Credit product personalized recommendation method and system based on enterprise credit assessment

Abstract

The present application relates to product personalized recommendation technical field, particularly to credit product personalized recommendation method and system based on enterprise credit evaluation, wherein the method comprises: obtaining the basic data of the enterprise to be evaluated and the known enterprise and the known social relationship data sent by the client to the server; the server extracts the features of the basic data to obtain the basic feature data; constructing the heterogeneous knowledge graph based on the basic feature data of the enterprise to be evaluated and the known enterprise and the known social relationship, then inputting the heterogeneous knowledge graph into the trained social relationship reasoning model to obtain the comprehensive social relationship data of the enterprise to be evaluated; the server inputs the basic feature data and the comprehensive social relationship data of the enterprise to be evaluated into the trained analysis network to obtain the credit level evaluation result of the enterprise to be evaluated; determining the credit product matched with the enterprise, and feeding back the recommended credit product to the client. The present application can recommend personalized credit product for enterprises.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of product personalized recommendation, in particular to a credit product personalized recommendation method and system based on enterprise credit evaluation. BACKGROUND

[0002] Before recommending a credit product to an enterprise user group, a bank needs to understand the attribute characteristics of each enterprise, which can be achieved through information collection or questionnaire survey.

[0003] The inventor found that in the existing credit product recommendation, the questionnaire survey may be incomplete or incorrect due to the negligence of the enterprise user, and some enterprise users do not cooperate, resulting in inaccurate classification of the type of enterprise user, and thus the credit product cannot be accurately recommended. For users who are not interested in the product, invalid advertisement information will cause interference, thereby reducing the user's experience of the product. For users with potential purchase intention, due to the inaccurate classification of advertisements, users need to reorganize and screen all products, which not only requires a lot of effort, but also requires strong professional identification ability; thereby reducing the matching degree between the products provided by the service platform and the customer demand. SUMMARY

[0004] In order to solve the problems of the prior art, the present application provides a credit product personalized recommendation method and system based on enterprise credit evaluation; the method not only takes the basic information of the enterprise itself as an important index for evaluating the credit level, but also considers important external factors such as industry and market environment, industry situation, etc. At the same time, the present application also comprehensively considers the social relationship of the enterprise and its shareholders in the credit activity, so as to realize the comprehensive and accurate evaluation of the credit level of the enterprise, reflect the financial risk, and then recommend the personalized credit product that is more suitable for the development stage and financing ability of the enterprise, thereby improving the pertinence and effectiveness of financial services.

[0005] On the one hand, a credit product personalized recommendation method based on enterprise credit evaluation is provided, which is applied to a server side, comprising:

[0006] Obtaining the basic data and known social relationship data of the enterprise to be evaluated and known enterprises sent by the client to the server, the basic data comprising internal data and external data; the server extracts features from the basic data to obtain basic feature data;

[0007] Based on the basic feature data of the enterprise to be evaluated and known enterprises and the known social relationship, a heterogeneous knowledge graph is constructed, and then the heterogeneous knowledge graph is input into the trained social relationship reasoning model to obtain the comprehensive social relationship data of the enterprise to be evaluated;

[0008] The server end inputs the basic feature data and the comprehensive social relationship data of the enterprise to be evaluated into the trained analysis network to obtain the credit rating evaluation result of the enterprise to be evaluated; and further based on the evaluation result, determines the credit product matched with the enterprise, and feeds back the recommended credit product to the client end.

[0009] In another aspect, a credit product personalized recommendation system based on enterprise credit evaluation is provided, comprising: a client end and a server end.

[0010] The server end obtains the basic data and known social relationship data of the enterprise to be evaluated and known enterprises sent by the client end to the server end, wherein the basic data comprises internal data and external data; the server end performs feature extraction on the basic data to obtain basic feature data;

[0011] The server end constructs a heterogeneous knowledge graph based on the basic feature data of the enterprise to be evaluated and known enterprises and known social relationships, and then inputs the heterogeneous knowledge graph into a trained social relationship reasoning model to obtain comprehensive social relationship data of the enterprise to be evaluated;

[0012] The server end inputs the basic feature data and the comprehensive social relationship data of the enterprise to be evaluated into the trained analysis network to obtain the credit rating evaluation result of the enterprise to be evaluated; and further based on the evaluation result, determines the credit product matched with the enterprise, and feeds back the recommended credit product to the client end.

[0013] The above technical solution has the following advantages or beneficial effects:

[0014] (1) The present application provides a credit product personalized recommendation method and system based on enterprise credit evaluation. This method and system not only take the basic information of the enterprise itself as an important indicator for evaluating its credit rating, but also consider important external factors such as industry and market environment, industry situation, etc. At the same time, the present application also considers the social relationships of the enterprise and its shareholders in credit activities, and through comprehensive and accurate evaluation of the credit rating of the enterprise, promotes the personalized and accurate matching of credit products and enterprise needs.

[0015] (2) The present application provides a multi-subject relationship inference method suitable for credit activities, which can deeply infer the complex relationship between multiple subjects such as enterprises, shareholders, etc. This method combines artificial intelligence technology, can effectively mine the potential social relationships of the enterprise to be evaluated and its shareholders, and reveal key information that is difficult to capture in traditional data analysis. This multi-dimensional relationship inference helps to improve the accuracy of enterprise credit evaluation, enables credit institutions to obtain more comprehensive reference in the decision-making process, and promotes the personalized and accurate matching of credit products and enterprise needs. BRIEF DESCRIPTION OF DRAWINGS

[0016] The accompanying drawings, which form a part of this specification, are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification. The embodiments of the application, and their

[0017] Figure 1 The method flowchart of Example One. DETAILED DESCRIPTION

[0018] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0019] The known enterprise in the present application refers to an enterprise that has been registered and exists in the client system. Credit evaluation is crucial for an enterprise, which not only determines whether the enterprise can obtain key resources such as financing and credit, but also directly affects supply chain cooperation, subsidy application, and market trust, etc. In the prior art credit evaluation method, attention is usually paid to the internal information of the enterprise itself, while the influence of external factors is ignored, and the accuracy of credit evaluation is still limited. In addition, the prior art credit evaluation method ignores two types of factors that are very important for credit evaluation of an enterprise: one type is the external factors of the enterprise such as industry and market environment, industry situation, etc.; the other type is the social relationship of the credit activity related to the enterprise and its shareholders (the social relationship between the enterprise and its shareholders and other enterprises or shareholders of other enterprises). For an enterprise whose information is not completely open, the existing method cannot comprehensively and objectively reflect the true credit status of the enterprise. Therefore, relying only on the internal information of the enterprise while ignoring the influence of the related social relationship and external information, it is often difficult to accurately evaluate the credit level of the enterprise, which not only weakens the accuracy of personalized recommendation, but also affects the financing efficiency and rationality of resource allocation.

[0020] Example One

[0021] The present embodiment provides a credit product personalized recommendation method based on enterprise credit evaluation;

[0022] As shown in Figure 1 The credit product personalized recommendation method based on enterprise credit evaluation is applied to a server side, and comprises the following steps:

[0023] S101: Obtain the basic data of a known enterprise and an enterprise to be evaluated sent by a client to a server, and known social relationship data, wherein the basic data comprises internal data and external data; the server extracts features from the basic data to obtain basic feature data;

[0024] S102: Construct a heterogeneous knowledge graph based on the basic feature data of the enterprise to be evaluated and the known enterprise and the known social relationship, and then input the heterogeneous knowledge graph into the trained social relationship reasoning model to obtain comprehensive social relationship data of the enterprise to be evaluated;

[0025] S103: The server inputs the basic feature data and the comprehensive social relationship data of the enterprise to be evaluated into the trained analysis network to obtain the credit rating evaluation result of the enterprise to be evaluated; further based on the evaluation result, determine the credit product matched with the enterprise, and feed back the recommended credit product to the client.

[0026] Further, the internal data includes enterprise shareholder structure, professional personnel configuration, number of authorized patents, enterprise cash flow, enterprise asset-liability ratio, and enterprise innovation capability, etc.

[0027] Further, the external data includes industry competition pattern and industry situation, etc.

[0028] Further, the social relationship data includes investment, shareholding, senior management, holding, relatives, common address, telephone sharing, job intersection, and common legal person, etc.

[0029] Illustratively, under the premise of ensuring legal and compliant data processing, clear authorization, and controllable privacy, internal data such as enterprise shareholder structure, professional personnel configuration, number of authorized patents, enterprise cash flow, asset-liability ratio, and innovation capability are collected for evaluating enterprise credit rating by using artificial intelligence technologies such as intelligent crawling and data fusion; wherein the innovation capability can be reflected by professional personnel configuration, number of authorized patents, etc.

[0030] Illustratively, under the premise of legal and compliant authorization and controllable privacy, external data such as industry competition pattern and industry situation are collected for evaluating enterprise credit rating by using artificial intelligence technologies such as intelligent crawling and data fusion.

[0031] The industry competition pattern can be used for evaluation by using index data that can reflect the competitiveness of the industry. Under the premise of legal compliance, intelligent crawling and other artificial intelligence technologies can be used to automatically extract key information data such as the number of patents, patent types, and technical fields of related inventions in the industry from intellectual property websites to reflect the industry competition pattern.

[0032] The industry situation can be obtained from news websites by using intelligent crawling and other artificial intelligence technologies under the premise of legal compliance. Then, with the help of a large language model, relevant information is summarized to reflect the industry situation.

[0033] The acquisition of the above-mentioned external data of enterprises is very important. The credit rating of a merchant is not only affected by its own operating performance, but also closely related to the confidence and expectations of the external environment. Even if the financial data of an enterprise is stable and the operating condition is good, the sudden change of external factors may quickly change the market's evaluation of its credit value. Therefore, the present application focuses on important external factors such as industry competition pattern and industry situation that affect enterprise credit evaluation. In addition, the present application also considers the social relations of the enterprise and its shareholders.

[0034] Further, the server extracts features from the basic data to obtain basic feature data, including:

[0035] The structured data in the internal data is standardized to obtain the first type of feature data;

[0036] The unstructured data in the internal data is subjected to semantic vector extraction to obtain the second type of feature data;

[0037] The external data is subjected to semantic vector extraction to obtain the third type of feature data;

[0038] The first, second and third types of feature data are fused to obtain the basic feature data.

[0039] It should be understood that the structured data (such as cash flow, asset-liability ratio, number of authorized patents, and other financial and operating indicators) in the internal data is subjected to standardization and feature engineering processing, and at the same time, the OpenAI Embeddings technology in the field of natural language processing (NLP) is used to extract semantic vectors from the unstructured data (such as shareholder structure, professional personnel allocation, innovation ability, etc.) in the internal data and the external data. Finally, all the features are fused to construct a unified high-dimensional feature vector to form the basic feature data of the enterprise.

[0040] The beneficial effects of the above technical solution are: the method fuses structured and unstructured data, processes through standardization and semantic representation technologies such as OpenAI Embeddings, and automatically constructs a unified high-dimensional feature vector. Compared with the traditional method relying on manual feature design, this process greatly reduces the dependence on manual intervention and expert experience, improves the efficiency and consistency of feature extraction, significantly improves the comprehensiveness and depth of information expression, reduces the dependence on manual feature construction, improves the efficiency and consistency of feature extraction, and provides high-quality, automated, and scalable input basis for subsequent analysis and decision-making.

[0041] Further, S102: constructing a heterogeneous knowledge graph based on the basic feature data of the to-be-evaluated enterprise and the known enterprise and the known social relationship, and then inputting the heterogeneous knowledge graph into the trained social relationship reasoning model to obtain comprehensive social relationship data of the to-be-evaluated enterprise, wherein the comprehensive social relationship data includes: known social relationship data and potential social relationship data obtained based on the social relationship reasoning model.

[0042] Further, the trained social relationship reasoning model is implemented by using a relational graph convolutional network (R-GCN). The relational graph convolutional network takes the heterogeneous knowledge graph established based on the basic data of the enterprise and the known enterprise and the known social relationship data as input, and through continuous iterative learning, infers the social relationship between the enterprise and its shareholders.

[0043] The relational graph convolutional network (in the training process, the graph convolution operation of neighbor sampling and relationship distinguishing is adopted, the node representation with semantic expression ability is constructed by using the adjacent node features of different relationship types, and the embedding representation of each node in the graph is further extracted and updated layer by layer; then, in combination with a relationship scoring function (such as DistMult), the given entity pair and relationship type triple are scored to determine whether they are established in the graph. Through positive and negative sample training, the model can infer the social relationship between unknown enterprises and their shareholders, and finally a social relationship reasoning model for inferring the social relationship between enterprises and their shareholders is trained. The relational graph convolutional network extracts attribute features of each node in the graph, wherein: is a node set, each node has an initial feature vector , i.e. , (the superscript 0 represents the 0th layer of the network), is an edge set, is a relationship type set, .

[0044] In the training process of the relational graph convolutional network, a subgraph is selected by node sampling, and the associated nodes of each node under various relationships are obtained by neighbor sampling. For each relationship type , the node sampling neighbor set is:

[0045] ;

[0046] wherein, represents the edge node and the node Previously had relational type , Represents a node Correspondence The set of neighboring nodes;

[0047] Define independent definitions for each relation type. 3D real weight matrix ,Right now , used in the The model performs neighbor aggregation and relationship differentiation at each layer, thereby enhancing its ability to express multiple semantics. During graph convolution, the model comprehensively considers the node's own features and the structural information of its multiple relational neighbors, updating the node embedding representation layer by layer.

[0048] ;

[0049] in, Represents a set of relations All relationships The summation operation, Indicates a node Correspondence The set of neighboring nodes All neighboring nodes The summation operation, Indicates the first Layer nodes Embedded representation, Indicates the first Layer nodes neighboring nodes In the Embedded representation in layers, It is a node Its own information weight matrix, Normalization coefficient (usually) or its square root). It is a nonlinear function (such as ReLU); the final node embedding is:

[0050] ;

[0051] in, Indicates the first Layer nodes The final embedding representation, This represents the total number of layers in the graph convolution;

[0052] Furthermore, for any triple to be predicted Using relational scoring functions Estimate its probability of existence. For example, using the DistMult model, the scoring function is:

[0053] ;

[0054] wherein, denotes the transpose of and denote the nodes and in the layer, respectively, and denotes the final embedding representation, denotes transforming the vector representation of the relation type into a diagonal matrix.

[0055] It should be understood that each relation (such as "investment", "holding", "partnership") has a dedicated , which represents the semantic weight of the relation in the embedding space;

[0056] Further, the score is mapped to the existence probability by the Sigmoid function:

[0057] ;

[0058] wherein is the Sigmoid activation function, and the parameters are constantly optimized by forward propagation and back propagation, so that the prediction result gradually approaches the true label ;

[0059] Finally, the social relationship reasoning model for inferring the social relationship between enterprises and their shareholders is trained.

[0060] Further, the trained social relationship reasoning model, the training process includes:

[0061] A first data set is constructed, which includes: enterprise basic feature data with complete known social relationship data; known complete social relationship data, including: existence of investment relationship, existence of holding relationship, existence of holding relationship, existence of kinship, existence of common residence, existence of common telephone, existence of intersection of positions, and existence of common legal person, etc.; the complete social relationship data further includes: all real social relationship information between enterprise shareholders.

[0062] Based on the first data set, a heterogeneous knowledge graph is constructed, and all real existing relationship triples between enterprises are extracted as a positive sample set ; at the same time, a negative sample set is generated by negative sampling method (such as replacing entity or relation), which is used to construct a supervised learning training set;

[0063] The heterogeneous knowledge graph is input into the social relationship reasoning model, and the model is trained based on positive and negative samples;

[0064] During the training process, the model parameters are continuously optimized by maximizing the positive sample score and minimizing the negative sample score.

[0065] When the loss function value no longer decreases, or the number of training iterations exceeds the preset upper limit, the training is stopped, and the trained social relationship reasoning model is obtained.

[0066] Further, the loss function of the social relationship reasoning model can be a binary classification cross-entropy loss function:

[0067] ;

[0068] Wherein, represents the summation operation of the relationship triplets of all nodes and node in the positive sample set , represents the summation operation of the relationship triplets of all nodes and node in the negative sample set , represents the scoring function of node and node relationship , represents the scoring function of node and node relationship , is an activation function. Further, the construction process of the first data set comprises: data cleaning and feature extraction are performed on the sample enterprise basic data to obtain basic feature data; the data cleaning comprises: missing value processing, abnormal value detection and standardization processing.

[0069] Further, the construction of the heterogeneous knowledge graph based on the first data set comprises:

[0070] Further, the nodes of the heterogeneous knowledge graph are enterprises and shareholders.

[0071] The edges of the heterogeneous knowledge graph are the association relationships between the enterprises, between the enterprises and the shareholders, and between the shareholders; the association relationships include investment, shareholding, senior management, holding, kinship, common address, telephone sharing, job intersection, and common legal person.

[0072] The edges of the heterogeneous knowledge graph are the association relationships between the enterprises, between the enterprises and the shareholders, and between the shareholders; the association relationships include investment, shareholding, senior management, holding, kinship, common address, telephone sharing, job intersection, and common legal person.

[0073] ​The beneficial effects of the above technical solution are: the social relationship reasoning model can use the basic data of the enterprise to infer the relevant social relationships of the enterprise and its shareholders, which helps to conduct a more comprehensive and accurate assessment of the enterprise's credit rating, thereby helping to match credit products with the personalized and accurate needs of the enterprise.

[0074] Further, S103: The server inputs the basic characteristic data and comprehensive social relationship data of the enterprise to be evaluated into the trained analysis network to obtain the credit rating assessment result of the enterprise to be evaluated; the specific training process of the trained analysis network includes:

[0075] Construct a second dataset, which includes: basic characteristic data and complete social relationship data of known enterprises for which credit rating results have been obtained; credit rating results, including: low risk level (1-3 points), medium risk level (4-6 points) and high risk level (7-10 points).

[0076] The second dataset is input into the analysis network to train it, resulting in the trained analysis network.

[0077] It should be understood that the analysis network is implemented using a fully connected neural network, taking basic feature data and social relationship data as input, and outputting the credit rating prediction results of enterprises.

[0078] When the analysis network is used to perform credit rating assessment training, the loss function is calculated through forward propagation and the loss is minimized before backpropagation to continuously optimize the model parameters.

[0079] Furthermore, the specific expression for the "single-sample loss" function of the analysis network is as follows:

[0080] ;

[0081] Among them, superscript Representing the There are 1 sample, and its true category is 1 , for the The predicted probability of the "true class" of each sample is ,parameter This is a focus factor used to adjust the model's attention to difficult-to-classify samples. The class weights are used to further mitigate the impact of uneven class distribution on training performance. Finally, the network's loss function is analyzed as follows:

[0082] ;

[0083] in This represents the total number of samples.

[0084] Compared with the traditional method, the present application does not need to rely on the comprehensive collection and artificial analysis of complex information, significantly reduces the artificial intervention and human input, and reduces the operation cost. At the same time, through the automatic data processing and model reasoning mechanism, the evaluation of the enterprise credit level can be quickly completed, the evaluation efficiency and processing speed are greatly improved, and efficient and reliable support is provided for financing decision.

[0085] Further, the S103 comprises:

[0086] if the credit level evaluation result is low risk, recommending a low-risk credit product;

[0087] if the credit level evaluation result is medium risk, recommending a medium-risk credit product;

[0088] if the credit level evaluation result is high risk, recommending a high-risk credit product.

[0089] Further, the low-risk credit product refers to a credit product provided by a financial institution for enterprise clients with higher credit level and stronger repayment ability, having the characteristics of low default probability and small risk exposure. The medium-risk credit product and the high-risk credit product correspond to clients with medium or low credit level respectively, having corresponding risk control strategies and product characteristics.

[0090] The beneficial effects of the above technical solution are: based on the enterprise credit evaluation result, the present application automatically matches and recommends the credit product with corresponding risk level, avoids the traditional process of relying on artificial review and complex information analysis, significantly reduces the artificial intervention and operation cost, improves the efficiency and accuracy of individualized recommendation of credit products, enhances the intelligent level of credit decision, and greatly improves the evaluation efficiency and processing speed.

[0091] It should be understood that the "credit product" described in the present application refers to the loan service provided by a financial institution according to the business needs of an enterprise, mainly including working capital loans, fixed asset loans, project financing, trade financing, bill discounting, etc., and is widely used in daily operation, fixed asset purchase, import and export trade, and capital turnover of enterprises.

[0092] Embodiment two

[0093] The present embodiment provides a credit product individualized recommendation system based on enterprise credit evaluation, comprising: a client and a server;

[0094] The credit product individualized recommendation system based on enterprise credit evaluation comprises: a client and a server;

[0095] The server end obtains the basic data and known social relationship data of the to-be-evaluated enterprise and the known enterprise sent by the client to the server end, wherein the basic data comprises internal data and external data; the server end performs feature extraction on the basic data to obtain basic feature data;

[0096] The server end constructs a heterogeneous knowledge graph based on the basic feature data of the to-be-evaluated enterprise and the known enterprise and the known social relationship, and then inputs the heterogeneous knowledge graph into the trained social relationship reasoning model to obtain comprehensive social relationship data of the to-be-evaluated enterprise;

[0097] The server end inputs the basic feature data and the comprehensive social relationship data of the to-be-evaluated enterprise into the trained analysis network to obtain a credit rating evaluation result of the to-be-evaluated enterprise; and further determines a credit product matched with the enterprise based on the evaluation result, and feeds back the recommended credit product to the client.

[0098] The above merely describes preferred embodiments of the present application but is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A credit product personalized recommendation method based on enterprise credit assessment, characterized in that, Applied to the server side, comprising: Obtaining the basic data of the enterprise to be evaluated and the known enterprise and the known social relationship data sent by the client to the server side, wherein the basic data comprises internal data and external data; the server side extracts features from the basic data to obtain basic feature data; under the premise of legality and compliance, the intelligent crawling and data fusion artificial intelligence technology is used to collect enterprise external data for evaluating the credit level of the enterprise; Based on the basic feature data of the enterprise to be evaluated and the known enterprise and the known social relationship, a heterogeneous knowledge graph is constructed, and then the heterogeneous knowledge graph is input into the trained social relationship reasoning model to obtain comprehensive social relationship data of the enterprise to be evaluated, wherein the comprehensive social relationship data comprises known social relationship data and potential social relationship data obtained based on the social relationship reasoning model; the trained social relationship reasoning model is realized by using a relationship graph convolution network; The server side inputs the basic feature data and the comprehensive social relationship data of the enterprise to be evaluated into the trained analysis network to obtain the credit level evaluation result of the enterprise to be evaluated; further based on the evaluation result, a credit product matched with the enterprise is determined, and the recommended credit product is fed back to the client; Wherein, the training process of the trained social relationship reasoning model comprises: constructing a first data set, wherein the first data set comprises enterprise basic feature data of known complete social relationship data; the known complete social relationship data comprises investment relationship, holding relationship, holding relationship, kinship, common residence, common telephone, job intersection and common legal person; the complete social relationship data further comprises all real social relationship information between enterprise shareholders; Based on the first data set, a heterogeneous knowledge graph is constructed, and all real existing enterprise relationship triples are extracted as a positive sample set ; at the same time, a negative sample set is generated by negative sampling , which is used to construct a supervised learning training set; The heterogeneous knowledge graph is input into the social relationship reasoning model, and the model is trained based on positive and negative samples; in the training process, the model parameters are continuously optimized by maximizing the positive sample score and minimizing the negative sample score; when the loss function value no longer decreases or the training iteration number exceeds the preset upper limit, the training is stopped, and the trained social relationship reasoning model is obtained; The server side extracts features from the basic data to obtain basic feature data, comprising: Standardizing the structured data in the internal data to obtain first-type feature data; Extracting semantic vectors from the unstructured data in the internal data to obtain second-type feature data; Extracting semantic vectors from the external data to obtain third-type feature data; Fusing the first, second and third-type feature data to obtain the basic feature data; The relationship graph convolution network selects a subgraph through node sampling in a training process, acquires associated nodes of each node under each type of relationship through neighbor sampling, and acquires a neighbor set of each node under each type of relationship , and samples the neighbor set ； wherein represents an edge node and a node having a relationship type , represents a node corresponding relationship neighbor node set; independent weight matrix is defined for each relation type respectively weight matrix , , neighbor aggregation and relation distinguishing are implemented at the first layer, so as to enhance the expression ability of the network to multiple semantics; in the process of graph convolution, the network comprehensively considers the structural information of the node itself and its multi-relation neighbors, and updates the node embedding representation layer by layer: ； wherein, denotes a summation operation over all relations in the relation set , denotes a summation operation over all neighbor nodes of the corresponding relation , denotes the embedding representation of node in the -th layer, denotes the embedding representation of node in the -th layer, in the -th layer, is the information weight matrix of node itself, is a normalization coefficient, is a non-linear function; and the final node embedding is:​​​ ； in, Indicates the first Layer nodes The final embedding representation, This represents the total number of layers in the graph convolution; For any triple to be predicted using a relation scoring function estimate its probability of existence; with the DistMult model, the scoring function is: ； wherein, denotes the transpose of and denote the nodes and in the layer, respectively, denotes the final embedding representation, denotes transforming the vector representation of the relation type into a diagonal matrix; Map the score to the existence probability through the Sigmoid function: wherein is a sigmoid activation function, the parameters are constantly optimized by forward propagation and backpropagation, and the prediction result gradually approaches the true label ; Finally, the social relationship reasoning model for inferring the social relationship between the enterprise and its shareholders is obtained.

2. The method of claim 1, wherein the credit product is recommended based on the credit evaluation of the enterprise. The loss function of the social relationship reasoning model can be a binary classification cross-entropy loss function: ； wherein, denotes a summation operation over all relation triples of nodes and nodes in the positive sample set , denotes a summation operation over all relation triples of nodes and nodes in the negative sample set , denotes a summation operation over all relation triples of nodes and nodes , denotes a scoring function for a relation between node and node , denotes a scoring function for a relation between node and node is an activation function.

3. The method of claim 1, wherein the credit product is recommended based on the credit evaluation of the enterprise. The construction of the heterogeneous knowledge graph based on the first data set comprises: The nodes of the heterogeneous knowledge graph are enterprises and shareholders; The edges of the heterogeneous knowledge graph are the association relationships between the enterprises, between the enterprises and the shareholders, and between the shareholders; the association relationships include investment, shareholding, senior management, holding, kinship, co-address, shared phone, job intersection, and common legal person.

4. The personalized recommendation method for credit products based on enterprise credit assessment as described in claim 1, characterized in that, The server inputs the basic feature data and the comprehensive social relationship data of the enterprise to be evaluated into the trained analysis network to obtain a credit rating evaluation result of the enterprise to be evaluated. The specific training process of the trained analysis network includes: A second data set is constructed, which includes the basic feature data and the complete social relationship data of the known enterprises that have obtained credit rating evaluation results; the credit rating evaluation results include low-risk level, medium-risk level, and high-risk level. The second data set is input into the analysis network to train the analysis network to obtain the trained analysis network.

5. The method of claim 4, wherein the credit product is recommended to the enterprise based on the credit evaluation. The specific expression of the "single sample loss function" of the analysis network is: ​ ； in, Representing the There are 1 sample, and its true category is 1 , for the The predicted probability of the "true class" of each sample is ,parameter This is a focus factor used to adjust the model's attention to difficult-to-classify samples. The class weight coefficients are used to further mitigate the impact of uneven class distribution on training performance; finally, the loss function of the network is analyzed as follows: ； wherein N is the total number of samples.

6. A credit product personalized recommendation system based on enterprise credit assessment, characterized in that, The specific expression of the "single sample loss function" of the analysis network is: The client and the server; The server obtains the basic data and the known social relationship data of the enterprise to be evaluated and the known enterprises sent by the client to the server; the basic data includes internal data and external data; the server extracts features from the basic data to obtain basic feature data; under the premise of legality and compliance, the server collects enterprise external data for evaluating the credit rating of the enterprise by using intelligent crawling and data fusion artificial intelligence technology; The server constructs a heterogeneous knowledge graph based on the basic feature data and the known social relationship data of the enterprise to be evaluated and the known enterprises, and then inputs the heterogeneous knowledge graph into the trained social relationship reasoning model to obtain comprehensive social relationship data of the enterprise to be evaluated, wherein the comprehensive social relationship data includes the known social relationship data and the potential social relationship data obtained based on the social relationship reasoning model; the trained social relationship reasoning model is implemented by using a relationship graph convolution network; The server inputs the basic feature data and the comprehensive social relationship data of the enterprise to be evaluated into the trained analysis network to obtain a credit rating evaluation result of the enterprise to be evaluated; further based on the evaluation result, a credit product matched with the enterprise is determined, and the recommended credit product is fed back to the client; The training process of the trained social relationship reasoning model includes: constructing a first data set, which includes the basic feature data of the enterprises with known complete social relationship data; the known complete social relationship data includes investment relationship, shareholding relationship, holding relationship, kinship, co-address, shared phone, job intersection, and common legal person; the complete social relationship data further includes all real social relationship information between the shareholders of the enterprise; Based on the first data set, a heterogeneous knowledge graph is constructed, and all real existing enterprise relationship triples are extracted as a positive sample set ; at the same time, a negative sample set is generated by negative sampling , which is used to construct a supervised learning training set; The heterogeneous knowledge graph is input into the social relationship reasoning model, and the model is trained based on positive and negative samples; during the training process, the model parameters are constantly optimized by maximizing the positive sample score and minimizing the negative sample score; when the loss function value no longer decreases or the training iteration number exceeds a preset upper limit, the training is stopped, and the trained social relationship reasoning model is obtained; The server extracts features from the basic data to obtain basic feature data, including: standardizing the structured data in the internal data to obtain first-type feature data; extracting semantic vectors from the unstructured data in the internal data to obtain second-type feature data; extracting semantic vectors from the external data to obtain third-type feature data; fusing the first-type, second-type and third-type feature data to obtain the basic feature data; The relationship graph convolution network selects a subgraph through node sampling in a training process, acquires associated nodes of each node under each type of relationship through neighbor sampling, and samples a neighbor set for each type of relationship , for a node ​ ； wherein represents an edge node and a node having a relationship type , represents a node corresponding relationship neighbor node set; independent weight matrix is defined for each relation type respectively weight matrix , , for neighbor aggregation and relation distinguishing in the first layer, so as to enhance the expression ability of the network to multiple semantics; in the process of graph convolution, the network comprehensively considers the structural information of the node itself and its multi-relation neighbors, and updates the node embedding representation layer by layer: ； wherein, denotes a summation operation over all relations in the relation set , denotes a summation operation over all neighbor nodes of the node corresponding relation in the neighbor node set , denotes the embedding representation of the node in the -th layer, denotes the embedding representation of the neighbor node of the node in the -th layer, is the information weight matrix of the node itself, is a normalization coefficient, is a non-linear function; and the final node embedding is:​​ ； wherein, represents the final embedding of the node in the layer represents the final embedding of the node is the total number of layers of graph convolution. For any triple to be predicted using a relation scoring function estimate its probability of existence; with the DistMult model, the scoring function is: ； wherein, denotes the transpose of and denote the nodes and in the layer the final embedding representation, denotes transforming the vector representation of the relation type into a diagonal matrix; mapping the score into an existence probability through a Sigmoid function: wherein is a sigmoid activation function, the parameters are constantly optimized through forward propagation and backpropagation, and the prediction result gradually approaches the true label ; finally training a social relationship reasoning model for inferring social relationships of an enterprise and its shareholders.

Images