Medium, small and micro-sized enterprise financing guarantee limit measuring and calculating method and system based on fusion of large model and machine learning model

Abstract

The invention discloses a method and a system for measuring and calculating financing guarantee limits of small, medium and micro enterprises based on fusion of a large model and a machine learning model, and relates to the technical field of limit evaluation, and the method comprises the following steps: S1, data collection; s2, training a limit calculation model; s3, constructing an RAG external experience knowledge base; and S4, constructing a large model Agent framework. According to the method, a large model Agent framework and a machine learning model KAN are adopted, the limit calculation precision is remarkably improved, the soft information of an enterprise and a large amount of evaluation experience of a project manager and an asset evaluator are fully utilized, the method is combined with a traditional structured machine learning prediction model, the defect that the traditional machine learning model can only receive structured data is overcome, and the accuracy of limit calculation is improved. A large amount of expert experience and knowledge accumulation in the field are perfectly applied, the prediction accuracy and logic rationality are remarkably improved, complete review suggestions can be given, and end-to-end calculation of financing guarantee limits of small and medium-sized and micro enterprises is realized.

Description

Technical Field

[0001] The present invention relates to the field of credit limit assessment technology, and more specifically to a method and system for calculating the financing guarantee credit limit for small and medium-sized enterprises based on the fusion of a large model and a machine learning model. Background Art

[0002] Small, medium and micro enterprises occupy a pivotal position on the global economic stage. They are not only the solid foundation for economic stability and growth of various countries, but also an indispensable force in promoting global economic prosperity and diversified development. As the source of economic vitality, small, medium and micro enterprises play a vital role in driving economic growth, promoting employment and social stability, leading innovation and technological revolution, and enhancing the resilience of the economic system.

[0003] However, despite their significant contributions, small and medium-sized enterprises often face financing challenges due to inherent problems such as their own scale limitations, information asymmetry and imperfect credit system, which seriously restricts their further development and expansion. Therefore, building an efficient and accurate financing guarantee quota assessment model has become the key to solving the financing difficulties of small and medium-sized enterprises, effectively matching their operating capabilities with funding needs, and reducing guarantee risks.

[0004] Financing guarantee for small and medium-sized enterprises refers to a guarantee or security measure provided for the financing of small and medium-sized enterprises to reduce loan risks and promote financing activities. At present, the industry has developed a variety of advanced and diverse means and methods for the assessment of the financing guarantee quota of small and medium-sized enterprises. These methods aim to comprehensively and accurately reveal the credit status and repayment ability of small and medium-sized enterprises by comprehensively considering multiple dimensions such as the financial status, operating performance, market prospects, industry status, and management team quality of the enterprise, combining quantitative analysis and qualitative judgment, and provide scientific and reliable decision-making support for financial institutions. Specifically, it includes but is not limited to the comprehensive rating method, which scores and rates enterprises through a systematic and standardized evaluation system; the expert consultation method, which uses the rich experience and professional knowledge of experts in the field to conduct in-depth analysis; the data model method, which uses big data and machine learning technology to explore the value behind the data and predict credit risks; and the counter-collateral risk rating model, which specifically conducts a detailed assessment of the value and risk of collateral.

[0005] However, the current setting of guarantee quotas for small and medium-sized enterprises generally relies on an expert consultation mechanism. Although this method has its value, it inevitably introduces subjective judgment and potential relationship guarantee risks. At the same time, banks and financing guarantee institutions have accumulated a large amount of guarantee enterprise data in their long-term operations. These data treasures contain the complex relationship between the operating conditions of small and medium-sized enterprises and the appropriate guarantee quotas, indicating the great potential for building a data-driven guarantee quota assessment model. Unfortunately, despite the broad prospects, research on using historical guarantee data to build quantitative assessment models is still insufficient.

[0006] Furthermore, the report submitted by the enterprise when applying for a guarantee contains valuable information about the enterprise's survival status, operating capabilities and development prospects, which is essential for a comprehensive and objective assessment of the enterprise's qualifications. However, current research and practice have not yet fully explored how to effectively integrate this text information into the evaluation system of the financing guarantee quota for small and medium-sized enterprises, and there is currently no research on quantifying the previous expert review experience and combining it with machine learning.

[0007] Therefore, how to make up for the current lack of relevant research on determining the matching financing guarantee amount for small and medium-sized enterprises based on the existing operating information of the enterprises, and solve the shortcomings such as low efficiency and subjectivity of expert review in the traditional financing guarantee field are technical problems that technical personnel in this field urgently need to solve. Summary of the invention

[0008] In view of this, the present invention provides a method and system for calculating the financing guarantee amount of small and medium-sized enterprises based on the fusion of a large model and a machine learning model, which solves the problems existing in the background technology.

[0009] In order to achieve the above object, the present invention provides the following technical solutions:

[0010] A method for calculating the financing guarantee amount of small and medium-sized enterprises based on the fusion of a large model and a machine learning model, comprising the following steps:

[0011] S1. Data collection: Obtain multi-source data from the project review reports of various small and medium-sized enterprises, and organize the model training database, soft information database and experience database based on the evaluation experience of senior project managers and asset appraisers of the enterprises;

[0012] S2. Credit calculation model training: After pre-processing the model training database, it is input into the machine learning model KAN for model training to obtain a credit calculation model;

[0013] S3, build RAG external experience knowledge base: use the large model vectorization tool M3E to vectorize the experience database and soft information database respectively and build them into RAG vector knowledge base;

[0014] S4. Build a large model Agent framework: Use the quota calculation model and RAG vector knowledge base as the tool set, set prompts, and use tools in combination with the large model to complete the operations of receiving data, processing data, predicting quotas with models, giving suggestions and supervision based on the experience database, and retrieving database information.

[0015] Optionally, the data collection of S1 specifically includes the following steps:

[0016] The text information of the enterprise status recorded in the project review report of each small and medium-sized enterprise is extracted. The soft information data record of each sample is Q = {q1, q2, ..., q i}, where q i For a certain text soft information of the enterprise; organize the complete text soft information into a soft information database;

[0017] Quantify the non-financial characteristics of the enterprise in the text soft information and extract the soft information structure data. The soft information structure characteristics of each sample are N = {N1, N2, ..., N i}, where N i A non-financial indicator characteristic of an enterprise;

[0018] The financial structure data is extracted from the financial table in the project review report. The financial structure feature of each sample is X = {X1, X2, ..., X i}, where X i A financial indicator characteristic of an enterprise;

[0019] The financial structure data of each sample X={X1,X2,...,X i} and soft information structure data N = {N1, N2, ..., N i} to form a model training database Z = {X1, X2, ..., X i ,N1,N2,...,N i};

[0020] Communicate with senior project managers and asset appraisers of enterprises to obtain appraisal experience, summarize the appraisal experience and organize it into an experience database.

[0021] Optional, textual soft information includes: actual controller’s work experience, family status, counter-guarantee measures, etc.;

[0022] Soft information structure data include: the number of children of the actual controller, the number of credit inquiries, marital status, etc.;

[0023] Financial structure data include: enterprise working capital, debt-to-asset ratio, sales-to-liability ratio, personal guarantees, enterprise guarantees, etc.;

[0024] Evaluation experience includes: the cross-reference relationship between different data, the impact of the age of the actual controller of the enterprise on the repayment risk, etc.

[0025] Optionally, the S2 quota calculation model training includes the following steps:

[0026] Perform missing value processing, outlier processing and data standardization on the feature data in the model training database to unify the data format of each sample;

[0027] The preprocessed model training database is normalized, and the KAN model is used to train the normalized model training database and adjust the parameters to adapt to different data sets. The trained model is saved to obtain the credit limit calculation model.

[0028] Optionally, the specific operations for missing value processing are:

[0029] For a sample with only some features as missing values, the missing values ​​of each feature data in the original data set are filled with the mean of each feature data, where the missing value of the jth sample of the i-th feature data in the original data set is the value x' after the missing value processing is filled. missing ( i,j) for:

[0030]

[0031] Where: X ij represents the jth sample under the i-th feature in the original data set, j = 1, 2, ..., m, and m represents the number of samples.

[0032] Optionally, the specific operations for outlier processing are:

[0033] For each feature in the original data set, filter out the outliers whose difference from the mean of the current feature is greater than three times the standard deviation of the feature data, and adjust the outliers to the sum of the mean of the feature data and three times the standard deviation of the current feature data;

[0034] Screen out the outliers whose difference with the mean of the current feature is less than the opposite of the standard deviation of three times the feature data, and adjust the outliers to the difference between the mean of the feature data and the standard deviation of three times the current feature data;

[0035] Among them, the value x' of the outlier value of the i-th feature data of the j-th sample in the original data set after outlier processing adjustment outlier(i,j) for:

[0036]

[0037] Where: x ij Represents the value of the jth sample under the i-th feature data in the original data set, μi represents the mean of the i-th feature data in the original data set, σ i Represents the standard deviation of the i-th feature data in the original data set.

[0038] Optionally, the specific operations of data standardization are:

[0039] The StandardScaler method is used to replace the value of each feature data in the original data set with the z-score of each feature data, where the value of the i-th feature data of the j-th sample in the original data set is replaced by the value x' (i,j) for:

[0040]

[0041] Where: x ij Represents the value of the jth sample under the i-th feature data in the original data set, μ i represents the mean of the i-th feature data in the original data set, σ i Represents the standard deviation of the i-th feature data in the original data set.

[0042] Optionally, S3 builds a RAG external experience knowledge base, specifically:

[0043] The large model vectorization tool M3E is used to perform text embedded expression on the experience database and the soft information database respectively, and vectors with dimensions determined by the large model and the data set are obtained, and stored in the local Chroma vector database to obtain the experience knowledge vector database and the soft information knowledge vector database respectively.

[0044] Optionally, S4 builds a large model Agent framework, which specifically includes the following steps:

[0045] Instantiate a language model object in the Agent, call the online API or deploy a large model locally;

[0046] Define the toolsets jingyan_RAG, wenben_RAG, pinggu_model, and baidu_API, and use natural language to describe the functions, inputs, and outputs of the tools respectively;

[0047] Among them, jingyan_RAG uses the Chroma vector database of empirical data to learn from experience, combines user input, and judges whether the input enterprise information is wrong from the perspective of a project reviewer, and gives suggestions based on evaluation experience;

[0048] wenben_RAG uses the soft information knowledge Chroma vector database and the user's input to automatically determine the user's needs. If the user wants to query the information in the knowledge base, it searches for the relevant information of the corresponding enterprise through similarity and returns it. If the user wants to evaluate the data of a certain enterprise, it learns the previous similar evaluation data based on the data entered by the user as an evaluation reference, and combines other tools to give the final evaluation results.

[0049] pinggu_model obtains the enterprise data to be evaluated input by the user, calls the trained quota calculation model to predict the enterprise data and returns the evaluation result;

[0050] baidu_API obtains the name of the enterprise to be evaluated and the name of the actual controller entered by the user and searches for it as a reference for evaluation;

[0051] Create an Agent method and pass in the set language model and tool set, and then pass in the prompt so that the big model can understand how to think about solving the problem;

[0052] Add advanced system prompts in the defined big model chat function. Before calling the big model each time, input the advanced system prompts to the big model to guide the big model on how to call and combine tools.

[0053] A system applied to the method for calculating the financing guarantee amount of small and medium-sized micro-enterprises based on the fusion of a large model and a machine learning model as described in any one of the above items, comprising: a user interaction interface and an Agent processing module;

[0054] User interaction interface, used to implement end-to-end interactive response between users and large model systems, including user deployment access and result output;

[0055] The Agent processing module includes the process of calling the tool set and the big model. It is used to receive questions input by users, automatically call the tools in the tool set and combine them to give feedback. The big model combines the language understanding ability and logical reasoning ability of the big model based on user needs and prompt words, and gives the final output results in the form of natural language.

[0056] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a method and system for calculating the financing guarantee amount of small and medium-sized enterprises based on the fusion of a large model and a machine learning model, which has the following beneficial effects:

[0057] (1) Improving the accuracy of credit limit calculation: By integrating the big model and the machine learning model KAN, the present invention can simultaneously process structured data and text soft information, and comprehensively evaluate the financial indicators, non-financial indicators and text information of the enterprise, which not only improves the model's ability to process complex data, but also significantly improves the accuracy of financing guarantee limit calculation;

[0058] (2) Effective use of expert experience and knowledge: The present invention incorporates the assessment experience of senior project managers and asset appraisers of enterprises into the RAG external experience knowledge base of the large language model. Through the large model Agent framework, the system can call on experience knowledge during the amount calculation process, supervise and optimize the model prediction results, and give full play to the important role of expert experience in risk assessment;

[0059] (3) Solving the limitations of traditional models: Traditional machine learning models can usually only process structured data and cannot effectively utilize text soft information and expert experience. The present invention breaks through this limitation and successfully integrates soft information and review experience into the model, realizing an end-to-end financing guarantee amount calculation method, which not only improves the logic and rationality of the prediction, but also can provide complete review opinions, thereby improving the efficiency and quality of the evaluation work. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0061] Figure 1 A flow chart of the method for calculating the financing guarantee amount of small and medium-sized enterprises provided by the present invention;

[0062] Figure 2 A block diagram for calculating the financing guarantee amount for small and medium-sized enterprises provided by the present invention. DETAILED DESCRIPTION

[0063] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0064] The embodiment of the present invention discloses a method for calculating the financing guarantee amount of small and medium-sized enterprises based on the fusion of a large model and a machine learning model. Figure 1 , Figure 2 As shown, the following steps are included:

[0065] S1. Data collection: Obtain multi-source data from the project review reports of various small, medium and micro enterprises, and organize the model training database, soft information database and experience database based on the evaluation experience of senior project managers and asset appraisers of the enterprises; among them, multi-source data includes the internal structured information of the enterprise such as the balance sheet and income statement of the enterprise, and textual soft information such as the business model and family situation of the enterprise;

[0066] S2. Credit calculation model training: After pre-processing the model training database, it is input into the machine learning model KAN for model training to obtain a credit calculation model;

[0067] S3, build RAG external experience knowledge base: use the large model vectorization tool M3E to vectorize the experience database and soft information database respectively and build them into RAG vector knowledge base;

[0068] S4. Build a large model Agent framework: Use the quota calculation model and RAG vector knowledge base as the tool set, set prompts, and use tools in combination with the large model to complete the operations of receiving data, processing data, predicting quotas with models, giving suggestions and supervision based on the experience database, and retrieving database information.

[0069] Furthermore, the data collection of S1 specifically includes the following steps:

[0070] The text information of the enterprise status recorded in the project review report of each small and medium-sized enterprise is extracted. The soft information data record of each sample is Q = {q1, q2, ..., q i}, where q i is a text soft information of the enterprise; the complete text soft information is sorted into a soft information database; in this embodiment, the text soft information includes: the actual controller's work experience, family status, counter-guarantee measures, etc.;

[0071] Quantify the non-financial characteristics of the enterprise in the text soft information and extract the soft information structure data. The soft information structure characteristics of each sample are N = {N1, N2, ..., N i}, where N i is a non-financial indicator feature of the enterprise; in this embodiment, the soft information structure data includes: the number of children of the actual controller, the number of credit inquiries, marital status, etc.;

[0072] The financial structure data is extracted from the financial table in the project review report. The financial structure feature of each sample is X = {X1, X2, ..., X i}, where X iis a certain financial indicator characteristic of the enterprise; in this embodiment, the financial structure data includes: enterprise working capital, asset-liability ratio, sales-liability ratio, personal guarantee, enterprise guarantee, etc.;

[0073] The financial structure data of each sample X={X1,X2,...,X i} and soft information structure data N = {N1, N2, ..., N i} to form a model training database Z = {X1, X2, ..., X i ,N1,N2,...,N i};

[0074] Communicate with senior project managers and asset appraisers of the enterprise to obtain evaluation experience, summarize the evaluation experience and organize it into an experience database by numbering; in this embodiment, the evaluation experience includes: the cross-reference relationship between different data, the impact of the age of the actual controller of the enterprise on the repayment risk, etc.

[0075] Furthermore, the quota calculation model training of S2 specifically includes the following steps:

[0076] Perform missing value processing, outlier processing and data standardization on the feature data in the model training database to unify the data format of each sample;

[0077] The preprocessed model training database is normalized, the normalized model training database is trained using the KAN model and the parameters are adjusted to adapt to different data sets, the values ​​predicted by the trained KAN are denormalized to obtain reference values, the trained model is saved, and the credit estimation model is obtained. The adjusted parameters include the training step size steps, the number of grid intervals of KAN gird (a set of discrete points representing the spline function used for fitting), and the k-order polynomial representation used when fitting the function in each interval.

[0078] Furthermore, the specific operation of missing value processing is: delete samples with 1 / 3 or more of the indicators being empty;

[0079] For a sample with only some features as missing values, the feature mean method is used to fill the missing values ​​of each feature data in the original data set with the mean of each feature data, where the missing value of the jth sample of the i-th feature data in the original data set is the value x' after the missing value processing is filled. missing(i,j) for:

[0080]

[0081] Where: X ijrepresents the jth sample under the i-th feature in the original data set, j = 1, 2, ..., m, and m represents the number of samples.

[0082] Furthermore, the specific operations of outlier processing are:

[0083] For each feature in the original data set, the three-times standard deviation method is used to screen out outliers whose difference from the mean of the current feature is greater than three times the standard deviation of the feature data, and the outliers are adjusted to the sum of the mean of the feature data and three times the standard deviation of the current feature data;

[0084] Screen out the outliers whose difference with the mean of the current feature is less than the opposite of the standard deviation of three times the feature data, and adjust the outliers to the difference between the mean of the feature data and the standard deviation of three times the current feature data;

[0085] Among them, the value x' of the outlier value of the i-th feature data of the j-th sample in the original data set after outlier processing adjustment outlier(i,j) for:

[0086]

[0087] Where: x ij Represents the value of the jth sample under the i-th feature data in the original data set, μ i represents the mean of the i-th feature data in the original data set, σ i Represents the standard deviation of the i-th feature data in the original data set.

[0088] Furthermore, the specific operations of data standardization are:

[0089] The StandardScaler method is used to replace the value of each feature data in the original data set with the z-score of each feature data, where the value of the i-th feature data of the j-th sample in the original data set is replaced by the value x' (i,j) for:

[0090]

[0091] Where: x ij Represents the value of the jth sample under the i-th feature data in the original data set, μ i represents the mean of the i-th feature data in the original data set, σ i Represents the standard deviation of the i-th feature data in the original data set.

[0092] Furthermore, S3 constructs the RAG external experience knowledge base, specifically:

[0093] The large model vectorization tool M3E is used to perform text embedded expression on the experience database and the soft information database, respectively, to obtain vectors with dimensions determined by the large model and the data set, and store them in the local Chroma vector database, respectively obtaining an experience knowledge vector database and a soft information knowledge vector database. In this embodiment, the text embedded expression can be a 1*1536-dimensional vector, and the specific vector dimension can vary depending on the large model and the data set.

[0094] Furthermore, the construction of the large model Agent framework of S4 specifically includes the following steps:

[0095] Instantiate a language model object in the Agent and call the online API or deploy a large model locally. The online API is ChatGPT. If you want to achieve privacy localization, you can deploy large models such as Llama3 and ChatGLM locally.

[0096] Define the tool sets jingyan_RAG, wenben_RAG, pinggu_model, and baidu_API, and use natural language to describe the functions, inputs, and outputs of the tools, respectively. The auxiliary functions called when the proxy processes user requests;

[0097] Among them, jingyan_RAG: uses the Chroma vector database of experience data, learns the experience in the database, combines the user's input, and judges whether the input corporate information is wrong from the perspective of a project reviewer (such as the cross-checking relationship between financial information), and gives suggestions based on the evaluation experience (such as reminding that the actual controller is too old and the book standards should be tightened, etc.); in addition, the experience in the database can be increased at any time as the evaluation experience accumulates and will not be forgotten;

[0098] wenben_RAG: Using the soft information knowledge Chroma vector database and the user's input, the Agent will automatically determine the user's needs; if the user wants to query the information in the knowledge base, the Agent will search for the relevant information of the corresponding enterprise through similarity and return it; if the user wants to evaluate the data of a certain enterprise, the Agent will learn similar evaluation data from the past based on the data entered by the user as an evaluation reference, and combine other tools to give the final evaluation result;

[0099] pinggu_model: obtains the enterprise data to be evaluated input by the user, calls the trained quota calculation model to predict the enterprise data and returns the evaluation result;

[0100] baidu_API: Get the name of the company to be evaluated and the name of the actual controller entered by the user, and call Baidu API for retrieval as a reference for evaluation; if necessary, you can also call Qichacha API to check the company's credibility, or retrieve other information about the company from the Internet as a reference for evaluation;

[0101] Create an Agent method and pass in the set language model and tool set, and then pass in the prompt so that the big model can understand how to think about solving the problem;

[0102] The specific implementation of setting the advanced prompt in combination with the big model is: add advanced system prompts in the defined big model chat function, and before each call to the big model, input the advanced system prompts to the big model to guide the big model on how to call and combine the tools.

[0103] It should be noted that this embodiment provides the definition method of the above four tool sets, but according to the technical solution provided by the present invention, other tool sets can be defined as appropriate. For example: if there is content that you want to use as an external database, you can continue to define new tools. The overall idea is the same, that is, through the external database and the method of defining the assessment model as a tool, they complement each other, and through the content of the external database and the model assessment results, the financing guarantee assessment task is performed and the final feedback is output. Therefore, using the technical ideas of the present invention, the definitions of other tool sets that continue to extend on the basis of this framework can be protected, not limited to the tools mentioned above.

[0104] Here is a specific example: you are an AI assistant named "Small and Medium Enterprise Asset Valuation Assistant", focusing on small and medium-sized enterprise asset valuation predictions and querying enterprise-related issues from the enterprise database. The data entered by the user will be separated by commas.

[0105] Your task is to:

[0106] 1. Based on the user's question, call the appropriate tool (such as RAG tool or API) to provide the most accurate enterprise asset forecast and query information. If the user wants to conduct an asset evaluation, in addition to informing the user, it is also necessary to call Baidu API for relevant information on the enterprise on the Internet as a reference based on the enterprise's information and its own knowledge reserves and previous evaluation experience in RAG, and comprehensively provide users with suggestions, that is, whether to lend to the enterprise or whether the loan amount is appropriate or what issues of the enterprise need to be focused on.

[0107] 2. When answering, keep your tone gentle and friendly, and provide useful suggestions; explain complex query results step by step, and explain background knowledge in detail when necessary; ensure that the answer is concise and clear, but can be detailed if necessary.

[0108] 3. Handle special situations: If data is unavailable or the result cannot be determined, please politely inform the user of the reason and suggest that the user provide more information (such as specific company ID); if the user inputs an incorrect format, please kindly remind them and provide the correct input format; in each answer, please choose the appropriate tool based on the user's question and explain the results in a way that the user can understand.

[0109] In addition, this embodiment also provides a system for calculating the financing guarantee amount of small and medium-sized micro enterprises based on the fusion of a large model and a machine learning model, which is applied to the above method, and includes: a user interaction interface and an Agent processing module;

[0110] The user interaction interface is used to realize the end-to-end interactive response between the user and the large model system, including the user's deployment access and result output; specifically, the user inputs the required questions (such as evaluating the financing guarantee amount of a certain enterprise data and giving suggestions or querying the financial or text information of a certain enterprise in the database), receives the questions and passes them to the Agent processing module, which is processed by the module and output in natural language and fed back to the user at the terminal;

[0111] The Agent processing module includes the process of calling the toolset and the big model. It receives the questions input by the user and passes them to the Agent processing module. The Agent automatically calls the tools in the toolset and combines them to give feedback and pass it to the big model. The big model receives the feedback and gives the final output results in the form of natural language based on the user's needs and prompt words combined with the language understanding ability and logical reasoning ability of the big model.

[0112] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0113] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for calculating the financing guarantee amount of small and medium-sized enterprises based on the fusion of a large model and a machine learning model, characterized in that: The following steps are involved: S1. Data collection: Obtain multi-source data from the project review reports of various small and medium-sized enterprises, and organize the model training database, soft information database and experience database based on the evaluation experience of senior project managers and asset appraisers of the enterprises; S2. Credit calculation model training: After pre-processing the model training database, it is input into the machine learning model KAN for model training to obtain a credit calculation model; S3, build RAG external experience knowledge base: use the large model vectorization tool M3E to vectorize the experience database and soft information database respectively and build them into RAG vector knowledge base; S4. Build a large model Agent framework: Use the quota calculation model and RAG vector knowledge base as the tool set, set prompts, and use tools in combination with the large model to complete the operations of receiving data, processing data, predicting quotas with models, giving suggestions and supervision based on the experience database, and retrieving database information.

2. According to claim 1, a method for calculating the financing guarantee amount of small and medium-sized enterprises based on the fusion of a large model and a machine learning model is characterized in that: The data collection of S1 specifically includes the following steps: The text information of the enterprise status recorded in the project review report of each small and medium-sized enterprise is extracted. The soft information data record of each sample is Q = {q1, q2, ..., q i }, where q i For a certain text soft information of the enterprise; organize the complete text soft information into a soft information database; Quantify the non-financial characteristics of the enterprise in the text soft information and extract the soft information structure data. The soft information structure characteristics of each sample are N = {N1, N2, ..., N i }, where N i A non-financial indicator characteristic of an enterprise; The financial structure data is extracted from the financial table in the project review report. The financial structure feature of each sample is X = {X1, X2, ..., X i }, where X i A financial indicator characteristic of an enterprise; The financial structure data of each sample X={X1,X2,...,X i } and soft information structure data N = {N1, N2, ..., N i } to form a model training database Z = {X1, X2, ..., X i ,N1,N2,...,N i }; Communicate with senior project managers and asset appraisers of enterprises to obtain appraisal experience, summarize the appraisal experience and organize it into an experience database.

3. According to claim 2, a method for calculating the financing guarantee amount of small and medium-sized enterprises based on the fusion of a large model and a machine learning model is characterized in that: Textual soft information includes: the actual controller’s work experience, family status, and counter-guarantee measures; Soft information structure data include: the number of children of the actual controller, the number of credit inquiries, and marital status; Financial structure data include: enterprise working capital, debt-to-asset ratio, sales-to-debt ratio, personal guarantees, and enterprise guarantees; Evaluation experience includes: the cross-reference relationship between different data and the impact of the age of the actual controller of the enterprise on repayment risk.

4. According to claim 1, a method for calculating the financing guarantee amount of small and medium-sized enterprises based on the fusion of a large model and a machine learning model is characterized in that: The training of the S2 quota calculation model includes the following steps: Perform missing value processing, outlier processing and data standardization on the feature data in the model training database to unify the data format of each sample; The preprocessed model training database is normalized, and the KAN model is used to train the normalized model training database and adjust the parameters to adapt to different data sets. The trained model is saved to obtain the credit limit calculation model.

5. According to claim 4, a method for calculating the financing guarantee amount of small and medium-sized enterprises based on the fusion of a large model and a machine learning model is characterized in that: The specific operations for missing value processing are: For a sample with only some features as missing values, the missing values ​​of each feature data in the original data set are filled with the mean of each feature data, where the missing value of the jth sample of the i-th feature data in the original data set is filled with the value x' missing(i,j) for: Where: X ij represents the jth sample under the i-th feature in the original data set, j = 1, 2, ..., m, and m represents the number of samples.

6. According to claim 4, a method for calculating the financing guarantee amount of small and medium-sized enterprises based on the fusion of a large model and a machine learning model is characterized in that: The specific operations of outlier processing are: For each feature in the original data set, filter out the outliers whose difference from the mean of the current feature is greater than three times the standard deviation of the feature data, and adjust the outliers to the sum of the mean of the feature data and three times the standard deviation of the current feature data; Screen out the outliers whose difference with the mean of the current feature is less than the opposite of the standard deviation of three times the feature data, and adjust the outliers to the difference between the mean of the feature data and the standard deviation of three times the current feature data; Among them, the value x' of the outlier value of the i-th feature data of the j-th sample in the original data set after outlier processing adjustment outlier(i,j) for: Where: x ij Represents the value of the jth sample under the i-th feature data in the original data set, μ i represents the mean of the i-th feature data in the original data set, σ i Represents the standard deviation of the i-th feature data in the original data set.

7. According to claim 4, a method for calculating the financing guarantee amount of small and medium-sized enterprises based on the fusion of a large model and a machine learning model is characterized in that: The specific operations of data standardization are: The StandardScaler method is used to replace the value of each feature data in the original data set with the z-score of each feature data, where the value of the i-th feature data of the j-th sample in the original data set is replaced by the value x′ after the standardization process. (i,j) for: Where: x ij Represents the value of the jth sample under the i-th feature data in the original data set, μ i represents the mean of the i-th feature data in the original data set, σ i Represents the standard deviation of the i-th feature data in the original data set.

8. According to claim 1, a method for calculating the financing guarantee amount of small and medium-sized enterprises based on the fusion of a large model and a machine learning model is characterized in that: S3 builds the RAG external experience knowledge base, specifically: The large model vectorization tool M3E is used to perform text embedded expression on the experience database and the soft information database respectively, and vectors with dimensions determined by the large model and the data set are obtained, and stored in the local Chroma vector database to obtain the experience knowledge vector database and the soft information knowledge vector database respectively.

9. According to claim 1, a method for calculating the financing guarantee amount of small and medium-sized enterprises based on the fusion of a large model and a machine learning model is characterized in that: S4 builds a large model Agent framework, which specifically includes the following steps: Instantiate a language model object in the Agent, call the online API or deploy a large model locally; Define the toolsets jingyan_RAG, wenben_RAG, pinggu_model, and baidu_API, and use natural language to describe the functions, inputs, and outputs of the tools respectively; Among them, jingyan_RAG uses the Chroma vector database of empirical data to learn from experience, combines user input, and judges whether the input enterprise information is wrong from the perspective of a project reviewer, and gives suggestions based on evaluation experience; wenben_RAG uses the soft information knowledge Chroma vector database and the user's input to automatically determine the user's needs. If the user wants to query the information in the knowledge base, it searches for the relevant information of the corresponding enterprise through similarity and returns it. If the user wants to evaluate the data of a certain enterprise, it learns the previous similar evaluation data based on the data entered by the user as an evaluation reference, and combines other tools to give the final evaluation results. pinggu_model obtains the enterprise data to be evaluated input by the user, calls the trained quota calculation model to predict the enterprise data and returns the evaluation result; baidu_API obtains the name of the enterprise to be evaluated and the name of the actual controller entered by the user and searches for it as a reference for evaluation; Create an Agent method and pass in the set language model and tool set, and then pass in the prompt so that the big model can understand how to think about solving the problem; Add advanced system prompts in the defined big model chat function. Before calling the big model each time, input the advanced system prompts to the big model to guide the big model on how to call and combine tools.

10. A system applied to the method for calculating the financing guarantee amount of small and medium-sized micro enterprises based on the fusion of a large model and a machine learning model as described in any one of claims 1 to 9, characterized in that: include: User interaction interface and Agent processing module; User interaction interface, used to implement end-to-end interactive response between users and large model systems, including user deployment access and result output; The Agent processing module includes the process of calling the tool set and the big model. It is used to receive questions input by users, automatically call the tools in the tool set and combine them to give feedback. The big model combines the language understanding ability and logical reasoning ability of the big model based on user needs and prompt words, and gives the final output results in the form of natural language.

Images