Deep learning-based risk prediction method, device, equipment and medium
By constructing a supply chain network and a graph convolutional neural network, and by filtering and supplementing the information features of SMEs, the problem of low accuracy in risk prediction in existing technologies has been solved, and efficient and accurate prediction of risks for SMEs has been achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- PING AN TECH (SHENZHEN) CO LTD
- Filing Date
- 2023-09-20
- Publication Date
- 2026-07-21
AI Technical Summary
Existing methods for predicting risks in SMEs simply combine the credit characteristics of core enterprises and SMEs, resulting in low accuracy in risk prediction.
By constructing a supply chain network, we can obtain supply and enterprise information among companies, screen out target companies with missing information features, use information from similar companies to complete the information features of the target companies, and use graph convolutional neural networks for risk prediction.
It improves the accuracy of risk prediction for SMEs, especially for SMEs with incomplete information disclosure, and can obtain more accurate prediction results.
Smart Images
Figure CN117172912B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of deep learning technology, and in particular to a risk prediction method, apparatus, electronic device, and storage medium based on deep learning. Background Technology
[0002] In the context of supply chain finance, SMEs leverage the credit of core enterprises to obtain financing. However, predicting the risk of SMEs is crucial for the success of supply chain finance. Currently, there are many methods for predicting SME risk, such as corporate credit risk prediction. These methods combine the credit characteristics of SMEs and core enterprises, employing techniques like SVM, AdaBoost, and XGboost to predict SME risk. However, simply merging the credit characteristics of core and SMEs based on disclosed corporate features to aggregate SME credit risk achieves the goal of risk prediction. This approach, however, yields low accuracy in risk prediction. Summary of the Invention
[0003] Therefore, it is necessary to provide a risk prediction method, device, electronic device, and storage medium based on deep learning to address the aforementioned technical problems and improve the accuracy of risk prediction.
[0004] A risk prediction method based on deep learning includes:
[0005] Determine the supply chain network in which the enterprise to be predicted is located, and obtain supply information among the enterprises in the supply chain network, including supply relationships and supply proportions;
[0006] Obtain enterprise information corresponding to each enterprise in the supply chain network;
[0007] Based on the enterprise information, enterprises with missing information features are screened out, and these enterprises with missing information features are identified as target enterprises.
[0008] For each target enterprise, based on the supply information among the enterprises in the supply chain network, the similarity between the target enterprise and any enterprise with complete information features is calculated to determine the similar enterprises of the target enterprise;
[0009] Based on the enterprise information of the similar enterprises, the missing information features of the target enterprise are determined to complete the enterprise information of each enterprise in the supply chain network;
[0010] Based on the complete supply chain network of enterprise information, risk prediction is performed on the enterprise to be predicted, and the risk prediction result of the enterprise to be predicted is obtained.
[0011] In this embodiment of the application, the step of filtering out enterprises with missing information features based on the enterprise information and identifying the enterprises with missing information features as target enterprises includes: generating an information disclosure vector corresponding to each enterprise according to the disclosure status of enterprise information corresponding to each enterprise in the supply chain network, wherein the parameters in the information disclosure vector are used to characterize the disclosure status of different information features; calculating the information missing rate of the enterprise based on the information disclosure vector; filtering out enterprises with missing information features based on the information missing rate and identifying the enterprises with missing information features as target enterprises.
[0012] In this embodiment, the step of calculating the similarity between the target enterprise and any enterprise with complete information features based on the supply information between enterprises in the supply chain network for each target enterprise, in order to determine the similar enterprises of the target enterprise, includes: determining a weighted directed graph corresponding to the supply chain network, wherein the nodes in the graph represent enterprises in the supply chain network, the directional edges in the graph represent the supply relationship, and the weight of the edge represents the supply proportion; for each node in the weighted directed graph, determining the set of adjacent nodes and the weight set corresponding to the adjacent node set based on the structure of the weighted directed graph, wherein the weight set is the set composed of the weights of the edges between the node and its corresponding adjacent nodes; and calculating the similarity between the node corresponding to the target enterprise and the node corresponding to any enterprise with complete information features based on the set of adjacent nodes and the weight set, in order to determine the similar enterprises of the target enterprise.
[0013] In this embodiment of the application, determining the set of adjacent nodes and the weight set corresponding to the set of adjacent nodes for each node in the weighted directed graph, based on the structure of the weighted directed graph, includes: determining the set of adjacent nodes of the node based on the edge structure and the direction of the edges in the weighted directed graph, wherein the set of adjacent nodes includes the node's access nodes and outgoing nodes; and determining the weight set based on the weight of the edge between the node and the access node, and the weight of the edge between the node and the outgoing node.
[0014] In this embodiment of the application, calculating the similarity between the node corresponding to the target enterprise and the node corresponding to any enterprise with complete information features based on the neighbor node set and the weight set includes: performing the following similarity calculation based on the neighbor node set and the weight set:
[0015]
[0016] Wherein, A is the node corresponding to the target enterprise, and B is the node corresponding to the enterprise with complete information features.both Let B be the set of adjacent nodes of node A. both W is the set of adjacent nodes of node B. Aboth Let A be the set of adjacent nodes. both The corresponding weight set, where n is the weight set W. Aboth The number of parameters in W Bboth Let B be the set of adjacent nodes. both The corresponding weight set, where m is the weight set W. Bboth The number of parameters in the J w (A,B) represents the similarity between node A and node B.
[0017] In this embodiment of the application, the enterprise information corresponds to multiple information dimensions, and each information dimension includes multiple information features. The step of determining the missing information features of the target enterprise based on the enterprise information of the similar enterprises to complete the enterprise information of each enterprise in the supply chain network includes: determining the information dimension in which the missing information features of the target enterprise are located; extracting the information features on the information dimension from the enterprise information of the similar enterprises; and determining the missing information features of the target enterprise based on preset information completion rules and using the information features extracted from the enterprise information of the similar enterprises.
[0018] In this embodiment of the application, the step of performing risk prediction on the enterprise to be predicted based on the supply chain network with completed enterprise information to obtain the risk prediction result of the enterprise to be predicted includes: determining the weighted directed graph corresponding to the supply chain network and the feature matrix corresponding to each node in the graph based on the supply chain network with completed enterprise information; performing convolution processing on the weighted directed graph and the feature matrix corresponding to each node in the graph based on a graph convolutional neural network to obtain the risk label of the enterprise to be predicted, so as to obtain the risk prediction result of the enterprise to be predicted.
[0019] A risk prediction device based on deep learning, comprising:
[0020] The network construction module is used to determine the supply chain network in which the enterprise to be predicted is located, and to obtain supply information between enterprises in the supply chain network, including supply relationships and supply proportions.
[0021] The information acquisition module is used to acquire enterprise information corresponding to each enterprise in the supply chain network;
[0022] The determination module is used to filter out enterprises with missing information features based on the enterprise information, and determine the enterprises with missing information features as target enterprises;
[0023] The similarity calculation module is used to calculate the similarity between each target enterprise and any enterprise with complete information features, based on the supply information between the enterprises in the supply chain network, so as to determine the similar enterprises of the target enterprise.
[0024] The information completion module is used to determine the missing information features of the target enterprise based on the enterprise information of the similar enterprises, so as to complete the enterprise information of each enterprise in the supply chain network;
[0025] The risk prediction module is used to predict the risk of the enterprise to be predicted based on the supply chain network after the enterprise information is completed, and to obtain the risk prediction result of the enterprise to be predicted.
[0026] In this embodiment of the application, the determining module is further configured to: generate an information disclosure vector corresponding to each enterprise in the supply chain network based on the disclosure status of enterprise information, wherein the parameters in the information disclosure vector are used to characterize the disclosure status of different information features; calculate the information missing rate of the enterprise based on the information disclosure vector; and filter out enterprises with missing information features based on the information missing rate, and determine the enterprises with missing information features as the target enterprises.
[0027] In this embodiment, the similarity calculation module is further configured to: determine a weighted directed graph corresponding to the supply chain network, wherein nodes in the graph represent enterprises in the supply chain network, directional edges in the graph represent supply relationships, and the weights of the edges represent supply proportions; for each node in the weighted directed graph, determine a set of adjacent nodes and a set of weights corresponding to the adjacent node set based on the structure of the weighted directed graph, wherein the weight set is a set composed of the weights of the edges between the node and its corresponding adjacent nodes; and calculate the similarity between the node corresponding to the target enterprise and any node corresponding to any enterprise with complete information features based on the adjacent node set and the weight set, thereby determining similar enterprises to the target enterprise.
[0028] In this embodiment of the application, the similarity calculation module is further configured to: determine the set of adjacent nodes of the node according to the edge structure and the direction of the edge in the weighted directed graph, wherein the set of adjacent nodes includes the access node and the outgoing node of the node; and determine the weight set according to the weight of the edge between the node and the access node, and the weight of the edge between the node and the outgoing node.
[0029] In this embodiment of the application, the similarity calculation module is further configured to: perform the following similarity calculation based on the set of adjacent nodes and the set of weights:
[0030]
[0031] Wherein, A is the node corresponding to the target enterprise, and B is the node corresponding to the enterprise with complete information features. both Let B be the set of adjacent nodes of node A. both W is the set of adjacent nodes of node B. Aboth Let A be the set of adjacent nodes. both The corresponding weight set, where n is the weight set W. Aboth The number of parameters in W Bboth Let B be the set of adjacent nodes. both The corresponding weight set, where m is the weight set W. Bboth The number of parameters in the J w (A,B) represents the similarity between node A and node B.
[0032] In this embodiment of the application, the enterprise information corresponds to multiple information dimensions, and each information dimension includes multiple information features. The information completion module is further used to: determine the information dimension in which the missing information feature of the target enterprise is located; extract the information feature on the information dimension from the enterprise information of similar enterprises; and determine the missing information feature of the target enterprise based on preset information completion rules and by using the information feature extracted from the enterprise information of similar enterprises.
[0033] In this embodiment of the application, the risk prediction module is further used to: determine the weighted directed graph corresponding to the supply chain network and the feature matrix corresponding to each node in the graph based on the supply chain network after the enterprise information is completed; and perform convolution processing on the weighted directed graph and the feature matrix corresponding to each node in the graph based on the graph convolutional neural network to obtain the risk label of the enterprise to be predicted, so as to obtain the risk prediction result of the enterprise to be predicted.
[0034] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the deep learning-based risk prediction method described above.
[0035] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the deep learning-based risk prediction method described above.
[0036] In summary, this application proposes a risk prediction method, device, electronic device, and storage medium based on deep learning. This application utilizes the supply chain network of the enterprise to be predicted to extract the relationships between the enterprise and its upstream and downstream enterprises, and even indirectly related enterprises. This relationship information, i.e., supply information, is applied to the risk prediction of the enterprise. Based on the relationships between enterprises, similar enterprises are selected, and the enterprise information of similar enterprises is used to supplement the information of enterprises with missing information features. The enterprise risk is predicted by using the supplemented enterprise information and the node relationships shown by the supply chain network, which improves the accuracy of risk prediction. More accurate prediction results can also be obtained for SMEs with incomplete information disclosure. Attached Figure Description
[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0038] Figure 1 This is a flowchart illustrating a deep learning-based risk prediction method according to an exemplary embodiment of this application;
[0039] Figure 2 This is a schematic diagram of an exemplary supply chain network;
[0040] Figure 3 This is a flowchart illustrating a deep learning-based risk prediction method according to another exemplary embodiment of this application;
[0041] Figure 4 This is a flowchart illustrating a deep learning-based risk prediction method according to another exemplary embodiment of this application;
[0042] Figure 5 This is a flowchart illustrating a deep learning-based risk prediction method according to another exemplary embodiment of this application;
[0043] Figure 6 This is a flowchart illustrating a deep learning-based risk prediction method according to another exemplary embodiment of this application;
[0044] Figure 7 This is a flowchart illustrating a deep learning-based risk prediction method according to another exemplary embodiment of this application;
[0045] Figure 8This is a schematic diagram of a two-layer graph convolutional neural network feedforward network in a deep learning-based text information extraction method according to an exemplary embodiment of this application;
[0046] Figure 9 This is another exemplary supply chain network diagram;
[0047] Figure 10 This is a block diagram illustrating a deep learning-based risk prediction device according to an exemplary embodiment of this application;
[0048] Figure 11 This is a schematic diagram of an electronic device structure according to an exemplary embodiment of this application. Detailed Implementation
[0049] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The embodiments described with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.
[0050] Figure 1 This is a flowchart illustrating a deep learning-based risk prediction method according to an exemplary embodiment of this application, such as... Figure 1 As shown, this risk prediction method includes the following steps:
[0051] S101, determine the supply chain network in which the enterprise to be predicted is located, and obtain supply information between the enterprises in the supply chain network, including supply relationships and supply share.
[0052] In some embodiments, a supply chain network can be constructed by identifying the upstream and downstream enterprises of the enterprise to be predicted, and other upstream and downstream enterprises of the upstream and downstream enterprises. Furthermore, supply information among the various enterprises included in the supply chain network is obtained based on this network, wherein the supply information includes, but is not limited to, the supply relationships and supply proportions between enterprises. Here, a supply relationship can be understood as the supply relationship between a product supplier and a supplied enterprise, and a supply proportion can be understood as the percentage of the supply business between the product supplier and the supplied enterprise relative to the total supply business of the product supplier or the supplied enterprise.
[0053] For example, the supply share can be the proportion of a product supplier's supply business in the overall supply business of the supplied enterprise (i.e., the overall demand of the supplied enterprise).
[0054] Figure 2This is a schematic diagram of an exemplary supply chain network. Figure 2 As shown in the figure, A represents the core enterprise, B represents small and medium-sized enterprises (5% of the supply), and suppliers 1, 2 and 3 are the largest supplier (40% of the supply), the second largest supplier (30% of the supply), and the third largest supplier (25% of the supply), respectively.
[0055] In addition, supply share can also refer to the proportion of a supplier's business to a supplier within the supplier's overall supply business. For example, if SME 4 supplies to both SME C and SME D, with 40% of its supply business to SME C and 60% to SME D, then in the supply chain network, the supply share between SME C and SME 4 can be 40%, and the supply share between SME D and SME 4 can be 60%.
[0056] When constructing a supply chain network, the supply chain data used can be obtained from publicly disclosed corporate supply data (e.g., corporate financial reports, bidding announcements, and corporate prospectuses) to determine the supply relationships of various companies, identifying the upstream and downstream companies of a particular company, as well as other upstream and downstream companies of the upstream companies and other upstream and downstream companies of the downstream companies. Based on this data, a corporate supply chain network is constructed using knowledge graph construction methods.
[0057] In practice, since there are no official mandatory requirements for the disclosure of enterprise supply data, a lot of information is usually hidden or anonymized. To ensure the effectiveness of the information processed in subsequent steps, this step can perform multiple rounds of filtering and screening of enterprise supply information, such as filtering and screening information through multi-source verification of data from the same period.
[0058] S102, Obtain enterprise information corresponding to each enterprise in the supply chain network.
[0059] This involves acquiring enterprise information for each company within the supply chain network. In this embodiment, enterprise information may include, but is not limited to, financial information, basic information, and risk information. This information can be categorized into different dimensions or types based on its content; for example, financial information, basic information, and risk information each correspond to an information dimension. Each information dimension may include multiple information features. For instance, financial information may include, but is not limited to, enterprise operating profit, cash flow, asset structure, growth capacity, solvency, operational capacity, and profitability. Basic information may include, but is not limited to, the enterprise's main business, industry, operating scale, and nature. Risk information may include, but is not limited to, enterprise litigation risk, operational anomalies, industry trends, and whether the enterprise has recently been designated as a Special Treatment (ST) stock or has engaged in financial manipulation. ST stands for Special Treatment, a policy implemented by the Shanghai and Shenzhen stock exchanges to warn listed companies with abnormal financial or other conditions by adding the letter "ST" before their stock name, commonly known as "wearing a cap," to warn investors to exercise caution when investing in such stocks.
[0060] It should be noted that the content of enterprise information can be screened and added according to the actual needs and predicted risks, and this application does not impose any restrictions.
[0061] S103, based on enterprise information, filter out enterprises with missing information features, and identify enterprises with missing information features as target enterprises.
[0062] In some embodiments, the information characteristics of core enterprises with relatively complete information disclosure in the supply chain network can be used as the inspection benchmark. By comparing and examining whether the information characteristics of each enterprise are disclosed and the extent of disclosure, the completeness of the enterprise's information can be determined, thereby filtering out enterprises with missing information characteristics. Alternatively, based on the risk to be predicted (such as credit risk, financial risk, etc.), the necessary information related to the risk and the validity period of the information can be determined. By judging whether each enterprise's enterprise information contains this necessary information and determining the validity of the information based on the validity period, it can be determined whether the enterprise is missing information characteristics.
[0063] For ease of description, this application refers to enterprises lacking information features as target enterprises.
[0064] The target companies selected in this application may be one or more.
[0065] In some embodiments, the target enterprise may include the enterprise to be predicted. In reality, SMEs, due to their shorter establishment time and slower development, often disclose less information, resulting in incomplete information features. When the enterprise to be predicted is an SME with limited information disclosure, it will also be screened out as a target enterprise with missing information features. For example, ... Figure 2 In the supply chain network diagram, the information features of company A, suppliers 1, 2, and 3 are relatively comprehensive, while the information features of company B are fewer. When it is necessary to predict the risk of company B, company B is the company to be predicted and also serves as the target company in the screening.
[0066] In addition, the target companies may include not only those companies to be predicted, but also companies with missing information features.
[0067] S104. For each target enterprise, based on the supply information between enterprises in the supply chain network, calculate the similarity between the target enterprise and any enterprise with complete information features, so as to determine the similar enterprises of the target enterprise.
[0068] In this embodiment, the similarity between enterprises is determined based on their supply relationships and supply proportions. The similarity between the target enterprise and any other enterprise besides all target enterprises is calculated. The other enterprises in the supply chain network besides all target enterprises are those with complete information features, determined after screening. In some embodiments, for each target enterprise, the similarity between that target enterprise and each enterprise with complete information features is calculated. Similar enterprises to the target enterprise are then selected based on the similarity values.
[0069] When filtering similar companies, the similarity scores can be sorted, and the company with the highest similarity score can be selected and recorded as a similar company to the target company; alternatively, a preset number of companies can be selected based on their similarity ranking and recorded as similar companies to the target company. It should be noted that the number of similar companies can be set as needed, and this application does not impose any limit on it.
[0070] S105, Based on the enterprise information of similar enterprises, determine the missing information features of the target enterprise in order to complete the enterprise information of each enterprise in the supply chain network.
[0071] Based on the enterprise information of similar companies, the missing information features of the target company can be supplemented. As a feasible implementation method, companies with the highest similarity to the target company can be selected, and the missing information of the target company can be determined based on the enterprise information of those companies.
[0072] As another feasible implementation method, multiple similar enterprises are selected based on similarity, and the corresponding similarity is used as a weight. The enterprise information of multiple similar enterprises is combined to determine the missing information features of the target enterprise.
[0073] S106, based on the supply chain network after the enterprise information is completed, performs risk prediction on the enterprise to be predicted and obtains the risk prediction result of the enterprise to be predicted.
[0074] In this embodiment, a classification model can be built based on deep learning to process the supply information and the enterprise information of each enterprise after information completion in the supply chain network, determine whether the enterprise to be predicted has a preset type of risk, and output the risk prediction result. The preset type of risk can be information risk, financial risk, etc., and credit risk, financial risk, etc., are predicted based on the supply chain network and the complete enterprise information of each enterprise during risk prediction.
[0075] As a feasible implementation method, a graph convolutional neural network is used to perform convolution processing on the supply chain network after the enterprise information is completed and the enterprise information corresponding to each node, so as to transmit the risk of the enterprise associated with the enterprise to be predicted and obtain the final risk prediction result of the enterprise to be predicted.
[0076] This application supplements the missing information features of enterprises in the supply chain network before risk prediction, and performs risk prediction based on complete enterprise information, avoiding model overfitting. It also fully utilizes the relationship information in the supply chain network to facilitate risk transmission, thereby enhancing the accuracy of risk prediction. In summary, this application proposes a deep learning-based risk prediction method. This method uses the supply chain network of the enterprise to be predicted to extract the relationships between the enterprise and its upstream and downstream enterprises, and even indirectly related enterprises. This relationship information, i.e., supply information, is applied to the risk prediction of the enterprise. Based on the relationships between enterprises, similar enterprises are selected, and their information is used to supplement the missing information features of the enterprises. The supplemented enterprise information and the node relationships shown by the supply chain network are used to predict enterprise risk, improving the accuracy of risk prediction. More accurate prediction results can also be obtained for SMEs with incomplete information disclosure.
[0077] Based on the above embodiments, such as Figure 3 As shown, in this embodiment of the application, step S103, "based on enterprise information, filtering out enterprises with missing information features, and identifying enterprises with missing information features as target enterprises," includes the following steps:
[0078] S301, based on the disclosure of enterprise information corresponding to each enterprise in the supply chain network, generate an information disclosure vector for each enterprise. The parameters in the information disclosure vector are used to characterize the disclosure status of different information features.
[0079] When screening for companies with missing information features, an information disclosure vector can be constructed based on the disclosure status of each company in the supply chain network. The parameters in this vector can be used to characterize the disclosure status of different information features.
[0080] For example, the disclosure status is used to characterize whether information has been disclosed and whether it was recently disclosed. "Recently" can be a preset time period, such as within one year. Numerical values can be used to represent the disclosure status; for example, 0 represents no disclosure, 1 represents non-recent disclosure, and 2 represents recent disclosure.
[0081] Based on the disclosure status of each information feature in the enterprise information, an information disclosure vector corresponding to that enterprise information is constructed. For example, the information disclosure vector corresponding to enterprise A is A(a1, a2, a3, ..., a...). n ), a∈(0,1,2), n is the number of information features, which is also the total number of parameters in the information disclosure vector.
[0082] It should be noted that this application does not limit the specific numerical value corresponding to the disclosure status.
[0083] S302, Calculate the information missing rate of an enterprise based on the information disclosure vector.
[0084] For each enterprise in the supply chain network, the information missing rate of that enterprise is calculated based on its corresponding information disclosure vector. The information missing rate can be calculated using the following formula:
[0085]
[0086] Where P is the information missing rate of the enterprise, a i represents the parameters in the information disclosure vector, and n is the total number of parameters in the information disclosure vector.
[0087] S303: Based on the information missing rate, screen out enterprises with missing information features and identify them as target enterprises.
[0088] Based on the value of P, we can filter and identify companies with missing information features, thereby limiting the scope of target companies whose information needs to be supplemented.
[0089] In this embodiment of the application, a time-series feature is introduced in the verification of whether enterprise information is disclosed, so as to increase the verification and screening of the timeliness of enterprise information, thereby ensuring the effectiveness of enterprise information and enhancing the accuracy of risk prediction.
[0090] Based on the above embodiments, such as Figure 4As shown, step S104, "For each target enterprise, calculate the similarity between the target enterprise and any enterprise with complete information features based on the supply information among the enterprises in the supply chain network, so as to determine the similar enterprises of the target enterprise," further includes the following steps:
[0091] S401, determine the weighted directed graph corresponding to the supply chain network. The nodes in the graph represent enterprises in the supply chain network, the directional edges in the graph represent supply relationships, and the weight of the edge represents the supply share.
[0092] A directed graph is constructed based on the supply relationships between enterprises in the supply chain network. Each enterprise is treated as a node in the graph. Edges are set between enterprise nodes with supply relationships, with the direction of the supply relationship (e.g., from the supplier to the supplier, or from the supplier to the supplier) serving as the direction of the edges between nodes. This yields the directed graph corresponding to the supply chain network. The proportion of supply between enterprises is used as the weight of the edges between corresponding nodes, thus obtaining the weighted directed graph corresponding to the supply chain network.
[0093] S402, for each node in the weighted directed graph, based on the structure of the weighted directed graph, determine the set of adjacent nodes of the node and the weight set corresponding to the set of adjacent nodes. The weight set is the set composed of the weights of the edges between the node and its corresponding adjacent nodes.
[0094] Based on the connection relationships and directions between nodes in a weighted directed graph, determine the set of adjacent nodes corresponding to any node, and construct the weight set corresponding to the set of adjacent nodes based on the weights between the node and its adjacent nodes.
[0095] S403, based on the set of adjacent nodes and the set of weights, calculate the similarity between the node corresponding to the target enterprise and the node corresponding to any enterprise with complete information features, so as to determine the similar enterprises of the target enterprise.
[0096] For each target company, calculate the similarity between the node corresponding to that target company and the nodes corresponding to all other companies besides all target companies. This can also be understood as determining the similarity between the target company and any other company with complete information features. Identify similar companies to the target company from among the companies with complete information.
[0097] In some embodiments, when calculating the similarity between a target enterprise and any enterprise with complete information based on the set of adjacent nodes and the set of weights, the following similarity calculation is performed:
[0098]
[0099] Where A is the node corresponding to the target enterprise, and B is the node corresponding to the enterprise with complete information features. bothLet B be the set of adjacent nodes of node A. both Let B be the set of adjacent nodes. Let A be the set of adjacent nodes. both The corresponding weight set, where n is the weight set. The number of parameters in the middle, Let B be the set of adjacent nodes. both The corresponding weight set, m is the weight set The number of parameters in J w (A,B) represents the similarity between node A and node B.
[0100] It's not hard to understand, J w The closer the value of (A,B) is to 1, the more similar the two nodes are.
[0101] In this embodiment, when determining the similarity between two nodes, a weight set is introduced simultaneously with the similarity between the sets of adjacent nodes corresponding to the two nodes. The supply ratio between enterprises is also included as a factor in the similarity assessment, allowing for more accurate filtering of similar enterprises to the target enterprise. This application performs correlation assessments in two dimensions: first, determining whether the adjacent nodes corresponding to the two nodes are similar; and second, determining whether the weights between the two nodes and their adjacent nodes are similar. Furthermore, the correlation assessments from different dimensions are integrated into a single correlation coefficient, making it easier to determine the degree of similarity between enterprises.
[0102] Based on the above embodiments, such as Figure 5 As shown, step S402, "For each node in the weighted directed graph, based on the structure of the weighted directed graph, determine the set of adjacent nodes and the weight set corresponding to the adjacent node set, where the weight set is the set of weights of the edges between the node and its corresponding adjacent nodes," further includes the following steps:
[0103] S501. Based on the edge structure and edge direction in the weighted directed graph, determine the set of adjacent nodes of a node. The set of adjacent nodes includes the node's access nodes and outgoing nodes.
[0104] In this embodiment, for each node, a set of adjacent nodes is determined based on whether there are connecting edges between the node and other nodes and the direction of the edges. The set of adjacent nodes includes the node's accessing nodes and its exiting nodes, where the accessing node can be understood as the adjacent node where the beginning of the edge direction is located, and the exiting node can be understood as the adjacent node where the end of the edge direction is located.
[0105] For example, by determining whether there are edges between the node corresponding to enterprise A and other nodes, and whether the endpoint of the edge is located at the node corresponding to enterprise A, the access node of the node corresponding to enterprise A can be determined, thus obtaining A. inVector {c i1 ,c i2 ,c i3 ,…,c ih By determining whether there are edges between the node corresponding to enterprise A and other nodes, and whether the starting point of the edge is located at the node corresponding to enterprise A, the outgoing node of the node corresponding to enterprise A is determined, thus obtaining A. out Vector {c o1 ,c o2 ,c o3 ,…,c oj}. Where c ih and c oj For enterprise nodes, h is A in The number of parameters in the vector, i.e., the number of access nodes; j is A out The number of parameters in the vector, i.e. the number of outgoing nodes.
[0106] This application simultaneously writes the access node and the outgoing node into the adjacent node set. By setting the direction coefficient, A is respectively... in Vector and A out After multiplying the vectors by their corresponding direction coefficients, they are merged to obtain the set of adjacent nodes.
[0107] For example, when obtaining the set of adjacent nodes of a given node, the direction coefficient for connecting to that node is set to 1, and the direction coefficient for connecting out of that node is set to -1. For example, when enterprise node c is in A... in and A out If both exist in A, then in the merged A both There are two parameters, c and -c. It should be noted that the direction coefficient is set to distinguish the direction of edges between nodes, and this application does not limit the specific value of the direction coefficient.
[0108] S502, determine the weight set based on the weight of the edge between the node and the access node, and the weight of the edge between the node and the outgoing node.
[0109] In this embodiment of the application, edge weights are introduced when calculating the similarity between nodes. That is, the supply ratio corresponding to the access node and the supply ratio corresponding to the outgoing node are recorded in the weight set, so that the weight parameters in the weight set correspond one-to-one with the neighboring nodes in the neighboring node set.
[0110] Introducing the direction and weight of edges between nodes when constructing the adjacency node matrix and weight matrix can make fuller use of the relationships between enterprises represented in the supply chain network, determine the similarity between enterprises, and enhance the accuracy of similarity judgment.
[0111] Based on the above embodiments, such as Figure 6As shown, step S105 above, "Based on the enterprise information of similar enterprises, determine the missing information features of the target enterprise to complete the enterprise information of each enterprise in the supply chain network," includes:
[0112] S601, determine the information dimension in which the missing information features of the target enterprise are located.
[0113] In this embodiment of the application, the enterprise information corresponds to multiple information dimensions, and each information dimension includes multiple information features.
[0114] For example, in the application scenario of supply chain networks, information can be divided into three main dimensions: basic information, financial information, and risk information. Basic information can include multiple information features such as the company's main business, industry, operating scale, and nature of the company. Risk information can include multiple information features such as the company's litigation risk, abnormal operations, and industry popularity. Financial information can include multiple information features such as the company's operating profit, cash flow, and asset structure.
[0115] It should be noted that the information dimensions corresponding to enterprise information and the information characteristics contained in each dimension can be adjusted according to the actual application scenario, and this application does not impose any restrictions.
[0116] S602, extract information features in the information dimension from the enterprise information of similar enterprises.
[0117] In this embodiment of the application, information features in the above-mentioned information dimensions are extracted from the enterprise information corresponding to one or more similar enterprises.
[0118] S603, based on preset information completion rules, uses information features extracted from the enterprise information of similar enterprises to determine the missing information features of the target enterprise.
[0119] Rules for information completion should be formulated based on actual needs. In practice, basic information among similar enterprises has certain commonalities, and risk information has a certain degree of transmissibility, while financial information is generally more enterprise-specific. For enterprise information that has both commonalities and transmissibility, it can be predicted and completed based on the enterprise information of similar enterprises.
[0120] The information completion rules can be as follows: determine the missing information features of the target company based on the information features extracted from the most similar companies; or use the similarity of each of the multiple similar companies as weights, and fuse the information features extracted from the multiple similar companies based on the weights to determine the missing information features of the target company.
[0121] In addition, by judging the degree of transmissibility or commonality of the content corresponding to each information feature, different completion weights can be set for different information features in different dimensions. Based on the completion weights, the missing information features of the target enterprise can be determined according to the similar enterprises with the highest similarity or multiple similar enterprises.
[0122] In practice, for two companies that have high similarity in related enterprises (i.e., adjacent nodes in a weighted directed graph) and also have high similarity in weights with related enterprises, their main business and industry characteristics can also be considered to have high similarity. As a feasible implementation method, when determining the missing information features of the target company, if the missing information features are in the basic information dimension, the information features of similar enterprises can be directly used as the information features of the target company. If the missing information features are in the risk information dimension, the risk of similar enterprises can be directly transmitted to predict the risk information features of the target company.
[0123] This application extracts the relationships between various enterprises in the supply chain network. Based on these relationships, it supplements missing information about incomplete enterprises, ensuring the completeness of enterprise information for each entity. This allows for risk prediction based on comprehensive enterprise information, improving the accuracy of predictions.
[0124] Based on the above embodiments, such as Figure 7 As shown, step S106 above, "Based on the supply chain network with complete enterprise information, perform risk prediction on the enterprise to be predicted, and obtain the risk prediction result of the enterprise to be predicted," further includes the following steps:
[0125] S701, based on the supply chain network after the enterprise information is completed, determine the weighted directed graph corresponding to the supply chain network and the feature matrix corresponding to each node in the graph.
[0126] The process of determining the weighted directed graph corresponding to the supply chain network can be found in the description of the above embodiments, and will not be repeated here. The feature matrix corresponding to each node in the graph is a feature matrix generated based on the enterprise information of the enterprise corresponding to each node.
[0127] S702, based on graph convolutional neural networks, performs convolution processing on a weighted directed graph and the feature matrix corresponding to each node in the graph to obtain the risk label of the enterprise to be predicted, so as to obtain the risk prediction result of the enterprise to be predicted.
[0128] As a feasible implementation method, a graph convolutional neural network is pre-constructed and trained to obtain a trained graph convolutional neural network. A weighted directed graph and a feature matrix consisting of the enterprise information corresponding to each node in the graph are input into the graph convolutional neural network. Convolution is performed on the weighted directed graph and the feature matrix corresponding to each node to classify the risk of the enterprise to be predicted, obtaining the risk label of the enterprise to be predicted, thereby determining the risk prediction result of the enterprise to be predicted.
[0129] As another feasible implementation, a prediction model can be constructed using a semi-supervised graph convolutional neural network for risk prediction. For example, a prediction model can be constructed based on a two-layer graph convolutional neural network, whose feedforward network is as follows: Figure 8 As shown, this model uses the features of all nodes to predict the credit risk of the enterprise to be predicted, where W (0) and W (1) These are the weight parameters that need to be trained.
[0130] In some embodiments, a supply chain network graph G = (V, E) is input to the model. This network graph is the weighted directed graph described above. An adjacency matrix A and a degree matrix D are constructed based on the input graph. When constructing the adjacency matrix, the supply share between enterprises can be used as weights. When constructing the degree matrix, the number of edges connecting nodes is used as the degree. A model based on a graph convolutional neural network is used to classify nodes, thereby determining whether there is risk associated with the corresponding node of the enterprise to be predicted.
[0131] In some embodiments, for a model built on a semi-supervised graph convolutional network, the process of classifying nodes using the model includes multiple iterative training processes: after sampling nodes, the feature matrix of each node is obtained; convolution is performed based on the feature matrix to propagate the information of neighboring nodes; then node classification is performed; the loss function of the model is calculated based on the set of labeled nodes; the model is updated according to the loss function; and finally, the iterative training process is repeated based on the updated model.
[0132] In some embodiments, the Adam algorithm is used for iterative optimization. Since graph convolution has high computational performance requirements and a large amount of data operation, the optimizer has a fast operation speed and small memory footprint, which can better meet the requirements.
[0133] In some embodiments, when evaluating a model, the ROC curve may be used to assess the model's training performance due to the class imbalance characteristic of credit risk labels.
[0134] Taking a semi-supervised graph convolutional neural network as an example, the process of convolution processing based on a weighted directed graph and the feature matrix corresponding to each node in the graph is described:
[0135] Construct adjacency and degree matrices based on the weighted directed graph corresponding to the supply chain network, for example, Figure 9 For a supply chain network containing three enterprise nodes, its adjacency matrix and degree matrix are as follows:
[0136]
[0137]
[0138] When using semi-supervised graph convolutional networks for node prediction, a feedforward network is constructed as follows: Figure 8 The two-layer graph convolutional neural network shown is subjected to the following convolution processing:
[0139]
[0140]
[0141] in, A is the adjacency matrix of the supply chain network, D is the degree matrix of the supply chain network, and W... (0) and W (1) These are the weight parameters to be trained, I N Let X be the identity matrix and X be the characteristic matrix. i represents the feature variables in the feature matrix. Z represents the node classification result, i.e., the risk prediction result for each node.
[0142] Based on the node classification results, a cross-entropy loss function is calculated for the set of labeled nodes. The model is then updated according to the loss function. After multiple iterations of training and model updates, a trained model is obtained. Based on this trained model, the nodes corresponding to the enterprise to be predicted are classified to obtain the enterprise's risk label, thereby determining the risk prediction result for that enterprise.
[0143] In summary, this application supplements the missing information features of enterprises in the supply chain network before using a graph convolutional network model for risk prediction. Risk prediction is then based on complete enterprise information, avoiding model overfitting. Graph convolutional networks can fully explore the potential relationship information of "neighbor nodes" and learn the feature aggregation functions of "neighbor" enterprises. This method can effectively solve the problems of incomplete features and inaccurate risk assessment for SMEs. Complex relationships and interdependencies exist among enterprises in the supply chain network. This application comprehensively considers the complex relationships between enterprise nodes, which, compared to existing algorithms, can improve the accuracy and rationality of model predictions to a certain extent.
[0144] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.
[0145] Figure 10This is a block diagram illustrating a deep learning-based risk prediction device according to an exemplary embodiment of this application, such as... Figure 10 As shown, the device 1000 includes: a network construction module 1001, an information acquisition module 1002, a determination module 1003, a similarity calculation module 1004, an information completion module 1005, and a risk prediction module 1006.
[0146] The network construction module 1001 is used to determine the supply chain network in which the enterprise to be predicted is located, and to obtain supply information between enterprises in the supply chain network, including supply relationships and supply proportions.
[0147] Information acquisition module 1002 is used to acquire enterprise information corresponding to each enterprise in the supply chain network;
[0148] The determination module 1003 is used to filter out enterprises with missing information features based on enterprise information, and to identify enterprises with missing information features as target enterprises.
[0149] The similarity calculation module 1004 is used to calculate the similarity between each target enterprise and any enterprise with complete information features based on the supply information between various enterprises in the supply chain network, so as to determine the similar enterprises of the target enterprise.
[0150] The information completion module 1005 is used to determine the missing information features of the target enterprise based on the enterprise information of similar enterprises, so as to complete the enterprise information of each enterprise in the supply chain network.
[0151] The risk prediction module 1006 is used to predict the risk of the enterprise to be predicted based on the supply chain network after the enterprise information is completed, and obtain the risk prediction result of the enterprise to be predicted.
[0152] In this embodiment of the application, the determining module is further configured to: generate an information disclosure vector corresponding to each enterprise based on the disclosure status of enterprise information corresponding to each enterprise in the supply chain network, wherein the parameters in the information disclosure vector are used to characterize the disclosure status of different information features; calculate the information missing rate of the enterprise based on the information disclosure vector; and screen out enterprises with missing information features based on the information missing rate, and determine the enterprises with missing information features as target enterprises.
[0153] In this embodiment, the similarity calculation module is further configured to: determine a weighted directed graph corresponding to the supply chain network, where nodes represent enterprises in the supply chain network, directional edges represent supply relationships, and edge weights represent supply proportions; for each node in the weighted directed graph, determine the set of adjacent nodes and the weight set corresponding to the adjacent node set based on the structure of the weighted directed graph, where the weight set is the set of weights of the edges between the node and its corresponding adjacent nodes; and calculate the similarity between the node corresponding to the target enterprise and the node corresponding to any enterprise with complete information features based on the adjacent node set and the weight set, so as to determine the similar enterprises of the target enterprise.
[0154] In this embodiment of the application, the similarity calculation module is further used to: determine the set of adjacent nodes of a node based on the edge structure and the direction of the edge in the weighted directed graph, wherein the set of adjacent nodes includes the node's access node and the node's outgoing node; and determine the weight set based on the weight of the edge between the node and the access node, and the weight of the edge between the node and the outgoing node.
[0155] In this embodiment of the application, the similarity calculation module is further used to: perform the following similarity calculation based on the set of adjacent nodes and the set of weights:
[0156]
[0157] Where A is the node corresponding to the target enterprise, and B is the node corresponding to the enterprise with complete information features. both Let B be the set of adjacent nodes of node A. both Let B be the set of adjacent nodes. Let A be the set of adjacent nodes. both The corresponding weight set, where n is the weight set. The number of parameters in the middle, Let B be the set of adjacent nodes. both The corresponding weight set, m is the weight set The number of parameters in J w (A,B) represents the similarity between node A and node B.
[0158] In this embodiment of the application, the enterprise information corresponds to multiple information dimensions, and each information dimension includes multiple information features. The information completion module is further used to: determine the information dimension in which the missing information features of the target enterprise are located; extract information features on the information dimension from the enterprise information of similar enterprises; and determine the missing information features of the target enterprise based on preset information completion rules and by using the information features extracted from the enterprise information of similar enterprises.
[0159] In this embodiment, the risk prediction module is further used to: determine the weighted directed graph corresponding to the supply chain network and the feature matrix corresponding to each node in the graph based on the supply chain network after the enterprise information is completed; and perform convolution processing on the weighted directed graph and the feature matrix corresponding to each node in the graph based on the graph convolutional neural network to obtain the risk label of the enterprise to be predicted, so as to obtain the risk prediction result of the enterprise to be predicted.
[0160] In summary, the deep learning-based risk prediction device proposed in this application utilizes the supply chain network of the enterprise to be predicted to extract the relationships between the enterprise and its upstream and downstream enterprises, and even indirectly related enterprises. This relationship information, i.e., supply information, is applied to the risk prediction of the enterprise. Based on the relationships between enterprises, similar enterprises are selected, and the enterprise information of similar enterprises is used to supplement the information of enterprises with missing information features. The enterprise risk is predicted by using the supplemented enterprise information and the node relationships shown by the supply chain network, which improves the accuracy of risk prediction. More accurate prediction results can also be obtained for SMEs with incomplete information disclosure.
[0161] To implement the above embodiments, this application also proposes an electronic device 1100, such as... Figure 11 As shown, the electronic device 1100 may specifically include: a memory 1101, a processor 1102, and a computer program stored on the memory 1101 and executable on the processor 1102. When the processor 1102 executes the program, it implements the steps of the risk prediction method based on deep learning as shown in the above embodiments.
[0162] In summary, this application proposes an electronic device that utilizes the supply chain network of the enterprise to be predicted to extract the relationships between the enterprise and its upstream and downstream enterprises, and even indirectly related enterprises. This relationship information, i.e., supply information, is applied to the risk prediction of the enterprise. Based on the relationships between enterprises, similar enterprises are selected, and the enterprise information of similar enterprises is used to supplement the information of enterprises with missing information features. The supplemented enterprise information and the node relationships shown by the supply chain network are used to predict enterprise risk, thereby improving the accuracy of risk prediction. More accurate prediction results can also be obtained for SMEs with incomplete information disclosure.
[0163] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.
[0164] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.
[0165] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.
Claims
1. A risk prediction method based on deep learning, characterized in that, include: Determine the supply chain network in which the enterprise to be predicted is located, and obtain supply information among the enterprises in the supply chain network, including supply relationships and supply proportions; Obtain enterprise information corresponding to each enterprise in the supply chain network; Based on the enterprise information, enterprises with missing information features are screened out, and these enterprises with missing information features are identified as target enterprises. For each target enterprise, based on the supply information among the enterprises in the supply chain network, the similarity between the target enterprise and any enterprise with complete information features is calculated to determine the similar enterprises of the target enterprise; Based on the enterprise information of the similar enterprises, the missing information features of the target enterprise are determined to complete the enterprise information of each enterprise in the supply chain network; Based on the supply chain network after the enterprise information is completed, risk prediction is performed on the enterprise to be predicted, and the risk prediction result of the enterprise to be predicted is obtained. For each target enterprise, based on the supply information among enterprises in the supply chain network, the similarity between the target enterprise and any enterprise with complete information features is calculated to determine the similar enterprises of the target enterprise, including: A weighted directed graph corresponding to the supply chain network is determined, wherein the nodes in the graph represent enterprises in the supply chain network, the directional edges in the graph represent the supply relationships, and the weight of the edge represents the supply proportion; For each node in the weighted directed graph, based on the structure of the weighted directed graph, determine the set of adjacent nodes of the node and the weight set corresponding to the set of adjacent nodes. The weight set is the set of weights of the edges between the node and its corresponding adjacent nodes. Based on the set of adjacent nodes and the set of weights, the similarity between the node corresponding to the target enterprise and the node corresponding to any enterprise with complete information features is calculated to determine the similar enterprises of the target enterprise. The step of calculating the similarity between the node corresponding to the target enterprise and the node corresponding to any enterprise with complete information features, based on the neighboring node set and the weight set, includes: Based on the set of adjacent nodes and the set of weights, the following similarity calculation is performed: Wherein, A is the node corresponding to the target enterprise, and B is the node corresponding to the enterprise with complete information features. Let A be the set of adjacent nodes. Let B be the set of adjacent nodes. For the set of adjacent nodes The corresponding weight set, where n is the weight set. The number of parameters in the middle, the For the set of adjacent nodes The corresponding weight set, where m is the weight set. The number of parameters in the middle, the This represents the similarity between node A and node B.
2. The method according to claim 1, characterized in that, The step of filtering out enterprises with missing information features based on the enterprise information, and identifying the enterprises with missing information features as target enterprises, includes: Based on the disclosure of enterprise information corresponding to each enterprise in the supply chain network, an information disclosure vector corresponding to the enterprise is generated, and the parameters in the information disclosure vector are used to characterize the disclosure status of different information features. Calculate the information missing rate of the enterprise based on the information disclosure vector; Based on the information missing rate, enterprises with missing information features are screened out, and these enterprises with missing information features are identified as the target enterprises.
3. The method according to claim 1, characterized in that, For each node in the weighted directed graph, determining the set of adjacent nodes and the weight set corresponding to the set of adjacent nodes, based on the structure of the weighted directed graph, includes: Based on the edge structure and the direction of the edge in the weighted directed graph, the set of adjacent nodes of the node is determined, and the set of adjacent nodes includes the access node and the outgoing node of the node. The weight set is determined based on the weight of the edge between the node and the access node, and the weight of the edge between the node and the outgoing node.
4. The method according to claim 1, characterized in that, The enterprise information corresponds to multiple information dimensions, and each information dimension includes multiple information features. The step of determining the missing information features of the target enterprise based on the enterprise information of the similar enterprises, in order to complete the enterprise information of each enterprise in the supply chain network, includes: Determine the information dimension in which the missing information features of the target enterprise reside; Extract information features from the enterprise information of the similar enterprises in the information dimension; Based on preset information completion rules, the missing information features of the target enterprise are determined by using the information features extracted from the enterprise information of the similar enterprises.
5. The method according to claim 1, characterized in that, The supply chain network based on the completed enterprise information is used to perform risk prediction on the enterprise to be predicted, and the risk prediction result of the enterprise to be predicted is obtained, including: Based on the supply chain network after the enterprise information is completed, determine the weighted directed graph corresponding to the supply chain network and the feature matrix corresponding to each node in the graph; Based on a graph convolutional neural network, the weighted directed graph and the feature matrix corresponding to each node in the graph are convolved to obtain the risk label of the enterprise to be predicted, so as to obtain the risk prediction result of the enterprise to be predicted.
6. A risk prediction device based on deep learning, characterized in that, include: The network construction module is used to determine the supply chain network in which the enterprise to be predicted is located, and to obtain supply information between enterprises in the supply chain network, including supply relationships and supply proportions. The information acquisition module is used to acquire enterprise information corresponding to each enterprise in the supply chain network; The determination module is used to filter out enterprises with missing information features based on the enterprise information, and determine the enterprises with missing information features as target enterprises; The similarity calculation module is used to calculate the similarity between each target enterprise and any enterprise with complete information features, based on the supply information between the enterprises in the supply chain network, so as to determine the similar enterprises of the target enterprise. The information completion module is used to determine the missing information features of the target enterprise based on the enterprise information of the similar enterprises, so as to complete the enterprise information of each enterprise in the supply chain network; The risk prediction module is used to perform risk prediction on the enterprise to be predicted based on the supply chain network after the enterprise information is completed, and to obtain the risk prediction result of the enterprise to be predicted. For each target enterprise, based on the supply information among enterprises in the supply chain network, the similarity between the target enterprise and any enterprise with complete information features is calculated to determine the similar enterprises of the target enterprise, including: A weighted directed graph corresponding to the supply chain network is determined, wherein the nodes in the graph represent enterprises in the supply chain network, the directional edges in the graph represent the supply relationships, and the weight of the edge represents the supply proportion; For each node in the weighted directed graph, based on the structure of the weighted directed graph, determine the set of adjacent nodes of the node and the weight set corresponding to the set of adjacent nodes. The weight set is the set of weights of the edges between the node and its corresponding adjacent nodes. Based on the set of adjacent nodes and the set of weights, the similarity between the node corresponding to the target enterprise and the node corresponding to any enterprise with complete information features is calculated to determine the similar enterprises of the target enterprise. The step of calculating the similarity between the node corresponding to the target enterprise and the node corresponding to any enterprise with complete information features, based on the neighboring node set and the weight set, includes: Based on the set of adjacent nodes and the set of weights, the following similarity calculation is performed: Wherein, A is the node corresponding to the target enterprise, and B is the node corresponding to the enterprise with complete information features. Let A be the set of adjacent nodes. Let B be the set of adjacent nodes. For the set of adjacent nodes The corresponding weight set, where n is the weight set. The number of parameters in the middle, the For the set of adjacent nodes The corresponding weight set, where m is the weight set. The number of parameters in the middle, the This represents the similarity between node A and node B.
7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the deep learning-based risk prediction method as described in any one of claims 1-5.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the deep learning-based risk prediction method as described in any one of claims 1-5.