Credit risk prediction model training method and credit risk prediction method

Abstract

The embodiment of the invention provides a credit risk prediction model training method and a credit risk prediction method, and the model training method comprises the steps: training an initial credit risk prediction model constructed based on a time sequence diagram network through employing the historical credit transaction record data of a user and the historical social record data of the user; according to a target data difference between a predicted credit evaluation result output by the credit risk prediction model and a real credit overdue result of the user corresponding to each training sample data, adjusting a model parameter until the target data difference is smaller than a preset data difference threshold value; and determining the credit risk prediction model of which the target data difference is smaller than a preset data difference threshold as a trained target credit risk prediction model. Therefore, the target credit risk prediction model can comprehensively give a more accurate risk prediction result according to the historical credit transaction record data of the user and the historical social relation record data of the user, and the accuracy of credit risk assessment can be improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a credit risk prediction model training method and a credit risk prediction method. Background Art

[0002] With the continuous development of artificial intelligence technology, artificial intelligence technology has gradually been applied to the application scenario of financial credit, such as using artificial intelligence network models to predict credit risks, thereby avoiding bad debt losses caused by overdue credit. Existing common credit risk prediction methods include: credit risk assessment schemes based on statistical models, credit risk assessment schemes based on machine learning, credit risk assessment schemes based on deep learning, credit risk assessment schemes based on ensemble learning, etc.

[0003] However, judging from the actual application effects of various credit risk prediction schemes, existing credit risk assessment schemes are usually unable to effectively capture long-term dependencies in time series, have difficulty in handling complex time characteristics, and have the problem of low prediction accuracy. Summary of the invention

[0004] In view of this, embodiments of the present application provide a credit risk prediction model training method and a credit risk prediction method to improve the prediction accuracy of the credit risk prediction model.

[0005] In a first aspect, an embodiment of the present application provides a method for training a credit risk prediction model, wherein the method comprises: obtaining a training sample data set, and inputting the training sample data set into an initial credit risk prediction model, wherein the training sample data set comprises various training sample data, and each of the training sample data comprises: historical credit transaction record data of a user, and historical social relationship record data of a user; the credit risk prediction model is a neural network model constructed based on a time series graph network;

[0006] The initial credit risk prediction model processes each piece of training sample data and outputs a predicted credit assessment result corresponding to each piece of training sample data, wherein the predicted credit assessment result is a predicted credit overdue probability of each of the users under different credit cycles;

[0007] According to the target data difference between the predicted credit assessment result and the actual credit overdue result of the user corresponding to each training sample data, the model parameters of the initial credit risk prediction model are adjusted until the target data difference is less than a preset data difference threshold;

[0008] The credit risk prediction model corresponding to the target data difference being less than the preset data difference threshold is determined as the trained target credit risk prediction model.

[0009] In some possible embodiments, the credit risk prediction model includes: a feature encoder, a sequence model, a sequence embedding module, and a risk assessment module, wherein the initial credit risk prediction model performs data processing on each piece of training sample data and outputs a predicted credit assessment result corresponding to each piece of training sample data, including:

[0010] The feature encoder performs feature extraction based on each of the training sample data to obtain a first transaction behavior feature of each user in a time dimension and a second transaction behavior feature of each user in a social relationship dimension; wherein the first transaction behavior feature and the second transaction behavior feature are both represented by feature vectors;

[0011] Inputting the first transaction behavior feature and the second transaction behavior feature into the sequence model, and obtaining the time dependency and dynamic change of each feature vector from the sequence model;

[0012] The sequence embedding module converts the time dependency and dynamic change of each feature vector into a sequence embedding vector, and the sequence embedding vector is used to represent the comprehensive credit behavior characteristics of the user;

[0013] The comprehensive credit behavior characteristics are input into the risk assessment module, and the risk assessment module outputs the predicted credit assessment result based on the comprehensive credit behavior characteristics.

[0014] In some possible embodiments, the first transaction behavior feature includes: transaction frequency μ, transaction amount standard deviation σ, and obtaining the first transaction behavior feature of each user in the time dimension includes:

[0015] The transaction frequency μ is obtained by the feature encoder based on the following formula:

[0016]

[0017] Where n is the number of transactions of the user within time t, x t is the transaction amount within time t;

[0018] The transaction amount standard deviation σ is extracted by the feature encoder based on the following formula:

[0019]

[0020] In some possible embodiments, the second transaction behavior feature includes: a relationship feature Ru between user u and associated user v within a preset time period, and obtaining the second transaction behavior feature of each user in the social relationship dimension includes:

[0021] The feature encoder extracts the relationship feature Ru between the user and the associated users within a preset time period based on the following formula:

[0022]

[0023] Among them, f(v) is the transaction behavior feature of user v, w v is the characteristic coefficient of the transaction behavior characteristics of user v.

[0024] In some possible embodiments, the sequence model includes: a long short-term memory network model; the first transaction behavior feature and the second transaction behavior feature are input into the sequence model, and the time dependency and dynamic change of each feature vector are obtained by the sequence model; including:

[0025] The sequence model obtains the time dependency and dynamic changes in each feature vector based on the following formula:

[0026] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0027] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0028]

[0029] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0030] h t =o t tanh(C t )

[0031] Among them, f t For the forget gate, i t is the input gate, o t is the output gate, C t is the state of the memory unit, h t is the hidden state, is the candidate memory cell state, σ() is the sigmoid function, x t is the input of the sequence model at time t, W f is the weight parameter of the forget gate, Wo is the weight parameter of the output gate, W C is the weight parameter of the state of the memory unit, W i is the weight parameter of the input gate, b f is the forget gate bias, b i is the output gate bias, b c is the memory unit bias, b o Bias for the output gate.

[0032] In some possible embodiments, the credit risk prediction model further includes: a relationship graph updating module, which adjusts the model parameters of the initial credit risk prediction model according to the target data difference between the predicted credit assessment result and the actual credit overdue result of the user corresponding to each training sample data, including:

[0033] The relationship graph updating module is used to update the weight coefficients between neighbor nodes in the graph network of the credit risk prediction model according to the target data difference between the predicted credit assessment result output by the credit risk prediction model in each round and the actual credit overdue result.

[0034] In some possible embodiments, the outputting of the predicted credit assessment result by the risk assessment module based on the comprehensive credit behavior characteristics includes:

[0035] The risk assessment module calculates the predicted credit overdue probability of each user based on the sequence embedding vector and using a preset weighted calculation rule.

[0036] In some possible embodiments, the credit risk prediction model further includes: a fusion module, wherein the fusion module integrates the output of the time series graph network and the LSTM through a self-attention mechanism, and specifically integrates the output of the time series graph network and the LSTM through the following formula:

[0037]

[0038] Where Q is the latent vector of each time step of the temporal network, K is the key vector, which is used to calculate the similarity of the target vector with the latent vector, V is the weight vector, which is used to extract the target output vector based on the target vector similarity vector, and d k is the characteristic dimension of the latent vector and the key vector.

[0039] In a second aspect, an embodiment of the present application provides a credit risk prediction method, wherein the method comprises:

[0040] Acquire the credit information of the target user to be processed, wherein the credit information of the target user includes: historical transaction records and historical spatiotemporal relationship records of the target user;

[0041] Inputting the target user's credit information into a target credit risk prediction model, and having the target credit risk prediction model perform data processing based on the target user's credit information to output the target user's credit overdue probability under different credit cycles;

[0042] The target credit risk prediction model is a credit risk prediction model obtained by training in advance using the credit risk prediction model training scheme described in the first aspect.

[0043] In a third aspect, an embodiment of the present application provides a credit risk prediction model training device, wherein the device comprises:

[0044] An acquisition module is used to acquire a training sample data set and input the training sample data set into an initial credit risk prediction model, wherein the training sample data set includes various training sample data, and each of the training sample data includes: a user's historical credit transaction record data and a user's historical social relationship record data; the credit risk prediction model is a neural network model constructed based on a time series graph network;

[0045] A model training module, used to process each piece of training sample data using the initial credit risk prediction model, and output a predicted credit assessment result corresponding to each piece of training sample data, wherein the predicted credit assessment result is a predicted credit overdue probability of each of the users under different credit cycles;

[0046] A model parameter adjustment module, used to adjust the model parameters of the initial credit risk prediction model according to the target data difference between the predicted credit assessment result and the actual credit overdue result of the user corresponding to each training sample data, until the target data difference is less than a preset data difference threshold;

[0047] The model determination module is used to determine the credit risk prediction model corresponding to the target data difference being less than a preset data difference threshold as the trained target credit risk prediction model.

[0048] In a fourth aspect, an embodiment of the present application provides a credit risk prediction device, wherein the device comprises:

[0049] A credit information acquisition module, used to acquire the credit information of the target user to be processed, wherein the credit information of the target user includes: the historical transaction records and historical spatiotemporal relationship records of the target user;

[0050] A credit information processing module, used to input the credit information of the target user into a target credit risk prediction model, and the target credit risk prediction model performs data processing based on the credit information of the target user, and outputs the credit overdue probability of the target user under different credit cycles;

[0051] The target credit risk prediction model is a credit risk prediction model obtained by training in advance using the credit risk prediction model training scheme described in the first aspect.

[0052] In a fifth aspect, an embodiment of the present application provides an electronic device, wherein the electronic device includes:

[0053] Processor; and

[0054] Memory for storing programs,

[0055] Wherein, the program includes instructions, which, when executed by the processor, cause the processor to execute the credit risk prediction model training method described in the first aspect or the credit risk prediction method described in the second aspect.

[0056] In a fourth aspect, an embodiment of the present application provides a non-transitory computer-readable storage medium storing computer instructions, characterized in that the computer instructions are used to enable a computer to execute the credit risk prediction model training method described in the first aspect or the credit risk prediction method described in the second aspect.

[0057] Beneficial effects of this application:

[0058] The embodiments of the present application provide a credit risk prediction model training method and a credit risk prediction method, wherein the credit risk prediction model training method uses the user's historical credit transaction record data and the user's historical social record data to train an initial credit risk prediction model constructed based on a time-series graph network, and the initial credit risk prediction model constructed based on the time-series graph network processes data for each training sample, outputs a predicted credit assessment result corresponding to each training sample, and finally adjusts the model parameters of the initial credit risk prediction model according to the target data difference between the predicted credit assessment result and the actual credit overdue result of the user corresponding to each training sample data, until the target data difference between the two is less than a preset data difference threshold, and then determines the credit risk prediction model whose target data difference is less than the preset data difference threshold as the trained target credit risk prediction model.

[0059] Since the user's credit behavior is not only affected by his own credit behavior, but also by the user's social relations, economic activities, and living habits, the embodiment of the present application is selected to reflect the changes in the social relations, economic activities, and living habits of the credit user through historical social records, and the user's historical social record data is included in the training sample data. The target credit risk prediction model obtained by training in this way can comprehensively consider multi-dimensional credit assessment factors such as time and user relationship dimensions, so that the target credit risk prediction model obtained by training can comprehensively give a more accurate risk prediction result based on the user's historical credit transaction record data and the user's historical social relationship record data, which helps to improve the accuracy of credit risk assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Further details, features and advantages of the present application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which:

[0061] Figure 1 A flow chart of a credit risk prediction model training method provided in an embodiment of the present application is shown;

[0062] Figure 2 A schematic diagram of a network topology of a timing diagram network provided in an embodiment of the present application is shown;

[0063] Figure 3 A schematic diagram of a model topology structure of a credit risk prediction model provided by an embodiment of the present application is shown;

[0064] Figure 4 Another flow chart of a method for training a credit risk prediction model provided in an embodiment of the present application is shown;

[0065] Figure 5 A schematic diagram of a process of a credit risk prediction method provided by an embodiment of the present application is shown;

[0066] Figure 6 A logical structure diagram of a credit risk prediction model training device provided in an embodiment of the present application is shown;

[0067] Figure 7 A logical structure diagram of a credit risk prediction device provided by an embodiment of the present application is shown;

[0068] Figure 8 A structural block diagram of an exemplary electronic device that can be used to implement an embodiment of the present application is shown. DETAILED DESCRIPTION

[0069] The embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be construed as being limited to the embodiments described herein. Instead, these embodiments are provided to provide a more thorough and complete understanding of the present application. It should be understood that the drawings and embodiments of the present application are only for exemplary purposes and are not intended to limit the scope of protection of the present application.

[0070] It should be understood that the various steps described in the method implementation of the present application can be performed in different orders and / or performed in parallel. In addition, the method implementation may include additional steps and / or omit the steps shown. The scope of the present application is not limited in this respect.

[0071] The term "including" and its variations used in this document are open inclusions, that is, "including but not limited to". The term "based on" means "based at least in part on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one other embodiment"; the term "some embodiments" means "at least some embodiments". Relevant definitions of other terms will be given in the description below. It should be noted that the concepts of "first", "second", etc. mentioned in this application are only used to distinguish different devices, modules or units, and are not used to limit the order or interdependence of the functions performed by these devices, modules or units.

[0072] It should be noted that the modifications of "one" and "plurality" mentioned in the present application are illustrative rather than restrictive, and those skilled in the art should understand that unless otherwise clearly indicated in the context, it should be understood as "one or more".

[0073] Before describing in detail the credit risk prediction model training method and the credit risk prediction method provided by this application, the professional terms involved in this application will be explained here:

[0074] Temporal Graph Convolutional Network (TGCN): A model that combines a graph convolutional network (GCN) and a time series model, wherein the time series model can be a network model with time dependencies such as the LSTM (Long Short-Term Memory, LSTM) long short-term memory network model. In an embodiment of the present application, the temporal graph convolutional network TGCN is used to capture the spatiotemporal dependencies in the data, specifically by extracting the spatial features of the data of the input model through the graph convolution layer, and then capturing the temporal features of the data through the time series model, thereby comprehensively considering the spatiotemporal dynamic changes in the data.

[0075] Graph Convolutional Network: A neural network suitable for processing graph structured data. It extracts high-level features of nodes by performing convolution operations between nodes and their neighbors. GCN has a wide range of applications in social network analysis, knowledge graph construction, recommendation systems, and other fields.

[0076] Long Short-Term Memory (LSTM) is a special recursive neural network that is suitable for processing and predicting long-term dependencies in time series data. LSTM effectively solves the problems of gradient vanishing and gradient exploding in traditional recursive neural network models by introducing forget gates, input gates, and output gates.

[0077] Self-Attention Mechanism: A method used to enhance the ability to express features in deep learning models. It summarizes the correlation weights between each element and all other elements in the sequence, highlights important features, and suppresses irrelevant features. The self-attention mechanism has been widely used in Transformer network models.

[0078] In the embodiment of the present application, credit risk assessment refers to predicting the future default risk of a user by analyzing his or her credit behavior data. Credit risk assessment is usually applied to financial services such as bank credit, credit card approval, and insurance underwriting to reduce the bad debt risk of financial institutions.

[0079] Among them, common credit risk assessment solutions include:

[0080] Solution 1: Credit risk assessment solution based on statistical models. Common statistical models include: logistic regression model and discriminant analysis model. The logistic regression model is the most classic binary classification model, which is often used in credit scoring models. It assesses the user's credit risk by predicting the probability of default. Discriminant analysis models include: linear discriminant analysis (LDA) and quadratic discriminant analysis (QDA), both of which are commonly used in credit risk assessment and are classified by maximizing intra-group differences.

[0081] Solution 2: Credit risk assessment solution based on machine learning. Common credit risk assessment models based on machine learning include: decision tree model, random forest model, and support vector machine network model. Among them, decision tree can be used for credit risk assessment. By building a tree-structured decision process, users can be scored for credit. Random forest improves the robustness and prediction accuracy of the model by integrating multiple decision trees. It is often used for credit scoring and risk prediction. Support vector machine (SVM) separates data by finding the optimal hyperplane and is used for credit risk classification.

[0082] Solution 3: Credit risk assessment solution based on deep learning. Among them, the credit risk assessment model based on deep learning includes: multi-layer perceptron network model, recurrent neural network model, convolutional neural network model, autonomous mechanism model, etc. Among them, multi-layer perceptron (Multi-Layer Perceptron, MLP) is a feedforward neural network that predicts credit risk through multi-layer nonlinear transformation. Recurrent Neural Network (RNN) and its variants LSTM and GRU (Gate Recurrent Unit, gated recurrent unit network model) are often used to process time series data and capture the time dependency in user transaction behavior. Convolutional Neural Network (CNN) extracts local features through convolution operations and has also been used in pattern recognition of financial data in recent years. Self-Attention Mechanism (Self-Attention Mechanism) and Transformer: The Transformer model processes long time series data through the self-attention mechanism and performs well in credit risk assessment.

[0083] Solution 4: Credit risk assessment solution based on ensemble learning. Ensemble learning improves the stability and accuracy of credit risk assessment by combining the prediction results of multiple models. Common ensemble methods include bagging (such as random forest), boosting (such as gradient boosting machine) and stacking (stacked generalization).

[0084] Among the above solutions, the traditional statistical models and simple machine learning models of Solution 1 and Solution 2 are usually unable to effectively capture the long-term dependencies in time series and have difficulty in processing complex time features. Solution 3 relies heavily on a single model with limited expressive power. Some models lack an attention mechanism, which makes it difficult for the model to effectively distinguish important features and is easily disturbed by irrelevant features, resulting in low prediction accuracy for the entire model.

[0085] In view of this, the present application provides a credit risk prediction model training method, a credit risk prediction method, a device and an electronic device. In the first aspect, the present application provides a credit risk prediction model training method, which is applied to any electronic device with a credit risk prediction model training function, including but not limited to a personal mobile terminal, a computer or a server, etc. Figure 1 As shown, the method comprises the following steps:

[0086] S11. Obtain a training sample data set, and input the training sample data set into an initial credit risk prediction model.

[0087] The training sample data set includes various pieces of training sample data, each piece of the training sample data includes: historical credit transaction record data of the user, historical social relationship record data of the user, and the credit risk prediction model is a neural network model constructed based on a time series graph network.

[0088] S12. The initial credit risk prediction model processes each piece of training sample data and outputs a predicted credit assessment result corresponding to each piece of training sample data.

[0089] The predicted credit assessment result is the predicted credit overdue probability of each user in different credit cycles.

[0090] S13. According to the target data difference between the predicted credit assessment result and the actual credit overdue result of the user corresponding to each training sample data, the model parameters of the initial credit risk prediction model are adjusted until the target data difference is less than a preset data difference threshold.

[0091] S14. Determine the credit risk prediction model corresponding to the target data difference being less than a preset data difference threshold as the trained target credit risk prediction model.

[0092] Since the user's credit behavior is not only affected by his own credit behavior, but also by the user's social relations, economic activities, and living habits, the embodiment of the present application is selected to reflect the changes in the social relations, economic activities, and living habits of the credit user through historical social records, and the user's historical social record data is included in the training sample data. The target credit risk prediction model obtained by training in this way can comprehensively consider multi-dimensional credit assessment factors such as time and user relationship dimensions, so that the target credit risk prediction model obtained by training can comprehensively give a more accurate risk prediction result based on the user's historical credit transaction record data and the user's historical social relationship record data, which helps to improve the accuracy of credit risk assessment.

[0093] The above steps S11 to S14 will be described in detail below with reference to specific examples:

[0094] In an embodiment of the present application, a training sample data set refers to a data set that contains several pieces of data used to train a network model, wherein the data used to train the network model is the training sample data. In an embodiment of the present application, in order to verify the trained credit risk prediction model later, all sample data involved in the generation process of the entire credit risk prediction model can be divided and subdivided into a training set and a validation set (or test set). Among them, the sample data in the training set is used to train the credit risk prediction model, and the sample data in the validation set is used to verify the performance of the trained credit risk prediction model, both of which aim to ensure that the predicted credit assessment results output by the trained credit risk prediction model are close to the actual credit overdue results.

[0095] In an embodiment of the present application, sample data comes from the user's historical credit transaction record data and the user's historical social record data. Among them, the historical credit transaction record data can specifically refer to the user's personal bill. From the machine execution level, the historical credit transaction record data is a personal bill sequence, which specifically includes: the amount, transaction date, transaction type, etc. of each credit transaction of the user, which can be subdivided into borrowing records and repayment records. The historical social record data can specifically refer to the social relationship data between users. From the machine execution level, the historical social record data is a personal social relationship sequence, which specifically includes: the association relationship between users, which can be friends, family, colleagues, merchants and customers, etc. In addition, the historical credit transaction record data and the historical social record data include timestamp data, which includes: the specific time when the credit transaction occurred, and the specific time when the social relationship changed.

[0096] On this basis, when executing step S11, as an implementation method, the credit product history record can be obtained through a designated data interface to obtain the historical credit transaction record data. As another implementation method, the historical credit transaction record data for model training can be directly provided by a model developer or tester, and the historical credit transaction record data for model training can be read specifically through a designated path.

[0097] Similarly, when executing step S11, as an implementation method, the historical social record data can be obtained through the designated data interface by obtaining the user's address book using the social software or the relationship information table filled in by the user, and the association relationship between each user can be determined from it. As another implementation method, the historical social record data is the historical social record data directly provided by the model developer or tester for model training, and the historical social record data indicates the association relationship between each user and the change of the association relationship. Among them, the change of the association relationship also includes the time node when the association relationship changes.

[0098] In an embodiment of the present application, in the process of obtaining a training sample data set, the data processing method of data cleaning can also be used to clean the data of each sample data in the training sample data set according to the preset data screening rules, and delete the sample data with missing items and abnormal items to ensure the data quality of the sample data participating in the model training. Among them, the preset data screening rules can be flexibly set according to the actual data requirements, and this application does not make strict restrictions. For example, for example, the sample data must completely include the transaction amount. If there is a sample data that lacks the transaction amount, it indicates that there are missing items in the sample data. At this time, the sample data can be deleted to ensure the data quality of the sample data participating in the model training.

[0099] As an implementation method, the historical social record data may be a data table generated based on the communication records of the social software, in which the association relationship between each user and the strength of the association relationship between users are indicated. When executing step S11, a row of data corresponds to a piece of historical social record data. For example, the data table may be shown in Table 1 below:

[0100] Table 1. A schematic table of historical social record data

[0101] Serial number Specify User Associated Users Relationship Degree of association 1 Zhang San Li Si Father and Son 1 2 Zhang San Wang Wu colleague 0.2 3 Zhang San Zhao Liu colleague 0.2

[0102] The degree of association indicates the strength of the association relationship between users. The larger the value of the degree of association, the stronger the association relationship between users.

[0103] As another implementation, the historical social record data may be a relationship network diagram generated based on the communication records of social software. The relationship network diagram corresponding to the historical social record data may be as follows: Figure 2As shown in the figure, the relationship network diagram shows the relationship network between credit users and their surrounding users. Each node in the diagram represents a credit user, and the edge between nodes represents the degree of association between credit users, which can be represented by the connection weight. Specifically, each letter node in the nodes (A, B, C, D, E, F, G, H, I) represents a credit user. Each edge (w1, w2, ..., w10) represents the relationship strength or connection weight between two users. For example, w1 represents the relationship strength between user D and user C.

[0104] Among them, the contact weight represents the degree of connection between credit users. The closer the connection between credit users, the higher the contact weight. In the embodiment of the present application, the relationship between users is determined based on constraints such as the distance of geographical location, co-existing transaction behaviors or social behaviors.

[0105] In the embodiment of the present application, the initial credit risk prediction model and the target credit risk prediction model are essentially neural network models constructed based on a time series graph network, and are collectively referred to as credit risk prediction models, wherein the initial credit risk prediction model refers to an initially constructed time series network model that has not yet been trained with training sample data, and the target credit risk prediction model is a time series graph network model that is fully trained with training sample data on the basis of the initial credit risk prediction model and can accurately output credit risk assessment results. During the execution of step S13, the model parameters of the initial credit risk prediction model are adjusted, specifically, the model parameters of the initial credit risk prediction model can be directly adjusted, or the model parameters of the credit risk prediction model of the previous round of execution calculation can be adjusted according to the results of this round of training.

[0106] As an implementation method, the model architecture of the time series graph network model corresponding to the credit risk prediction model can be as follows: Figure 3 As shown, it includes the following parts:

[0107] Feature Encoder, Sequence Model, Sequence Embedding, and Risk Assessment Module.

[0108] Based on this, in some embodiments, Figure 4 As shown, when executing step S12, the following steps S12-1 to S12-4 are implemented:

[0109] S12-1. The feature encoder performs feature extraction based on each of the training sample data to obtain a first transaction behavior feature of each user in a time dimension and a second transaction behavior feature of each user in a social relationship dimension; wherein the first transaction behavior feature and the second transaction behavior feature are both represented by feature vectors.

[0110] As described above, the historical credit transaction record data is a personal bill sequence, and the historical social record data is a personal social relationship sequence. Based on this, when executing step S12-2, the feature encoder is specifically used to encode the input data sequence into a feature vector. Since the input sample data includes historical credit transaction record data and historical social record data, the personal bill sequence and the personal social relationship sequence are respectively encoded into vector representations that the model can process.

[0111] As an implementation method, the input historical credit transaction record data and historical social record data can be encoded according to the two dimensions of time dimension and social relationship dimension, and the first transaction behavior characteristics in the time dimension and the second transaction behavior characteristics in the social relationship dimension can be output.

[0112] Among them, in some embodiments, the first transaction behavior feature includes the statistical value of the transaction behavior in the time dimension, specifically including the transaction frequency μ and the transaction amount standard deviation σ. Based on this, when executing the above step S12-1, the feature encoder can obtain the transaction frequency μ through the following formula:

[0113]

[0114] Where n is the number of transactions of the user within time t, x t is the transaction amount within time t.

[0115] On this basis, the feature encoder can also obtain the standard deviation σ of the transaction amount through the following formula:

[0116]

[0117] Among them, μ is the above transaction frequency, x t is the transaction amount within time t. The values ​​of n and t can be determined according to the specific statistical time length, and this application does not make strict restrictions.

[0118] In some embodiments, the second transaction behavior feature includes: a relationship feature Ru between the user u and the associated user v within a preset time period t, wherein the feature encoder extracts the relationship feature Ru between the user and the associated user within the preset time period t from the historical social record data based on the following formula:

[0119]

[0120] Among them, f(v) is the transaction behavior feature of user v, w v is the characteristic coefficient of the transaction behavior characteristics of user v. Wherein, f(v) can be calculated by obtaining the historical transaction records of user v in the credit institution. As an implementation method, f(v) can be a feature extracted by the RFM paradigm. For example, it can be extracted by executing the code logic as follows:

[0121] [mean / max / mean / sum]*Recent[1 / 3 / 6 / 9 / 12m]*[Number of transactions, transaction type, quota utilization rate, etc.]

[0122] In the embodiment of the present application, w v is the characteristic coefficient of the transaction behavior characteristics of user v, wherein the larger the characteristic coefficient is, the greater the influence of user v on user u is.

[0123] By selecting an embodiment of the present application, the user's personal bill sequence and social relationship sequence are feature extracted through a feature encoder, thereby obtaining the user's transaction behavior characteristics in the time dimension and the behavior characteristics in the social relationship dimension, thereby processing complex data into feature vectors that can be recognized and processed by the model, and further helping the model to autonomously learn the correlation between the user's personal bill sequence, social relationship sequence and credit overdue according to the feature vector obtained by the feature encoder, thereby facilitating the subsequent output of a target credit risk prediction model that can accurately predict credit risk.

[0124] S12-2. Input the first transaction behavior feature and the second transaction behavior feature into the sequence model, and use the sequence model to obtain the time dependency and dynamic change of each feature vector.

[0125] Wherein, in some embodiments, the sequence model includes: a long short-term memory network model LSTM or other sequence models. The sequence model uses a gate control unit to deal with the problems of gradient vanishing and gradient exploding encountered in long sequences, and can better capture the time dependency in time series data. In this article, the time dependency is referred to as a time dependency relationship, and the time dependency relationship specifically refers to the chronological relationship of credit record data in the time dimension, specifically, there is a dependency relationship between a data point and another data point located before it, and there is also a dependency relationship between a data point and another data point located after it. The dependency relationship between the data point and the data point in the time dimension can be obtained through the sequence model. Based on this, the embodiment of the present application can capture the time dependency and dynamic changes in the feature vector output by the feature encoder through the sequence model.

[0126] In the embodiment of the present application, the gate control unit introduced by LSTM specifically includes: a forget gate, an input gate and an output gate, wherein each gate control unit satisfies the following formula, that is, the sequence model obtains the time dependency and dynamic change between each feature vector based on the following formula:

[0127] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0128] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0129]

[0130] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0131] h t =o t tanh(C t )

[0132] Among them, f t For the forget gate, i t is the input gate, o t is the output gate, C t is the state of the memory unit, h t is the hidden state, is the candidate memory cell state, σ() is the sigmoid function, x t is the input of the sequence model at time t, W f is the weight parameter of the forget gate, W o is the weight parameter of the output gate, W C is the weight parameter of the state of the memory unit, v i is the weight parameter of the input gate, b f is the forget gate bias, b i is the output gate bias, b C is the memory unit bias, b o Bias for the output gate.

[0133] Among them, the hidden state specifically refers to the hidden state, which is an internal state in the LSTM network model used to store and transfer information between time steps, which is conducive to LSTM to effectively process and predict complex patterns and long-term dependencies in sequence data. In the embodiment of the present application, the forget gate is used to determine the information lost in the memory cell state of the previous time step, and the input gate is used to obtain what new information from the current input and the hidden state of the previous time step, and combine the new information obtained with the output of the forget gate to update the memory cell state of the current time step. The output gate is used to determine what information is output from the current memory cell state as the hidden state of the current time step.

[0134] An embodiment of the present application is selected to introduce an LSTM long short-term memory network into a credit risk prediction model constructed based on a time-series graph network. The LSTM long short-term memory network captures the time dependency and dynamic changes in the feature vectors corresponding to the user's credit bill and the user's social relationship. The impact of changes in the user's personal bill and changes in the user's social relationship on credit overdue can be determined more accurately, thereby facilitating training to obtain a target credit risk prediction model that can accurately output corresponding credit overdue prediction results based on changes in the user's personal bill and changes in social relationships.

[0135] S12-3. The sequence embedding module converts the time dependency and dynamic change of each feature vector into a sequence embedding vector, and the sequence embedding vector is used to represent the comprehensive credit behavior characteristics of the user.

[0136] In the embodiment of the present application, the timing diagram network model is based on the timing diagram for model training. The timing diagram is as follows Figure 2 In the graph structure shown, in addition to nodes and edges, each edge is also associated with a corresponding timestamp. For example, user A and user B conduct a transaction at a certain point in time. At this time, the time point corresponds to a timestamp, which is also recorded in the time series graph as an influencing factor of a social relationship and associated with the corresponding edge. Based on this, the sequence embedding module is used to incorporate time information into the graph for representation, and its core is to capture the dynamic changes of nodes and edges in the graph.

[0137] In this way, when executing step S12-3, the time dependency and dynamic changes captured by the LSTM long short-term memory network model can be converted into corresponding sequence embedding vectors through secondary vector transformation according to the transaction timestamps of each transaction behavior in the historical transaction record data and the time sequence of the social timestamps in the historical social record data. The sequence embedding vectors can then be used to represent the user's comprehensive credit behavior characteristics, so as to integrate the user's personal bill sequence with the personal social relationship sequence.

[0138] S12-4, inputting the comprehensive credit behavior characteristics into the risk assessment module, and having the risk assessment module output the predicted credit assessment result based on the comprehensive credit behavior characteristics.

[0139] In an embodiment of the present application, the risk assessment module learns and mines the implicit rules between the comprehensive credit behavior characteristics and credit overdue in each credit cycle based on the comprehensive credit behavior characteristics, and then calculates the credit overdue probability in different credit cycles based on the learned and mined rules. The credit overdue probability is used to characterize the probability that a user may default in different credit cycles.

[0140] Specifically, as an implementation method, the risk assessment module calculates the predicted credit overdue probability of each user based on the sequence embedding vector using a preset weighted calculation rule. The preset weighted rules can be flexibly modified according to actual business needs, and this application does not strictly limit them. As a preferred implementation method, the preset weighted calculation rules can meet the following conditions:

[0141] The initial mean alpha = 1 / N if N>0else alpha = 0, where N is the number of neighbors.

[0142] By selecting the embodiment of the present application, the risk assessment module is used to calculate the overdue probability based on the comprehensive credit behavior characteristics, so that the model can autonomously learn the rules between the user's personal bill sequence, personal social relationship sequence and credit overdue, and output the corresponding overdue probability, which is helpful to continuously adjust the model parameters of the credit risk prediction model according to the difference between the results output by the risk assessment module and the actual overdue situation, and finally obtain a network model that can accurately predict credit risks.

[0143] In some embodiments, a loss function, such as a logarithmic loss function, a cross-entropy loss function, etc., can be constructed based on the predicted credit assessment results and the actual credit overdue results. By continuously adjusting the model parameters of the initial credit risk prediction model until the loss function converges, it indicates that the credit risk prediction model has been trained and can accurately predict the user's credit overdue situation.

[0144] On this basis, in some embodiments, the credit risk prediction model further includes: a relationship graph updating module. Thus, when executing step S13, it can be implemented through the following steps:

[0145] The relationship graph updating module is used to update the weight coefficients between neighbor nodes in the graph network of the credit risk prediction model according to the target data difference between the predicted credit assessment result output by the credit risk prediction model in each round and the actual credit overdue result.

[0146] Specifically, the relationship graph update module can be a functional module constructed by a graph network such as a graph convolutional network or a graph attention network. As an implementation method, the relationship graph update module uses the predicted credit assessment results output in each round to perform reverse propagation iterations to update the weight coefficients between neighbor nodes in the graph network in the credit risk prediction model.

[0147] In addition, in some embodiments, the credit risk prediction model further includes: a fusion module, which is specifically used to integrate the outputs of the time series graph network and the long short-term memory network with the help of a self-attention mechanism, and specifically integrates the outputs of the time series graph network and the LSTM through the following formula:

[0148]

[0149] Where Q is the sequence embedding that incorporates neighbor information. Specifically, Q is the latent vector of each time step of the temporal network, and a query vector is obtained by encoding the encoding layer. K is the key vector, which is used to calculate the similarity of the target vector with the latent vector. V is the weight vector, which is used to extract the target output vector based on the target vector similarity vector. k is the characteristic dimension of the latent vector and the key vector.

[0150] Exemplarily, the query vector is the vector of the encoded input information, with three dimensions (a, b, c). The key vector is also the vector of the encoded input information, with three dimensions (a, b, c). Then, the query vector and the key vector are matrix-operated to obtain the similarity matrix (a, b, b). Among them, the parameters of the abc dimensions can be learned dynamically, and finally the similarity matrix (a, b, b) is used to perform weighted summation with the value (a, b, c) weight vector to obtain the final target output vector.

[0151] By selecting the embodiment of the present application and using the self-attention mechanism of the above-mentioned fusion module to process long-term credit bills and long-term social relationship changes, it is beneficial for the target credit risk prediction model obtained by subsequent training to cover both short-term credit risk prediction and long-term credit risk prediction.

[0152] In the second aspect, the present application provides a credit risk prediction method, such as Figure 5 As shown, the method comprises the following steps:

[0153] S51, obtaining credit information of a target user to be processed, wherein the credit information of the target user includes: historical transaction records and historical spatiotemporal relationship records of the target user;

[0154] S52: input the target user's credit information into a target credit risk prediction model, and the target credit risk prediction model performs data processing based on the target user's credit information to output the target user's credit overdue probability under different credit cycles.

[0155] The target credit risk prediction model is a credit risk prediction model trained in advance using the credit risk prediction model training method described in the first aspect.

[0156] In an embodiment of the present application, the target user to be processed refers to a user who needs to undergo credit risk prediction, specifically a user who applies for a credit product, or a user who has already applied for a credit product but needs to assess the probability of overdue payment in the subsequent loan repayment process.

[0157] The target user's credit information can be a credit report or daily transaction flow record provided by the user. The daily transaction flow record contains the historical spatiotemporal relationship records between other users who have transaction relationships with the user. The credit information is input into the target credit risk prediction model, which can automatically perform feature extraction and probability calculation based on the input data, and finally output the credit prediction probability under different credit cycles.

[0158] By selecting the embodiment of the present application, since the target credit risk prediction model is pre-trained using historical credit transaction record data and historical social relationship record data, the target credit risk prediction model can accurately combine the user's personal bill information and social relationship information to comprehensively evaluate the user's credit overdue probability, thereby obtaining an accurate credit overdue prediction result, which helps to solve the problem of poor accuracy of existing credit risk prediction methods.

[0159] In a third aspect, the present application provides a credit risk prediction model training device, wherein, Figure 6 As shown, the device 60 includes:

[0160] The acquisition module 601 is used to acquire a training sample data set and input the training sample data set into an initial credit risk prediction model, wherein the training sample data set includes various training sample data, and each training sample data includes: a user's historical credit transaction record data and a user's historical social relationship record data; the credit risk prediction model is a neural network model constructed based on a time series graph network;

[0161] The model training module 602 is used to process each piece of training sample data using the initial credit risk prediction model, and output a predicted credit assessment result corresponding to each piece of training sample data, wherein the predicted credit assessment result is a predicted credit overdue probability of each user under different credit cycles;

[0162] A model parameter adjustment module 603 is used to adjust the model parameters of the initial credit risk prediction model according to the target data difference between the predicted credit assessment result and the actual credit overdue result of the user corresponding to each training sample data, until the target data difference is less than a preset data difference threshold;

[0163] The model determination module 604 is used to determine the credit risk prediction model corresponding to the target data difference being less than a preset data difference threshold as the trained target credit risk prediction model.

[0164] In a fourth aspect, the present application provides a credit risk prediction device, which can be Figure 7 As shown, the device 70 includes the following parts:

[0165] The credit information acquisition module 701 is used to acquire the credit information of the target user to be processed, wherein the credit information of the target user includes: the historical transaction records and historical spatiotemporal relationship records of the target user;

[0166] The credit information processing module 702 is used to input the credit information of the target user into the target credit risk prediction model; the target credit risk prediction model performs data processing based on the credit information of the target user and outputs the credit overdue probability of the target user under different credit cycles;

[0167] The target credit risk prediction model is a credit risk prediction model obtained by training in advance using the credit risk prediction model training scheme described in the first aspect.

[0168] Among them, the collection, storage, use, processing, transmission, provision and disclosure of user personal information involved in this application are in compliance with the relevant laws and regulations and do not violate public order and good morals.

[0169] The names of the messages or information exchanged between multiple devices in the embodiments of the present application are only used for illustrative purposes and are not used to limit the scope of these messages or information.

[0170] In a third aspect, the exemplary embodiments of the present application further provide an electronic device, comprising: at least one processor; and a memory connected to the at least one processor in communication. The memory stores a computer program that can be executed by the at least one processor, and the computer program is used to enable the electronic device to perform a method according to an embodiment of the present application when executed by the at least one processor.

[0171] The exemplary embodiment of the present application further provides a non-transitory computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor of a computer, is used to cause the computer to perform a method according to an embodiment of the present application.

[0172] The exemplary embodiments of the present application further provide a computer program product, including a computer program, wherein when the computer program is executed by a processor of a computer, it is used to enable the computer to execute the method according to the embodiment of the present application.

[0173] refer to Figure 8 , the structural block diagram of the electronic device 800 that can be used as the server or client of the present application will now be described, which is an example of the hardware device that can be applied to various aspects of the present application. The electronic device is intended to represent various forms of digital electronic computer equipment, such as laptop computers, desktop computers, workbenches, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices and other similar computing devices. The components shown herein, their connections and relationships, and their functions are only examples, and are not intended to limit the implementation of the present application described and / or required herein.

[0174] like Figure 8 As shown, the electronic device 800 includes a computing unit 801, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 802 or a computer program loaded from a storage unit 808 into a random access memory (RAM) 803. In the RAM 803, various programs and data required for the operation of the electronic device 800 can also be stored. The computing unit 801, the ROM 802, and the RAM 803 are connected to each other via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0175] A plurality of components in the electronic device 800 are connected to the I / O interface 805, including: an input unit 806, an output unit 807, a storage unit 808, and a communication unit 809. The input unit 806 may be any type of device capable of inputting information to the electronic device 800, and the input unit 806 may receive input digital or character information, and generate key signal inputs related to user settings and / or function control of the electronic device. The output unit 807 may be any type of device capable of presenting information, and may include, but is not limited to, a display, a speaker, a video / audio output terminal, a vibrator, and / or a printer. The storage unit 808 may include, but is not limited to, a disk, an optical disk. The communication unit 809 allows the electronic device 800 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks, and may include, but is not limited to, a modem, a network card, an infrared communication device, a wireless communication transceiver, and / or a chipset, such as a Bluetooth™ device, a WiFi device, a WiMax device, a cellular communication device, and / or the like.

[0176] The computing unit 801 may be various general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 801 performs the various methods and processes described above. For example, in some embodiments, the aforementioned credit risk prediction model training method and credit risk prediction method may be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit 808. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 800 via the ROM 802 and / or the communication unit 809. In some embodiments, the computing unit 801 may be configured to perform the aforementioned credit risk prediction model training method and credit risk prediction method in any other appropriate manner (e.g., by means of firmware).

[0177] The program code for implementing the method of the present application can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, implements the functions / operations specified in the flow chart and / or block diagram. The program code can be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0178] In the context of the present application, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0179] As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus, and / or device (e.g., disk, optical disk, memory, programmable logic device (PLD)) for providing machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0180] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0181] The systems and techniques described herein may be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), and the Internet.

[0182] A computer system may include clients and servers. Clients and servers are generally remote from each other and usually interact through a communication network. The relationship of client and server is generated by computer programs running on respective computers and having a client-server relationship to each other.

Claims

1. A credit risk prediction model training method, characterized in that: The method comprises: Acquire a training sample data set, and input the training sample data set into an initial credit risk prediction model, wherein the training sample data set includes various training sample data, and each of the training sample data includes: historical credit transaction record data of a user, and historical social relationship record data of a user; the credit risk prediction model is a neural network model constructed based on a time series graph network; The initial credit risk prediction model processes each piece of training sample data and outputs a predicted credit assessment result corresponding to each piece of training sample data, wherein the predicted credit assessment result is a predicted credit overdue probability of each of the users under different credit cycles; According to the target data difference between the predicted credit assessment result and the actual credit overdue result of the user corresponding to each training sample data, the model parameters of the initial credit risk prediction model are adjusted until the target data difference is less than a preset data difference threshold; The credit risk prediction model corresponding to the target data difference being less than the preset data difference threshold is determined as the trained target credit risk prediction model.

2. The method according to claim 1, characterized in that: The credit risk prediction model includes: a feature encoder, a sequence model, a sequence embedding module, and a risk assessment module, wherein the initial credit risk prediction model processes each piece of training sample data and outputs a predicted credit assessment result corresponding to each piece of training sample data, including: The feature encoder performs feature extraction based on each of the training sample data to obtain a first transaction behavior feature of each user in a time dimension and a second transaction behavior feature of each user in a social relationship dimension; wherein the first transaction behavior feature and the second transaction behavior feature are both represented by feature vectors; Inputting the first transaction behavior feature and the second transaction behavior feature into the sequence model, and obtaining the time dependency and dynamic change of each feature vector from the sequence model; The sequence embedding module converts the time dependency and dynamic change of each feature vector into a sequence embedding vector, and the sequence embedding vector is used to represent the comprehensive credit behavior characteristics of the user; The comprehensive credit behavior characteristics are input into the risk assessment module, and the risk assessment module outputs the predicted credit assessment result based on the comprehensive credit behavior characteristics.

3. The method according to claim 2, characterized in that The first transaction behavior feature includes: transaction frequency μ, transaction amount standard deviation σ, and the first transaction behavior feature of each user in the time dimension is obtained, including: The transaction frequency μ is obtained by the feature encoder based on the following formula: Where n is the number of transactions of the user within time t, x t is the transaction amount within time t; The transaction amount standard deviation σ is extracted by the feature encoder based on the following formula:

4. The method according to claim 2, characterized in that: The second transaction behavior feature includes: a relationship feature Ru between user u and associated user v within a preset time period. The second transaction behavior feature of each user in the social relationship dimension is obtained, including: The feature encoder extracts the relationship feature Ru between the user and the associated users within a preset time period based on the following formula: Among them, f(v) is the transaction behavior feature of user v, w v is the characteristic coefficient of the transaction behavior characteristics of user v.

5. The method according to claim 2, characterized in that: The sequence model includes: a long short-term memory network model; the first transaction behavior feature and the second transaction behavior feature are input into the sequence model, and the time dependency and dynamic change of each feature vector are obtained by the sequence model; including: The sequence model obtains the time dependency and dynamic changes in each feature vector based on the following formula: f t =σ(W f ·[h t-1 ,x t ]+b f ) i t =σ(W i ·[h t-1 ,x t ]+b i ) the t =σ(W o ·[h t-1 ,x t ]+b o ) h t =o t ·tanh(C t ) Among them, f t For the forget gate, i t is the input gate, o t is the output gate, C t is the state of the memory unit, h t is the hidden state, is the candidate memory cell state, σ() is the sigmoid function, x t is the input of the sequence model at time t, W f is the weight parameter of the forget gate, W o is the weight parameter of the output gate, W C is the weight parameter of the state of the memory unit, W i is the weight parameter of the input gate, b f is the forget gate bias, b i is the output gate bias, b C is the memory unit bias, b o Bias for the output gate.

6. The method according to claim 2, characterized in that The credit risk prediction model further includes: a relationship graph updating module, which adjusts the model parameters of the initial credit risk prediction model according to the target data difference between the predicted credit assessment result and the actual credit overdue result of the user corresponding to each training sample data, including: The relationship graph updating module is used to update the weight coefficients between neighbor nodes in the graph network of the credit risk prediction model according to the target data difference between the predicted credit assessment result output by the credit risk prediction model in each round and the actual credit overdue result.

7. The method according to claim 2, characterized in that The outputting of the predicted credit assessment result by the risk assessment module based on the comprehensive credit behavior characteristics includes: The risk assessment module calculates the predicted credit overdue probability of each user based on the sequence embedding vector and using a preset weighted calculation rule.

8. The method according to claim 2, characterized in that: The credit risk prediction model further includes: a fusion module, wherein the fusion module integrates the outputs of the time series graph network and the LSTM through a self-attention mechanism, and specifically integrates the outputs of the time series graph network and the LSTM through the following formula: Where Q is the latent vector of each time step of the temporal network, K is the key vector, which is used to calculate the similarity of the target vector with the latent vector, V is the weight vector, which is used to extract the target output vector based on the target vector similarity vector, and d k is the characteristic dimension of the latent vector and the key vector.

9. A credit risk prediction method, characterized in that: The method comprises: Acquire the credit information of the target user to be processed, wherein the credit information of the target user includes: historical transaction records and historical spatiotemporal relationship records of the target user; Inputting the target user's credit information into a target credit risk prediction model, and having the target credit risk prediction model perform data processing based on the target user's credit information to output the target user's credit overdue probability under different credit cycles; Wherein, the target credit risk prediction model is a credit risk prediction model trained in advance using the credit risk prediction model training scheme according to any one of claims 1 to 8.

10. An electronic device, characterized in that: The electronic device comprises: Processor; and Memory for storing programs, The program includes instructions, which, when executed by the processor, cause the processor to perform the method according to any one of claims 1 to 9.

11. A non-transitory computer-readable storage medium storing computer instructions, characterized in that: The computer instructions are used to make a computer execute the method according to any one of claims 1-9.

Images