Enterprise credit recommendation method and device based on multi-model integration

Abstract

The invention relates to the technical field of internet finance, in particular to an enterprise credit recommendation method and device based on multi-model integration. The recommendation method comprises the following steps: step 1, data collection; 2, data preprocessing; step 3, feature extraction; 4, model training; 5, credit product matching; and 6, recommending and outputting. According to the invention, by collecting and analyzing the multi-dimensional data of the enterprise, combining the characteristics and requirements of credit products issued by multiple banks, and utilizing machine learning and data mining technologies, personalized and accurate credit product recommendation services are provided for the enterprise. The technology not only improves the accuracy and efficiency of credit product recommendation, but also provides more scientific decision support for banks and financial institutions, optimizes the credit approval process and resource configuration, reduces the credit risk, and improves the overall operation efficiency. And meanwhile, through visual recommendation result display, enterprise users are helped to understand and select credit products more clearly.

Description

Technical Field

[0001] The present invention relates to the field of Internet financial technology, and in particular to a method and device for recommending corporate credit based on multi-model integration. Background Art

[0002] Existing methods for recommending corporate credit products mainly rely on manual analysis and empirical judgment, which are inefficient and inaccurate. Traditional credit recommendations are usually based on limited financial statements and credit scores, which provide a one-sided understanding and evaluation of the company and fail to fully utilize the company's multi-dimensional data. In addition, there are many types of credit products, and the credit products of various banks differ significantly in terms of features, requirements, bonus items, processing cycles, and difficulty. Companies face great challenges in choosing suitable credit products. Therefore, some credit product recommendation processing methods have emerged on the market to provide companies with more accurate and personalized credit product recommendations.

[0003] For example, a credit financial product recommendation processing method and device (publication number: CN112231550B, publication date: 2024-05-17), by obtaining the product elements of each credit financial product to be processed, including the historical loan approval rate, the maximum loan amount, the maximum loan term, the average annualized interest rate and the repayment method; and obtaining the credit loan amount of the credit financial product to be processed for the target user; obtaining the product element weight matrix; calculating the first recommendation score of the credit financial product to be processed; obtaining the weight value for the credit loan amount; calculating the second recommendation score of the credit financial product to be processed for the target user; using the first recommendation score and the second recommendation score, processing to obtain the recommendation index of the credit financial product to be processed for the target user, so as to recommend it to the target user.

[0004] The application focuses on the differences in product elements, allowing users to match several high-quality products and allowing applicant companies to more efficiently discover the "best" financial products. However, the application lacks the ability to systematically integrate the company's multi-dimensional data and credit product characteristics for information elements of credit financial products, and is unable to provide companies with efficient and accurate credit product recommendation services. Summary of the invention

[0005] In view of the technical defects existing in the background technology, the present invention proposes a corporate credit recommendation method and device based on multi-model integration, which solves the above technical problems and meets the actual needs. The specific technical solution is as follows: A corporate credit recommendation method based on multi-model integration includes the following steps: Step 1: Data collection: obtain multidimensional data of the enterprise; Step 2: converting the multidimensional data obtained in step 1 into a format suitable for machine learning, including digitizing categorical data, normalizing numerical data, processing missing values, data cleaning, data standardization, and data fusion feature extraction to form a unified original data set; Step 3: Feature extraction: Extract features from the original data set in step 2 to obtain a feature-extracted data set, which includes basic features, credit features, tax features, financial features, and water and electricity features; Step 4: Model training: Build and train multiple machine learning models, and obtain the final prediction results through ensemble learning; Step 4.1: Dataset division: The data set in step 3 is divided into a training set, a validation set and a test set; Step 4.2: construct multiple machine learning models, wherein the multiple machine learning models include a decision tree model, a random forest model, a support vector machine (SVM) model, and a deep learning model; Step 4.3: Model training: using the training set of step 4.1 to train the multiple machine learning models of step 4.2 to obtain multiple preliminary risk assessment models; Step 4.3: Model evaluation: Use the validation set of step 4.1 to evaluate the parameters and structures of the multiple preliminary risk assessment models after fine-tuning, and select one or more of the evaluation indicators selected from cross-validation, AUC, precision, and recall, and adjust the hyperparameters of the preliminary risk assessment model according to the feedback indicators to obtain the adjusted further risk assessment model, wherein the algorithm call involved in the training is one or more of the decision tree algorithm, random forest algorithm, gradient boosting algorithm, and deep learning algorithm; Step 4.4: Testing: Use the test set from step 3 to evaluate the predictive ability of the further risk assessment model adjusted in step 4.3 on unseen data, and obtain multiple trained risk assessment models; Step 4.5: Ensemble learning: combining the multiple risk assessment models in step 4.4 through an ensemble learning method to obtain the final prediction result; Step 5: Credit product matching: According to the prediction results of step 4, the enterprise is matched with credit products of different banks, and suitable credit products are recommended to the enterprise in a personalized manner; Step 6: Recommendation output: Output the recommended credit products and their detailed information to the enterprise, and provide a visual display of the recommendation results; It also includes an acquisition module, a matching module and a recommendation output module; The acquisition module is used to acquire multi-dimensional data of the target enterprise; The matching module is used to input the multi-dimensional data of the target enterprise, obtain the prediction results of the target enterprise, and make personalized recommendations of suitable credit products to the target enterprise based on the prediction results of the target enterprise; The output module outputs the recommended credit products and their detailed information to the enterprise and provides a visual display of the recommendation results.

[0006] As a further technical solution of the present invention, the multidimensional data of step 1 includes enterprise basic information data, enterprise multi-platform credit data, enterprise tax data, enterprise operating data and enterprise water and electricity data. The weight ratio of the enterprise multi-platform credit data and enterprise operating data in the original data set is higher than the weight ratio of the enterprise basic information data, enterprise tax data and water and electricity data in the original data set.

[0007] As a further technical solution of the present invention, the enterprise basic information data includes enterprise registration information, business scope and legal person information; the enterprise multi-platform credit data includes the credit score and credit record of the target enterprise on different platforms; the enterprise tax data includes tax payment records and tax ratings; the enterprise operating data includes financial statements, revenue and profit; the enterprise water and electricity data includes water and electricity fee payment and water and electricity consumption.

[0008] As a further technical solution of the present invention, the step 2 comprises the following steps: Step 2.1: Data cleaning: Clean the multidimensional data collected in step 1 to remove noise, errors, duplications and irrelevant data, including null value processing, illegal value processing, inconsistent data processing and duplicate data processing, to obtain the cleaned multidimensional data; Step 2.2: Data standardization: standardize the cleaned multidimensional data to obtain standardized multidimensional data; Step 2.3: Data fusion: Integrate the data from different sources and formats in the standardized multidimensional data in step 2.2 into a unified data set to form a unified original data set.

[0009] As a further technical solution of the present invention, the feature extraction in step 3 includes a principal component analysis (PCA) method, which converts the original data set into a set of linearly unrelated principal components through an orthogonal transformation, and the principal components are features extracted from the original data set.

[0010] As a further technical solution of the present invention, the model training process of the deep learning model also includes introducing LoRA technology for migration fine-tuning.

[0011] As a further technical solution of the present invention, the step 5 comprises the following steps: Step 5.1: Collect information: Collect detailed information of credit products from multiple banks, including the loan amount, interest rate, loan term, application conditions, processing period, bonus points and difficulty assessment of the credit products. The difficulty assessment of the credit products is the application approval rate based on the historical data of the credit products; Step 5.2: Preliminary screening: Based on the prediction results of step 4, first exclude those credit products that obviously do not meet the basic conditions or application conditions of the enterprise, so as to reduce the amount of data to be processed later and improve matching efficiency; Step 5.3: Comprehensive scoring: Calculate a comprehensive score for each credit product that passes the initial screening to comprehensively assess its match with the enterprise's needs; Step 5.4: Sorting and recommendation: According to the comprehensive score of step 5.3, the credit products are sorted from high to low, and the credit products that best meet their needs are recommended to the enterprise. The enterprise is matched with the credit products of different banks for personalized recommendation. Among them, when the debt ratio of the enterprise is greater than 75%, the credit product is not recommended; when 40%> the debt ratio of the enterprise is less than 75%, the approval rate of the recommended credit product is low; when the debt ratio of the enterprise is ≤40%, the approval rate of the recommended credit product is high.

[0012] The beneficial effects of the present invention are: The present invention provides enterprises with personalized and accurate credit product recommendation services by collecting and analyzing multi-dimensional data of enterprises, combining the characteristics and requirements of credit products launched by multiple banks, and using machine learning and data mining technology. This technology not only improves the accuracy and efficiency of credit product recommendations, but also provides more scientific decision-making support for banks and financial institutions, optimizes credit approval processes and resource allocation, reduces credit risks, and improves overall operational efficiency. At the same time, through the visual display of recommendation results, it helps corporate users understand and choose credit products more clearly.

[0013] The present invention adopts multiple model ensemble learning to effectively improve the accuracy and stability of overall prediction, improve the model's adaptability to abnormal data, and enhance the model's robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 The figure is a flowchart of a corporate credit recommendation method based on multi-model integration. DETAILED DESCRIPTION

[0015] The following is combined with Figure 1 The implementation modes of the present invention are described with relevant embodiments, but the implementation modes of the present invention are not limited to the following embodiments, and the present invention relates to relevant necessary components in this technical field, which should be regarded as the known technology in this technical field and can be known and mastered by the technicians in this technical field.

[0016] Existing methods for recommending corporate credit products mainly rely on manual analysis and empirical judgment, which are inefficient and inaccurate. Traditional credit recommendations are usually based on limited financial statements and credit scores, which provide a one-sided understanding and evaluation of the company and fail to fully utilize the company's multi-dimensional data. In addition, there are many types of credit products, and the credit products of various banks differ significantly in terms of features, requirements, bonus items, processing cycles, and difficulty. Companies face great challenges in choosing suitable credit products.

[0017] However, in the field of corporate credit recommendation, there is still a lack of a method and device that can systematically integrate multi-dimensional data and credit product characteristics and use intelligent models for recommendation. Therefore, the present invention proposes a corporate credit recommendation method and device based on multi-model integration, which provides enterprises with efficient and accurate credit product recommendation services through comprehensive data collection and intelligent model analysis, filling the gap in this field. The steps and device are as follows.

[0018] The present invention provides a corporate credit recommendation method based on multi-model integration, such as Figure 1 As shown, the following steps are included: Step 1: Data collection: obtain multidimensional data of the enterprise; Step 2: Data preprocessing: converting the multidimensional data obtained in step 1 into a format suitable for machine learning, including digitizing categorical data, normalizing numerical data, processing missing values, data cleaning, data standardization, and data fusion feature extraction to form a unified original data set; Step 3: Feature extraction: Extract features from the original data set in step 2 to obtain a feature-extracted data set, which includes basic features, credit features, tax features, financial features, and water and electricity features; Step 4: Model training: Build and train multiple machine learning models, and obtain the final prediction results through ensemble learning; Step 5: Credit product matching: Based on the prediction results of step 4, the enterprise is matched with credit products of different banks, and personalized recommendations of suitable credit products are given to the enterprise; Step 6: Recommendation output: Output the recommended credit products and their detailed information to the enterprise, and provide a visual display of the recommendation results.

[0019] The present invention provides an enterprise credit recommendation device based on multi-model integration, including an acquisition module, a matching module and a recommendation output module; Acquisition module: used to obtain multi-dimensional data of target enterprises; Matching module: used to input the multi-dimensional data of the target enterprise, obtain the prediction results of the target enterprise, and make personalized recommendations of suitable credit products to the target enterprise based on the prediction results of the target enterprise; Output module: outputs recommended credit products and their detailed information to enterprises, and provides a visual display of recommendation results.

[0020] The properties of each model are different, so the advantages are also different.

[0021] Taking the tree model as an example, its advantages are: first, it is a nonlinear model and can properly handle continuous features; second, it can discover multiple distinguishing feature combinations. Its disadvantages are: First, it cannot achieve good processing results for high-dimensional features, and this situation is prone to occur when dividing leaf nodes, that is, only a few strong features can be used as division features; second, if all features are considered each time a division operation is performed, good performance cannot be obtained; third, it is actually a mechanical memory of sample features, and the probability of overfitting is high.

[0022] Taking the generalized linear model as an example, its advantages are: first, it is suitable for parallel computing; second, it is very advantageous when processing massive discrete features. Its disadvantages are: first, it cannot directly process continuous features. Taking the LR model as an example, it needs to use the sigmoid function to process continuous features, which is prone to feature distortion; second, if you want to introduce nonlinear factors, you must complete the feature crossover work through manual means. In this process, not only does it cost a lot of manpower, but it is also impossible to exhaust the feature combination; third, the added value of its cross dimension is proportional to the complexity of the model.

[0023] It can be seen that when a single model is used to simulate the actual distribution of samples, the simulation results will be in error due to the nature of the model. In theory, the greater the difference between different individual learners, the higher their own accuracy, and the more ideal the prediction results of the integrated model. This application adopts an integrated algorithm, which takes each classification algorithm as the bottom layer, combines the advantages of each underlying algorithm, and exerts the best effect. The idea of ​​the integrated algorithm is to bring together multiple underlying classification methods to improve the accuracy of classification. If a single classifier is compared to a decision maker, the method of integrated learning is equivalent to multiple decision makers making a decision together.

[0024] The present invention provides enterprises with personalized and accurate credit product recommendation services by collecting and analyzing multi-dimensional data of enterprises, combining the characteristics and requirements of credit products launched by multiple banks, and using machine learning and data mining technology. The present invention adopts multiple model ensemble learning to effectively improve the accuracy and stability of overall prediction, improve the model's adaptability to abnormal data, and enhance the model's robustness. The present invention has the following advantages: Improve the accuracy of recommendations: Through multi-source data fusion and intelligent model analysis, the accuracy of the assessment of the actual situation of the enterprise is improved to ensure that the recommended credit products are more in line with the needs and conditions of the enterprise. This method can comprehensively and accurately assess the credit status and financing needs of the enterprise by comprehensively analyzing the basic information, credit data, tax data, operating data, water and electricity data of the enterprise and other multi-dimensional data.

[0025] Improve recommendation efficiency: Utilize automated data processing and model calculation to quickly screen and match credit products suitable for enterprises, shortening the recommendation and application cycle. Traditional credit assessment methods rely on manual analysis and are inefficient, while this invention greatly improves recommendation efficiency and the approval rate of credit applications through intelligent means.

[0026] Optimize resource allocation: Help banks and financial institutions make credit decisions more scientifically, optimize credit resource allocation, and reduce credit risks. Through accurate risk assessment models, banks can allocate credit resources more reasonably, reduce bad debt rates, and improve overall operational efficiency.

[0027] Personalized service: Provide customized credit product recommendations based on the specific needs and conditions of the enterprise to meet the financing needs of the enterprise at different stages of development. The intelligent model can recommend the most suitable credit products based on the different characteristics of the enterprise to help the enterprise grow and develop.

[0028] Enhanced user experience: Through the visual display of recommendation results, it helps corporate users to understand and select credit products more clearly, improving user experience and satisfaction. Users can more easily understand the reasons for recommendation and product advantages through intuitive charts and analysis results, so as to make more informed decisions.

[0029] The following are examples of the present invention.

[0030] Example 1

[0031] The present invention provides an enterprise credit recommendation device based on multi-model integration, including an acquisition module, a matching module and a recommendation output module; Acquisition module: used to obtain multi-dimensional data of the target enterprise, including basic enterprise information, credit data, tax data, operating data, water and electricity data, etc.

[0032] Matching module: used to input the multi-dimensional data of the target enterprise and obtain the prediction results of the target enterprise. Based on the prediction results of the target enterprise and the characteristics and requirements of the credit products launched by multiple banks, machine learning and data mining technology are used to make personalized recommendations of suitable credit products to the target enterprise.

[0033] Output module: Outputs recommended credit products and their detailed information to enterprises, and provides a visual display of recommendation results. Visual display not only improves user experience, but also increases the transparency and explainability of recommendation results.

[0034] A corporate credit recommendation method based on multi-model integration includes the following steps: Step 1: Data collection: obtain multidimensional data of the enterprise; Multidimensional data includes basic enterprise information data, enterprise multi-platform credit data, enterprise tax data, enterprise operation data, and enterprise water and electricity data; Basic enterprise information data includes enterprise registration information, business scope, enterprise scale, industry, operating years and legal person information; Enterprise multi-platform credit data includes the target enterprise’s credit score on different platforms, credit records including historical repayment techniques, etc.; Corporate tax data includes tax records, tax ratings, etc.

[0035] Corporate financial data includes financial statements, revenue, profit margin, debt ratio, etc.; Other enterprise data include the personal information of the business owner, the payment of the enterprise’s water and electricity bills, and the water and electricity consumption.

[0036] The weight of enterprise multi-platform credit data and enterprise operation data in the original data set is higher than that of enterprise basic information data, enterprise tax data, and water and electricity data in the original data set. According to the business logic or the importance of the data, you can manually assign weights to each sample to automatically adjust the sample weights or generate new synthetic samples to balance the data set to set the weight ratio.

[0037] Among them, there is a profound internal connection between tax data and other enterprise data, because tax data can largely reflect the financial status, operating results and compliance of the enterprise; the tax amount, tax rate and other information in the tax data are directly related to the company's operating income, profit and other financial indicators, and can verify the financial health of the enterprise. Corporate income tax declaration data can be compared with the profit data in the financial statements, reflecting the profitability of the company and the compliance of its accounting treatment; value-added tax declaration data can reflect the frequency and scale of the company's sales activities, and is closely related to the company's sales data, inventory data and other operating indicators. The company's energy consumption data, such as electricity consumption and water consumption, may be indirectly related to the company's production scale and business status. These data can indirectly verify the tax data; tax ratings, tax payment records, etc. can reflect the company's compliance and credit status, and complement the company's credit data such as credit scores and loan history; certain indicators in the tax data may be related to the market environment and industry development trends, providing a reference for the company's market positioning and strategic decision-making.

[0038] Step 2: Data preprocessing: Convert the multidimensional data obtained in step 1 into a format suitable for machine learning, including digitizing categorical data, normalizing numerical data, processing missing values, data cleaning, data standardization, and data fusion feature extraction to form a unified original data set. Since tax data is usually different from other data formats, data cleaning and conversion techniques are needed to unify data formats and solve data inconsistencies.

[0039] Step 2.1: Data cleaning: Clean the multidimensional data collected in step 1 to remove noise, errors, duplications and irrelevant data, including null value processing, illegal value processing, inconsistent data processing and duplicate data processing, to obtain cleaned multidimensional data.

[0040] Step 2.2: Data standardization: Standardize the cleaned multidimensional data to obtain standardized multidimensional data; the standardization formula is: , is the original eigenvalue, is the feature mean, is the characteristic standard deviation.

[0041] Step 2.3: Data fusion: Integrate the data from different sources and formats in the standardized multidimensional data in step 2.2 into a unified data set to form a unified original data set.

[0042] Data fusion refers to the integration and merging of data from different data sources to produce more comprehensive and accurate data. Data fusion helps eliminate information silos and improve the overall value of data and analytical results. The merging methods can include appending, connecting, and aggregation, depending on the characteristics of the data and the analysis requirements.

[0043] Step 3: Feature extraction: Perform feature extraction on the original data set in step 2 to obtain a data set after feature extraction. The data set includes basic features, credit features, tax features, financial features, and water and electricity features. Use principal component analysis (PCA) to reduce dimensions and extract useful features, especially to extract predictive features from tax data.

[0044] Feature extraction includes the principal component analysis (PCA) method, which converts the original data set into a set of linearly uncorrelated principal components through orthogonal transformation. The principal components are the features extracted from the original data set. The feature extraction process of PCA is to reduce the dimension of the data from n dimensions to k dimensions and sort the axis list in descending order of divergence.

[0045] First, based on the construction of the covariance matrix, we need to calculate the discreteness and covariance of the initial features. Assume that X has a data set consisting of m data, each of which is composed of n-dimensional column vectors. Arrange X into a matrix of n rows and m columns: , Then centralize: .

[0046] Afterwards, the coordinate transformation is required , w is an orthogonal basis, where W T W is a unit vector. If we reduce the n-dimensional space to the k-dimensional space, some coordinates in the coordinate system will be deleted, and the new coordinates will be , the projection of X in k dimension is , the original data is: , where w is the standard normal basis matrix, which can be rearranged to The largest eigenvector is taken out and normalized to form the matrix W. Each sample in the sample set is transformed to obtain the output sample set Z′.

[0047] Step 4: Model training: Build and train multiple machine learning models, and obtain the final prediction results through ensemble learning; Step 4.1: Dataset division: The dataset in step 3 is divided into training set, validation set and test set; Step 4.2: Build multiple machine learning models, including decision tree models, random forest models, support vector machine (SVM) models, and deep learning models; Decision tree model: Decision trees (DT) are a classification of decisions in the form of a tree. A tree itself consists of roots, branches, and leaves. The root of a decision tree is the root node, also known as the parent node. The branches of a decision tree are decision nodes. Each decision node can be regarded as another tree. The leaves of a decision tree are leaf nodes, which are very numerous. Decision trees make decisions based on tree structures. For the left subtree in the figure, decision node A is the parent node of leaf nodes B and C. If the analysis involves time series data, such as changes in tax amounts over time, the ARIMA model is used. This model is a necessary model in this method and is used to determine the credit rating or default risk of an enterprise.

[0048] Random forest model: It can be seen as a super collection of decision tree algorithms. It is an ensemble learning method based on multiple decision trees. The final prediction is made by voting on the results of multiple decision trees. In this method, random forest constructs multiple decision trees by randomly selecting data samples and features, and improves the accuracy of prediction by voting (classification problems) or averaging (regression problems). Diversity can bring complementarity between models and improve the overall prediction ability and robustness.

[0049] Support Vector Machine Model: Support Vector Machines (SVM) is a statistical learning algorithm that uses convex optimization theory to find the optimal hyperplane. The function of the optimal hyperplane is to distinguish between two types of samples. When new data is input into the optimal hyperplane, it will be brought to the optimal hyperplane and then classified according to the data in the optimal hyperplane. Support vector machines can not only effectively classify data, but also solve many other problems, such as the complexity of training models, training and optimization of large-scale data sets, etc.

[0050] The deep learning model is a machine learning method based on artificial neural networks. It simulates the connection mode of human brain neurons to build a network structure with multiple hidden layers, so as to achieve efficient feature learning and classification of complex data. For complex data relationships, deep learning models such as convolutional neural networks (CNN) are used to extract deeper features. By building a deep neural network and optimizing parameters using the back propagation algorithm, the layer-by-layer feature extraction and classification of input data can be achieved. During the training process, the model continuously adjusts the network parameters to gradually reduce the error between the output result and the true label, and finally achieves accurate prediction of new data.

[0051] Compared with multi-model integration, the single random forest model of this solution has the following characteristics, which enable this solution to better handle complex corporate credit recommendation problems and provide more accurate and reliable recommendation services.

[0052] First, this method adopts the bagging integration strategy, which focuses on reducing overfitting and improving the generalization ability of the model; Second, during the training process, a single random forest will randomly select a subset of data and a subset of features, which helps reduce the model's dependence on a specific data distribution; Third, it already has a high generalization ability because its construction process includes the randomness of the data; Fourth, it can provide stable prediction results in most cases. Although it may not be accurate enough on some complex or unstable data sets, since this method is used for comprehensive enterprise data with a large number of sample types, it can effectively reduce the error range of a single model. Fifth, if this method uses a single random forest model, its computational cost is relatively low because its implementation is relatively simple and has a certain degree of interpretability because the structure and feature importance of each decision tree can be observed.

[0053] If multi-model integration is required, a weighted integration strategy is used to assign different weights to the models according to the performance of each model (such as evaluation indicators), and a weighted voting method is used for prediction, giving higher weights to models with better performance.

[0054] Step 4.3: Model training: Use the training set of step 4.1 to train the multiple machine learning models of step 4.2 to obtain multiple preliminary risk assessment models. The evaluation indicators are selected from one or more of cross-validation, AUC, precision, and recall, and the hyperparameters of the preliminary risk assessment model are adjusted according to the feedback indicators to obtain an adjusted further risk assessment model. The above evaluation indicators refer to the application principles of similar methods to adjust the evaluation accuracy, deploy the trained integrated model to the production environment, implement a monitoring mechanism to track the performance of the model, and retrain or adjust it when necessary; (1) The basic idea of ​​a decision tree is to first determine the root node, which is the feature that best reflects the difference in data. After the root node is determined, it is screened step by step to select the partitioning points with obvious differences other than the root node for branching. The partitioning process is to find the feature with the largest difference value. Its fundamental goal is to cultivate a decision tree with strong generalization ability and integrate the decision-making of the "divide and conquer" idea. As the partitioning attributes are continuously carried out, the better the decision tree is, the more accurate the partitioning categories are, and the partitioning points are crucial to the partitioning of the decision tree.

[0055] The first step of the decision tree is to find the root node, which is the most representative. Then, the child nodes of the division are regarded as the root node to find the next division point. For dividing the data set, it is important to find the most characteristic root. Information entropy is an important indicator for evaluating the quality of decision trees. Assuming that the data set is D and there are k ways to divide the data, it can be regarded as C.k What needs to be considered when dividing data is whether the boundary of the division is clear or not. The boundary can also be calculated mathematically. For example, a root node has n attributes, one of which is A. The information entropy when it is not divided is calculated to be H(D): .

[0056] Assume that the data set divided by A becomes D i , where D ik Indicates D i The information entropy of A is k-way partitioning. : .

[0057] The CART algorithm uses the Gini coefficient as the dividing point. The Gini coefficient refers to randomly extracting data from a data set. The smaller the Gini coefficient, the higher the discrimination of the data set. Gini formula: .

[0058] (2) The training process of random forest requires four major steps: 1) First, randomly select the decision tree. For example, if there are n data, select one from the n data each time, put it back after sampling, and continue sampling n times. The n types of samples randomly selected are used as the root of the decision tree, and then the decision tree is divided and trained.

[0059] 2) Randomly select attributes. Take the root of the decision tree just randomly drawn as the main directory, and continue to divide it down. Assuming that each number has K attribute values, each time we draw k attributes according to the ones just put back. The k attribute is much smaller than the K attribute. Among these k attributes, use a certain strategy of the decision tree (such as the Gini coefficient) to select 1 partition point as the split point of the node.

[0060] 3) Repeat the second step until the tree cannot be divided any further. In a random forest, each tree root is random, the branches (attribute values) of the tree are random, and the leaves from the branches are also random. Each division is based on randomness, without pruning, until the leaves cannot be divided any further.

[0061] 4) Repeat steps 1, 2, and 3; build multiple random decision trees to form a random forest.

[0062] (3) Definition of normal vector of support vector machine , which determines the direction of the hyperplane, where d represents the number of eigenvalues, X represents the training sample, and b represents a displacement term that affects the distance between a hyperplane and an origin. The following formula can be used to represent the partitioning hyperplane: T X + b = 0.

[0063] This vector can be used to describe the direction and characteristics of the hyperplane, which can help us better understand the performance of the hyperplane at different time and space scales. Support vector machine (SVM) is a classification method that has been widely used in recent years. By constructing the concept of minimum risk, this classification method can effectively improve the generalization performance of the classifier and enable it to obtain good classification results under the condition of a small number of samples. SVM is based on the optimized classifier surface and can also show good classification results under small sample conditions. SVM is a method based on the optimal classification surface. Its goal is to transform a nonlinear separable problem into a linear separable problem, so that the optimal classification surface not only has the ability to distinguish between two categories, but also has the largest classification interval. In higher dimensions, the optimal classification line becomes the optimal classification surface. Support vector machine is an effective learning algorithm that can handle nonlinear separable problems and high-dimensional data space, thereby improving classification results.

[0064] (4) Depending on the network structure and application scenario, deep learning models can be divided into many types, including but not limited to the following: Convolutional Neural Network (CNN): A neural network specifically designed to process image data. It automatically extracts image features through convolution operations and performs classification or recognition.

[0065] The core of deep learning models lies in their powerful feature extraction capabilities, which can automatically learn useful information from raw data, avoiding the tedious process of manually designing features in traditional machine learning methods. Deep learning models have multiple hidden layers. By increasing the number of hidden layers, the model can learn more complex and abstract features, thereby improving the accuracy of predictions. Each neuron receives the output of the previous layer and performs a weighted sum, and then performs a nonlinear transformation through an activation function, which enables the model to handle complex nonlinear relationships.

[0066] In this step, the deep learning model introduces LoRA technology for migration fine-tuning. LoRA adds a side network to all fully connected layers in the pre-trained large model. This side network uses the product of two rank decomposition matrices to approximate the update value of the fully connected layer weights in domain adaptation.

[0067] Specifically, LoRA adds two structures A and B next to the pre-trained model structure. The parameters of these two structures are initialized to Gaussian distribution and 0, respectively, so the additional parameters are 0 at the beginning of training.

[0068] The input dimension of A and the output dimension of B are the same as the input and output dimensions of the original model, while the output dimension of A and the input dimension of B are much smaller than the input and output dimensions of the original model. During training, only the parameters of A and B are updated, and the pre-trained model parameters are fixed. During inference, the idea of ​​reparametrization can be used to merge AB and W, so that no additional calculations are introduced during inference. Moreover, for different downstream tasks, only A and B need to be retrained based on the pre-trained model, which can also speed up the training pace of large models.

[0069] Step 4.3: Model evaluation: Use the validation set in step 4.1 to evaluate the parameters and structures of multiple fine-tuned preliminary risk assessment models, and adjust the hyperparameters of the preliminary risk assessment models to obtain adjusted further risk assessment models, wherein the algorithm calls involved in the training are one or more of the decision tree algorithm, random forest algorithm, gradient boosting algorithm, and deep learning algorithm; Specifically, the decision tree algorithm is used to establish a classification or regression model. In this method, the split node is determined mainly through information gain; the random forest algorithm is used in this method to improve the accuracy of product prediction; the gradient boosting algorithm (such as XGBoost) is used in this method to minimize the loss function by gradually building a model in order to correct the output results; the deep learning algorithm is used in this method to extract high-level abstract features from the comprehensive information of the enterprise and make predictions.

[0070] Step 4.4: Testing: Use the test set from step 3 to evaluate the predictive ability of the further risk assessment model adjusted in step 4.3 on unseen data, and obtain multiple trained risk assessment models; Step 4.5: Ensemble learning: The multiple risk assessment models in step 4.4 are combined through the ensemble learning method to obtain the final prediction result.

[0071] Specifically, in step 4.5, the matching analysis between the enterprise and the credit product is carried out in the following ways: Prediction: Use the trained model to predict the characteristic data of the enterprise to obtain the credit score or default probability of the enterprise when applying for a specific credit product.

[0072] Application of rules in forecasting: Apply a set of predefined rules, taking into account the characteristics of credit products (such as interest rate, limit, term, repayment method, etc.) and the specific needs of the enterprise.

[0073] Specific output sorting and scoring: Based on the prediction results, multiple credit products are sorted to find the credit product that best meets the needs of the enterprise.

[0074] Matching algorithm: Similarity-based algorithms, such as cosine similarity, Euclidean distance, etc., may be used to measure the degree of match between corporate characteristics and credit product requirements.

[0075] Ensemble learning is expressed as underlying model + ensemble strategy; when the preferred underlying models can be perfectly combined, a model with higher accuracy or stronger robustness can be obtained, and good experimental results can be obtained. And combine them to obtain better results.

[0076] Among them, ensemble learning includes the Bagging algorithm. Bagging algorithm: Assume that the number of samples in the data set D is m, then randomly extract m samples with replacement and put them into the sampling set. Finally, T sample sets that have been sampled m times can be obtained. The base learner is trained for each sample set, and then they are integrated to obtain the ensemble learner.

[0077] Step 5: Credit product matching: Based on the prediction results of step 4, the enterprise is matched with credit products of different banks, and personalized recommendations of suitable credit products are given to the enterprise; Step 5.1: Collect information: Collect detailed information on credit products from multiple banks. The detailed information includes the loan amount, interest rate, loan term, application conditions, processing period, bonus points and difficulty assessment of the credit products. The difficulty assessment of the credit products is the application approval rate based on the historical data of the credit products. Step 5.2: Preliminary screening: Based on the prediction results of step 4, first exclude those credit products that obviously do not meet the basic conditions or application conditions of the enterprise, so as to reduce the amount of data to be processed later and improve matching efficiency; Step 5.3: Comprehensive scoring: Calculate a comprehensive score for each credit product that passes the initial screening to comprehensively assess its match with the enterprise's needs; Step 5.4: Sorting and recommendation: According to the comprehensive score of step 5.3, the credit products are sorted from high to low, and the credit products that best meet the needs of the enterprise are recommended to the enterprise. The enterprise is matched with the credit products of different banks and personalized recommendations are made; among them, when the debt ratio of the enterprise is greater than 75%, no credit product is recommended; when 40%> the debt ratio of the enterprise is less than 75%, the approval rate of the recommended credit product is low; when the debt ratio of the enterprise is ≤40%, the approval rate of the recommended credit product is high.

[0078] Step 6: Recommendation output: Output the recommended credit products and their detailed information to the enterprise, and provide a visual display of the recommendation results.

[0079] The present invention collects multi-dimensional data of enterprises, including basic enterprise information, credit data, tax data, operating data, water and electricity data, etc., and standardizes and integrates these data to form a unified data set. This process ensures the comprehensiveness and consistency of the data, and provides a solid data foundation for subsequent multi-model analysis.

[0080] We use machine learning and data mining techniques to build multiple intelligent models (such as decision trees, random forests, support vector machines, and deep learning models) to analyze enterprise data and conduct risk assessment. The combination of multiple models improves the accuracy and robustness of the assessment, making the recommendation results more reliable.

[0081] According to the analysis results of the intelligent model, the enterprise is matched with the credit products of different banks. Personalized recommendations are made taking into account the characteristics, requirements, bonus items, processing cycle and difficulty of the credit products. This process ensures that the recommended credit products can meet the actual needs and conditions of the enterprise and improve the approval rate of credit applications.

[0082] Specifically, the present invention performs matching analysis between enterprises and credit products through the following steps: Step 6.1 Prediction: Use the trained model to predict the enterprise’s characteristic data to obtain the enterprise’s credit score or default probability when applying for a specific credit product.

[0083] Step 6.2 Sorting and scoring: Sort multiple credit products according to the prediction results to find the credit product that best meets the needs of the enterprise.

[0084] Step 6.3 Rule application: Apply a set of predefined rules, taking into account the characteristics of the credit product (such as interest rate, limit, term, repayment method, etc.) and the specific needs of the enterprise.

[0085] Step 6.4 Matching algorithm: Use similarity-based algorithms, such as cosine similarity, Euclidean distance, etc., to measure the degree of match between enterprise characteristics and credit product requirements.

[0086] Output the list of credit products that best match the enterprise and can provide the specific value or percentage of the matching degree. Through such a process, the computer can help enterprises quickly and accurately find the credit products that best suit their needs, while helping financial institutions reduce credit risks. The recommended credit products and their detailed information are displayed to users in a visual way, helping users to understand and choose credit products more clearly. Visual display not only improves the user experience, but also increases the transparency and explainability of the recommendation results.

[0087] Through comprehensive data analysis and the application of intelligent models, the credit approval process of banks and financial institutions has been optimized, the approval efficiency has been improved, and labor costs have been reduced. The optimized process not only improves the operational efficiency of banks, but also reduces credit risks.

[0088] The present invention provides scientific decision support for banks and financial institutions, helping them to allocate credit resources more reasonably and conduct accurate risk control and management. Through risk assessment of intelligent models, banks can better judge the credit risk of enterprises and make more intelligent credit decisions.

[0089] According to the specific needs and conditions of the enterprise, we provide personalized and accurate credit product recommendations to meet the financing needs of enterprises at different stages of development. Personalized recommendations not only improve the satisfaction of enterprises, but also help them grow and develop.

[0090] The above is only a preferred embodiment of the present invention. It should be pointed out that ordinary technicians in this technical field can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A corporate credit recommendation method based on multi-model integration, characterized in that: The following steps are involved: Step 1: Data collection: obtain multidimensional data of the enterprise; Step 2: Data preprocessing: converting the multidimensional data obtained in step 1 into a format suitable for machine learning, including digitizing categorical data, normalizing numerical data, processing missing values, data cleaning, data standardization, and data fusion feature extraction to form a unified original data set; Step 3: Feature extraction: Extract features from the original data set in step 2 to obtain a feature-extracted data set, which includes basic features, credit features, tax features, financial features, and water and electricity features; Step 4: Model training: Build and train multiple machine learning models, and obtain the final prediction results through ensemble learning; Step 4.1: Dataset division: The data set in step 3 is divided into a training set, a validation set and a test set; Step 4.2: construct multiple machine learning models, wherein the multiple machine learning models include a decision tree model, a random forest model, a support vector machine (SVM) model, and a deep learning model; Step 4.3: Model training: Use the training set of step 4.1 to train the multiple machine learning models of step 4.2 to obtain multiple preliminary risk assessment models, wherein the algorithm calls involved in the training are one or more of a decision tree algorithm, a random forest algorithm, a gradient boosting algorithm, and a deep learning algorithm; Step 4.3: Model evaluation: Use the validation set of step 4.1 to evaluate the parameters and structures of the multiple preliminary risk assessment models after fine-tuning, and select one or more of the evaluation indicators such as cross-validation, AUC, precision, and recall. Adjust the hyperparameters of the preliminary risk assessment model according to the feedback indicators to obtain the adjusted further risk assessment model. Step 4.4: Testing: Use the test set from step 3 to evaluate the predictive ability of the further risk assessment model adjusted in step 4.3 on unseen data, and obtain multiple trained risk assessment models; Step 4.5: Ensemble learning: combining the multiple risk assessment models in step 4.4 through an ensemble learning method to obtain the final prediction result; Step 5: Credit product matching: According to the prediction results of step 4, the enterprise is matched with credit products of different banks, and suitable credit products are recommended to the enterprise in a personalized manner; Step 6: Recommendation output: Output the recommended credit products and their detailed information to the enterprise, and provide a visual display of the recommendation results; It also includes an acquisition module, a matching module and a recommendation output module; The acquisition module is used to acquire multi-dimensional data of the target enterprise; The matching module is used to input the multi-dimensional data of the target enterprise, obtain the prediction results of the target enterprise, and make personalized recommendations of suitable credit products to the target enterprise based on the prediction results of the target enterprise; The output module outputs the recommended credit products and their detailed information to the enterprise in the form of a list, and provides a visual display of the recommendation results including specific values ​​or percentages.

2. The enterprise credit recommendation method based on multi-model integration according to claim 1 is characterized in that: The multidimensional data in step 1 includes enterprise basic information data, enterprise multi-platform credit data, enterprise tax data, enterprise operating data and enterprise water and electricity data. The weight proportion of the enterprise multi-platform credit data and enterprise operating data in the original data set is higher than the weight proportion of the enterprise basic information data, enterprise tax data and water and electricity data in the original data set.

3. The enterprise credit recommendation method based on multi-model integration according to claim 2 is characterized in that: The enterprise basic information data includes enterprise registration information, business scope and legal person information; the enterprise multi-platform credit data includes the target enterprise's credit score and credit record on different platforms; the enterprise tax data includes tax payment records and tax ratings; the enterprise operating data includes financial statements, revenue and profit; the enterprise water and electricity data includes water and electricity fee payment and water and electricity consumption.

4. The enterprise credit recommendation method based on multi-model integration according to claim 1 is characterized in that: The step 2 comprises the following steps: Step 2.1: Data cleaning: Clean the multidimensional data collected in step 1 to remove noise, errors, duplications and irrelevant data, including null value processing, illegal value processing, inconsistent data processing and duplicate data processing, to obtain the cleaned multidimensional data; Step 2.2: Data standardization: standardize the cleaned multidimensional data to obtain standardized multidimensional data; Step 2.3: Data fusion: Integrate the data from different sources and formats in the standardized multidimensional data in step 2.2 into a unified data set to form a unified original data set.

5. The enterprise credit recommendation method based on multi-model integration according to claim 1 is characterized in that: The feature extraction in step 3 includes a principal component analysis (PCA) method, which converts the original data set into a set of linearly unrelated principal components through orthogonal transformation, and the principal components are the features extracted from the original data set.

6. The enterprise credit recommendation method based on multi-model integration according to claim 1 is characterized in that: The model training process of the deep learning model also includes introducing LoRA technology for migration fine-tuning.

7. The enterprise credit recommendation method based on multi-model integration according to claim 1 is characterized in that: The step 5 comprises the following steps: Step 5.1: Collect information: Collect detailed information of credit products from multiple banks, including the loan amount, interest rate, loan term, application conditions, processing period, bonus points and difficulty assessment of the credit products. The difficulty assessment of the credit products is the application approval rate based on the historical data of the credit products; Step 5.2: Preliminary screening: Based on the prediction results of step 4, first exclude those credit products that obviously do not meet the basic conditions or application conditions of the enterprise, so as to reduce the amount of data to be processed later and improve matching efficiency; Step 5.3: Comprehensive scoring: Calculate a comprehensive score for each credit product that passes the initial screening to comprehensively assess its match with the enterprise's needs; Step 5.4: Sorting and recommendation: According to the comprehensive score of step 5.3, the credit products are sorted from high to low, and the credit products that best meet their needs are recommended to the enterprise. The enterprise is matched with the credit products of different banks for personalized recommendation. Among them, when the debt ratio of the enterprise is greater than 75%, the credit product is not recommended; when 40%> the debt ratio of the enterprise is less than 75%, the approval rate of the recommended credit product is low; when the debt ratio of the enterprise is ≤40%, the approval rate of the recommended credit product is high.

Images