Risk prediction model training and risk prediction method

Abstract

The invention provides a risk prediction model training and risk prediction method. The risk prediction model training method comprises the steps of obtaining a sample business undirected graph associated with a to-be-trained risk prediction model; determining an initial service feature of each sample node according to sample service data of a plurality of specified dimensions of a sample service entity indicated by each sample node in the sample service undirected graph; inputting the initial business characteristics of each sample node into a risk prediction model to obtain a risk probability prediction value of each sample node output by the risk prediction model; and training the risk prediction model based on the risk probability prediction value and the risk probability annotation value corresponding to each sample node, so that the risk prediction model input is more comprehensive and accurate, the model is helped to learn a more accurate risk assessment rule, the prediction reliability and stability are improved, and the risk prediction efficiency is improved. Therefore, the risk can be identified and managed more effectively in an actual service scene.

Description

Technical Field

[0001] The present disclosure relates to the field of risk control technology, and in particular, to a method for training a risk prediction model and a risk prediction method. Background Art

[0002] With the development of the financial market, the scope of financial transaction services has been continuously expanding, including various forms such as loans, foreign exchange transactions, cross-border financing, and derivative transactions. Each service is accompanied by potential threats in multiple aspects such as credit risk, market risk, liquidity risk, and operational risk. At the same time, global economic fluctuations and changes in policies and regulations have further exacerbated the instability of the financial environment. To ensure the security and stability of financial transactions, when an enterprise applies for a financial transaction service, risk prediction can be carried out on it to identify and quantify in advance the risks that the enterprise may face, so as to provide a basis for formulating effective countermeasures. Therefore, how to perform risk prediction on an enterprise when it applies for a financial transaction service is very important. Summary of the Invention

[0003] The present disclosure provides a method for training a risk prediction model and a risk prediction method to at least solve one of the technical problems in the related art to a certain extent. The technical solution of the present disclosure is as follows:

[0004] According to the first aspect of the embodiments of the present disclosure, a method for training a risk prediction model is provided, including: obtaining an undirected sample service graph associated with the risk prediction model to be trained; wherein, the sample nodes of the undirected sample service graph are used to indicate sample service entities, and the edges between the sample nodes are used to indicate the association relationships between the sample service entities; determining initial service features of each sample node according to sample service data of multiple specified dimensions of the sample service entities indicated by each sample node in the undirected sample service graph; inputting the initial service features of each sample node into the risk prediction model to obtain risk probability prediction values of each sample node output by the risk prediction model; and training the risk prediction model based on the risk probability prediction values and risk probability annotation values corresponding to each sample node.

[0005] According to a second aspect of the embodiments of the present disclosure, a risk prediction method is provided, including: obtaining a target business undirected graph associated with a target business entity; wherein, the target business undirected graph includes a first node for indicating the target business entity and a second node of a reference business entity having an association relationship with the target business entity; determining an initial business feature of the first node and an initial business feature of the second node according to target business data of multiple specified dimensions of the first node and reference business data of multiple specified dimensions of the second node; inputting the initial business feature of the first node and the initial business feature of the second node into a trained risk prediction model to obtain a risk probability prediction value of the first node output by the risk prediction model; wherein, the risk probability prediction value of the first node is used to indicate the risk degree of the target business entity handling a specified business; determining a risk level of the target business entity handling the specified business according to the risk probability prediction value of the first node.

[0006] According to a third aspect of the embodiments of the present disclosure, a training device for a risk prediction model is provided, including: an obtaining module, configured to obtain a sample business undirected graph associated with a risk prediction model to be trained; wherein, sample nodes of the sample business undirected graph are used to indicate sample business entities, and edges between the sample nodes are used to indicate association relationships between the sample business entities; a determining module, configured to determine an initial business feature of each sample node according to sample business data of multiple specified dimensions of the sample business entities indicated by the sample nodes in the sample business undirected graph; an input module, configured to input the initial business feature of each sample node into the risk prediction model to obtain a risk probability prediction value of each sample node output by the risk prediction model; a training module, configured to train the risk prediction model based on the risk probability prediction values corresponding to the sample nodes and risk probability annotation values.

[0007] According to a fourth aspect of the embodiments of the present disclosure, a risk prediction device is provided, including: an acquisition module configured to acquire a target business undirected graph associated with a target business entity; wherein, the target business undirected graph includes a first node for indicating the target business entity and a second node of a reference business entity having an association relationship with the target business entity; a first determination module configured to determine an initial business feature of the first node and an initial business feature of the second node according to target business data of a plurality of specified dimensions of the first node and reference business data of a plurality of specified dimensions of the second node; an input module configured to input the initial business feature of the first node and the initial business feature of the second node into a trained risk prediction model to obtain a risk probability prediction value of the first node output by the risk prediction model; wherein, the risk probability prediction value of the first node is used to indicate the risk degree of the target business entity handling a specified business; a second determination module configured to determine a risk level of the target business entity handling the specified business according to the risk probability prediction value of the first node.

[0008] According to a fifth aspect of the embodiments of the present disclosure, an electronic device is provided, including: a processor; a memory for storing executable instructions of the processor; wherein, the processor is configured to execute the instructions to implement the training method of the risk prediction model as described in the first aspect embodiment of the present disclosure, or to implement the risk prediction method as described in the second aspect embodiment of the present disclosure.

[0009] According to a sixth aspect of the embodiments of the present disclosure, a computer-readable storage medium is provided, which when the instructions in the computer-readable storage medium are executed by a processor of an electronic device, enables the electronic device to execute the training method of the risk prediction model as described in the first aspect embodiment of the present disclosure, or to execute the risk prediction method as described in the second aspect embodiment of the present disclosure.

[0010] According to a seventh aspect of the embodiments of the present disclosure, a computer program product is provided, including: a computer program, which when executed by a processor, implements the training method of the risk prediction model as described in the first aspect embodiment of the present disclosure, or implements the risk prediction method as described in the second aspect embodiment of the present disclosure.

[0011] The technical solutions provided by the embodiments of the present disclosure at least bring the following beneficial effects:

[0012] In this technical solution, initial business features are constructed based on the sample business undirected graph and its multi-dimensional sample business data, making the input of the risk prediction model more comprehensive and accurate, fully capturing the complex characteristics of business entities and their associated relationships. Furthermore, the initial business features are input into the risk prediction model to generate risk probability prediction values, and supervised training is carried out in combination with risk probability annotation values, which helps the model learn more accurate risk assessment rules, improving the reliability and stability of prediction, and thus more effectively identifying and managing risks in actual business scenarios. Among them, the risk probability prediction value of each sample node is obtained by the risk prediction model using at least one graph convolutional layer to update the initial business features of each sample node to obtain the target business features of each sample node, and then using the prediction layer of the risk prediction model to perform risk prediction on the target business features of each sample node, achieving that each node not only based on its own initial business features, but also can integrate the relevant information of its neighbor nodes, making the generated target business features more comprehensive and accurate, and performing risk prediction based on the updated target business features, improving the accuracy and comprehensiveness of risk prediction. In addition, by extracting features for any specified dimension of the sample business entity indicated by any sample node in the business undirected graph, representative sub-business features are refined from complex business data, ensuring that the data input into the model has a high degree of relevance and accuracy. Then, a linear transformation is performed on it using the weight matrix and attention mechanism of the sub-business features of the specified dimension, and a non-linear transformation is performed through an activation function to generate attention weights, so as to perform weighted summation on the sub-business features of each dimension according to the attention weights of each specified dimension to obtain the initial business features, avoiding biases or omissions that may be brought by a single dimension, providing a more comprehensive and accurate understanding of the enterprise or business entity, providing a comprehensive and refined data basis for the training of the risk prediction model, and enhancing the expressiveness and prediction accuracy of the model.

[0013] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and do not limit the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The accompanying drawings here are incorporated into the specification and form a part of this specification, showing embodiments consistent with the present disclosure, and are used together with the specification to explain the principles of the present disclosure, and do not constitute an improper limitation of the present disclosure.

[0015] Figure 1 is a schematic flowchart of the training method of the risk prediction model shown in the first embodiment of the present disclosure;

[0016] Figure 2 is a schematic flowchart of the process of generating risk probability prediction values shown in the second embodiment of the present disclosure;

[0017] Figure 3 It is a schematic flowchart of a method for training a risk prediction model shown in the third embodiment of the present disclosure;

[0018] Figure 4 It is a schematic diagram of the principle of a method for training a risk prediction model shown in the embodiments of the present disclosure;

[0019] Figure 5 It is a schematic flowchart of the processing of a multi-layer perceptron integrating an attention mechanism shown in the embodiments of the present disclosure;

[0020] Figure 6 It is a schematic flowchart of the processing of the GCN algorithm shown in the embodiments of the present disclosure;

[0021] Figure 7 It is a schematic flowchart of a risk prediction method shown in the fourth embodiment of the present disclosure;

[0022] Figure 8 It is a schematic structural diagram of a device for training a risk prediction model shown in the fifth embodiment of the present disclosure;

[0023] Figure 9 It is a schematic structural diagram of a risk prediction device shown in the sixth embodiment of the present disclosure;

[0024] Figure 10 It is a schematic structural diagram of an electronic device shown in an exemplary embodiment of the present disclosure. Detailed implementation manners

[0025] In order to enable those of ordinary skill in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0026] It should be noted that in the specification and claims of the present disclosure and the above-mentioned accompanying drawings, the terms "first", "second", etc. are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that such data used can be interchanged under appropriate circumstances so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The implementation manners described in the following exemplary embodiments do not represent all implementation manners consistent with the present disclosure. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0027] It should be noted that in the technical solutions of the present disclosure, the collection, storage, use, processing, transmission, provision, and disclosure of the user's personal information, etc. are all carried out on the premise of obtaining the user's consent, and all comply with the provisions of relevant laws and regulations and do not violate public order and good customs.

[0028] In the related art, by integrating real-time business data, clustering algorithms, and an optimized neural network model, efficient enterprise financial risk identification and early warning have been achieved. However, there are still some deficiencies in this method. For example, it may ignore the complexity of dynamic market conditions and the external economic environment. At the same time, non-financial interactions between enterprises may not be fully considered, limiting the comprehensiveness of risk assessment.

[0029] In view of at least one of the above problems, the present disclosure proposes a method for training a risk prediction model and risk prediction.

[0030] The following describes the method for training a risk prediction model and risk prediction according to an embodiment of the present disclosure with reference to the accompanying drawings.

[0031] Figure 1 FIG. is a schematic flowchart of a method for training a risk prediction model according to a first embodiment of the present disclosure. It should be noted that the execution subject of the embodiment of the present disclosure may be a training device for a risk prediction model, and the training device for the risk prediction model can be applied to any electronic device with computing capabilities, so that the electronic device can execute the training function of the risk prediction model.

[0032] As Figure 1 shown, the method for training the risk prediction model includes the following steps:

[0033] Step 101, obtain a sample business undirected graph associated with the risk prediction model to be trained.

[0034] Among them, the sample nodes of the sample business undirected graph are used to indicate sample business entities, and the edges between the sample nodes are used to indicate the association relationships between the sample business entities.

[0035] In order to provide structured data support for the training of the risk prediction model, so that the risk prediction model can better capture complex patterns and potential risks in the business network, in the embodiment of the present disclosure, a sample business undirected graph is constructed based on multiple sample business entities and the association relationships between multiple sample business entities. Among them, the sample nodes in the sample business undirected graph represent sample business entities, the edges between the sample nodes are used to indicate the association relationships between the sample business entities, and the sample business entities can be business institutions (enterprises), customers, accounts, etc.

[0036] Step 102, determine the initial business characteristics of each sample node according to the sample business data of multiple specified dimensions of the sample business entities indicated by each sample node in the sample business undirected graph.

[0037] In order to improve the accuracy and comprehensiveness of the risk prediction model, in the embodiments of the present disclosure, multiple specified dimensions of sample business data of the sample business entity indicated by each sample node can be obtained, and through data preprocessing and feature extraction, and the features extracted from the sample business data of each sample node are fused to generate the initial business features of each sample node.

[0038] It should be noted that the multiple specified dimensions of sample business data of the sample business entity refer to some representative business instances (sample business entities) selected for a specific business scenario, and the actual data (sample business data) obtained according to several predefined angles or attributes (specified dimensions).

[0039] For example, taking the sample business entity as an enterprise, in order to fully consider the complexity of the dynamic market conditions and the external economic environment, the multiple specified dimensions may include but are not limited to: financial dimension, industry dimension, and shareholder dimension, and the sample business data may include but are not limited to: financial data, shareholder information, industry data. Among them, the financial data includes but is not limited to key indicators such as the current ratio, debt ratio, and return on net assets of the enterprise, and the financial data is used to reflect the financial health status of the enterprise; the shareholder information includes but is not limited to the shareholding ratio of major shareholders, the transparency of the ultimate beneficiary, and the frequency of shareholder changes, etc., and the shareholder information is used to analyze the stability of the company's shareholder structure; the industry data includes but is not limited to industry average financial indicators, market share, and growth rate, etc., and the industry data is used to provide the background and comparison benchmark of the industry.

[0040] Step 103: Input the initial business features of each sample node into the risk prediction model to obtain the risk probability prediction values of each sample node output by the risk prediction model.

[0041] In the embodiments of the present disclosure, the initial business features (feature vectors) of each sample node are used as inputs and passed to the risk prediction model. The risk prediction model calculates based on the input feature vectors and generates a risk probability prediction value for each sample node. It should be noted that the risk probability prediction value is a numerical value between 0 and 1, indicating the degree of risk of the sample node handling the specified business.

[0042] Step 104: Train the risk prediction model based on the risk probability prediction values and risk probability annotation values corresponding to each sample node.

[0043] In order to improve the prediction accuracy of the risk prediction model, in the embodiments of the present disclosure, a loss function value is generated based on the risk probability prediction values and risk probability annotation values corresponding to each sample node, and the risk prediction model is trained based on the loss function value.

[0044] In summary, by constructing initial business features based on the sample business undirected graph and its multi-dimensional sample business data, the risk prediction model input is more comprehensive and accurate, fully capturing the complex characteristics of business entities and their relationships. Then, the initial business features are input into the risk prediction model to generate risk probability prediction values, and supervised training is performed in combination with the risk probability annotation values, which helps the model learn more accurate risk assessment rules and improve the reliability and stability of the prediction, thereby more effectively identifying and managing risks in actual business scenarios.

[0045] Based on the above embodiments, Figure 2 As shown, the risk probability prediction value of each sample node is generated by the risk prediction model using the following steps 201 to 202, which are as follows:

[0046] Step 201, using at least one graph convolution layer in the risk prediction model to update the initial business features of each sample node to obtain the target business features of each sample node.

[0047] It should be understood that relying solely on the characteristics of the sample nodes themselves may not be sufficient to fully capture the risk information of the nodes, because the associations between the sample nodes (such as cooperative relationships and resource exchanges between enterprises) may also have an important impact on the risk. Therefore, in the embodiment of the present disclosure, the graph convolution layer in the risk prediction model is used to update the initial business characteristics of the sample nodes, so that it can integrate the characteristic information of the neighboring nodes, thereby generating more comprehensive target business characteristics.

[0048] As an example, the first graph convolution layer in at least one graph convolution layer is used to aggregate the initial business features of any sample node and the initial business features of the neighboring nodes of any sample node to obtain the updated business features of any sample node output by the first graph convolution layer; the non-first graph convolution layer in at least one graph convolution layer is used to aggregate the updated business features of any sample node output by the previous graph convolution layer and the updated business features of the neighboring nodes to obtain the updated business features of any sample node output by the non-first graph convolution layer; the updated business features of any sample node output by the last graph convolution layer in at least one graph convolution layer are used as the target business features of any sample node.

[0049] That is to say, the graph convolution layer updates the features of the sample nodes by aggregating the features of the sample nodes themselves and the features of their neighboring nodes. During the feature updating process, the output of each layer serves as the input of the next layer, and the features of the sample nodes are updated layer by layer, and finally the target business features of the sample nodes are obtained.

[0050] For the first-layer graph convolution layer, the first-layer graph convolution layer aggregates the initial business features of the current node with the initial business features of its neighboring nodes, which can be specifically expressed as the following formula:

[0051]

[0052] Among them, is the updated service feature of the i-th sample node at the first layer, N(i) is the set of neighbor nodes of sample node i, and W is the weight matrix for linear transformation; d i is the degree of node i, d j is the degree of neighbor node j, and σ is the activation function.

[0053] For non-first-layer graph convolutional layers, the non-first-layer graph convolutional layers aggregate the updated service features of any sample node output by the previous layer of graph convolutional layer and the updated service features of neighbor nodes to obtain the updated service features of any sample node output by the non-first-layer graph convolutional layer, which can be specifically expressed as the following formula:

[0054]

[0055] Among them, is the updated service feature of the i-th sample node at the l+1 layer, N(i) is the set of neighbor nodes of sample node i, and W is the weight matrix for linear transformation; d i is the degree of node i, d j is the degree of neighbor node j, and σ is the activation function.

[0056] Furthermore, the last layer of graph convolutional layer completes the last aggregation operation to generate the feature representation of the final sample node, that is, the updated service feature of any sample node output by the last layer of graph convolutional layer in at least one graph convolutional layer is used as the target service feature of any sample node.

[0057] Step 202: Use the prediction layer in the risk prediction model to perform risk prediction on the target service features of each sample node to obtain the risk probability prediction values of each sample node.

[0058] To improve the accuracy and comprehensiveness of the risk prediction model, in the embodiments of the present disclosure, after passing through multiple graph convolutional layers, each sample node has an updated and richer feature representation, that is, the target service feature. The target service feature contains information about the sample node itself and its neighbor nodes, and has undergone multiple layers of abstraction and aggregation, and can better represent information related to risks. Use the prediction layer in the risk prediction model to perform risk prediction on the target service features of each sample node, and output the risk probability prediction values of each sample node. Among them, the prediction layer can use a fully connected layer and a Sigmoid activation function to calculate the output probability of each sample node.

[0059] In summary, at least one graph convolutional layer in the risk prediction model is used to update the initial business features of each sample node to obtain the target business features of each sample node; the prediction layer in the risk prediction model is used to perform risk prediction on the target business features of each sample node to obtain the risk probability prediction value of each sample node. Thus, through the graph convolutional layer, each node can not only be based on its own initial business features, but also fuse the relevant information of its neighbor nodes, making the generated target business features more comprehensive and accurate. Risk prediction based on the updated target business features improves the accuracy and comprehensiveness of risk prediction.

[0060] To clearly illustrate how to determine the initial business features of each sample node according to the sample business data of multiple specified dimensions of the sample business entities indicated by each sample node in the sample business undirected graph in the above embodiments, the present disclosure proposes another training method for the risk prediction model.

[0061] Figure 3 It is a schematic flowchart of the training method of the risk prediction model shown in the third embodiment of the present disclosure.

[0062] As Figure 3 shown, the training method of the risk prediction model includes the following steps:

[0063] Step 301, obtain a sample business undirected graph associated with the risk prediction model to be trained.

[0064] Among them, the sample nodes of the sample business undirected graph are used to indicate sample business entities, and the edges between the sample nodes are used to indicate the association relationships between the sample business entities.

[0065] To accurately reflect the association relationships between each sample business entity, as a possible implementation, construct a sample business undirected graph for showing the association relationships between each sample business entity.

[0066] In the embodiments of the present disclosure, the sample business entities include: business institutions, and among multiple business institutions, determine the business institutions that share the same resources; where the resources include: suppliers and / or service objects; according to the business institutions that share the same resources, establish nodes with association relationships in the sample business undirected graph.

[0067] That is to say, add each business institution as a node to the undirected graph, and each node represents a specific business institution. For each pair of business institutions that share the same resources, establish an edge between their corresponding nodes, and this edge indicates that there is a certain form of association relationship (such as sharing suppliers or service objects) between these two business institutions. Among them, it should be noted that a business institution refers to an enterprise or unit participating in market activities, and sharing resources means that two or more business institutions jointly use resources.

[0068] Step 302: For any specified dimension of the sample business entity indicated by any sample node in the business undirected graph, extract features from the sample business data of any specified dimension to obtain sub-business features of any specified dimension.

[0069] In the embodiments of the present disclosure, by extracting features from the sample business data of each specified dimension of the sample business entities indicated by each sample node, sub-business features of each specified dimension can be obtained.

[0070] For example, taking the sample business entity as an enterprise, when the specified dimension is the "financial dimension", the sub-business features of the "financial dimension" include, but are not limited to: current ratio, quick ratio, debt ratio, interest coverage ratio, cash flow ratio, return on net assets, asset turnover rate, and operating leverage coefficient; when the specified dimension is the "shareholder dimension", the sub-business features of the "shareholder dimension" include, but are not limited to: the shareholding ratio of the major shareholder, the transparency of the ultimate beneficiary, the frequency of shareholder changes, etc.; when the specified dimension is the "industry dimension", the sub-business features of the "industry dimension" include, but are not limited to: industry average financial indicators, market share, industry growth rate, etc.

[0071] Step 303: Use the weight matrix and attention mechanism of the sub-business features of any specified dimension to perform a linear transformation on the sub-business features of any specified dimension to obtain the linearly transformed sub-business features of any specified dimension.

[0072] In the embodiments of the present disclosure, data of different dimensions may have different importance. To better capture these differences and improve the performance of the risk prediction model, a weight matrix and an attention mechanism can be used to perform a linear transformation on the sub-business features of the specified dimension. Among them, it should be noted that the weight matrix is a matrix used to perform a linear transformation on the input features. By multiplying the weight matrix, the spatial representation form of the original features can be changed, so that some features become more prominent or hidden; the attention mechanism is used to dynamically focus on more important parts when the model processes data.

[0073] Step 304: Use an activation function to perform a non-linear transformation on the linearly transformed sub-business features of any specified dimension to generate the attention weights of any specified dimension.

[0074] In the embodiments of the present disclosure, the LeakyReLU activation function is used to perform a non-linear transformation on the shareholder information features, industry data features, and financial features to generate attention scores.

[0075] For example, the formula for calculating the feature attention score of the shareholder information of the i-th node:

[0076] e ig= LeakyReLU(A i ·W g );

[0077] Among them, e ig represents the attention score of the shareholder information feature, A i represents the shareholder information feature of the i-th node, and W g represents the weight matrix corresponding to A i ;

[0078] The calculation formula for the attention score of the industry data feature of the i-th node:

[0079] e il = LeakyReLU(B i ·W l );

[0080] Among them, e il represents the attention score of the industry data feature, B i represents the industry data feature of the i-th node, and W l represents the weight matrix corresponding to B i ;

[0081] The calculation formula for the attention score of the financial data feature of the i-th node:

[0082] e ix = LeakyReLU(X i ·W x );

[0083] Among them, e ix represents the attention score of the financial data feature, X i represents the financial data feature of the i-th node, and W x represents the weight matrix corresponding to X i ;

[0084] Furthermore, use the softmax function to normalize the attention scores to obtain the attention weights;

[0085] For the shareholder information feature:

[0086] For the industry data feature:

[0087] For the financial data feature:

[0088] Step 305, according to the attention weights of each specified dimension of any sample node, perform weighted summation on the sub-business features of each specified dimension of any sample node to obtain the initial business feature of any sample node.

[0089] As an example, in order to effectively integrate information of multiple specified dimensions into a unified feature, weighted summation is performed on the sub-business features of each specified dimension of any sample node based on the attention weights of each specified dimension to obtain the initial business feature of any sample node. Specifically, it can be expressed by the following formula:

[0090]

[0091] Among them, represents the initial business feature of the i-th node; for example, A i represents the shareholder information feature of the i-th node, such as the shareholding ratio of major shareholders, UBO transparency, and the frequency of shareholder changes, etc., B i represents the industry data feature of the i-th node, such as industry average financial indicators, market share, and industry growth rate, etc., X i represents the financial data feature of the i-th node, current ratio, debt ratio, and ROE, etc., W g 、W l W x represent the weight matrices of each specified dimension, a g 、a l 、a x represent the attention weights of the corresponding specified dimensions.

[0092] Step 306: Input the initial business features of each sample node into the risk prediction model to obtain the risk probability prediction values of each sample node output by the risk prediction model.

[0093] Step 307: Train the risk prediction model based on the risk probability prediction values and risk probability annotation values corresponding to each sample node.

[0094] In order to improve the prediction accuracy of the risk prediction model, in the embodiments of the present disclosure, based on the risk probability prediction values and risk probability annotation values corresponding to each sample node, a target loss function value is generated, and based on the target loss function value, the model parameters of the risk prediction model are adjusted until the set stop training condition is satisfied.

[0095] As an example, based on the risk probability prediction value and risk probability annotation value corresponding to any sample node, the sub-loss function value of any sample node is determined; according to the sub-loss function values of each sample node, the target loss function value is determined; according to the target loss function value, the risk prediction model is trained in multiple rounds of iteration, and it is judged whether the current round of iterative training satisfies the set stop training condition; if so, the training process of the risk prediction model is ended, if not, the risk prediction model is continued to be trained until the set stop training condition is satisfied.

[0096] That is to say, for each sample node, its sub-loss function value is calculated based on its risk probability prediction value and risk probability annotation value. Based on the sub-loss function values of all sample nodes, the target loss function value is determined. Furthermore, based on the target loss function value, the model parameters of the risk prediction model are adjusted until the set stop training condition is satisfied.

[0097] For example, for each sample node, based on the difference between the risk probability prediction value and the risk probability annotation value of each sample node, the sub-loss function value of this sample node is generated. The sub-loss function values of all sample nodes are added together and divided by the total number of all sample nodes to obtain the target loss function value. Furthermore, based on the target loss function value, multiple rounds of iteration are performed on the risk prediction model. In each round of iteration, the risk prediction model adjusts its parameters (such as the weight matrix) according to the current target loss function value to reduce the target loss function value in the next round of iteration. When the iterative training in the current round satisfies the set stop training condition, the training process of the risk prediction model ends; if the iterative training in the current round does not satisfy the set stop training condition, the risk prediction model continues to be trained until the set stop training condition is satisfied. It should be noted that the set stop training conditions include but are not limited to: reaching a preset maximum number of iterations (such as 1000), the change in the target loss function value being less than a certain threshold, etc.

[0098] In summary, by extracting features for any specified dimension of the sample business entity indicated by any sample node in the business undirected graph, representative sub-business features are refined from complex business data, ensuring that the data input into the model has a high degree of relevance and accuracy; then, the weight matrix and attention mechanism of the sub-business features of this specified dimension are used to perform a linear transformation on it, and a non-linear transformation is performed through an activation function to generate attention weights, enabling the model to adaptively identify and emphasize the feature dimensions that are most important for the prediction task, thereby improving the flexibility and adaptability of the model in the face of different business scenarios; finally, the sub-business features of each specified dimension are weighted and summed according to the attention weights of each specified dimension to obtain the initial business features, avoiding biases or omissions that may be brought about by a single dimension, providing a more comprehensive and accurate understanding of the enterprise or business entity, providing a comprehensive and refined data basis for the training of the risk prediction model, and enhancing the expressiveness and prediction accuracy of the model.

[0099] Based on any one of the embodiments of the present disclosure, an embodiment of the present disclosure further proposes a system for implementing the training method of any one of the above risk prediction models. Among them, the principle of this system for implementing the training method of the risk prediction model can be as Figure 4As shown in the figure, the system mainly includes the following two modules: a multi-layer perceptron module integrating an attention mechanism and a Graph Convolutional Network (GCN) algorithm module. First, based on the multi-layer perceptron module integrating the attention mechanism, the weights of features are dynamically adjusted, and the financial and non-financial indicators of each company are feature-fused to form node labels. Then, the labels are substituted into the GCN algorithm module to obtain the trained labels. Among them,

[0100] (1) Multi-layer perceptron module integrating an attention mechanism

[0101] As Figure 5 shown, in the data preparation stage, historical data (sample business data) related to enterprises (sample business entities) is collected. For example, the historical data includes: shareholder information, financial data, and industry data; the financial data contains key indicators such as the current ratio, debt ratio, and return on net assets of the company to reflect the financial health of the company; the shareholder information indicators include the shareholding ratio of major shareholders, the transparency of the ultimate beneficiary, and the frequency of shareholder changes, which are used to analyze the stability of the company's shareholder structure; in addition, the average industry financial indicators, market share, and growth rate are also collected to provide the industry background and comparison benchmark; then, in the feature fusion stage, features from different sources are integrated to generate the initial feature representation of the nodes. These initial features integrate the company's financial condition, shareholder information, and industry background, enabling the node features to more comprehensively reflect the position and characteristics of the company in the graph; furthermore, in the feature fusion stage, first enter the attention score calculation stage, where the node features are processed through an activation function to calculate the feature importance of each node; further, the attention mechanism is applied to determine the weights of each feature, with higher weights paying more attention to important features and lower weights ignoring unimportant features; finally, through these steps, features from multiple sources are fused by weight into a comprehensive node representation, laying a foundation for subsequent model training and analysis;

[0102] (2) GCN algorithm module

[0103] As Figure 6As shown, first, the edges in the graph are defined by checking the supply chain relationships between nodes; if there is an overlap in the supply chain between two enterprises (sample nodes), such as sharing the same suppliers or customers, an edge is created between these two sample nodes, indicating their association in the supply chain. After the definition of the edges in the graph is completed, it enters the information aggregation stage. The model integrates the neighbor feature information of each sample node to update the node's embedding representation. The aggregation operation uses a function specific to GCN to integrate the feature information of neighbor nodes, and then introduces non-linear features through the activation function ReLU to ensure that the features of the sample nodes not only reflect their own attributes but also contain the relationship information of the surrounding nodes; among them, in the feature extraction and update steps, the model further processes the aggregated node features to generate the final embedding representation of the sample nodes, enabling them to better reflect the position and attributes of the nodes in the entire graph structure. Next, in the output layer stage, the model calculates the risk probability value of each node through a fully connected layer and the Sigmoid activation function. The generated risk probability value represents the likelihood of a risk occurring at the sample node; subsequently, in the loss function calculation step, the model uses the binary cross-entropy (BCE) loss function to quantify the difference between the prediction result and the true value; it should be noted that this loss value provides the direction for model optimization, helping the model adjust parameters to minimize the prediction error; finally, in the parameter update stage, the model optimizes each parameter through the gradient descent method; at this time, the model updates key parameters such as the weight matrix, bias term, and attention weights, gradually reducing the loss value. Through multiple iterations and optimizations, the model improves the prediction accuracy during continuous adjustment and obtains a stable performance after final convergence. By adding the method of defining edges based on supply chain overlap relationships, the entire GCN process from data preparation to edge definition, then to feature aggregation and output prediction constitutes a complete risk assessment and prediction process.

[0104] As an example, the risk prediction model based on GCN mainly includes the following two steps:

[0105] 1. Construct a multi-layer perceptron module integrating an attention mechanism to fuse financial indicators and non-financial indicators;

[0106] 1.1. Collect relevant data from data sources, where the relevant data includes, for example, financial data, shareholder information, and industry data.

[0107] Feature extraction is performed on financial data to obtain financial data features, which may include, but are not limited to, current ratio, quick ratio, debt ratio, interest coverage ratio, cash flow ratio, return on equity, asset turnover ratio, and degree of operating leverage. The current ratio (CR) is used to measure a company's short-term debt repayment ability; the quick ratio (QR) is used to further measure the short-term debt repayment ability, excluding assets with poor liquidity such as inventory; the debt ratio (DR) is used to indicate the proportion of a company's assets that come from debt; the interest coverage ratio (ICR) is used to indicate the company's ability to cover interest with profits; the return on equity (ROE) is used to reflect the income level of shareholders' equity; the asset turnover ratio (ATR) is used to reflect the utilization efficiency of a company's assets; the degree of operating leverage (DOL) is used to evaluate the impact of a company's business on profits; the cash flow ratio (CFR) is used to measure the company's ability to repay short-term debts with cash flows generated from operating activities.

[0108] Among them,

[0109]

[0110] In the embodiments of the present disclosure, the shareholder information features corresponding to the shareholder information include: the shareholding ratio of major shareholders, the transparency of the ultimate beneficiary, the frequency of shareholder changes, etc. Among them, the shareholding ratio of major shareholders (MSR) is used to measure the shareholder concentration; the transparency (UBOT) of the ultimate beneficiary (UBO) is used to measure the transparency of the shareholders; the frequency of shareholder changes (SF) is used to reflect the stability of the shareholders.

[0111]

[0112] UBOT = the score or measure of the public disclosure of UBO information;

[0113]

[0114] In the embodiments of the present disclosure, the industry data features corresponding to the industry data include: industry average financial indicators, market share, industry growth rate, etc. Among them, the industry average financial indicators use the average financial performance of companies in the industry as a benchmark; the market share (MS) is used to represent the market occupancy rate of a company; the industry growth rate (GR industry ) is used to represent the overall growth of the industry and serves as the external economic background.

[0115] Among them,

[0116]

[0117]

[0118] 1.2. Feature Fusion

[0119] In the feature fusion stage, features from different data sources are weighted and combined to generate the initial embedded feature representation (initial business feature) of the node, which is specifically expressed by the following formula:

[0120]

[0121] Wherein, represents the initial embedded feature of the i-th node, A i represents the shareholder information feature of the i-th node, such as the shareholding ratio of major shareholders, UBO transparency, and the frequency of shareholder changes, B i represents the industry data feature of the i-th node, such as industry average financial indicators, market share, and industry growth rate, X i represents the financial data feature of the i-th node, such as current ratio, debt ratio, and ROE, W g , W l W x represent the weight matrices corresponding to the features, a g , a l , a x represent the corresponding attention weights; wherein, a g , a l , a x are obtained by performing non-linear transformation on the shareholder information feature, industry data feature, and financial feature using the LeakyReLU activation function to generate attention scores, and normalizing the activated attention scores using the softmax function.

[0122] 2. Construct the GCN algorithm module to output accurate full customer group classification labels

[0123] 2.1. Information Aggregation

[0124] Integrate the feature information of neighbor nodes with its own features through an aggregation function. The aggregated result is non-linearly transformed through the ReLU activation function to enhance the expression ability of the model; the features output by the last layer are used as the final embedded features of the node,

[0125] 2.2. Calculate the output probability of each node through a fully connected layer and the Sigmoid activation function to obtain the risk probability value p i of the i-th node, which can be specifically expressed by the following formula:

[0126]

[0127] Wherein, is the final embedded representation of node i, W p is the weight matrix of the output layer, bp is the bias term.

[0128] 2.3. Use binary cross-entropy (BCE) as the loss function to measure the difference between the model's predicted value and the true value, which can be specifically expressed by the following formula:

[0129]

[0130] where N is the total number of samples, y i is the true label. Among them, the occurrence of risk in the enterprise's handling of the specified business is defined as 1, and the non-occurrence of risk in the enterprise's handling of the specified business is defined as 0; is the risk probability predicted by the model, that is, p i .

[0131] 2.4. Parameter update

[0132] Optimize the model parameters through the gradient descent method to minimize the loss function L BCE . The formula for parameter update is where η represents the learning rate, which is used to control the update amplitude, X represents the learnable parameters, represents the gradient of the loss function with respect to X; among them, it should be noted that in order to prevent overfitting and reduce unnecessary computational complexity, the parameter optimization will stop when one of the following stopping conditions is met:

[0133] (1) The number of iterations reaches the set upper limit: The training reaches the set maximum number of iterations, and the maximum number in this model is set to 1000;

[0134] (2) When the change in the loss function is less than a set threshold, that is, the model training converges. The threshold in this model is set to 0.001.

[0135] The above is the embodiment corresponding to the training method of the risk prediction model. The present disclosure also proposes an application method of the model, that is, a risk prediction method, Figure 7 which is the schematic flowchart of the risk prediction method shown in the fourth embodiment of the present disclosure.

[0136] As Figure 7 shown, the risk prediction method includes the following steps:

[0137] Step 701, obtain the target business undirected graph associated with the target business entity.

[0138] Among them, the target business undirected graph includes a first node for indicating the target business entity and a second node of a reference business entity having an association relationship with the target business entity.

[0139] In order to improve the accuracy of risk prediction, in the embodiments of the present disclosure, based on the relationship between a business entity (such as an enterprise) and reference business entities (partners, suppliers, customers, etc.), a target business undirected graph reflecting these association relationships is constructed. Based on the target business undirected graph, the potential risks and opportunities of the target business entity can be evaluated more accurately.

[0140] Among them, it should be noted that the first node is used to indicate the target business entity itself; the second node is used to indicate other business entities having a certain association relationship with the target business entity; in the target business undirected graph, there is an edge between the first node (target business entity) and each second node (associated business entity), indicating the association relationship between them.

[0141] Step 702: Determine the initial business characteristics of the first node and the initial business characteristics of the second node according to the target business data of multiple specified dimensions of the first node and the reference business data of multiple specified dimensions of the second node.

[0142] In the embodiments of the present disclosure, the target business data of the first node in multiple specified dimensions and the reference business data of the second node in multiple specified dimensions are collected. By performing data preprocessing and feature extraction on the target business data of the first node, and fusing the features extracted from the target business data of the first node, the initial business characteristics of the first node are generated; similarly, by performing data preprocessing and feature extraction on the reference business data of the second node, and fusing the features extracted from the reference business data of the second node, the initial business characteristics of the second node are obtained.

[0143] Among them, the multiple specified dimensions may include but are not limited to: financial dimension, industry dimension, and shareholder dimension. The sample business data may include but are not limited to: financial data, shareholder information, and industry data. Among them, the financial data includes but is not limited to key indicators such as the current ratio, debt ratio, and return on net assets of the enterprise. The financial data is used to reflect the financial health status of the enterprise; the shareholder information includes but is not limited to the shareholding ratio of major shareholders, the transparency of the ultimate beneficiary, and the frequency of shareholder changes. The shareholder information is used to analyze the stability of the company's shareholder structure; the industry data includes but is not limited to industry average financial indicators, market share, and growth rate. The industry data is used to provide the background and comparison benchmark of the industry.

[0144] Step 703: Input the initial business characteristics of the first node and the initial business characteristics of the second node into the trained risk prediction model to obtain the risk probability prediction value of the first node output by the risk prediction model.

[0145] Among them, the risk probability prediction value of the first node is used to indicate the risk degree of the target business entity handling the specified business.

[0146] In the embodiments of the present disclosure, the initial service characteristics of the first node and the initial service characteristics of the second node are input into a trained risk prediction model, and a risk probability prediction value of the first node output by the risk prediction model can be obtained.

[0147] As an example, at least one graph convolutional layer in the risk prediction model is used to update the initial service characteristics of the first node based on the initial service characteristics of the second node to obtain the target service characteristics of the first node; the prediction layer in the risk prediction model is used to perform risk prediction on the target service characteristics of the first node to obtain the risk probability prediction value of the first node.

[0148] That is to say, the graph convolutional layer updates the characteristics of the sample node by aggregating the characteristics of the sample node itself and the characteristics of its neighbor nodes. During the process of feature update, the output of each layer is used as the input of the next layer to gradually update the characteristics of the sample node, and finally the target service characteristics of the sample node are obtained; furthermore, the prediction layer in the risk prediction model is used to perform risk prediction on the target service characteristics of the first node to obtain the risk probability prediction value of the first node.

[0149] Among them, the risk prediction model can be trained by Figures 1 to 6 the training method of the risk prediction model described in the embodiments.

[0150] Step 704, determine the risk level of the target business entity handling the specified business according to the risk probability prediction value of the first node.

[0151] For the convenience of risk control, as a possible implementation manner, the risk probability prediction value of the first node output by the risk prediction model is mapped to a predefined risk probability interval to determine the risk level of the target business entity handling the specified business.

[0152] As an example, obtain multiple reference risk probability intervals; according to the risk probability prediction value of the first node, determine the target risk probability interval from the multiple reference risk probability intervals; wherein, the risk probability prediction value of the first node is within the target risk probability interval; according to the risk level indicated by the target risk probability interval, determine the risk level of the target business entity handling the specified business.

[0153] In the embodiments of the present disclosure, the reference risk probability interval is a set of pre-defined probability ranges for dividing different risk levels. For example, three reference risk probability intervals can be defined: low risk [0, 0.3), medium risk [0.3, 0.7), high risk [0.7, 1]. The target risk probability interval refers to the reference risk probability interval that contains the risk probability prediction value of the first node. According to the risk level indicated by the target risk probability interval, determine the risk level of the target business entity handling the specified business.

[0154] For example, assume that the predicted risk probability value of the first node output by the risk prediction model is 0.45, and multiple reference risk probability intervals are: low risk [0, 0.3), medium risk [0.3, 0.7), high risk [0.7, 1]. Since 0.45 falls within the medium risk interval [0.3, 0.7), the target risk probability interval corresponding to 0.45 is the medium risk interval. Therefore, the risk level for the target business entity to handle the specified business is determined to be "medium risk".

[0155] The risk prediction method according to the embodiments of the present disclosure realizes a comprehensive understanding of the target business entity and its business environment by obtaining a target business undirected graph associated with the target business entity. The target business undirected graph includes a first node for indicating the target business entity and a second node of a reference business entity having an association relationship with the target business entity, ensuring the information integrity in the risk assessment process. Furthermore, representative initial business features are extracted based on the multi-dimensional business data of the nodes in the target business undirected graph, and the initial business features are input into a trained risk prediction model, thereby accurately calculating the risk probability for the target business entity to handle the specified business. This not only improves the accuracy and reliability of risk prediction but also enables the rapid determination of the risk level for business handling according to the predicted risk probability value, providing strong data support for decision-making.

[0156] As described above Figures 1 to 3 Corresponding to the training method of the risk prediction model provided in the above embodiments, the present disclosure also provides a training device for the risk prediction model. Since the training device for the risk prediction model provided in the embodiments of the present disclosure corresponds to the training method of the risk prediction model provided in the above embodiments, the implementation manners of the training method of the risk prediction model are also applicable to the training device for the risk prediction model provided in the embodiments of the present disclosure and will not be described in detail in the embodiments of the present disclosure.

[0157] Figure 8 It is a schematic structural diagram of the training device for the risk prediction model shown in the fifth embodiment of the present disclosure.

[0158] As Figure 8 shown, the training device 800 for the risk prediction model includes: an acquisition module 810, a determination module 820, an input module 830, and a training module 840.

[0159] Among them, an acquisition module 810 is configured to acquire a sample business undirected graph associated with a risk prediction model to be trained. In the sample business undirected graph, sample nodes are used to indicate sample business entities, and edges between the sample nodes are used to indicate the association relationships between the sample business entities. A determination module 820 is configured to determine initial business features of each sample node according to sample business data of multiple specified dimensions of the sample business entities indicated by each sample node in the sample business undirected graph. An input module is configured to input the initial business features of each sample node into the risk prediction model to obtain risk probability prediction values of each sample node output by the risk prediction model. A training module 840 is configured to train the risk prediction model based on the risk probability prediction values and risk probability annotation values corresponding to each sample node.

[0160] As a possible implementation manner of an embodiment of the present disclosure, the risk probability prediction values of each sample node are predicted by the risk prediction model using the following modules: an update module and a prediction module.

[0161] Among them, the update module is configured to update the initial business features of each sample node by using at least one graph convolutional layer in the risk prediction model to obtain target business features of each sample node. The prediction module is configured to perform risk prediction on the target business features of each sample node by using a prediction layer in the risk prediction model to obtain risk probability prediction values of each sample node.

[0162] As a possible implementation manner of an embodiment of the present disclosure, the update module is configured to aggregate the initial business features of any sample node and the initial business features of the neighbor nodes of any sample node by using the first graph convolutional layer in at least one graph convolutional layer to obtain the updated business features of any sample node output by the first graph convolutional layer. Aggregate the updated business features of any sample node output by the previous graph convolutional layer and the updated business features of the neighbor nodes by using a non-first graph convolutional layer in at least one graph convolutional layer to obtain the updated business features of any sample node output by the non-first graph convolutional layer. Use the updated business features of any sample node output by the last graph convolutional layer in at least one graph convolutional layer as the target business features of any sample node.

[0163] As a possible implementation manner of an embodiment of the present disclosure, the determination module 820 is configured to perform feature extraction on the sample business data of any specified dimension of the sample business entity indicated by any sample node in the business undirected graph to obtain sub-business features of any specified dimension. Use an attention mechanism to determine the attention weight of any specified dimension according to the sub-business features of any specified dimension. Perform weighted summation on the sub-business features of each specified dimension of any sample node according to the attention weights of each specified dimension of any sample node to obtain the initial business features of any sample node.

[0164] As a possible implementation manner of the embodiments of the present disclosure, a determination module 820 is configured to perform a linear transformation on sub-business features of any specified dimension by using a weight matrix of sub-business features of any specified dimension and the attention mechanism to obtain sub-business features after linear transformation of any specified dimension; and perform a non-linear transformation on the sub-business features after linear transformation of any specified dimension by using an activation function to generate attention weights of any specified dimension.

[0165] As a possible implementation manner of the embodiments of the present disclosure, the sample business entity includes: business institutions, and an acquisition module 810 is configured to determine, from multiple business institutions, business institutions sharing the same resources; where the resources include: suppliers and / or service objects; and establish nodes with an association relationship in the sample business undirected graph according to the business institutions sharing the same resources.

[0166] As a possible implementation manner of the embodiments of the present disclosure, a training module 840 is configured to determine a sub-loss function value of any sample node based on a risk probability prediction value and a risk probability annotation value corresponding to any sample node; determine a target loss function value according to the sub-loss function values of each sample node; perform multiple rounds of iterative training on the risk prediction model according to the target loss function value, and determine whether the iterative training of the current round meets a set stop training condition; if so, end the training process of the risk prediction model, and if not, continue to train the risk prediction model.

[0167] The training device of the risk prediction model according to the embodiments of the present disclosure constructs initial business features based on the sample business undirected graph and its multi-dimensional sample business data, making the input of the risk prediction model more comprehensive and accurate, fully capturing the complex characteristics of business entities and their association relationships. Furthermore, inputting the initial business features into the risk prediction model to generate risk probability prediction values and performing supervised training in combination with risk probability annotation values helps the model learn more accurate risk assessment rules and improve the reliability and stability of prediction. Thereby, risks can be more effectively identified and managed in actual business scenarios.

[0168] Corresponding to the risk prediction method provided in the above Figure 7 embodiment, the present disclosure further provides a risk prediction device. Since the risk prediction device provided in the embodiments of the present disclosure corresponds to the risk prediction method provided in the above embodiment, the implementation manners of the risk prediction method are also applicable to the risk prediction device provided in the embodiments of the present disclosure and will not be described in detail in the embodiments of the present disclosure.

[0169] Figure 9 It is a schematic structural diagram of the risk prediction device shown in the sixth embodiment of the present disclosure.

[0170] As Figure 9As shown in the figure, the risk prediction device 900 includes: an acquisition module 910, a first determination module 920, an input module 930, and a second determination module 940.

[0171] Among them, the acquisition module 910 is configured to acquire a target business undirected graph associated with a target business entity; wherein, the target business undirected graph includes a first node for indicating the target business entity and a second node of a reference business entity having an association relationship with the target business entity; the first determination module 920 is configured to determine an initial business feature of the first node and an initial business feature of the second node according to target business data of multiple specified dimensions of the first node and reference business data of multiple specified dimensions of the second node; the input module 930 is configured to input the initial business feature of the first node and the initial business feature of the second node into a trained risk prediction model to obtain a risk probability prediction value of the first node output by the risk prediction model; wherein, the risk probability prediction value of the first node is used to indicate the risk degree of the target business entity handling a specified business; the second determination module 940 is configured to determine the risk level of the target business entity handling the specified business according to the risk probability prediction value of the first node.

[0172] As a possible implementation manner of the embodiment of the present disclosure, the risk probability prediction value of the first node is predicted by the risk prediction model using the following modules: an update module and a prediction module.

[0173] Among them, the update module is configured to update the initial business feature of the first node based on the initial business feature of the second node by using at least one graph convolutional layer in the risk prediction model to obtain a target business feature of the first node; the prediction module is configured to perform risk prediction on the target business feature of the first node by using a prediction layer in the risk prediction model to obtain a risk probability prediction value of the first node.

[0174] The risk prediction method of the embodiment of the present disclosure realizes a comprehensive understanding of the target business entity and its business environment by acquiring a target business undirected graph associated with the target business entity; wherein, the target business undirected graph includes a first node for indicating the target business entity and a second node of a reference business entity having an association relationship with the target business entity, ensuring the information integrity in the risk assessment process. Furthermore, representative initial business features are extracted based on the multi-dimensional business data of the nodes in the target business undirected graph, and the initial business features are input into a trained risk prediction model, thereby accurately calculating the risk probability of the target business entity handling a specified business, not only improving the accuracy and reliability of risk prediction, but also realizing the rapid determination of the risk level of business handling according to the risk probability prediction value, providing strong data support for decision-making.

[0175] In an exemplary embodiment, an electronic device is further proposed.

[0176] Among them, the electronic device includes:

[0177] A processor;

[0178] A memory for storing instructions executable by the processor;

[0179] Among them, the processor is configured to execute instructions to implement the training method or risk prediction method of the risk prediction model proposed in any of the foregoing embodiments.

[0180] As an example, Figure 10 is a schematic structural diagram of the electronic device 1000 shown in an exemplary embodiment of the present disclosure. As Figure 10 shown, the above-mentioned electronic device 1000 may further include:

[0181] A memory 1010 and a processor 1020, a bus 1030 connecting different components (including the memory 1010 and the processor 1020). The memory 1010 stores a computer program, and when the processor 1020 executes the program, it implements the training method or risk prediction method of the embodiment of the present disclosure.

[0182] The bus 1030 represents one or more of several types of bus structures, including a memory bus or a memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus using any bus structure in a variety of bus structures. For example, these architectures include, but are not limited to, Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MAC) bus, Enhanced ISA bus, Video Electronics Standards Association (VESA) local bus, and Peripheral Component Interconnect (PCI) bus.

[0183] The electronic device 1000 typically includes a variety of electronic device-readable media. These media can be any available media accessible by the electronic device 1000, including volatile and non-volatile media, removable and non-removable media.

[0184] The memory 1010 may further include a computer system-readable medium in the form of volatile memory, such as random access memory (RAM) 1040 and / or cache memory 1050. The server 1000 may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, the storage system 1060 may be used to read and write non-removable, non-volatile magnetic media ( Figure 10 not shown, commonly referred to as a "hard disk drive"). Although Figure 10Not shown in the figure, a disk drive for reading and writing to a removable non-volatile disk (such as a "floppy disk") and an optical disk drive for reading and writing to a removable non-volatile optical disk (such as a CD-ROM, DVD-ROM or other optical medium) can be provided. In these cases, each drive can be connected to the bus 1030 through one or more data medium interfaces. The memory 1010 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the various embodiments of the present disclosure.

[0185] A program / utility 1080 having a set (at least one) of program modules 1070 can be stored, for example, in the memory 1010. Such program modules 1070 include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include the implementation of a network environment. The program modules 1070 generally perform the functions and / or methods in the embodiments described in the present disclosure.

[0186] The electronic device 1000 can also communicate with one or more external devices 1090 (such as a keyboard, a pointing device, a display 1091, etc.), and can also communicate with one or more devices that enable a user to interact with the electronic device 1000, and / or communicate with any device that enables the electronic device 1000 to communicate with one or more other computing devices (such as a network card, a modem, etc.). Such communication can be carried out through the input / output (I / O) interface 1092. And the electronic device 1000 can also communicate with one or more networks (such as a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) through the network adapter 1093. As shown in the figure, the network adapter 1093 communicates with other modules of the electronic device 1000 through the bus 1030. It should be understood that although not shown in the figure, other hardware and / or software modules can be used in combination with the electronic device 1000, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, etc.

[0187] The processor 1020 executes various functional applications and data processing by running the programs stored in the memory 1010.

[0188] It should be noted that for the implementation process and technical principle of the electronic device in this embodiment, refer to the foregoing explanation of the training method or risk prediction method of the risk prediction model of the embodiments of the present disclosure, which will not be elaborated here.

[0189] In an exemplary embodiment, there is also provided a computer-readable storage medium including instructions, such as a memory including instructions, and the instructions can be executed by a processor of an electronic device to complete the training method or the risk prediction method of the risk prediction model proposed in any of the above embodiments. Optionally, the computer-readable storage medium may be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, and an optical data storage device, etc.

[0190] In an exemplary embodiment, there is also provided a computer program product including a computer program / instructions, characterized in that when the computer program / instructions are executed by a processor, the training method or the risk prediction method of the risk prediction model proposed in any of the above embodiments is implemented.

[0191] Those skilled in the art will readily conceive of other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. The present disclosure is intended to cover any variations, uses, or adaptations of the present disclosure, which follow the general principles of the present disclosure and include common general knowledge or conventional technical means in the technical field not disclosed by the present disclosure. The specification and embodiments are only to be regarded as exemplary, and the true scope and spirit of the present disclosure are pointed out by the following claims.

[0192] It should be understood that the present disclosure is not limited to the exact structures already described and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the present disclosure is only limited by the appended claims.

Claims

1. A training method for a risk prediction model, characterized in that: include: Acquire a sample business undirected graph associated with the risk prediction model to be trained; wherein the sample nodes of the sample business undirected graph are used to indicate sample business entities, and the edges between the sample nodes are used to indicate association relationships between the sample business entities; Determine the initial business characteristics of each sample node according to the sample business data of multiple specified dimensions of the sample business entity indicated by each sample node in the sample business undirected graph; Inputting the initial business characteristics of each of the sample nodes into the risk prediction model to obtain the risk probability prediction value of each of the sample nodes output by the risk prediction model; The risk prediction model is trained based on the risk probability prediction value and the risk probability annotation value corresponding to each of the sample nodes.

2. The method according to claim 1, characterized in that The risk probability prediction value of each sample node is predicted by the risk prediction model using the following steps: Using at least one graph convolutional layer in the risk prediction model to update the initial business features of each of the sample nodes to obtain the target business features of each of the sample nodes; The prediction layer in the risk prediction model is used to perform risk prediction on the target business characteristics of each sample node to obtain a risk probability prediction value of each sample node.

3. The method according to claim 2, characterized in that The adopting at least one graph convolution layer in the risk prediction model to update the initial business features of each of the sample nodes to obtain the target business features of each of the sample nodes includes: Aggregating the initial service features of any sample node and the initial service features of neighboring nodes of any sample node using the first graph convolution layer in the at least one graph convolution layer to obtain updated service features of any sample node output by the first graph convolution layer; A non-first graph convolution layer in the at least one graph convolution layer is used to aggregate the update service features of the any sample node output by the previous graph convolution layer and the update service features of the neighboring nodes to obtain the update service features of the any sample node output by the non-first graph convolution layer; The updated service feature of any sample node output by the last graph convolution layer in the at least one graph convolution layer is used as the target service feature of any sample node.

4. The method according to claim 1, characterized in that: The determining of the initial service characteristics of each sample node according to the sample service data of multiple specified dimensions of the sample service entity indicated by each sample node in the sample service undirected graph includes: For any specified dimension of the sample business entity indicated by any sample node in the business undirected graph, feature extraction is performed on the sample business data of the any specified dimension to obtain a sub-business feature of the any specified dimension; Adopting an attention mechanism, determining an attention weight of any specified dimension according to the sub-service characteristics of any specified dimension; According to the attention weights of the designated dimensions of any one of the sample nodes, weighted summation is performed on the sub-business features of the designated dimensions of any one of the sample nodes to obtain the initial business features of the any one of the sample nodes.

5. The method according to claim 4, characterized in that The adopting of the attention mechanism to determine the attention weight of any specified dimension according to the sub-service characteristics of any specified dimension includes: Using the weight matrix of the sub-service feature of any specified dimension and the attention mechanism, linearly transform the sub-service feature of any specified dimension to obtain the sub-service feature after linear transformation of any specified dimension; An activation function is used to perform nonlinear transformation on the linearly transformed sub-service features of any specified dimension to generate an attention weight of any specified dimension.

6. The method according to any one of claims 1 to 5, characterized in that The sample business entities include: operating institutions, The step of obtaining an undirected graph of sample businesses associated with the risk prediction model to be trained includes: Determine, from among a plurality of operating organizations, an operating organization that shares the same resources; wherein the resources include: suppliers and / or service objects; According to the business organizations that share the same resources, nodes with association relationships in the sample business undirected graph are established.

7. The method according to any one of claims 1 to 5, characterized in that The risk prediction model is trained based on the risk probability prediction value and the risk probability label value corresponding to each of the sample nodes, including: Determine the sub-loss function value of any sample node based on the risk probability prediction value and the risk probability annotation value corresponding to any sample node; Determine the target loss function value according to the sub-loss function value of each sample node; According to the target loss function value, the risk prediction model is trained for multiple rounds of iterations, and it is determined whether the current round of iteration training meets the set training stop condition; If so, the training process of the risk prediction model is terminated; if not, the training of the risk prediction model continues.

8. A risk prediction method, characterized in that: include: Acquire a target business undirected graph associated with a target business entity; wherein the target business undirected graph includes a first node for indicating the target business entity and a second node of a reference business entity having an association relationship with the target business entity; Determining, according to the target service data of the first node in multiple specified dimensions and the reference service data of the second node in multiple specified dimensions, an initial service feature of the first node and an initial service feature of the second node; Inputting the initial business features of the first node and the initial business features of the second node into a trained risk prediction model to obtain a risk probability prediction value of the first node output by the risk prediction model; wherein the risk probability prediction value of the first node is used to indicate the risk level of the target business entity in handling a designated business; The risk level of the target business entity in handling the designated business is determined according to the risk probability prediction value of the first node.

9. The method according to claim 8, characterized in that The risk probability prediction value of the first node is predicted by the risk prediction model using the following steps: Using at least one graph convolutional layer in the risk prediction model to update the initial business features of the first node based on the initial business features of the second node to obtain the target business features of the first node; The prediction layer in the risk prediction model is used to perform risk prediction on the target business characteristics of the first node to obtain a risk probability prediction value of the first node.

10. The method according to claim 8, characterized in that The step of determining the risk level of the target business entity handling the designated business according to the risk probability prediction value of the first node includes: Obtain multiple reference risk probability intervals; Determine a target risk probability interval from the multiple reference risk probability intervals according to the risk probability prediction value of the first node; wherein the risk probability prediction value of the first node is within the target risk probability interval; The risk level of the target business entity in handling the designated business is determined according to the risk level indicated by the target risk probability interval.

Images