A commercial credit assessment and supervision method based on multimodal co-evolutionary algorithm

Abstract

The present invention relates to a commercial credit evaluation and supervision method based on a multimodal co-evolutionary algorithm, which collects and cleans multimodal data, ensures data quality, uniformly stores it as a structured database, uses cross-modal comparative learning to extract features, embeds features into a dynamic multimodal knowledge graph, uses a modal sharing network to extract and fuse modal sharing and specific features, optimizes the consistency of modal feature distribution through adversarial learning, initializes a global co-evolutionary credit scoring model, uses a dynamic multi-objective algorithm to optimize model accuracy, robustness and interpretability, introduces a causal reasoning model to analyze the driving factors of corporate credit changes, establishes causal relationships, designs a closed-loop feedback system, updates the credit evaluation model, enhances adaptability, generates an interpretable credit evaluation report based on a generative adversarial network, and enhances transparency and trust. The method improves the accuracy and adaptability of credit evaluation and ensures the robustness and interpretability of evaluation results.

Description

Technical Field

[0001] The present invention belongs to the cross-technical field of financial technology and artificial intelligence, and involves technologies such as credit risk assessment, big data analysis and optimization algorithms. Specifically, it is a commercial credit assessment and supervision method based on a multimodal co-evolutionary algorithm. Background Art

[0002] Commercial credit assessment is an important basis for financial institutions and regulatory authorities to judge corporate credit risks and formulate management policies. As corporate business activities become increasingly complex and the degree of informatization continues to increase, traditional credit assessment methods have gradually exposed their limitations. Traditional assessment methods are usually based on limited structured data such as financial statements and credit records, ignoring the potential value of unstructured data such as market dynamics, market feedback, and social media. This reliance on a single data source limits the comprehensiveness and accuracy of credit assessment results.

[0003] With the rapid development of artificial intelligence and data mining technology, the application of deep learning and machine learning technology in the financial field has made significant progress. Compared with traditional statistical models, deep learning methods have the advantage of processing massive data and complex relationships. However, the "black box" attribute of such methods makes their results lack transparency and poses challenges to the interpretability and application of models. Especially in credit assessment, the processing ability of deep learning models for multimodal data has not been fully released. How to efficiently integrate and deeply mine heterogeneous data is still an important direction of current research.

[0004] In recent years, collaborative optimization algorithms have been widely used in optimization problems of complex systems due to their flexibility and dynamic adjustment capabilities. These algorithms achieve a balance between global search capabilities and local accuracy by simulating the collaboration and competition between different factors. However, existing research in the field of credit assessment mostly focuses on single model optimization, fails to fully explore the collaborative relationship between multimodal data, and is difficult to meet the high requirements of credit assessment for data comprehensiveness and result dynamics.

[0005] Although some studies have attempted to combine multimodal data fusion, collaborative optimization algorithms and real-time credit assessment mechanisms, they still face the following problems: First, the heterogeneity and complexity of multimodal data increase the difficulty of data processing and fusion. Most existing methods integrate data at the level of a single modality or simple weighted fusion, and fail to effectively capture the potential correlation between multimodal data; second, most credit assessment models are static models that lack dynamic adaptability and are difficult to reflect changes in corporate credit in real time; third, traditional models lack interpretability and cannot provide clear decision-making basis for regulators and enterprises. Summary of the invention

[0006] Based on the above problems, the present invention starts from the technical bottleneck of multimodal data fusion and combines the collaborative optimization algorithm to propose a commercial credit assessment and supervision method based on a multimodal co-evolutionary algorithm, aiming to achieve efficient integration and dynamic optimization of multi-source heterogeneous data, improve the comprehensiveness, accuracy and real-time performance of credit assessment, and provide more scientific support for risk warning and regulatory decision-making.

[0007] The purpose of the present invention is to provide a commercial credit assessment and supervision method based on a multimodal co-evolutionary algorithm, aiming to solve the problems of single data processing, poor dynamic adaptability and insufficient model interpretability in existing methods. The present invention realizes scientific, efficient and dynamic assessment of corporate credit through feature extraction and fusion of multimodal data, co-evolutionary optimization mechanism and dynamic credit risk monitoring. The method integrates structured, unstructured and semi-structured data, optimizes the fusion strategy of multimodal features, improves the accuracy and adaptability of credit assessment, especially in different scenarios and when data distribution changes, to ensure the robustness and interpretability of the assessment results.

[0008] In order to achieve the above object, the present invention provides a commercial credit evaluation and supervision method based on a multimodal co-evolutionary algorithm, the method comprising the following steps:

[0009] S1: Collect and clean multimodal data to ensure data quality and store them in a structured database.

[0010] S2: Use cross-modal contrastive learning to extract features and optimize the consistency of data in different modalities.

[0011] S3: embed features into a dynamic multimodal knowledge graph and update the relationships and node representations between enterprises.

[0012] S4: Extract and fuse modality sharing and specific features using modality sharing networks,

[0013] S5: Optimizing the consistency of modal feature distribution through adversarial learning,

[0014] S6: Initialize the global co-evolutionary credit scoring model to optimize accuracy, robustness and interpretability.

[0015] S7: Use dynamic multi-objective algorithms to optimize model accuracy, robustness, and interpretability.

[0016] S8: Introduce causal reasoning model to analyze the driving factors of corporate credit changes and establish causal relationships.

[0017] S9: Design a closed-loop feedback system to update the credit assessment model in real time and enhance adaptability.

[0018] S10: Generate explainable credit assessment reports based on generative adversarial networks to enhance transparency and credibility.

[0019] The specific process is as follows:

[0020] S1: Collect multimodal data related to commercial credit, including financial data, transaction records, market feedback and supply chain data. First, clean, normalize and fill missing values ​​of the collected multimodal data to ensure data quality. After these preprocessing steps, all data will be uniformly stored in a structured database to provide consistent input for subsequent processing and modeling;

[0021] S2 is as follows:

[0022] The cross-modal contrastive learning algorithm is used to extract features from data of different modalities, generate a unified embedding representation, and optimize modal consistency in the embedding space. Specifically, cross-modal contrastive learning ensures the similarity between data of different modalities by optimizing the following contrast loss function:

[0023]

[0024] in, The loss function representing the similarity of different modal data in the embedding space, For all samples Perform expectation calculations, and It is the data representation of different modes. express and The similarity between is the temperature coefficient, It represents the exponentially weighted sum of the similarities calculated for all possible samples. By minimizing this loss function, data from different modalities can maintain consistent representation in the same embedding space.

[0025] S3 is as follows: The construction steps of the dynamic multimodal knowledge graph are as follows: Graph initialization requires initializing the basic structure of the knowledge graph based on the enterprise and partnership information in the multimodal data: nodes are enterprise entities, edges are the relationships between them, and attributes include the embedded features extracted by S2. The weights of the edges are dynamically updated using the time window mechanism, and the weight update rules are defined as:

[0026]

[0027] in, Indicates at a point in time Time Node and nodes The edge weights between Indicates at a point in time Time Node and nodes The edge weights between is the attenuation factor, To increase the interaction intensity,

[0028] Then the GNN message passing mechanism is carried out. The node representation is learned through the graph neural network GNN, and the message passing mechanism based on time series is designed in combination with dynamics. The layer update formula is:

[0029]

[0030] in, It is The node table of the layer, is the activation function, Representation Node The neighbor set of It is The weight matrix of the layer, Representation Node In the Node representation of a layer.

[0031] S4 is as follows:

[0032] Based on the knowledge graph, the modality sharing network is used to extract modality shared features and modality specific features, and then merge them into collaborative representations through the fusion layer, which are optimized through the following loss function:

[0033]

[0034] in, is a modality-shared feature, are the specific characteristics of each mode, is the regularization parameter. This process optimizes the loss function so that shared features and specific features can be integrated in the same space.

[0035] S5 is as follows:

[0036] S5: Design an adversarial learning mechanism between the generator and the discriminator to optimize the distribution consistency between different modal features. Specifically, the generator and the discriminator The adversarial process between them is optimized by the following loss function:

[0037]

[0038] in, represents the adversarial loss function, which is used to optimize the adversarial learning process of the generator and the discriminator. Represents the generator, responsible for generating false data, the input is random noise , the output is the generated data ,

[0039] Represents the discriminator, which is used to distinguish whether the input data comes from the real data distribution , or the fake data generated by the generator, Represents real data From real data distribution , that is, the real data sample, Represents the discriminator for the real data The judgment value indicates that the discriminator believes is the probability of true data, : represents the input distribution from the generator The noise vector sampled in , It means that the generator will make noise Convert to Generate Data , The discriminator believes is the probability of generating data, The expected symbol indicates the expected value calculation of the sample. This mechanism ensures that the feature distribution between different modalities is consistent, and during the optimization process, the generated pseudo features can better align the distribution differences between modalities. The pseudo features of different modalities and the real features are fused and averaged to generate the final feature vector. .

[0040] S6: Based on the fused multimodal features, the global co-evolutionary credit scoring model is initialized. Through the multi-objective dynamic optimization algorithm, the model will be initially optimized on the three objectives of accuracy, robustness and interpretability, so that the model has basic evaluation capabilities.

[0041] S6 is as follows: initialization of the co-evolutionary credit scoring model, fusion feature vector As initial input, It contains cross-modal shared features and modality-specific features, with unified dimensions and aligned distribution through adversarial learning. Input a preliminary scoring model (Multi-layer Perceptron MLP) to generate a raw credit score , and then the loss function is defined based on this, and the mean square error (MSE) loss function is used to optimize the initial parameters of the model ,

[0042] The optimization goal is:

[0043]

[0044] in, is the loss value calculated when training the multi-layer perceptron (MLP) model in step S6, It is a sample Real credit score, is the total number of samples, are the initial parameters of the model,

[0045] is the predicted output of the multi-layer perceptron (MLP) model, representing the input feature vector A preliminary credit score generated after MLP calculation;

[0046] S7 is as follows:

[0047] S7: Introduce a dynamic multi-objective optimization algorithm, take the accuracy, robustness and interpretability of the model as optimization goals, use genetic algorithms to explore the optimal solution, and optimize through the following multi-objective optimization function:

[0048]

[0049] in, Indicates the prediction accuracy of the model (AUC indicator), represents the robustness of the model to data perturbations (measured by randomly adding noise), Represent the interpretability of the model (measured by feature importance score) and construct a comprehensive objective function , in, , , It is a dynamic weight, which adjusts the proportion according to the optimization needs at different stages;

[0050] S8 is specifically as follows: it uses a dynamic multimodal knowledge graph and combines it with a causal reasoning model to analyze the driving factors of changes in corporate credit and establish a causal relationship between credit changes and their upstream behaviors. The model infers the impact of different modal factors (such as market feedback, transaction relationships, etc.) on credit scores and realizes dynamic analysis.

[0051] Represent the company's market feedback events and transaction behaviors as potential variables and design a causal diagram ,in is a set of variables, including credit score, market feedback score, transaction frequency, is the causal relationship edge between variables. To quantify the contribution of different causal paths, the inverse probability weighted IPW method is used to calculate the direct effect DE and indirect effect, IE: ,When analyzing the impact of market feedback fluctuations on credit scores, the direct effect represents the direct effect of market feedback, and the indirect effect is reflected through mediating variables such as transaction behavior.,Combined with the dynamic multimodal knowledge graph DMKG, the edge weights and variable states in the causal model are updated in real time to reflect the latest association between corporate behavior and credit scores;

[0052] S9 is as follows:

[0053] S9: Design a closed-loop feedback system to dynamically update the credit assessment model by collecting corporate behavior data in real time and feeding the assessment results back to the model to enhance its adaptability. Specifically, feedback updates are performed through the following optimization formula:

[0054]

[0055] in, are the current model parameters, is the learning rate, is the gradient of the feedback loss function, and the optimization process dynamically adjusts the model parameters through this formula;

[0056] S10 is as follows:

[0057] S10: Based on the generative adversarial network (GAN), an explanatory report of credit assessment is generated, including risk sources, model decision paths and visualization results. Specifically, the explainable path of model decision is generated through the structure of the generative adversarial network, and it is presented to the user in the form of a report to enhance the trustworthiness of the model.

[0058] Furthermore, in order to ensure the quality and consistency of multimodal data in step S1, the specific operations include: first, data collection is performed, market feedback is obtained from public data sources (such as news and social media) through crawler technology, financial and transaction records are extracted from the internal ERP system of the enterprise using standardized interfaces, and logistics supply chain information is obtained in combination with API interfaces. Then, data cleaning is performed, regular expressions are used to clean non-standard characters and noise data, and principal component analysis (PCA) is combined to detect and remove outliers. Then, missing values ​​are filled, and corresponding methods are used for different modes, such as interpolation for time series data, statistical distribution inference for classified data, and KNN or prediction based on deep autoencoders are used to fill continuous values. Finally, normalization and storage are performed, the normalization process is integrated into a pipeline, and a unified structured database (PostgreSQL) is selected to store all preprocessed data to support subsequent batch query and analysis.

[0059] Furthermore, the specific implementation of cross-modal contrastive learning in step S2 requires the design of a feature extractor first, and the selection of a dedicated network architecture for different modal data, such as using CNN to extract spatial features of image modalities, Transformer to extract semantic features of text modalities, and GRU to extract dynamic features of time series modalities. Then, feature alignment is performed, and the relationship between modalities is modeled through shared parameters or self-attention mechanisms in the embedding layer, and shared layer weights are added to learn common representations of different modalities. Finally, the contrast loss is optimized, and combined with the online difficult sample mining mechanism, only the weight optimization is increased for sample pairs with low similarity to reduce redundant calculations of simple samples.

[0060] Furthermore, in the sharing and specific decomposition of modal features in step S4, when constructing the modal sub-network, a specific network is constructed for each modal data to generate modal specific features, and then shared features are extracted, and modal shared features are generated through an additional shared network, and the shared network structure is a multi-layer fully connected network. Finally, fusion optimization is performed, regularization constraints are introduced between shared and specific features, the fusion loss function is optimized, and the distribution difference between modal specific features and shared features is eliminated using the distribution alignment method.

[0061] Use the gradient descent method based on the Adam optimizer to update the model parameters , so that the model gradually learns to generate reasonable initial scoring results under multimodal feature input. The performance on the validation set is monitored through the early stopping mechanism to prevent overfitting. After the initialization phase is completed, the model scoring results are interactively verified with the multimodal dynamic knowledge graph (step S3) to ensure that the scoring output by the model is consistent with the behavioral characteristics of the associated nodes in the graph (such as counterparties, partners), so as to have collaborative perception capabilities.

[0062] Furthermore, the co-evolutionary credit scoring model in step S7 further improves its performance through a dynamic multi-objective optimization method, with particular attention paid to the accuracy, robustness and interpretability of the model. Three core objective functions are set, and the proportions are adjusted according to the optimization requirements of different stages. Then the genetic algorithm is implemented, and a genetic algorithm (GA) is used for global search. A set of random parameters is generated when the population is initialized, and it is iterated through the following steps: Selection: Select suitable individuals according to fitness. Crossover: Crossover operations are performed on individuals with high fitness to generate new solutions. Mutation: Random perturbations are made to some parameters to explore the global solution space. During the optimization process, the constraints (scoring deviations and risk tolerance) are dynamically adjusted to ensure that the searched solution not only meets the technical indicators, but also adapts to the requirements of the actual application scenarios. Finally, through dynamic multi-objective optimization, the model gradually converges to the Pareto optimal solution, achieving a balance between performance and practicality.

[0063] Furthermore, in step S8, in order to identify the driving factors of corporate credit changes, a causal reasoning model is introduced, and a causal analysis is performed in combination with a dynamic multimodal knowledge graph. The relationship between credit scores and their upstream behaviors is modeled using a structural causal model (SCM). The market feedback and transaction behaviors of the enterprise are represented as potential variables, and a causal graph is designed. ,in is a set of variables, including credit score, market feedback score, transaction frequency, is the causal relationship edge between variables. To quantify the contribution of different causal paths, the inverse probability weighting (IPW) method is used to calculate the direct effect (DE) and indirect effect (IE): When analyzing the impact of market feedback fluctuations on credit scores, direct effects represent the direct effects of market feedback, and indirect effects are reflected through mediating variables such as transaction behavior. Combined with the dynamic multimodal knowledge graph (DMKG), the edge weights and variable states in the causal model are updated in real time to reflect the latest associations between corporate behavior and credit scores. The causal reasoning results are visualized to generate a clear causal path diagram to help users understand which factors have the greatest impact on credit scores and their mechanisms of action. Through causal analysis, the causal relationship between complex multimodal data can be revealed, providing in-depth insights into credit changes.

[0064] Furthermore, in step S9, a closed-loop feedback mechanism is designed to dynamically adjust the credit assessment model so that it can respond to changes in corporate behavior in real time. An automated data collection pipeline is built to capture real-time behavioral data of enterprises (such as new transaction records and the latest market feedback). Low-latency transmission of data streams is ensured to support real-time updates of the model. : Using an incremental learning framework, the model parameters are updated using the following formula:

[0065]

[0066] in, are the current model parameters, is the learning rate, is the gradient of the feedback loss function, and the optimization process dynamically adjusts the model parameters through this formula; the goal of feedback learning is to maximize the adaptability of the model to new data while retaining the knowledge of historical data. Introduce dynamic monitoring indicators (model prediction deviation, data distribution drift detection) to analyze in real time whether the model performance reaches the set threshold. If the model performance degrades, trigger a larger-scale retraining operation. Collect expert feedback through the user interface, integrate human evaluation results into model adjustments, and further improve the credibility and applicability of the model. The closed-loop feedback mechanism ensures the long-term stability and flexibility of the credit assessment model, enabling it to cope with complex dynamic environments.

[0067] Furthermore, in step S10, in order to enhance the transparency of the model and user trust, a generative adversarial network (GAN) is introduced to generate an explanatory report for credit assessment. The SHAP value (Shapley Additive Explanations) is used to calculate the contribution of multimodal features to credit scoring, generate a feature importance ranking diagram, and reveal key risk factors (such as transaction frequency and market feedback negativity). The decision-making process of the model is simulated by the generator of GAN, and a path diagram is generated to show the logical flow of credit scoring from input features to output scores. The discriminator is used to verify the credibility of the path diagram to ensure that the generated explanation has real meaning. The analysis results are presented in a multimodal form (charts, text, heat maps), so that users can intuitively understand the basis for scoring and risk distribution, and use sankey diagrams to show the causal flow between features. The report content is dynamically adjusted according to user needs, and scenario-based explanations are added. Finally, the explanatory report enhances the transparency of the model and improves users' trust and acceptance of the evaluation results.

[0068] Compared with the prior art, the present invention has the following advantages:

[0069] This paper proposes a corporate credit assessment method based on multimodal co-evolution, which makes full use of the multimodal characteristics of corporate financial data, transaction records, market feedback and supply chain data, and overcomes the processing limitations of single-modal analysis methods on data complexity and diversity. Through cross-modal comparative learning, the unified optimization of embedding representation of different modal data is achieved, and the consistency modeling between multimodal data is realized, thereby improving the reliability and accuracy of credit assessment without relying on the quality of a single modality;

[0070] This paper proposes a collaborative credit analysis framework based on a dynamic multimodal knowledge graph, and constructs a graph structure of dynamic associations between enterprises. By introducing the message passing mechanism of the graph neural network (GNN), the node representation of the enterprise can be dynamically updated, so that the model has the ability to handle the dynamic behavior and relationship changes of the enterprise. Compared with the traditional static credit analysis method, the dynamic multimodal knowledge graph can capture the potential relationship between credit scores and the behavior of upstream and downstream enterprises, thereby providing more interpretable and traceable credit assessment results;

[0071] This paper designs a feature alignment fusion mechanism through a modality sharing network, balances the sharing characteristics between multimodal data and the modality independence characteristics, and optimizes the representation ability of embedded features. Compared with traditional direct splicing or weighted fusion methods, the network effectively avoids information loss and modality conflict problems in the feature space, greatly improving the robustness and generalization ability of the evaluation model;

[0072] The present invention proposes a modal distribution consistency optimization method based on generative adversarial learning. Through the adversarial training mechanism of the generator and the discriminator, the generation of pseudo features and the alignment of the distribution between modalities can be carried out simultaneously, thereby effectively solving the problem of uneven distribution of multimodal data. Compared with the existing methods, the adversarial optimization mechanism of the present invention not only improves the global consistency of the model, but also significantly reduces the risk of deviation in feature representation;

[0073] This paper designs a collaborative evolution model based on dynamic multi-objective optimization, takes the accuracy, robustness and interpretability of credit assessment as optimization goals, and dynamically adjusts model parameters through the global search capability of genetic algorithms. Compared with single-objective optimization methods, dynamic multi-objective optimization can balance multiple performance requirements in complex scenarios, improving the applicability of the model in real business environments;

[0074] This invention combines causal reasoning models to conduct in-depth analysis of the driving factors of corporate credit changes. Through the dynamic analysis capabilities of causal reasoning, the model can infer the causal relationship between corporate behavior and credit scores, thereby identifying the key factors that affect credit changes. Compared with traditional correlation analysis methods, causal reasoning can avoid pseudo-correlation problems and provide more reliable data support for credit decisions;

[0075] The present invention designs a closed-loop feedback mechanism that combines the real-time collected corporate behavior data with the model's credit scoring results to build a dynamically updated adaptive credit assessment system. Through incremental learning and real-time feedback, the model can continuously optimize its predictive capabilities, significantly improve its adaptability to rapidly changing business environments, and reduce maintenance costs;

[0076] This invention generates an explanatory report of credit assessment through generative adversarial networks (GAN), including risk sources, decision paths, and visual analysis results. Different from traditional credit assessment models, this method emphasizes transparency and user trust, and can provide clear and understandable credit analysis reports for corporate managers and financial institutions, thereby enhancing the acceptability of assessment results;

[0077] The multimodal co-evolutionary algorithm and machine learning method proposed in this invention are real-time, flexible and adaptive, and can adapt to complex business environments and changing corporate behavior characteristics. Through the combination of model optimization and knowledge graph, the whole process from data processing to credit scoring is realized, providing a novel and efficient solution for corporate credit analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 This is the overall flow chart of the evaluation and supervision method of the present invention.

[0079] Figure 2 A schematic diagram of the feature extractor design provided by the present invention,

[0080] Figure 3 A detailed flow chart of the consistency of the adversarial learning optimization modal feature distribution provided by the present invention,

[0081] Figure 4 A schematic diagram of the multimodal feature interactive verification results provided by the present invention,

[0082] Figure 5 This is a schematic diagram of the enterprise credit assessment results provided by the present invention. DETAILED DESCRIPTION

[0083] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0084] Example: Company A is a large global manufacturing company with a business network covering suppliers, distributors and end customers, with a large transaction volume and complex processes. In order to obtain more efficient financing support and optimize risk management, Company A needs to accurately assess its credit status. However, this task faces multiple challenges, including supply chain complexity, transaction diversity, market feedback volatility, etc. Company A's supply chain includes hundreds of suppliers, logistics companies and distributors. Due to its wide geographical distribution, the supply chain faces problems such as transportation delays and payment arrears. In addition, supply chain data is discrete and in various formats, making it difficult to integrate effectively. Company A processes millions of transaction records every month, which come from ERP systems, supply chain management platforms and external cooperation systems. Companies are susceptible to market feedback due to their industry characteristics and global layout. For example, negative news may lead to a decline in credit ratings, while good news may promote financing facilitation. In this context, traditional credit assessment models face many technical limitations and are difficult to meet the actual needs of Company A. Traditional credit assessment models often only use financial data (such as balance sheets, profit and loss statements) or transaction data (such as collection rates, contract performance) for analysis. Multimodal data such as market feedback data and supply chain information have not been effectively integrated, resulting in a lack of a global perspective on credit assessment results. Many traditional models cannot capture dynamic changes in enterprises in real time, especially in terms of transaction frequency and market feedback fluctuations. Model updates rely on manual operations, resulting in evaluation results lagging behind actual conditions. Past models lacked the ability to deeply mine multimodal data, making it difficult to uniformly represent structured and unstructured data. The features between different data sources were isolated from each other and could not form a comprehensive analysis capability. The patent of this invention overcomes the limitations of traditional credit assessment models through innovative technical solutions and provides a dynamic credit assessment method based on multimodal data. Its core technical advantages include: proposing a cross-modal comparative learning method to uniformly represent financial data, transaction records, market feedback information and supply chain data. By aligning different modal features through a shared embedding space, the problem of data heterogeneity is eliminated, and a multimodal joint feature representation is constructed. Based on the dynamic multimodal knowledge graph (DMKG), the relationship dynamics between an enterprise and its supply chain partners can be captured in real time. Through the edge weight update algorithm, dynamic changes such as transaction intensity and market feedback propagation effects are accurately reflected. Causal reasoning analysis is introduced, and the structural causal model (SCM) is used to identify the causal relationship between features, thereby improving the transparency of credit scoring. The generated explanation report can clearly show the direct and indirect effects of key features, supporting decision makers to quickly identify the source of risk. A dynamic optimization framework based on incremental learning is proposed, which realizes the continuous updating of the model by capturing new data (such as transaction records and market feedback) in real time. This mechanism significantly improves the timeliness and robustness of credit assessment results.This patent breaks through the limitations of traditional credit assessment technology through deep integration and dynamic modeling of multimodal data, and provides an accurate, efficient and transparent credit assessment solution for Enterprise A. This patent not only helps enterprises cope with complex supply chain environments, but also provides new ideas and methods for future credit risk management.

[0085] like Figure 1 As shown in the figure, it is a flow chart of the present invention, and the commercial credit evaluation and supervision method based on the multimodal co-evolutionary algorithm comprises the following steps:

[0086] S1: When preparing for the credit assessment model, Enterprise A collects, cleans, fills, normalizes and stores multimodal data. The specific goal is to ensure the quality and consistency of various data through precise steps, and provide high-quality input data for subsequent modeling and analysis.

[0087] S1.1. Data collection

[0088] In the business scenario of Enterprise A, data comes from multiple sources, including financial data, transaction records, market feedback, and supply chain data. These data must be fully collected and uniformly managed for subsequent processing.

[0089] (1) Financial data and transaction records collection: Financial data and transaction records are usually stored in the enterprise's ERP system and supply chain management platform. In order to extract this data, a standardized interface can be used to regularly extract the following information by connecting to the internal system:

[0090] Table 1 Financial data collection

[0091]

[0092] Table 2 Transaction record data collection

[0093]

[0094] This data can provide a detailed record of a company's financial health and operating activities, helping to assess its creditworthiness.

[0095] (2) Collection of market feedback information: Market feedback information is crucial to the credit assessment of enterprises, because public sentiment has a direct impact on the credit score of enterprises. Market feedback data can be collected from multiple public data sources through crawler technology, including: news websites (industry-related media, etc.); social media platforms; industry forums and consumer evaluation websites. Use the crawler tool (Scrapy), set keywords and crawling rules, and crawl news, comments and social media posts related to the enterprise from these platforms. This data will help analyze public sentiment, especially the potential impact of negative news and comments on corporate credit;

[0096] (3) Supply chain data collection: Enterprise A’s supply chain consists of multiple suppliers, logistics companies, and distributors. In order to fully understand the operation of the supply chain, supply chain information can be obtained directly from the partner’s data system through the API interface, including:

[0097] Table 3 Supply chain data collection

[0098]

[0099] This data can reveal potential risks in a company's supply chain management and whether it can meet its financial obligations on time.

[0100] S1.2. Data cleaning

[0101] The collected data often has various problems, such as inconsistent formats, non-standard characters, noisy data, etc., so it needs to be cleaned. The goal of data cleaning is to improve the quality and consistency of the data.

[0102] (1) Non-standard character cleaning

[0103] Market feedback may contain HTML tags, emoticons or other non-standard characters from different platforms. By using regular expressions to clean up these useless characters, the standardization of text data can be ensured.

[0104] Algorithm 1 Non-standard character cleaning method

[0105]

[0106] (2) Outlier detection: In financial data and transaction records, outliers may be introduced due to input errors, system failures, or other reasons. To ensure the rationality of the data, the principal component analysis (PCA) method is used to identify and remove outliers. PCA can identify abnormal values ​​and remove them by analyzing the variance and covariance of the data.

[0107] Algorithm 2 PCA dimensionality reduction process method

[0108]

[0109] (3) Format uniformity: When data comes from different sources, inconsistent formats are a common problem. For example, the date format may be "YYYY-MM-DD" or "MM / DD / YYYY".

[0110] S1.3. Missing value imputation

[0111] Missing data is a common challenge in data cleaning. For different types of missing data, you can choose appropriate methods to fill them.

[0112] (1) Time series data filling: For time series data in financial or transaction records (monthly sales data, inventory levels, etc.), interpolation can be used to fill missing values. Commonly used interpolation methods include linear interpolation and polynomial interpolation;

[0113] (2) Categorical data filling: For some categorical data (whether payment was made on time), missing value filling usually relies on statistical inference;

[0114] (3) Continuous value filling: For the possible missing continuous values ​​(missing account amounts) in financial data, KNN (K-Nearest Neighbors) can be used;

[0115] Algorithm 3 KNN interpolation method for processing missing continuous values

[0116]

[0117] S1.4. Data normalization

[0118] When performing data modeling, different types of input data (such as financial amounts, sentiment scores, transaction frequencies, etc.) may have different dimensions or ranges. In order to ensure the consistency and comparability of the data, all data needs to be normalized.

[0119] (1) Normalization method selection Normalization methods include Min-Max normalization and Z-score normalization. Min-Max normalization can convert data to the [0, 1] interval, while Z-score normalization can make the mean of the data 0 and the standard deviation 1;

[0120] (2) Pipeline processing To ensure the efficiency of normalization processing, the normalization process can be designed as a pipeline to ensure that each batch of data is processed according to a unified standard and avoid data errors caused by inconsistent operations;

[0121] S1.5. Data Storage

[0122] All cleaned and normalized data needs to be stored in a structured database for subsequent query and analysis. In order to support subsequent data query and efficient processing, the PostgreSQL database is used.

[0123] Through the above steps, Company A can ensure that the financial data, transaction records, market feedback and supply chain data collected from multiple sources can be effectively cleaned, filled, normalized and stored. After careful processing, these data will provide a solid data foundation for the credit assessment model and support Company A's risk management and financing optimization in global operations;

[0124] S2 cross-modal contrastive learning can help extract useful features from different sources (such as financial data, transaction records, market feedback information, and supply chain data) and map them into the same embedding space for unified analysis and further modeling. Specifically, by optimizing the consistency between modalities through contrastive learning, it can ensure that data information from different sources can work together to enhance the accuracy of credit assessment.

[0125] S2.1. Feature Extractor Design

[0126] S2 first needs to design adaptive feature extractors for each data modality. These feature extractors need to be efficient and targeted to extract meaningful features from each modality and map these features to a unified embedding space to facilitate subsequent comparative learning and feature fusion.

[0127] S2.1.1 Financial data feature extraction

[0128] The financial data collected by S1 is structured numerical data, including key indicators such as profit, balance sheet, cash flow, etc. In order to capture the potential patterns and trends in these data, a fully connected neural network (FCN) is used as a feature extractor. The multi-layer perception mechanism of FCN can learn the deep characteristics of the financial health of the enterprise, such as long-term profitability, debt repayment ability and liquidity characteristics;

[0129] S2.1.2 Transaction record feature extraction

[0130] Transaction records are dynamic time-series data, usually containing information such as timestamps, transaction amounts, and transaction parties. This type of data is characterized by its time dependency and pattern volatility, and is processed using a gated recurrent unit (GRU). The GRU network can efficiently model dynamic changes in transaction records, capture short-term and long-term patterns in historical transactions, and identify seasonal fluctuations and abnormal behaviors;

[0131] S2.1.3 Market feedback information feature extraction

[0132] Market feedback data mainly comes from unstructured text data, such as news reports, social media comments, etc. Here, the Transformer network is used to extract semantic features in market feedback information. The Transformer's multi-head attention mechanism can effectively parse text content, extract key semantics and sentiment tendencies, and thus quantify the potential impact of public opinion on corporate credit;

[0133] S2.1.4 Supply Chain Data Feature Extraction

[0134] Supply chain data involves multi-dimensional dynamic information, such as order fulfillment, logistics progress, payment status, etc. In the feature extraction design, a GRU-based temporal network is used to process the dynamic changes of the supply chain and extract time-related features; for the spatial characteristics, a convolutional network (CNN) can be combined to capture the local correlation in the supply chain.

[0135] Figure 2 This is a schematic diagram of the feature extractor design of the present invention. The specific process is as follows:

[0136] S2.2. Feature alignment

[0137] In the process of cross-modal learning, data from different modalities have different feature spaces, so they need to be aligned into a unified embedding space. In this step, feature alignment is mainly performed through the following methods:

[0138] S2.2.1 Embedding layer alignment The data of each modality is converted into a high-dimensional vector representation through an independent feature extractor. Subsequently, the feature vectors of all modalities will be mapped to a common embedding space. Specifically, feature alignment is achieved through the following techniques: By designing shared network layer weights, the representations of different modalities are mapped to the same space. This means that the same parameters will be used to process financial data and transaction record features to ensure that they have consistent representations in the same space. The self-attention mechanism can effectively capture the correlation between different modalities. When processing market feedback data, the implicit relationship between market feedback and transaction records and financial data can be learned through the self-attention mechanism, thereby improving the effect of feature alignment.

[0139] S2.2.2.2 Embedding Space Alignment The feature vectors of all modalities are converted into low-dimensional embedding representations through a shared embedding layer. These embedding representations will be able to better reflect the similarities between data and lay the foundation for subsequent contrastive learning optimization.

[0140] S2.3. Optimizing contrastive loss. In this step, a contrastive loss function is used to optimize the similarity between modalities.

[0141] S2.3.1 Contrastive loss function design In contrastive learning, the loss function is usually used to quantify the similarity of sample pairs in the embedding space. The contrastive loss function used by Enterprise A is as follows:

[0142]

[0143] in, The loss function representing the similarity of different modal data in the embedding space, For all samples Perform expectation calculations, and It is the data representation of different modes. express and The similarity between is the temperature coefficient, Represents the exponentially weighted sum of the similarities calculated for all possible samples. By minimizing this loss function, data from different modalities can be consistently represented in the same embedding space.

[0144] S2.3.2 Dynamically Adjusting Temperature Parameters In cross-modal contrastive learning, the temperature parameter Controls the difficulty of similarity calculation. Enterprise A can set dynamic temperature parameters to adaptively adjust the temperature value according to the difficulty of different modal alignments. The temperature is higher in the early stage to increase the fault tolerance of inter-modal alignment. As the training progresses, the temperature is gradually lowered to strengthen the matching degree between modalities.

[0145] S2.4. Optimization and training. Finally, the entire model will be trained by minimizing the contrastive loss function. In this process, all data modalities will be trained as consistent representations in the same embedding space. The training steps include: inputting the data of each modality into the corresponding feature extractor, and calculating the embedded representation through network forward propagation. Use the contrastive loss function to calculate the similarity between different modalities. Update the network parameters through backpropagation, minimize the loss function, and optimize the modal consistency. Use the validation set to evaluate the model performance, and adjust the hyperparameters such as learning rate and temperature coefficient when necessary;

[0146] Through cross-modal contrastive learning, Enterprise A can effectively unify heterogeneous data from different sources (such as financial data, transaction records, market feedback, and supply chain data) into the same embedding space, thereby improving the accuracy of credit assessment;

[0147] S3 integrates the embedding representation generated by cross-modal contrastive learning (S2) into the dynamic multimodal knowledge graph (DMKG), which is an important step to achieve complex relationship visualization and in-depth analysis. By dynamically updating the node and edge weights of the knowledge graph, enterprise A can fully characterize its interaction network with suppliers, distributors and end customers, supporting accurate credit assessment and risk management;

[0148] S3.1. Initialization of the knowledge graph

[0149] The basic construction of the knowledge graph is based on the multimodal data of enterprise A, mainly including the definition of nodes and edges and attribute initialization. The nodes of the knowledge graph represent enterprise entities (such as enterprise A itself, suppliers, distributors and customers), and each node carries a multimodal embedding representation generated by S2 as the initial feature attribute. (Features of the node "Enterprise A": combined with financial data embedding, transaction behavior embedding and market feedback information embedding; features of the node "Supplier X": combined with transaction record embedding and logistics delay indicators) At the same time, the edges between nodes represent the relationship between enterprises, including transactions, cooperation, market feedback, etc. The initial edge weights are based on the interaction intensity extracted from the multimodal data. (The transaction frequency and amount between enterprise A and supplier X are used as edge weights; the edges in the direction of market feedback reflect the flow of information in the network)

[0150] The initial edge weight is calculated using the following formula:

[0151] ,

[0152] in, Indicates at a point in time Moment, Node and nodes The edge weights between . For Node and nodes Extracting interaction strength in multimodal data.

[0153] S3.2. Time window mechanism dynamically updates the graph

[0154] In order to reflect the dynamic characteristics of the enterprise network, the knowledge graph needs to be dynamically updated through the time window mechanism, especially the adjustment of edge weights. The transaction data and market feedback of enterprise A have obvious time dependence, so the data is sharded by time period. The edge weight update rule uses the following formula to dynamically update the edge weight:

[0155]

[0156] in, Indicates at a point in time Time Node and nodes The edge weights between Indicates at a point in time Time Node and nodes The edge weights between is the attenuation factor, To increase the interaction intensity,

[0157] S3.3. Graph Neural Network (GNN) Message Passing Mechanism

[0158] The node representation of the knowledge graph needs to be continuously optimized through the message passing mechanism of the graph neural network (GNN) to capture the complex dynamic relationships between enterprises.

[0159] Node feature update formula

[0160]

[0161] in, It is The node table of the layer, is the activation function, Representation Node The neighbor set of It is The weight matrix of the layer, Representation Node In the Node representation of a layer.

[0162] In the message passing process, in the first layer, the embedded features of S2 are used as the initial input to calculate the messages contributed by the neighbor nodes of each node. In the subsequent layers, the neighbor node information is summarized layer by layer, and higher-order feature representations are learned through weight matrices and activation functions. The node representation of enterprise A can be updated through the embedding of its supplier and distributor nodes to capture the complex relationships and dynamic changes of the supply chain. Dynamic knowledge graphs need to combine the time dimension and incorporate time series information when passing messages.

[0163] S3.4. Result output and application

[0164] After iterative updates of the graph neural network, the dynamic multimodal knowledge graph of enterprise A is constructed, and the following results are output:

[0165] Final feature representation of the node It reflects the comprehensive characteristics of the enterprise and can be used to further assess its credit risk. Potential high-risk suppliers can be discovered through cluster analysis. Through the dynamic changes of edge weights and node characteristics, enterprise A can monitor the health of its supply chain and customer network in real time. The knowledge graph also serves as the input of subsequent prediction models (risk prediction, anomaly detection), ensuring that the model reflects the latest status through dynamic updates.

[0166] Based on the dynamic multimodal knowledge graph (DMKG) of enterprise A extracted in S3, S4 further uses a modal sharing network to extract modal sharing features and modal specific features to achieve deep fusion of multimodal information. This process optimizes the embedded representation of the knowledge graph, enabling it to support more accurate credit assessment and risk prediction.

[0167] S4.1. For the multimodal data of enterprise A (transaction data, market feedback information, supply chain dynamics, etc.), first build an independent modality-specific sub-network for each modality to extract modality-specific features. The data of each modality is extracted from DMKG. The transaction data modality is the embedding feature of enterprise A's transaction records. The market feedback modality is the embedding of enterprise-related market feedback extracted from news and social media. The supply chain modality is the embedding feature of supplier relationships and logistics dynamics. These data are input into the corresponding specific sub-networks. The data of each modality is generated through a specific network to generate modality-specific features. These features retain the unique information of the modality and facilitate the subsequent fusion of shared features.

[0168] S4.2. Modality sharing network extracts shared features from data of all modalities , which is the common representation between modalities. The shared network uses a multi-layer fully connected network (MLP), and the input is a combination of specific features of each modality. Output as shared features The shared network generates shared features through training ,These features integrate the cross-features of public information of all modalities, transaction data and supply chain dynamics, reflecting the comprehensive operating status of the enterprise; the correlation between market feedback and transaction behavior.

[0169] S4.3. Specific characteristics and shared features It is necessary to optimize them into collaborative representations through the fusion layer to ensure that they are distributed aligned in the same embedding space. The fusion layer introduces regularization constraints and is optimized through the following loss function:

[0170]

[0171] in, is a modality-shared feature, are the specific characteristics of each mode, This process optimizes the loss function to integrate shared features and specific features in the same space.

[0172] The first term constrains the overall consistency of shared features and modality-specific features; the second term controls the difference between shared features and single-modality features through regularization. The final collaborative representation It is a comprehensive characterization of the multimodal data of Enterprise A.

[0173] S5.1 Optimizing the consistency of feature distribution between modalities through adversarial learning. In order to further optimize the consistency of feature distribution of multimodal data of enterprise A, a generator is introduced based on the previously generated collaborative representation. ) and the Discriminator ) through adversarial learning mechanism. Through adversarial training, the pseudo-feature distribution of multimodal features is aligned with the real feature distribution, thereby improving the adaptability of the model to credit assessment and risk prediction.

[0174] Generator and the discriminator The adversarial process between them is optimized by the following loss function:

[0175]

[0176] in, represents the adversarial loss function, which is used to optimize the adversarial learning process of the generator and the discriminator. Represents the generator, responsible for generating false data, the input is random noise , the output is the generated data ,

[0177] Represents the discriminator, which is used to distinguish whether the input data comes from the real data distribution , or the fake data generated by the generator, Represents real data From real data distribution , that is, the real data sample, Represents the discriminator for the real data The judgment value indicates that the discriminator believes is the probability of true data, : represents the input distribution from the generator The noise vector sampled in , It means that the generator will make noise Convert to generate data , The discriminator believes is the probability of generating data, The expected symbol indicates the calculation of the expected value of the sample. The pseudo features and real features of different modes are fused and averaged to generate the final feature vector. .

[0178] S5.2 Modal feature distribution alignment

[0179] In order to achieve consistency in feature distribution, an adversarial learning mechanism is used to adjust the pseudo feature distribution output by the generator to align it with the real feature distribution. . Use Wasserstein distance instead of the original binary cross entropy loss, the formula is:

[0180]

[0181] Wasserstein distance can more accurately quantify the distance between the true distribution and the generated distribution, while improving the stability of adversarial training. The generator and the discriminator are updated alternately, and the generator optimizes the target Make the pseudo features closer to the real feature distribution; the discriminator enhances the ability to distinguish between real features and pseudo features through continuous learning.

[0182] S5.3 Model training and validation

[0183] Adversarial training uses a step-by-step iterative approach to adjust the performance of the generator and the discriminator so that the entire system reaches the Nash equilibrium state: Training the discriminator , before each generator update, fix , through the input of real features and pseudo features, optimize the parameters of DDD. The loss function of training the discriminator is:

[0184]

[0185] Training the Generator ,After the discriminator is stable, fix , generate pseudo features through noise distribution, update Parameters.

[0186] The training generator loss function is:

[0187] After alternating training for multiple epochs, it is ensured that the pseudo features generated by the generator cannot be effectively distinguished as pseudo features in the discriminator.

[0188] Figure 3 The following is a detailed flow chart of optimizing the consistency of modal feature distribution through adversarial learning of the present invention. The relevant steps are as follows:

[0189] In order to help enterprise A evaluate the credit status in a complex supply chain, S6 builds and initializes a global co-evolutionary credit scoring model based on the fusion of multimodal features (step S5). This model uses a multi-objective dynamic optimization algorithm to perform preliminary optimization in terms of accuracy, robustness, and interpretability, forming a credit scoring mechanism with collaborative perception and dynamic adaptability.

[0190] S6.1. Input of multimodal features and construction of scoring model

[0191] The fused multimodal feature vector (output from step S5) is used as the initial input of the credit scoring model. It contains cross-modal shared features and modality-specific features, and the feature distribution has been optimized through adversarial learning. The preliminary scoring model uses a multi-layer perceptron (MLP) to map the input features into the original credit score. , the specific process is as follows: The input dimensions of the model are aligned to ensure the consistency and compatibility of features between multimodal data. The input features include supply chain transaction features, financial features, market feedback features, etc., which represent the factors affecting the global credit of the enterprise. The model consists of a multi-layer neural network: the input layer receives features The hidden layer is responsible for capturing complex nonlinear relationships, and the output layer generates a single-value credit score. The activation function is Activate to improve the model's nonlinear fitting ability.

[0192] Define the loss function and use the mean square error (MSE) loss function to optimize the initial parameters of the model ,

[0193] The optimization goal is:

[0194] ,

[0195] in, is the loss value calculated when training the multi-layer perceptron (MLP) model in step S6, It is a sample Real credit score, is the total number of samples, are the initial parameters of the model,

[0196] is the predicted output of the multi-layer perceptron (MLP) model, representing the input feature vector A preliminary credit score generated after MLP calculation.

[0197] S6.2. Model training and parameter optimization, using the gradient descent method based on the Adam optimizer to adjust the initial model parameters The Adam optimizer is suitable for processing multimodal complex data with its adaptive learning rate characteristics. The data set is divided into training set, validation set and test set, and the model performance is gradually optimized through iterative training. During the training process, the loss value on the validation set is monitored to prevent the model from overfitting. The early stopping mechanism will terminate the training when the performance of the validation set no longer improves, ensuring the generalization ability of the model.

[0198] S6.3. Dynamic knowledge graph interactive verification, initial scoring results Input the dynamic multimodal knowledge graph (constructed in step S3) for interactive verification: Check whether the scoring results are consistent with the behavioral characteristics of nodes (such as counterparties, suppliers) and edges (such as transaction records, partnerships) in the graph. If the high-frequency default behavior of a node conflicts with the scoring result, adjust the model weight to make it more consistent with the graph facts.

[0199] Figure 4 The following is a schematic diagram of the multi-modal feature interactive verification results of the present invention,

[0200] In order to further improve the performance of the co-evolutionary credit scoring model of Enterprise A, S7 takes accuracy, robustness and interpretability as optimization goals and uses a dynamic multi-objective optimization algorithm to globally optimize the model. The genetic algorithm (GA) is used to explore the optimal solution under the multi-objective optimization function, so that the model performance meets the actual needs of complex business networks.

[0201] S7.1 First, define the optimization objective function of the model , taking into account accuracy, robustness and interpretability:

[0202]

[0203] in, Indicates the prediction accuracy of the model (AUC indicator), represents the robustness of the model to data perturbations (measured by randomly adding noise), Represent the interpretability of the model (measured by feature importance score) and construct a comprehensive objective function , in, , , It is a dynamic weight, and the proportion is adjusted according to the optimization needs at different stages.

[0204] In the initial stage of optimization, priority is given to improving accuracy (higher ). During mid-term optimization, balance robustness and interpretability (adjust and ), in the later optimization, the weights are dynamically fine-tuned according to specific needs.

[0205] S7.2. Using genetic algorithms to implement multi-objective optimization and randomly generate a set of model parameters (individual), as the initial solution of the population. The fitness of each individual is determined by the comprehensive objective function Calculated.

[0206] Algorithm 4: Genetic algorithm optimization process:

[0207]

[0208] Through multi-objective optimization, the model gradually approaches the Pareto frontier and the trade-off optimization points between different objectives (such as the balance between AUC improvement and robustness enhancement). The final results cover multiple optimal solutions for actual business selection. The validation set is used to test the model's AUC, robustness score changes, and feature importance interpretation results to ensure that the model performance meets the expected level.

[0209] The final model is embedded in the supply chain management system of Enterprise A to assist in dynamic credit scoring.

[0210] S8 In order to further help Enterprise A understand the reasons behind the changes in credit scores, we use the dynamic multimodal knowledge graph (DMKG) combined with the causal reasoning model to analyze the driving factors of corporate credit changes. By building a structural causal model (SCM), we establish a causal relationship between credit scores and multimodal upstream behaviors (such as market feedback, transaction relationships, cooperation networks, etc.), thereby achieving dynamic and explainable credit score optimization.

[0211] S8.1. Build a causal reasoning framework and determine the set of causal variables and relationships. Extract key nodes and relationships from the dynamic multimodal knowledge graph, including: Nodes: the company's credit score (target variable), market feedback score, transaction frequency, cooperation intensity, etc.

[0212] Edges: Dynamic causal connections between these nodes (e.g. market feedback score affects transaction frequency, and transaction frequency affects credit score).

[0213] Model this information as a causal graph ,in: Represents a set of variables. Represents a causal edge.

[0214] In view of the relationship between different variables, the inverse probability weighting (IPW) method is introduced to quantify the causal path: Direct Effect (DE): The direct impact of a variable (such as market feedback score) on the target variable (credit score). Indirect Effect (IE): The role of a variable indirectly affecting the target variable through an intermediary variable (such as transaction behavior).

[0215] The effect formula is:

[0216] ,

[0217] S8.2. Obtain real-time updated data of corporate behavior from the knowledge graph (step S3): Market feedback events (news, social media) are converted into market feedback scores through sentiment analysis algorithms. Corporate transaction behavior generates transaction intensity by analyzing transaction frequency, amount, etc.

[0218] Changes in the cooperation network (new or interrupted cooperation) are reflected as changes in the intensity of cooperation. These features are constantly updated as the time window changes, ensuring that the causal analysis is based on the latest data. In the structural causal model, the weights of the edges of the causal graph (indicating the strength of the causal relationship) are updated based on the knowledge graph. If there is a surge in negative market feedback events in the recent period and the transaction frequency decreases, the weight of the "market feedback → credit score" edge will increase, and the weight of the "transaction frequency → credit score" edge will decrease.

[0219] S8.3. Causal Reasoning and Analysis Process

[0220] In the causal diagram, the structural equation model is used to analyze the causal relationship between variables. By calculating the edge weights on the key path, the total contribution of different paths to the credit score is quantified. The causal effect decomposition is achieved: the direct and indirect effects of each variable are clarified. Assuming that the credit score of enterprise A has recently shown a downward trend, the specific reasoning process is as follows:

[0221] Direct effect: the impact of market feedback events. Negative market feedback scores directly reduce credit scores. The inference results show ,

[0222] Indirect effect: Through the influence of trading behavior, market feedback events lead to a decrease in trading frequency, which indirectly reduces credit scores. The inference results show that .

[0223] Total effect calculation:

[0224]

[0225] By combining dynamic multimodal knowledge graphs with causal reasoning, Enterprise A successfully built a dynamic analysis framework that can identify the key drivers of credit changes and their action paths in real time. This analysis method not only provides enterprises with a clear explanation of causal relationships, but also helps them formulate precise risk management strategies and effectively respond to credit risk challenges in complex business networks.

[0226] S9In the complex business network of Enterprise A, credit assessment needs to have real-time response capabilities to cope with dynamic changes in corporate behavior and external environment. The core of the closed-loop feedback system is to achieve real-time data collection, dynamic model update and expert feedback integration, so that the credit assessment model can be continuously optimized, adapt to changes and operate stably.

[0227] S9.1. Build a real-time data collection and update pipeline, build a low-latency automated data collection pipeline, and capture the behavioral data of Enterprise A, including: Transaction records: new orders, payment status, credit limit changes. Market feedback: media reports, social media comments, etc. Partnerships: new or interrupted supplier and customer relationships. Data is transmitted to the credit assessment system in real time through stream processing technology (Kafka).

[0228] S9.2. Initialization of the incremental learning framework, using the existing model parameters As a basis, real-time data is input into the model for preliminary prediction. The predicted value is compared with the actual behavior data to calculate the feedback loss function , to measure the degree of deviation of the model's current performance.

[0229] The model parameters are updated dynamically and incrementally using the following formula:

[0230] ,

[0231] The optimization process dynamically adjusts the model parameters through this formula.

[0232] S9.2.3. Closed-loop feedback integrates expert evaluation, analyzes model prediction results and confidence intervals, and provides interpretability of data input features, such as the impact of transaction or market feedback data on scores. At the same time, a feedback interface is designed to allow credit risk analysts to verify model outputs and provide improvement suggestions. Collect analysts' corrections (such as credit score adjustments) as supervision signals to optimize model performance. Enhance the interpretability of the model and the credibility of decisions by converting human feedback into training data. When monitoring indicators show that model performance has deteriorated (such as increased prediction bias or changes in data distribution), start the retraining process. Use the latest and historical data, combined with knowledge graphs, and adopt a dynamic weighting strategy to balance the impact of recent and historical data to ensure model adaptability.

[0233] By designing a closed-loop feedback system, the credit assessment model of Enterprise A has achieved real-time dynamic adjustment, which can not only quickly respond to changes in corporate behavior, but also continuously improve its applicability and credibility through expert feedback. This mechanism provides stable and efficient credit risk management support for the company's global business network, while enhancing the long-term stability and flexibility of the model.

[0234] Figure 5 The schematic diagram of the enterprise credit assessment result provided by the present invention is as follows:

[0235] S10 Generates an explanatory report of credit assessment based on generative adversarial networks (GAN). The credit assessment model of Enterprise A has been optimized for multiple rounds and has achieved high accuracy and robustness. However, in a global business environment, transparency and user trust are crucial. In order to allow users to better understand the source and logic of credit scores, the assessment results are presented in a multimodal form by combining feature contribution analysis and decision path diagrams, which significantly improves the interpretability of the model and user trust.

[0236] S10.1. Construct an explanatory analysis framework. In calculating the feature contribution using the SHAP value, the input data preparation needs to use a multimodal feature vector (transaction records, market feedback data, etc.) as input to generate the current corporate credit score. The contribution analysis uses the SHAP (Shapley Additive Explanations) value to quantify the contribution of each feature to the final score. Output feature contribution ranking list: transaction frequency: 40%; number of partners: 25%; market feedback negative score: 20%; financial performance: 15%;

[0237] Based on this, a causal flow analysis diagram is established to mine the causal relationship between features based on a multimodal dynamic knowledge graph. A Sankey diagram is generated to show the causal flow of features (negative market feedback further reduces credit scores by affecting partnerships. Improved financial performance improves scores by increasing transaction frequency.)

[0238] S10.2. Generator via GAN Simulate the decision-making process of the model and generate a path diagram to show the logical flow of credit scoring from input features to output scores. Used to verify the credibility of the path map and ensure that the generated explanations are meaningful. The analysis results are presented in a multimodal form (charts, text, heat maps), so that users can intuitively understand the scoring basis and risk distribution, and use Sankey diagrams to show the causal flow between features. Dynamically adjust the report content according to user needs and add scenario-based explanations. Ultimately, the explanatory report enhances the transparency of the model and improves users' trust and acceptance of the evaluation results.

[0239] The above embodiments are only used to help understand the method and core idea of ​​the present invention; at the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A commercial credit assessment and supervision method based on a multimodal co-evolutionary algorithm, characterized in that: The method comprises the following steps: S1: Collect and clean multimodal data to ensure data quality and store them in a structured database. S2: Use cross-modal contrastive learning to extract features and optimize the consistency of data in different modalities. S3: embed features into a dynamic multimodal knowledge graph and update the relationships and node representations between enterprises. S4: Extract and fuse modality sharing and specific features using modality sharing networks, S5: Optimizing the consistency of modal feature distribution through adversarial learning, S6: Initialize the global co-evolutionary credit scoring model to optimize accuracy, robustness and interpretability. S7: Use dynamic multi-objective algorithms to optimize model accuracy, robustness, and interpretability. S8: Introduce causal reasoning model to analyze the driving factors of corporate credit changes and establish causal relationships. S9: Design a closed-loop feedback system to update the credit assessment model in real time and enhance adaptability. S10: Generate explainable credit assessment reports based on generative adversarial networks to enhance transparency and credibility; S1 is as follows: Collect multimodal data related to commercial credit, including financial data, transaction records, market feedback and supply chain data. First, clean, normalize and fill missing values ​​of the collected multimodal data to ensure data quality. After these preprocessing steps, all data will be uniformly stored in a structured database to provide consistent input for subsequent processing and modeling. Transaction records refer to transaction amounts. S2 is as follows: The cross-modal contrastive learning algorithm is used to extract features from data of different modalities, generate a unified embedding representation, and optimize modal consistency in the embedding space. Specifically, cross-modal contrastive learning ensures the similarity between data of different modalities by optimizing the following contrast loss function: Among them, L 对比 The loss function representing the similarity of different modal data in the embedding space, Indicates the expected calculation of all sample pairs (xi, xj), h i and h j It is the data representation of different modes, sim(h i ,h j ) means h i and h j The similarity between them, τ is the temperature coefficient, ∑ k It represents the exponentially weighted sum of the similarities calculated for all possible samples. By minimizing this loss function, data from different modalities can maintain consistent representation in the same embedding space.

2. According to the commercial credit evaluation and supervision method based on multimodal co-evolution algorithm according to claim 1, it is characterized in that: S3 is as follows: The construction steps of the dynamic multimodal knowledge graph are as follows: Graph initialization requires initializing the basic structure of the knowledge graph based on the enterprise and partnership information in the multimodal data: nodes are enterprise entities, edges are the relationships between them, and attributes include the embedded features extracted by S2. The weights of the edges are dynamically updated using the time window mechanism, and the weight update rules are defined as: w ij (t+1)=λ 衰减 w ij (t)+(1-λ)R ij Among them, w ij (t+1) represents the edge weight between node i and node j at time point t+1, w ij (t) represents the edge weight between node i and node j at time point t, λ 衰减 is the attenuation factor, R ij To increase the interaction intensity, Then the GNN message passing mechanism is carried out. The node representation is learned through the graph neural network GNN. The message passing mechanism based on time series is designed in combination with dynamics. The update formula of the kth layer is: in, is the node table of the kth layer, σ is the activation function, represents the neighbor set of node g, W (k) is the weight matrix of the kth layer, Represents the node representation of node u at the k-1th layer.

3. The commercial credit evaluation and supervision method based on multimodal co-evolutionary algorithm according to claim 1 is characterized in that: S4 is as follows: Based on the knowledge graph, the modality sharing network is used to extract modality sharing features and modality specific features, and then merge them into collaborative representations through the fusion layer, and optimize them through the following loss function: Among them, h 共享 is the modal shared feature, h i is the specific feature of each mode, λ 正则 This process optimizes the loss function to integrate shared features and specific features in the same space.

4. The commercial credit evaluation and supervision method based on multimodal co-evolutionary algorithm according to claim 1 is characterized in that: S5 is as follows: S5: Design the adversarial learning mechanism between the generator and the discriminator to optimize the distribution consistency between the features of different modalities. Specifically, the adversarial process between the generator G and the discriminator D is optimized by the following loss function: Among them, L 对抗 represents the adversarial loss function, which is used to optimize the adversarial learning process of the generator and the discriminator. G represents the generator, which is responsible for generating fake data. The input is the random noise vector z, and the output is the generated data G(z). D represents the discriminator, which is used to distinguish whether the input data comes from the real data distribution p dt , or the fake data generated by the generator, x~p dt Indicates that the real data x comes from the real data distribution p dt , that is, the real data sample, D(x) represents the judgment value of the discriminator on the real data x, which represents the probability that the discriminator believes that x is the real data, z~p G : represents the input distribution p from the generator G The random noise vector z sampled in is G(z), G(z) represents the conversion of the generator random noise vector z into the generated data G(z), D(G(z)) represents the probability that the discriminator considers G(z) to be the generated data, E is the expected symbol, which represents the expected value calculation of the sample, and the pseudo features and real features of different modes are fused and averaged to generate the final feature vector F.

5. The commercial credit evaluation and supervision method based on multimodal co-evolutionary algorithm according to claim 1 is characterized in that: S6 is as follows: Initialization of the co-evolutionary credit scoring model. The fused feature vector F is used as the initial input. F contains cross-modal shared features and modality-specific features. The dimensions are unified and the distribution is aligned through adversarial learning. F is input into a preliminary scoring model to generate the original credit score S 初始 , and then the loss function is defined based on this, and the mean square error loss function is used to optimize the initial parameters of the model Θ 初始 , The optimization goal is: Among them, L 初步 is the loss value calculated when training the multi-layer perceptron (MLP) model in step S6, y i is the true credit score of sample i, N is the total number of samples, Θ 初始 is the initial parameter of the model, S 初始 (F i ; Θ 初始 ) is the predicted output of the multi-layer perceptron MLP model, representing the input feature vector F i A preliminary credit score generated after MLP calculation.

6. The commercial credit evaluation and supervision method based on multimodal co-evolution algorithm according to claim 1 is characterized in that: S7 is as follows: S7: Introducing a dynamic multi-objective optimization algorithm, taking the accuracy, robustness and interpretability of the model as optimization goals, using a genetic algorithm to explore the optimal solution, and optimizing through the following multi-objective optimization function: F(Θ)=αf1(Θ)+βf2(Θ)+γf3(Θ) Among them, f1(Θ) represents the prediction accuracy of the model, f2(Θ) represents the robustness of the model to data perturbations, and f3(Θ) represents the interpretability of the model. A comprehensive objective function F(Θ) is constructed, where α, β, and γ are dynamic weights, and the proportions are adjusted according to the optimization requirements of different stages.

7. The commercial credit evaluation and supervision method based on multimodal co-evolutionary algorithm according to claim 1 is characterized in that: S8 is as follows: using dynamic multimodal knowledge graphs and combining causal reasoning models to analyze the driving factors of corporate credit changes, and establish a causal relationship between credit changes and their upstream behaviors. Represent the company's market feedback events and transaction behaviors as potential variables and design causal diagrams in is a set of variables, including credit score, market feedback score, and transaction frequency. ε is the causal relationship edge between variables. In order to quantify the contribution of different causal paths, the inverse probability weighted IPW method is used to calculate the direct effect DE and indirect effect IE: total effect = DE + IE. When analyzing the impact of market feedback fluctuations on credit scores, the direct effect represents the direct effect of market feedback, and the indirect effect is reflected through the mediating variable of transaction behavior. Combined with the dynamic multimodal knowledge graph DMKG, the edge weights and variable states in the causal model are updated in real time to reflect the latest association between corporate behavior and credit scores.

8. The commercial credit evaluation and supervision method based on multimodal co-evolutionary algorithm according to claim 1 is characterized in that: S9 is as follows: S9: Design a closed-loop feedback system to dynamically update the credit assessment model. By collecting corporate behavior data in real time, the assessment results are fed back to the model to enhance its adaptability. Specifically, feedback updates are performed through the following optimization formula: Among them, Θ t is the current model parameter, η is the learning rate, is the gradient of the feedback loss function, S10 is as follows: S10: Generate an explanatory report of credit assessment based on generative adversarial networks, including risk sources, model decision paths, and visualization results. Specifically, the interpretable path of model decision is generated through the structure of generative adversarial networks, and presented to users in the form of a report to enhance the trustworthiness of the model.

Images