a schematic diagram of the process for generating feature input sample data in one embodiment;
[0062] Figure 7 This is a flowchart illustrating the process of obtaining an account status model in one embodiment;
[0063] Figure 8 This is a flowchart illustrating the process of training an account state model in one embodiment;
[0064] Figure 9 This is a structural block diagram of an account status model construction device in one embodiment;
[0065] Figure 10 This is an internal structural diagram of a computer device in one embodiment;
[0066] Figure 11 This is a schematic diagram of all nodes associated with a user in one embodiment;
[0067] Figure 12 This is a schematic diagram of a user feature map in one embodiment;
[0068] Figure 13 This is a flowchart illustrating the process of predicting the risk of overdue repayments for a user account in one embodiment. Detailed Implementation
[0069] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0070] The account status model construction method provided in this application embodiment can be applied to, for example, Figure 1 In the application environment shown, computer device 102 acquires historical user feature data, constructs a user feature map based on this data, and then generates corresponding initial feature vectors for each node based on the node attribute features of each node in the user feature map and the initial values of the model training parameters. The initial values of the model training parameters are the initial values of the parameters to be trained in the account status model. An association matrix is obtained by fusing the initial node feature vectors. This association matrix is used to characterize the relationships between nodes and the importance of each node. Feature input sample data is obtained based on the user feature map, the association matrix, and the initial node feature vectors. An account status model is trained based on this feature input sample data. The account status model is used to evaluate the operational status of the target account. Specifically, computer device 102 may include, but is not limited to, various personal computers, laptops, servers, smartphones, tablets, smart cameras, and portable wearable devices.
[0071] In one embodiment, such as Figure 2 As shown, an account status model construction method is provided, which can be applied to... Figure 1 Taking computer device 102 as an example, the following steps are included:
[0072] Step S202: Obtain historical user characteristic data.
[0073] The historical user characteristic data consists of users' personal data, including date of birth, education level, affiliated companies, contact person and inviter information, etc. This data is necessary information that users fill in when applying for related activities on the Internet. When users operate on the client, the system also assigns them a corresponding identity identification number, which serves as a unique identifier for each user and will not be duplicated. The server can also obtain the device information used by users when logging into the client, such as device identification code, device brand, physical address, etc. Finally, after anonymization, missing value processing, and anomaly handling, the user data is divided into six categories: user identifier, identity characteristics, company characteristics, bank card characteristics, device characteristics, and contact information. Each user also has a corresponding account status label information, which is used as the standard output sample data for subsequent model training.
[0074] Step S204: Construct a user feature map based on historical user feature data.
[0075] Among them, the user identification node includes node attributes such as account creation channel, geographical location, and application status; the identity feature node includes node attributes such as identity status, education and job position, and credit history; the unit node includes node attributes such as enterprise type, geographical location, and registered capital; the bank card node includes node attributes such as account type and city of account opening; and the device node includes node attributes such as device identification code and device system.
[0076] Specifically, the computer device uses user identifiers, identity features, unit features, bank card features, device features, and contact information from historical user feature data as nodes, and connects the relationships between each node as edges to construct a user feature graph. Due to relationships such as inviters and contacts between users, multiple users belonging to the same unit, and multiple users using the same device or contact information, different user identifier nodes can be associated with other nodes to ultimately generate a user feature graph containing all users.
[0077] Step S206: Generate the corresponding initial feature vector of the node based on the node attribute features of each node in the user feature map and the initial value of the model training parameters. The initial value of the model training parameters is the initial value of the parameters to be trained in the account status model.
[0078] The initial feature vector of a node is a vector representing the node attributes of the corresponding node. In other words, the initial feature vectors of nodes can reflect the relationship between each node. For example, when the relationship between two nodes is high, the distance between the initial feature vectors of the two nodes is smaller. Furthermore, the initial feature vectors of each node are vectors with the same dimension. The initial values of the model training parameters are the initial values of the parameters to be trained in the account state model. As the model is trained, each iteration will generate new model training parameters. The initial values of the parameters to be trained in the account state model can be specified based on experience or determined by intelligent optimization algorithms such as genetic algorithms and particle swarm optimization algorithms.
[0079] Specifically, the computer device generates initial feature vectors for each node based on the node attribute features of each node in the user feature map and the initial values of the model training parameters. The initial feature vectors are then normalized, and the values of the training parameters after each iteration of the model are used as new model training parameters to reconstruct the initial feature vectors of each node. In other words, with each iteration of the model, the initial feature vectors of each node are continuously iterated, thereby iterating to obtain a set of initial feature vectors that best represent the current user feature map.
[0080] Step S208: Based on the initial feature vectors of the nodes, the association matrix is obtained by fusing them. The association matrix is used to characterize the association relationship of each node and the importance of each node.
[0081] The correlation matrix is used to characterize the relationship between nodes and the importance of each node. The row and column information of each matrix element in the correlation matrix is used to represent the position information of the corresponding node in the user feature map relative to the map. The numerical value of the matrix element is used to characterize the importance of the corresponding node in the user feature map. That is, the larger the value of the matrix element, the greater the importance of the node corresponding to the matrix element.
[0082] Specifically, the computer device divides the user feature map into sub-maps based on the user feature nodes in the user feature map, determines the row and column positions of the matrix elements corresponding to the target node in the correlation matrix by the position of the target node in the sub-map of the user feature map, and constructs the size of the matrix elements corresponding to each node based on the values of the initial feature vectors of each node.
[0083] Step S210: Obtain feature input sample data based on user feature map, correlation matrix and node initial feature vector.
[0084] Specifically, the computer device divides the user feature map into sub-maps based on the user feature nodes in the user feature map, and assigns a corresponding weight value to each sub-map to obtain the weight diagonal matrix of the entire user feature map. Then, based on the weight value of each sub-map, the degree matrix of the corresponding node and the degree matrix of the sub-map are calculated. Finally, the weight diagonal matrix, correlation matrix, node degree matrix and sub-map degree matrix of the entire user feature map are used together as the input sample matrix of the model.
[0085] Step S212: Train an account status model based on the feature input sample data. The account status model is used to evaluate the operational status of the target account.
[0086] The account status model is used to evaluate the operational status of a target account. For example, it can determine the risk status of the target account and whether there is any risk. The account status model can, but is not limited to, use various deep learning models and neural network models, such as wavelet neural networks, graph convolutional neural networks, and recurrent neural networks.
[0087] Specifically, the computer device inputs the feature input sample data determined in the aforementioned steps into the account status model for training until the error between the model output result and the standard output sample data in the sample data is less than or equal to a preset error threshold or the number of iterations is greater than a preset number of iterations, at which point training stops.
[0088] In this embodiment, by incorporating the initial values of the model training parameters into the process of generating the initial feature vectors of the nodes, new training parameters are continuously generated as the model iterates. The feature vectors of each node are then regenerated based on the new parameters generated in each iteration, accelerating the convergence of the model. Furthermore, the correlation matrix is obtained by fusing the initial feature vectors of the nodes. During model training, the importance weights assigned to each node are dynamically adjusted according to the importance of different nodes. This allows for better mining of the importance of different nodes based on specific application scenarios, enabling the trained model to better mine data features and effectively improve the accuracy of account status model prediction.
[0089] In one embodiment, such as Figure 3 As shown, based on the node attribute features of each node in the user feature map and the initial values of the model training parameters, the corresponding initial feature vectors of the nodes are generated, including:
[0090] Step S302: Obtain the nonlinear activation function.
[0091] Among them, nonlinear activation functions are tool functions used to map the attribute features of each node in the user graph to the same spatial dimension, such as the sigmoid function. The node attribute features are unique, non-repeating feature vectors with different vector dimensions that are manually set based on the attribute features of each node.
[0092] Step S304: Determine the original node characteristics of each node based on the node attributes of each node.
[0093] Specifically, the computer device sets the original node features of each node according to preset empirical rules based on the node attributes of each node. The original node features set for the same type of node have the same dimension, while the original node features corresponding to different types of nodes can be different.
[0094] Step S306: Generate the initial feature vector of each node based on the nonlinear activation function, the original node features, and the initial values of the model training parameters.
[0095] Specifically, the computer device generates a fusion factor based on the original node features and model training parameters, and then uses the fusion factor as the input of a nonlinear activation function to calculate the corresponding node initial vector features.
[0096] For example, for a node v with a specific attribute among the nodes... i ∈V card The original node features are defined according to the following formula 1. Transform into a uniform dimension of 1×d N New node initial feature vector
[0097]
[0098] Among them, V card This is the set of all nodes in the user feature map. For the 1×d card A vector of dimension W card and b card These are all parameters that need to be learned during model training.
[0099] In this embodiment, a nonlinear activation function is obtained, and the original node features of each node are determined according to the node attributes of each node. The original node features and model training parameters are fused to generate a fusion factor. The fusion factor is then used as the input of the nonlinear activation function to calculate the node initial vector feature vector corresponding to the same dimension of each node attribute. As the model iterates continuously, the generated node initial vector can better reflect the attributes between nodes, thus improving the accuracy of generating node initial vectors.
[0100] In one embodiment, such as Figure 4 As shown, based on the node attribute features of each node in the user feature map and the initial values of the model training parameters, the corresponding initial feature vectors of the nodes are generated, including:
[0101] Step S402: When the node is a contact information node, obtain the initial feature vector of the adjacent node corresponding to the adjacent node of the contact information node. The contact information node is a node whose node attribute is contact information.
[0102] The contact information node is a node whose node attribute is contact information, containing the corresponding user's contact information, such as the user's phone number or email address. When multiple users share the same contact information, the related users are connected to the same contact information node in the user feature graph. Furthermore, the contact information node is always associated with the corresponding user's identity feature node. That is, in any subgraph of the user feature graph, the identity feature node and the contact information node must be connected. In the same subgraph, the user feature node and the corresponding contact information node must be connected. In other words, the node directly connected to the contact information node in the user feature graph is the adjacent node of the contact information node.
[0103] Step S404: Determine the initial feature vector of the node corresponding to the contact method node based on the mean of the initial feature vectors of adjacent nodes.
[0104] Specifically, the computer device takes the average value of the initial feature vectors of the neighboring nodes of the current contact information node obtained in the aforementioned steps, and uses the average value vector as the initial feature vector of the current contact information node.
[0105] For example, the computer device calculates the initial feature vector of the current contact information node according to the following formula 2.
[0106]
[0107] Where k is the number of nodes adjacent to the current contact node. The initial feature vectors of the adjacent nodes.
[0108] In this embodiment, when the node is a contact information node, the initial feature vector of the adjacent node corresponding to the adjacent node of the contact information node is obtained, and then the initial feature vector of the node corresponding to the contact information node is determined according to the mean of the initial feature vectors of the adjacent nodes. This simplifies the generation steps of the initial feature vector of the node corresponding to the contact information node and improves the calculation efficiency of the initial feature vector of the node.
[0109] In one embodiment, such as Figure 5 As shown, the association matrix of each node is obtained by fusing the initial feature vectors of the nodes, including:
[0110] Step S502: Obtain the user identifier node in each node.
[0111] Step S504: Obtain the user's initial feature vector based on the user's account channel information, geographical information, and application status information of the user identifier node.
[0112] Specifically, the computer device categorizes and identifies the account channel information, geographical information, and application status information of the user identification node, defines several common methods for creating an account, divides the geographical information into different regions or by province and city, and divides the application status information into multiple tags such as pending application, application in progress, and application review completed. Different values are assigned to different tags in each attribute according to preset rules, thereby obtaining the initial feature vector of the user corresponding to each user identification node.
[0113] Step S506: The association matrix is obtained by fusing the user's initial feature vector with the node's initial feature vectors corresponding to other nodes.
[0114] Specifically, the computer device calculates each corresponding matrix element in the correlation matrix according to the following formula 3, and then obtains the correlation matrix from each matrix element:
[0115]
[0116] Where, α ij This represents the matrix element in the i-th row and j-th column of the correlation matrix. For the user's initial feature vector, This represents the initial feature vector of the nodes corresponding to other nodes, where tanh is the hyperbolic tangent function, and N... j Let v and W be the total number of nodes contained in the j-th subgraph. att Both U and U are parameters learned during the training of the model.
[0117] In this embodiment, user identification nodes are obtained from each node. The account channel information, geographical information, and application status information of the user identification nodes are categorized and identified. Several common methods for creating accounts are defined. Geographical information is divided into different regions or by province and city. Application status information is divided into multiple tags such as pending application, application in progress, and application review completed. Different values are assigned to different tags in each attribute according to preset rules, thereby obtaining the user initial feature vector corresponding to each user identification node. Finally, based on the model training parameters, the user initial feature vector of the user identification node, and the node initial feature vectors of other nodes, the matrix elements in the association matrix are determined. The association matrix is then obtained from each matrix element, realizing the dynamic updating of the association matrix and automatically iterating to obtain appropriate matrix elements, i.e., the importance of each node, thereby improving the accuracy of constructing the association matrix.
[0118] In one embodiment, such as Figure 6As shown, feature input sample data is obtained based on user feature maps, correlation matrices, and initial feature vectors of nodes, including:
[0119] Step S602: Determine the sub-maps in the user feature map based on the user identifier nodes in each node.
[0120] Specifically, the computer device uses the topological structure of six categories of nodes corresponding to the same user in the historical user feature data—user identifier, identity feature, unit feature, bank card feature, device feature, and contact information—as a subgraph. Based on the user identifier node, it determines the other five nodes corresponding to it, forming a subgraph.
[0121] Step S604: Obtain the user feature map weight matrix based on the sub-map.
[0122] Specifically, the computer device assigns a weight value to each sub-map according to a preset rule specified based on experience, and then the weight values of the sub-maps in the user feature map are used to form a user feature map weight matrix, wherein the user feature map weight matrix is a diagonal matrix.
[0123] Step S606: Obtain the node degree matrix by fusing the user feature map weight matrix and the correlation matrix.
[0124] Specifically, the computer device calculates the node degree matrix D according to the following formula 4. v The matrix element D in the i-th row and i-th column v_ii :
[0125]
[0126] Where M is the number of sub-maps in the user feature map, and W ∈∈ H is the matrix element in the ∈ ∈ row and ∈ ∈ column of the user feature map weight matrix. i∈ Let be the matrix element in the i-th row and ∈ column of the correlation matrix H.
[0127] Step S608: Obtain the subgraph degree matrix based on the subgraph nodes and the correlation matrix in the subgraph.
[0128] Specifically, the computer device calculates the sub-graph degree matrix D according to the following formula 5. e The matrix element D in the ∈ column. e_∈∈ Then by D e_∈∈ Determine the subgraph spectral degree matrix D e :
[0129]
[0130] Among them, D e_∈∈ D is the subgraph spectral degree matrix eThe matrix element in row ∈ and column ∈, where N is the number of nodes in the entire user feature map, and H... i∈ Let be the matrix element in the i-th row and ∈ column of the correlation matrix H.
[0131] Step S610: Obtain feature input sample data based on the user feature graph weight matrix, correlation matrix, node degree matrix, subgraph degree matrix, and node initial feature vector.
[0132] In this embodiment, sub-graphs in the user feature graph are determined based on the user identifier nodes in each node. The user feature graph weight matrix is obtained based on the sub-graphs. The node degree matrix is obtained by fusing the user feature graph weight matrix and the correlation matrix. The sub-graph degree matrix is obtained based on the sub-graph nodes and the correlation matrix. Feature input sample data is obtained based on the user feature graph weight matrix, correlation matrix, node degree matrix, sub-graph degree matrix, and node initial feature vectors. This achieves full mining of the nodes and topology of the user feature graph, improving the accuracy of constructing feature input sample data based on the user feature graph.
[0133] In one embodiment, such as Figure 7 As shown, the account status model is trained based on the feature input sample data, including:
[0134] Step S702: Normalize the feature input sample data to obtain standard input sample data;
[0135] Step S704: Input the standard input sample data into the graph convolutional neural network model to train and obtain the account status model.
[0136] Specifically, the computer device obtains the node feature propagation rules in the hypergraph convolution by fusing standard input sample data and graph convolutional neural network kernels:
[0137]
[0138] Among them, X (l) and X (l+1) This represents the node features at layers l and l+1, where W is the user feature map weight matrix determined according to step S604, and Θ... (l) H represents the convolutional kernel of layer l, which can transform the dimensionality of node features, and its weights can be trained. T D represents the transpose of H, where H is the incidence matrix determined according to step S506. v (node degree matrix) and D e The (subgraph spectral degree matrix) serves a normalization function. According to the definition of hypergraph convolution, this layer-by-layer propagation rule can achieve a "point-edge-point" transformation: firstly, node features are transformed through a learnable Θ... (l)To extract new dimensional features from nodes, and then with H T After multiplication, the sum of features of the nodes contained in each hyperedge is calculated, which is the feature of that hyperedge. Finally, by multiplying with H, the features of all hyperedges containing a node are aggregated to form node features with higher-order correlation.
[0139] In this embodiment, by inputting feature input sample data into a graph convolutional neural network model for training, an account state model is obtained, which can effectively capture high-order relationships and local clustering structures between nodes, thereby achieving efficient information propagation between nodes.
[0140] In one embodiment, such as Figure 8 As shown, the account status model is obtained by training a graph convolutional neural network model with standard input sample data, including:
[0141] Step S802: Input the standard input sample data into the graph convolutional neural network model to obtain the actual output value of the model.
[0142] In the untrained graph convolutional neural network model, the initial values of each training parameter can be randomly specified, set based on experience, or determined by intelligent optimization algorithms such as particle swarm optimization and genetic algorithms. No specific restrictions are imposed here.
[0143] Step S804: Classify the actual output value of the model according to the preset classification algorithm to obtain the actual estimated result of the account status corresponding to the input sample data.
[0144] Specifically, the computer equipment classifies user accounts into two categories: accounts at risk of delinquency and accounts without risk of delinquency. An XGBoost classifier is used, with the actual output value of the graph convolutional neural network input into the XGBoost classifier to obtain the actual predicted account status. The actual predicted account status Y is calculated according to the following formula 7. p :
[0145] Y p =XGBoost(X K ), formula 7
[0146] Where X K These are the actual output values of the K graph convolutional neural networks determined according to the aforementioned steps.
[0147] Step S806: When the error between the actual predicted account status and the preset standard predicted account status in the standard input sample data is greater than the preset error threshold, return to the step of inputting the standard input sample data into the graph convolutional neural network model to obtain the actual output value of the model.
[0148] Specifically, the computer device compares the actual account status prediction result obtained in the aforementioned steps with the preset standard account status prediction result in the standard input sample data to obtain the prediction result error value. Then, it compares the prediction result error value with a preset error threshold. When the prediction result error value is greater than the preset error threshold, it returns to the step of inputting the standard input sample data into the graph convolutional neural network model to obtain the actual output value of the model, and continues training. The computer device calculates the prediction result error value according to the following formula 8:
[0149]
[0150] Where Loss is the error value of the prediction result. To represent the actual estimated result of the account status of the i-th node, y i This is the pre-defined standard prediction result for the account status in the standard input sample data.
[0151] Step S808: When the error between the actual predicted account status and the preset standard predicted account status in the standard input sample data is less than or equal to the preset error threshold, the training is complete and the account status model is obtained.
[0152] In this embodiment, standard input sample data is input into a graph convolutional neural network model to obtain the actual output value of the model. The actual output value of the model is classified according to a preset classification algorithm to obtain the actual estimated account status corresponding to the input sample data. When the error between the actual estimated account status and the preset standard estimated account status in the standard input sample data is greater than a preset error threshold, the process returns to the step of inputting the standard input sample data into the graph convolutional neural network model to obtain the actual output value of the model. Then, it is determined that when the error between the actual estimated account status and the preset standard estimated account status in the standard input sample data is less than or equal to the preset error threshold, the training is completed, and the account status model is obtained. Through continuous iteration until the iteration stopping condition is met, the model training is completed, thus improving the accuracy of model training.
[0153] This application also provides an application scenario in which the above-mentioned account status model construction method is applied to a scenario of predicting the risk of user account delinquency. Specifically, the application of the account status model construction method in this application scenario is as follows:
[0154] Computer devices acquire users' personal data, including date of birth, education level, affiliated companies, contact information, and inviter information. This data is essential information that users fill in when applying for loans through the loan application client. When a user interacts with the client, the system assigns them a unique ID number, which serves as the user's unique identifier in subsequent graph construction. Furthermore, the server can also obtain device information used by the user when logging into the client, such as device identification code, device brand, and physical address. Finally, after anonymization, missing value and outlier processing, this user data can be categorized into six main types: user characteristics, identity characteristics, company characteristics, bank card characteristics, device characteristics, and contact information. A graph database is used to store the user data obtained in the aforementioned steps. The graph database's data structure includes node sets, edge sets, and node attributes; nodes with the same characteristic fields are interconnected. User ID, identity, company, bank card, device, and contact information are used as nodes, such as... Figure 11 The graph shown represents all nodes associated with a user, which can also be considered a hyperedge belonging to that user; therefore, it is a heterogeneous graph. User ID nodes include attributes such as account creation channel, geographic location, and application status; identity nodes include attributes such as identity status, education level, job title, and credit history; organization nodes include attributes such as company type, geographic location, and registered capital; bank card nodes include attributes such as account type and city of account opening; and device nodes include attributes such as device identification code and device system.
[0155] Because users may have relationships such as inviters or contacts, multiple users may belong to the same organization, or multiple users may use the same device or contact information, different user nodes can be linked together through other nodes. For example... Figure 12 As shown, User A and User B work in the same company, and User A and User C share the same device and contact information.
[0156] Let the hypergraph consisting of all users and their associated nodes be represented as G = (V, E), containing N nodes and M hyperedges. Clearly, the number of user nodes is also M, and N is much larger than M. Next, we will discuss the input part of the model. For each hyperedge e... j Each ∈E is assigned a weight. Therefore, a trainable hyperedge weight diagonal matrix W∈R can be obtained. M ×M According to the definition of a hypergraph, a hypergraph can also be represented by the incidence matrix H∈R. N×M To indicate that if node v i Belongs to hyperedge e j Then H ij =1, otherwise H ij =0.
[0157] The degree of a node is the sum of the weights of all its connected hyperedges. The node degree matrix D is calculated as follows. v ∈R N×N :
[0158]
[0159] The degree of a hyperedge is the number of nodes contained within it. The degree matrix D of the hyperedge is calculated. e ∈R M×M :
[0160]
[0161] After normalizing and performing feature processing on the node attributes, we can obtain the feature vector of each node, which, along with the hyperedge weight diagonal matrix W, the incidence matrix H, and the node degree matrix D of the hypergraph, can be used to construct the feature vector. v Hypermarginality matrix D e These serve as the input to the model. Since different types of nodes possess features of different dimensions, it is necessary to learn a function to map these different feature vectors to d. N A public space of dimensions. For example, for a bank card node v i ∈V card The original node features Transform into a new initial feature vector
[0162]
[0163] Where σ is a non-linear activation function; unless otherwise specified, the non-linear activation function used in this invention is the sigmoid function. Therefore, the transformation function can also be written as:
[0164]
[0165] It can be known that W card and b card These are all parameters that need to be learned during model training. The processing of device, identity, user, and unit node features is similar to that of a bank card. It's important to note that the "contact" node itself does not have any node attributes; therefore, its initial feature vector is the average of the initial feature vectors of the k connected nodes.
[0166]
[0167] At this point, the initial feature vectors of all nodes are represented as follows:
[0168] In practical graph structures, the relationships between nodes are usually not paired but rather higher-order. To capture these higher-order relationships and local clustering structures, this invention uses hypergraph convolution to achieve efficient information propagation between nodes. We define an L-layer hypergraph convolutional neural network, where the first layer's node features are X. (0) The layer-by-layer propagation rules for node features are defined as follows:
[0169]
[0170] X (l) and X (l+1) Θ represents the node features at layers l and l+1. (l) H represents the convolutional kernel of layer l, which can transform the dimensionality of node features, and its weights can be trained. T Represents the transpose of the incidence matrix H, and the degree matrix D. v and hypermarginality matrix D e It serves a normalization function. According to the definition of hypergraph convolution, this layer-by-layer propagation rule can achieve a "point-edge-point" transformation: first, node features are transformed through a learnable Θ... (l) To extract new dimensional features from nodes, and then with H T After multiplication, the sum of features of the nodes contained in each hyperedge is calculated, which is the feature of that hyperedge. Finally, by multiplying with H, the features of all hyperedges containing a node are aggregated to form node features with higher-order correlation.
[0171] However, in the original correlation matrix H, nodes within the same hyperedge have the same impact on the hyperedge features, which is unsuitable for hypergraphs in credit scenarios. This is because different types of nodes have different features and different importance. Based on experience with manual review, bank card nodes have more financial attributes; therefore, when determining whether a user has a risk of delinquency, the features of bank card nodes are more important than those of device nodes. Attention mechanisms can assign different weights to different nodes during feature aggregation. This invention incorporates an attention mechanism into the calculation of hyperedge features, establishing a new correlation matrix H. new If node v i Belongs to hyperedge e j Then H ij =α ij Otherwise H ij =0. α ij That is, node v i In the super-edge e j The attention coefficients corresponding to the feature representation. We define the hyperedge e. j The original feature vector Features equal to those of user nodes in the hyperedge α ijThe calculation formula is as follows:
[0172]
[0173] Among them, v and W att Both U and N are trainable parameters, tanh is the hyperbolic tangent function, and N is the hyperbolic tangent function. j Indicates the superedge e j All nodes included. At this point, the node feature propagation rule in hypergraph convolution is updated to the following formula:
[0174]
[0175] Finally, the node representation vector of the last layer of the hypergraph convolutional neural network is obtained as follows:
[0176] In the task of predicting overdue payment risk, users need to be classified into two categories: customers with overdue payment risk and those without. Essentially, this is a binary classification problem. This model uses the XGBoost classifier to classify the XGBoost data obtained from the preceding steps. final The final representation vector X of K known overdue and non-overdue tagged user nodes in the dataset. K As input to the classifier:
[0177] Y p =XGBoost(X K ), Formula 17
[0178] The probability Y of obtaining the user's predicted category p ∈R K×1 Then, the cross-entropy loss function, commonly used in binary classification problems, is used as the objective function. The training objective of the model is to minimize the objective function. The objective function Loss is shown below, where y i This represents the actual overdue label of the i-th node; it is 1 if it is, and 0 otherwise. This represents the probability that the i-th node is predicted to be overdue.
[0179]
[0180] After multiple training iterations, once the loss function reaches its optimal value, we obtain a model with optimal parameters. Then, by inputting the current user data into the model, we obtain a prediction result regarding whether the current user has a risk of defaulting on payments. Figure 13 As shown.
[0181] The aforementioned account status model construction method, apparatus, computer equipment, and storage medium acquire historical user feature data, construct a user feature map based on the historical user feature data, and then generate corresponding initial feature vectors for nodes based on the node attribute features of each node in the user feature map and the initial values of the model training parameters. The initial values of the model training parameters are the initial values of the parameters to be trained in the account status model. An association matrix is obtained by fusing the initial feature vectors of the nodes. The association matrix is used to characterize the association relationship and importance of each node. Feature input sample data is obtained based on the user feature map, the association matrix, and the initial feature vectors of the nodes. The account status model is trained based on the feature input sample data. The account status model is used to evaluate the operational status of the target account. In this way, by incorporating the initial values of the model training parameters into the process of generating the initial feature vectors of the nodes, new training parameters are continuously generated as the model iterates. The feature vectors of each node are then regenerated based on the new parameters generated in each iteration, accelerating the convergence of the model. Furthermore, the correlation matrix is obtained by fusing the initial feature vectors of the nodes. During model training, the importance weights assigned to each node are dynamically adjusted according to the importance of different nodes. This allows for better mining of the importance of different nodes based on specific application scenarios, enabling the trained model to better mine data features and effectively improve the accuracy of account status model prediction.
[0182] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.
[0183] In one embodiment, such as Figure 9 As shown, an account status model construction device is provided. This device can be a software module, a hardware module, or a combination of both integrated into a computer device. Specifically, the device includes: a data acquisition module 902, a sample data generation module 904, and a training module 906, wherein:
[0184] Data acquisition module 902 is used to acquire historical user characteristic data;
[0185] The sample data generation module 904 is used to construct a user feature map based on historical user feature data; generate corresponding initial feature vectors for nodes based on the node attribute features of each node in the user feature map and the initial values of the model training parameters, where the initial values of the model training parameters are the initial values of the parameters to be trained in the account status model; obtain an association matrix by fusing the initial feature vectors of the nodes, which is used to characterize the association relationship between each node and the importance of each node; and obtain feature input sample data based on the user feature map, the association matrix, and the initial feature vectors of the nodes.
[0186] Training module 906 is used to train an account status model based on feature input sample data. The account status model is used to evaluate the operational status of the target account.
[0187] The aforementioned account status model construction device acquires historical user feature data, constructs a user feature map based on the historical user feature data, and then generates corresponding initial feature vectors for each node based on the node attribute features of each node in the user feature map and the initial values of the model training parameters. The initial values of the model training parameters are the initial values of the parameters to be trained in the account status model. An association matrix is obtained by fusing the initial feature vectors of the nodes. The association matrix is used to characterize the association relationship between each node and the importance of each node. Feature input sample data is obtained based on the user feature map, the association matrix, and the initial feature vectors of the nodes. The account status model is trained based on the feature input sample data. The account status model is used to evaluate the operational status of the target account. In this way, by incorporating the initial values of the model training parameters into the process of generating the initial feature vectors of the nodes, new training parameters are continuously generated as the model iterates. The feature vectors of each node are then regenerated based on the new parameters generated in each iteration, accelerating the convergence of the model. Furthermore, the correlation matrix is obtained by fusing the initial feature vectors of the nodes. During model training, the importance weights assigned to each node are dynamically adjusted according to the importance of different nodes. This allows for better mining of the importance of different nodes based on specific application scenarios, enabling the trained model to better mine data features and effectively improve the accuracy of account status model prediction.
[0188] In one embodiment, the sample data generation module 904 is further configured to obtain a nonlinear activation function; determine the original node features of each node based on the node attributes of each node; and generate the initial node feature vector corresponding to each node based on the nonlinear activation function, the original node features, and the initial values of the model training parameters.
[0189] In one embodiment, the sample data generation module 904 is further configured to, when the node is a contact information node, obtain the initial feature vector of the adjacent node corresponding to the adjacent node of the contact information node, wherein the contact information node is a node whose node attribute is contact information; and determine the initial feature vector of the node corresponding to the contact information node based on the mean of the initial feature vectors of the adjacent nodes.
[0190] In one embodiment, the sample data generation module 904 is further configured to obtain user identification nodes in each node; obtain the user's initial feature vector based on the user identification node's account channel information, geographical information, and application status information; and obtain an association matrix by fusing the user's initial feature vector with the node's initial feature vector corresponding to other nodes.
[0191] In one embodiment, the sample data generation module 904 is further configured to determine the sub-graph in the user feature graph based on the user identifier node in each node; obtain the user feature graph weight matrix based on the sub-graph; obtain the node degree matrix by fusing the user feature graph weight matrix and the correlation matrix; obtain the sub-graph degree matrix based on the sub-graph nodes and the correlation matrix in the sub-graph; and obtain feature input sample data based on the user feature graph weight matrix, the correlation matrix, the node degree matrix, the sub-graph degree matrix, and the node initial feature vector.
[0192] In one embodiment, the training module 906 is further configured to normalize the feature input sample data to obtain standard input sample data; and input the standard input sample data into the graph convolutional neural network model for training to obtain the account status model.
[0193] In one embodiment, the training module 906 is further configured to input standard input sample data into a graph convolutional neural network model to obtain the actual output value of the model; classify the actual output value of the model according to a preset classification algorithm to obtain the actual estimated account status corresponding to the input sample data; when the error between the actual estimated account status and the preset standard estimated account status in the standard input sample data is greater than a preset error threshold, return to the step of inputting standard input sample data into a graph convolutional neural network model to obtain the actual output value of the model; when the error between the actual estimated account status and the preset standard estimated account status in the standard input sample data is less than or equal to the preset error threshold, the training is completed and the account status model is obtained.
[0194] Specific limitations regarding the account state model construction device can be found in the limitations of the account state model construction method described above, and will not be repeated here. Each module in the aforementioned account state model construction device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.
[0195] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 10As shown, the computer device includes a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program implements an account state model construction method. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.
[0196] Those skilled in the art will understand that Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0197] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.
[0198] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above method embodiments.
[0199] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps in the above method embodiments.
[0200] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.
[0201] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.
[0202] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0203] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.
Claims
1. A method for constructing an account status model, characterized in that, The method includes: Obtain historical user characteristic data; Construct a user feature map based on the historical user feature data; Based on the node attribute features of each node in the user feature map and the initial values of the model training parameters, a corresponding initial node feature vector is generated. The initial values of the model training parameters are the initial values of the parameters to be trained in the account status model. An association matrix is obtained by fusing the initial feature vectors of the nodes. The association matrix is used to characterize the association relationship between the nodes and the importance of each node. Obtaining feature input sample data based on the user feature map, the correlation matrix, and the initial feature vectors of the nodes includes: determining a sub-map in the user feature map based on the user identifier nodes in each node; obtaining a user feature map weight matrix based on the sub-map; fusing the user feature map weight matrix and the correlation matrix to obtain a node degree matrix; obtaining a sub-map degree matrix based on the sub-map nodes and the correlation matrix; and obtaining feature input sample data based on the user feature map weight matrix, the correlation matrix, the node degree matrix, the sub-map degree matrix, and the initial feature vectors of the nodes. An account status model is trained based on the feature input sample data, and the account status model is used to evaluate the operational status of the target account.
2. The method according to claim 1, characterized in that, The process of generating corresponding initial feature vectors for nodes based on the node attribute features of each node in the user feature map and the initial values of the model training parameters includes: Obtain the non-linear activation function; The original node characteristics of each node are determined based on the node attributes of each node; The initial feature vectors of each node are generated based on the nonlinear activation function, the original node features, and the initial values of the model training parameters.
3. The method according to claim 1, characterized in that, The process of generating corresponding initial feature vectors for nodes based on the node attribute features of each node in the user feature map and the initial values of the model training parameters includes: When the node is a contact information node, the initial feature vector of the adjacent node corresponding to the adjacent node of the contact information node is obtained, where the contact information node is a node whose node attribute is contact information; The initial feature vector of the node corresponding to the contact information node is determined based on the mean of the initial feature vectors of the adjacent nodes.
4. The method according to claim 1, characterized in that, The association matrix of each node obtained by fusing the initial feature vectors of the nodes includes: Obtain the user identifier node from each of the nodes; The user's initial feature vector is obtained based on the user identification node's account channel information, geographical information, and application status information; The association matrix is obtained by fusing the user's initial feature vector with the node's initial feature vectors corresponding to other nodes.
5. The method according to claim 1, characterized in that, The process of obtaining feature input sample data based on the user feature map, the correlation matrix, and the initial feature vectors of the nodes includes: The sub-map in the user feature map is determined based on the user identifier node in each node; The user feature map weight matrix is obtained based on the sub-map. The node degree matrix is obtained by fusing the user feature map weight matrix and the correlation matrix. The sub-graph degree matrix is obtained based on the sub-graph nodes in the sub-graph and the correlation matrix; Feature input sample data is obtained based on the user feature graph weight matrix, the correlation matrix, the node degree matrix, the subgraph degree matrix, and the node initial feature vector.
6. The method according to claim 1, characterized in that, The account status model trained based on the feature input sample data includes: The feature input sample data is normalized to obtain standard input sample data; The standard input sample data is input into the graph convolutional neural network model for training to obtain the account status model.
7. The method according to claim 6, characterized in that, The step of training the account status model by inputting the standard input sample data into the graph convolutional neural network model includes: The standard input sample data is input into the graph convolutional neural network model to obtain the actual output value of the model; The actual output values of the model are classified according to a preset classification algorithm to obtain the actual estimated account status corresponding to the input sample data; When the error between the actual estimated account status and the preset standard estimated account status in the standard input sample data is greater than the preset error threshold, return to the step of inputting the standard input sample data into the graph convolutional neural network model to obtain the actual output value of the model; When the error between the actual predicted account status and the preset standard predicted account status in the standard input sample data is less than or equal to the preset error threshold, the training is complete and the account status model is obtained.
8. An account status model construction apparatus, characterized in that, The device includes: The data acquisition module is used to acquire historical user characteristic data; The sample data generation module is used to construct a user feature map based on the historical user feature data; generate corresponding initial feature vectors for nodes based on the node attribute features of each node in the user feature map and the initial values of the model training parameters, wherein the initial values of the model training parameters are the initial values of the parameters to be trained in the account status model; fuse the initial feature vectors of the nodes to obtain an association matrix, wherein the association matrix is used to characterize the association relationship of each node and the importance of each node; and obtain feature input sample data based on the user feature map, the association matrix, and the initial feature vectors of the nodes. The training module is used to train an account status model based on the feature input sample data, and the account status model is used to evaluate the operational status of the target account.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.