Method and device for dynamic enterprise credit risk assessment and audit recommendation generation based on multi-modal machine learning

Abstract

The application discloses a kind of dynamic enterprise credit risk assessment and audit suggestion generation method and device based on multi-modal machine learning.The method constructs structured table, unstructured text and multi-source heterogeneous credit dataset such as associated graph, compensates sparse features using conditional generative adversarial network;Quantitative features, text features and graph topology embedding features are extracted respectively;Cross-modal fusion and industry adaptive weight reconstruction are carried out through Bayesian prior attention layer and grey correlation analysis;Then input classification network outputs risk level, and generates risk management and control suggestions combined with SHAP traceability results and audit suggestion library;The application can alleviate the influence brought by data sparsity and distribution drift, help to reduce cross-industry assessment bias, and improve the accuracy and explainability of risk assessment results.

Description

Technical Field

[0001] This invention relates to the intersection of artificial intelligence and financial technology, specifically to a method and apparatus for dynamic corporate credit risk assessment and audit recommendation generation using multi-source heterogeneous data, graph neural networks, and interpretable machine learning. Background Technology

[0002] Credit risk assessment is a crucial step in the daily operational decisions of financial institutions, auditing agencies, and core enterprises in the supply chain. The initial purpose of credit rating was to professionally predict and quantify a company's ability and willingness to fulfill its financial commitments. Looking at the development of the credit assessment field, early assessment models primarily originated from statistical and machine learning algorithms based on small sample data, such as the well-known Altman Z-score model. This model uses multivariate discriminant analysis techniques, relying on five core financial ratio features to construct a classifier to predict the probability of future bankruptcy and default. Subsequently, techniques such as the Merton structured model based on option pricing theory, Logistic Regression, and Decision Tree scoring cards have gradually become widely used. In traditional credit risk assessment practice, the "6C principle" is a widely adopted analytical framework, which expands upon the internationally accepted "5C principle," comprehensively evaluating borrowers from six dimensions: Character, Capacity, Capital, Collateral, Conditions, and Continuity. However, the traditional 6C evaluation relies heavily on human experience in scoring, and the relative importance of financial indicators in each C dimension varies greatly across different industries and macroeconomic cycles, lacking a data-driven dynamic weighting mechanism.

[0003] However, with the increasing complexity of global business networks and the rapid evolution of supply chain ecosystems, traditional credit assessment models have gradually exposed the following structural bottlenecks and technological shortcomings:

[0004] First, it is necessary to assess the limited scope of data dimensions and the lag in early warning. Traditional academic and industrial credit models are mostly based on structured tabular data such as numerical and categorical financial statements and business indicators. This type of static financial data inherently suffers from a lag in its generation cycle; for example, the release of quarterly and annual reports is often significantly delayed. When a company experiences a cash flow crisis or supply chain disruption, financial indicators may not show abnormalities for several months, causing traditional models to lag in warnings of some significant risks. Furthermore, these models often struggle to effectively utilize risk warnings embedded in unstructured long texts, such as qualified opinions in audit reports, significant policy documents and notes to annual reports, as well as early risk signals conveyed through interconnected networks.

[0005] Secondly, there is the challenge of "data sparsity" for SMEs. SMEs generally lack continuous and standardized financial history and comprehensive audit reports. Traditional assessment models that rely on historical structured records, such as logistic regression, often face data scarcity in high-dimensional feature spaces on such samples, leading to decreased model performance and an inability to provide highly accurate credit profiles.

[0006] Third, there is a lack of ability to capture dynamic correlation characteristics. Defaults by modern enterprises are rarely isolated events. Research shows that supply chain networks connecting upstream suppliers, manufacturers, and downstream retailers can both buffer risks and exacerbate the contagion of systemic risks. Traditional isolated node evaluation models cannot capture this complex topological spatial dependency.

[0007] Fourth, the misconception of applying a "one-size-fits-all" approach to assessment weights. Traditional models often use a fixed, static weight formula to evaluate different companies. However, the main drivers of credit risk differ across industries. For example, asset-heavy industries like steel rely heavily on debt ratios and leverage levels, while the retail industry, characterized by high turnover, places greater emphasis on business liquidity and market characteristics. Failure to differentiate weights based on industry segmentation can lead to systemic biases in cross-industry assessments.

[0008] Furthermore, in the actual business loop of credit risk assessment and auditing, it is not only necessary to accurately identify problems, but also to propose targeted and actionable risk mitigation or audit recommendations. Currently, the generation of recommendations usually relies on human experience, which is tedious and prone to errors. While general-purpose large-scale language models have strong creativity, they often face problems such as inaccurate content or insufficient basis in serious audit and risk control scenarios, and the recommendations they generate often lack authenticity and accurate legal and regulatory basis. Therefore, how to fully utilize the standardized audit recommendations accumulated in the past and automatically generate rigorous and professional risk control intervention recommendations based on the specific risks currently identified is a problem that the industry urgently needs to solve. Summary of the Invention

[0009] To address the technical problems existing in the prior art, such as limitations in assessment dimensions, data sparsity, smoothing due to missing graph network labels, assessment distortion due to macroeconomic distribution bias, and ecological fallacies due to micro-macro mapping, the purpose of this invention is to provide a dynamic enterprise credit risk assessment method with adaptive learning capabilities, robustness against cyclical fluctuations, identification of topological risk transmission, and good transparency and compliance adaptability.

[0010] To achieve the above objectives, the present invention provides a method and apparatus for dynamic enterprise credit risk assessment and audit recommendation generation based on multimodal machine learning, the method comprising the following mutually complementary steps:

[0011] Step S1: Construct a multi-source heterogeneous credit dataset to improve the limitations of a single data source and compensate for sparsity. This involves not only collecting structured financial data but also introducing unstructured text (such as long audit reports and public opinion) and enterprise equity / supply chain relationship graphs. Addressing the unique financial gaps and textual vacuum issues faced by SMEs, this invention proposes using Conditional Generative Adversarial Networks (CGANs) to simulate the distribution of potentially missing unstructured features of target enterprises. Furthermore, it suggests combining transfer learning to share deep network parameters, thereby mitigating the problem of insufficient high-dimensional feature information caused by sparse matrices from a data perspective.

[0012] Step S2: Construct a multi-head independent modality feature extraction network to achieve initial high-dimensional feature representation. Addressing the heterogeneous structural attributes of different modalities, this method deploys a combination of domain-specific models in parallel at the front end: an XGBoost module is configured to extract nonlinear core quantitative feature vectors of structured financial and business indicators; a dual-channel natural language processing architecture based on BERT combined with 1D-CNN is configured to extract global semantic sentiment polarity and local high-frequency risk phrases from long texts; and a multi-relationship graph neural network is configured to mine multi-dimensional topological matrix embeddings including supply chains, loan guarantees, and equity investments. Furthermore, a contrast enhancement mechanism is embedded in the graph feature extraction branch to extract deep graph representations highly correlated with default contagion.

[0013] Step S3: Cross-modal dynamic attention feature fusion. To address the data distribution shift trap caused by sudden macroeconomic changes, this model differs from simple feedforward concatenation. In the cross-modal fusion layer, a prior distribution of industry experts based on historical steady-state data is pre-embedded. During the model inference phase, the model uses a multilayer perceptron to calculate the attention scores of the current feature combinations across modalities, adjusting according to changes in the macroeconomic environment. It dynamically reduces the weight of financial features whose effectiveness decreases in specific scenarios, achieving robust feature fusion resistant to fluctuations.

[0014] Step S4: Construct a three-level grey coupling operator ("macro-meso-micro") to achieve multi-dimensional correlation reconstruction. To overcome the ecological fallacy of mixing national macro indicators with corporate financial data in regression, i.e., the aggregation bias caused by cross-dimensional feature mapping, this invention constructs a three-dimensional correlation evaluation system. The indicator sequences of the macro, meso, and micro levels are extracted, and the grey correlation coefficients between these three levels are calculated step by step based on grey system theory. A learnable parameter is used for fusion mapping, and a scaling operator is output to globally adaptively reconstruct and optimize the fused feature vector.

[0015] Step S5: Deep Classifier Prediction and Quantization Output. The high-dimensional comprehensive representation tensor, after multi-level macro-micro dynamic coupling and adaptive reconstruction, is input into the backend deep multilayer perceptual classification network or advanced tree model for forward propagation inference, and outputs the credit risk level of the target enterprise.

[0016] Step S6: Dynamic Threshold Early Warning and Multidimensional Interpretable Source Tracing Analysis Based on SHAP Game Theory. To improve the interpretability of the deep neural network output and provide a basis for compliance within a stringent financial audit system, this method incorporates a SHAP analysis algorithm engine based on cooperative game theory into the decision chain. Through reverse tracing calculations, the massive and complex tensor matrix operations are decoded into attribution paths that intuitively reflect the marginal contributions of each dimension of macro-environment, meso-industry, and micro-characteristic anomalies. Simultaneously, an adaptive dynamic threshold is set for the marginal fluctuations of the target company's characteristics; once a high-risk signal is detected crossing the safe range, a risk warning is triggered.

[0017] Step S7: Automatic generation of risk management and audit recommendations based on semantic learning of the problem. In conjunction with the built-in semantic recommendation module, based on the main risk causes and audit problem characteristics traced in Step S6, the semantic model maps the problem to the audit recommendation space, and automatically generates targeted risk management strategies and audit intervention recommendations by matching and combining them from the pre-built recommendation library.

[0018] This invention also discloses a dynamic enterprise credit risk assessment and audit suggestion generation device based on multimodal machine learning, comprising: a multi-source data acquisition and compensation unit for acquiring multimodal data of the target enterprise and performing feature simulation compensation using a conditional generative adversarial network; a multi-head feature extraction unit for extracting quantitative features, text features, and graph topology embedding features in parallel; a cross-modal feature fusion unit for calculating the weights of each modality feature based on a Bayesian prior attention fusion layer and generating a fused feature vector; a grey relational adaptive reconstruction unit for calculating coefficients using grey relational analysis and performing industry dynamic weight reconstruction; a risk prediction and interpretable tracing unit for predicting credit risk through a deep classification network and outputting an attribution path graph using a SHAP algorithm engine; and an audit question mapping and suggestion generation unit for semantic feature encoding of the extracted audit questions, mapping them through a multilayer perceptron, retrieving them from an audit suggestion library, and generating customized audit and risk control intervention suggestions.

[0019] Beneficial effects of the invention

[0020] 1. Overcoming the "sample imbalance" and "overly smooth topology" problems in graph network mining to improve evaluation accuracy: Combining the multi-adjacency matrix graph neural network (GNN) and graph relation contrastive learning mechanism in step S2, by sampling positive and negative samples from the target enterprise node and its supply chain / equity neighbors and optimizing the contrastive loss function, the model can effectively identify hidden associated topological anomalies in data environments lacking partial default labels. This technique helps improve the prediction accuracy of associated network contagion risk.

[0021] 2. Enhancing the model's resilience and robustness against data distribution drift: Combining the cross-modal dynamic attention feature fusion method from step S3, this method introduces a Bayesian prior attention layer and incorporates industry expert prior knowledge to calculate attention weight scores. This allows the model to dynamically adjust the weight ratios of each modal feature in scenarios where macroeconomic environments cause data distribution drift. This mechanism reduces interference from short-term financial indicator distortions, thereby improving the stability and robustness of the credit assessment model in cross-cycle applications.

[0022] 3. Construct a multi-level correlation analysis mechanism to reduce assessment bias across regions and industries: Combining the "macro-meta-micro" three-level grey coupling operator structure and grey relational analysis (GRA) from step S4, the grey relational coefficients between indicators at different levels are calculated, and a reconstruction factor is generated by combining industry-differentiated weight matrices to adaptively and dynamically reconstruct the fused feature vector. This mechanism quantifies the transmission effect of macro policies and meso-level industries on micro-level enterprise characteristics at the algorithmic level, helping to reduce the bias that may arise from single-dimensional assessment and improve the matching degree between rating results and the actual situation of various industries.

[0023] 4. Achieving an interpretable, traceable, and risk control closed loop in the assessment process: Combining the SHAP algorithm based on cooperative game theory in step S6 and the problem-based semantic learning mapping mechanism in step S7, this approach not only outputs a quantitative, multi-dimensional risk tracing report by calculating the weighted expected value of the marginal contribution of features, but also utilizes a pre-built audit suggestion library to vectorize the core issues discovered through tracing and map them to the suggestion space, automatically retrieving and combining corresponding audit and risk control intervention suggestions. This technological combination provides interpretable compliance references for the black-box decision-making of deep learning models, helping to form a business closed loop from risk warning to suggestion generation. Attached Figure Description

[0024] Figure 1 is an overall flowchart of the risk assessment and recommendation generation method described in this invention.

[0025] Figure 2 is a schematic diagram of the Bayesian prior attention fusion layer and its structure described in this invention.

[0026] Figure 3 is a schematic diagram of the macro-meta-micro three-level grey coupling operator structure described in this invention.

[0027] Figure 4 is a schematic diagram of positive / negative sample sampling for graph relation comparison learning according to the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to specific business embodiments and detailed algorithm formulas. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0029] Example 1: Construction and Sparsity Handling of Multi-Source Heterogeneous Datasets

[0030] like Figure 1 As shown, the overall process of the risk assessment and recommendation generation method described in this invention covers steps from multi-source data acquisition (S1) to feature fusion (S4) and finally to classification and early warning (S6, S7). In specific implementation, this method first constructs a high-throughput data capture pipeline to simultaneously acquire: structured data, unstructured text, and related network data. Addressing the typical "small sample" and "information-poor" problems of SMEs in the credit system, this invention specifically includes a CGAN processing module. Assuming a startup retail enterprise lacks detailed historical audit report texts, this method extracts its existing structured financial indicators as input conditions. And combined with random noise that follows a standard normal prior distribution Common Input Generator Discriminator The network is trained on a large sample set with complete data to distinguish between real and fake data. The minimax game objective function of the conditional generative adversarial network is as follows:

[0031]

[0032] in, For condition generator networks, For conditional discriminator networks, It is the value function of a minimax game. For the real sample vector, For given conditions The actual data feature distribution; This is the random noise vector input to the generator. The prior noise distribution is set; This represents the expected value of the computation. After reaching Nash equilibrium through adversarial training, the generator... It can output simulated unstructured dense vectors that match the business status of the enterprise, thereby helping to fill the high-dimensional sparse matrix of the input layer.

[0033] Example 2: Multi-channel feature extraction and collaborative computation with graph neural networks (GNN) and contrastive learning

[0034] Once the data is complete, different modalities are fed into their corresponding deep learning branch networks. For structured numerical data, the top 30 feature importance scores are extracted using the existing XGBoost model and used as quantized representation vectors. For long audit texts, the existing pre-trained BERT model is used to extract global semantic features, supplemented by a one-dimensional convolutional neural network (1D-CNN) to capture local high-frequency phrase features. The concatenated text features are represented as follows:

[0035]

[0036] in, For the input text data, and These represent the output feature mapping functions of the two functions, respectively. This represents the concatenation operation for feature vectors.

[0037] Combination Figure 4 As shown, in feature extraction of the interconnected network topology graph, this step not only constructs a multi-dimensional adjacency matrix but also introduces a graph relation contrastive learning positive / negative sample sampling mechanism. This is applied to target SME nodes. Centered on this model, samples are taken from its industry supply chain or equity neighbors as positive samples (hidden features are denoted as...). ), sampling competitors or cross-industry alternatives as negative samples (hidden features are denoted as Contrastive learning loss function Defined as:

[0038]

[0039] in, Represents the cosine similarity function. The temperature hyperparameter is used to control the smoothness of the distribution. This represents summing over the corresponding set of nodes. Based on this, the Graph Neural Network (GNN) layer uses a message-passing mechanism to aggregate the central node. and its neighboring nodes The state information. The graph node embedding update calculation formula for each layer is as follows:

[0040]

[0041] in, For enterprises The set of joint neighbor nodes, For the first Layer nodes The hidden feature vectors, These are the transformation matrix weights that this layer needs to learn. It is the regularization constant. This is a non-linear activation function. The final output vector of the model... This can characterize topological implicit features.

[0042] Example 3: Industry Dynamic Adaptation of Credit 6C Model and Grey Relational Analysis (GRA)

[0043] like Figure 3 As shown, in some embodiments of the present invention, a three-level gray coupling operator structure of macro-meta-micro is constructed, and dynamic index weighting of different industries is realized by using mathematical systems theory.

[0044] Step 1: Select the main indicators reflecting relative health to form the benchmark sequence for comparison, i.e., the parent sequence. The characteristic indicator sequence extracted from the target company to be evaluated is used to form a comparison subsequence. .

[0045] Step 2: Perform Z-standardization to reduce the impact of dimensional differences:

[0046]

[0047] in, For the standardized subsequence, The sample mean. This represents the standard deviation. Similarly, the parent sequence... After standardization, it is denoted as .

[0048] Step 3: Calculate the absolute difference sequence and the global extremum:

[0049]

[0050] in, and These represent the parent sequence and the child sequence at the th order. The values ​​of each data point. Find the global maximum difference across all comparison dimensions. minimum difference from the global minimum .

[0051] Step 4: As Figure 3 As shown in the middle section, the grey relational coefficient is calculated using the correlation operator:

[0052]

[0053] in, The resolution coefficient (usually taken as...) By averaging the correlation coefficients at each level, we obtain the following... Figure 3 The macroscopic, mesoscopic, and microscopic grey relational degrees shown ( ). and then combine Figure 3 Learnable weight parameters below (satisfy ), calculate the reconstruction factor Dynamically adjust the fusion features ( ).

[0054] Example 4: Bayesian Prior Cross-Modal Dynamic Attention Fusion and SHAP Explainable Source Tracing

[0055] Combination Figure 2 The diagram shown illustrates the Bayesian prior attention fusion layer structure, which integrates the extracted high-dimensional feature vectors ( The data converges to this fusion layer. The basic scores for each modality are calculated using a multilayer perceptron (MLP). Furthermore, an industry expert prior distribution constructed using historical steady-state data was introduced. Calculate the final attention weight allocation :

[0056]

[0057] The fusion vector is represented as ( These represent different modalities, namely quantized feature vectors, text feature vectors, and hidden graph topological embedding vectors. After the vectors are fed into the deep network, the predicted credit rating is output. Finally, the background automatically runs the SHAP algorithm to calculate the marginal contribution.

[0058]

[0059] in, Features The expected value of marginal contribution, The set of all input features. To remove the features under investigation A subset after that, and The number of elements in the set. This is the output function for the model's probability prediction of a specific combination of features.

[0060] Example 5: Automatic Generation of Audit and Risk Mitigation Suggestions Based on Semantic Space Mapping

[0061] This method targets the core risk points identified by SHAP, extracting structured elements such as the problem title, problem description, and qualitative basis. Through hierarchical feature learning, it performs deep semantic encoding on each element.

[0062] 1. Title and Basic Sentence Encoding: Input the word vectors of the words constituting the question sentence into the LSTM module sequentially:

[0063]

[0064] in, For LSTM at the current time The corresponding output hidden layer state vector, For the previous moment The hidden layer state vector, Indicates the input at the current time. Word embedding vectors of 1 word This represents the network weights and bias parameters that need to be learned in the LSTM module. Word weights are calculated using the similarity between the main information word and each word in the sentence, and then weighted after Softmax normalization to obtain the question sentence and title encoding. .

[0065] 2. Description Paragraph Aggregation Encoding: The sentence vectors in the problem description are sequentially fed into the gate-controlled recurrent unit (GRU), with the problem title as the starting point. Given the query vector, calculate the aggregate weight of each descriptive sentence. And output the latent embedding vector that describes the problem, weighted. :

[0066]

[0067]

[0068] in, For the GRU module at the current moment The hidden layer state output, For the previous moment The hidden layer state output, The input is a query vector derived from the encoding of the question title. These are the learnable network parameters for the GRU module. Furthermore, The first calculated for the model Attention aggregation weights for each descriptive sentence. For the corresponding number Output the hidden state vector of each sentence. The total number of sentences contained in the paragraph describing the problem.

[0069] 3. Qualitative basis coding and fusion: Utilizing the model to extract coding vectors for violations of regulations. and ordinary vectors And by calculating the mean of both, a qualitative basis vector is obtained after fusion. The specific calculation formula is as follows:

[0070]

[0071] Ultimately, the core risk is obtained by splicing the vectors together in the problem space, forming a holographic embedding vector. :

[0072]

[0073] in, This represents the concatenation operation between feature vectors.

[0074] The vector is mapped to the target proposal space using a multilayer perceptron:

[0075]

[0076] Model calculation By comparing the suggestions with existing vectors in the financial compliance audit suggestion library, professional advice statements with the closest distances are selected and combined to form risk response recommendations. This method provides credit managers with a clear and relatively well-defined path for subsequent risk control operations.

[0077] The above embodiments and four core figures fully illustrate the technical effects of this invention in alleviating heterogeneous data barriers, improving graph network smoothness, and adapting to cross-industry macroeconomic characteristics. It should be noted that the deep network architecture design, implementation details, and related algorithm steps mentioned above are merely preferred embodiments provided to explain the core principles of this invention and are not intended to limit the scope of protection of this invention. Any modifications, equivalent substitutions, and additions, subtractions, optimizations, or adjustments to local functions made within the spirit and principles of this invention, based on the aforementioned system's underlying flow mechanism and core mathematical logic, should fall within the scope of protection defined by the claims of this invention.

Claims

1. A method for dynamic enterprise credit risk assessment and audit recommendation generation based on multimodal machine learning, characterized in that, Includes the following steps: Step S1: Construct a multi-source heterogeneous credit dataset, continuously crawling structured tabular data, unstructured text data, and correlation graph data of the target companies. For data that may be missing or sparse for the target companies, conditional generative adversarial networks (CGANs) can be used to generate simulated data for supplementation or enhancement. Step S2: Construct a multi-head independent modality feature extraction network, configure domain-specific model combinations for different modality data to extract features independently, and output quantized feature vectors, text feature vectors and hidden graph topology embedding vectors respectively; Step S3: Cross-modal dynamic attention feature fusion. The feature vectors of each modality are mapped to the same dimensional space, and a Bayesian prior attention fusion layer is introduced, combining prior knowledge from industry experts. Calculate attention weight scores during model inference. Weights are assigned to each feature, and the resulting vectors are dynamically concatenated to generate a fused feature vector. ; Step S4: Introduce the "Credit 6C" model and Grey Relational Analysis (GRA) to obtain multi-level data sources at the macro, meso, and micro levels, and extract the "Credit 6C" evaluation indicators; calculate the grey relational coefficients between the multi-level data, and generate reconstruction factors by combining them with an industry-differentiated weight matrix library. And based on the reconstruction factor The fused feature vectors are reconstructed using industry-adaptive dynamic weights to obtain the final state representation. ; Step S5: Deep network classification and prediction, converting the reconstructed final state representation. Input a deep classification network for forward propagation inference, and output multidimensional credit risk quantification indicators and credit risk level; Step S6: Threshold dynamic monitoring and multidimensional interpretable tracing. Set dynamic thresholds for risk warning and calculate the marginal contribution weighted expected value of features based on the SHAP algorithm of cooperative game theory, and output a multidimensional tracing analysis report. Step S7: Automatic generation of risk management and audit recommendations based on semantic learning of the problem. The audit problem is encoded using a semantic model and mapped to the audit recommendation space. Several vectors that are closest to the problem are selected from the pre-built audit recommendation library, and the corresponding recommendation statements are automatically combined to generate risk management and audit recommendations.

2. The method according to claim 1, characterized in that, In step S1, the specific steps for using Conditional Generative Adversarial Network (CGAN) to simulate and compensate for sparse data are as follows: The structured financial indicators that the target company may lack are used as conditional inputs. and combined with random noise Input generator Generate simulated text feature embedding vectors ; The generator can be trained on the real data distribution of samples from companies in the same industry, and the minimum-maximum function of the Conditional Generative Adversarial Network (CGAN) can be optimized according to the training objective. The output simulated vectors can help fill in or enrich the information of high-dimensional sparse matrices and serve as inputs for subsequent multimodal feature fusion.

3. The method according to claim 1, characterized in that, In step S2, the specific steps for constructing the multi-head independent modality feature extraction network are as follows: Step S21: Use the XGBoost algorithm to construct a gradient boosting decision tree for the structured financial and business data. Extract nonlinear features by calculating the splitting information gain of feature nodes and output a quantized feature vector. ; Step S22: The pre-trained natural language model BERT is used to perform global contextual semantic embedding on long text data, and a one-dimensional convolutional neural network (1D-CNN) is connected in series. By setting convolutional kernels of different sizes, local high-frequency risk phrase features are captured. After pooling and concatenation operations, the text feature vector is output. ; Step S23: Construct a multi-dimensional adjacency matrix that includes the enterprise supply chain and equity investment. Utilize a graph neural network (GNN) through a message passing mechanism to aggregate the state information of the central enterprise node and its associated neighbor nodes layer by layer, and extract the hidden graph topology embedding vector containing network contagion risk. .

4. The method according to claim 1, characterized in that, In step S4, the specific steps for calculating the correlation coefficients between macroscopic, mesoscopic, and microscopic levels are as follows: Step S41: Obtain multi-level data from different sources: the macroeconomic indicator series comes from the National Macroeconomic Database, the mesoeconomic indicator series comes from public reports of industry associations, and the microeconomic indicator series is composed of corresponding indicators extracted from public financial statements of enterprises based on the "Credit 6C" classification framework. Step S42: Select historical steady-state data of high-quality enterprises with no default records in the industry as the benchmark sequence for comparison, i.e., the parent sequence. The feature sequences extracted from each level of the target enterprise to be evaluated are used as subsequences. ; Step S43: Perform Z-standardization on each sequence data with dimensional differences to make them dimensionless. in, The mean, The standard deviation is denoted as ; the relationship between the parent sequence and each child sequence on the th . Absolute difference sequence across each indicator dimension And extract the global maximum difference. and minimum difference ; combined with the resolution coefficient to resist outlier interference Calculate the grey relational coefficients for each indicator dimension: The macro-level correlation degree is obtained by averaging the correlation coefficients at each level. Meso-level correlation Micro-level correlation The macro-level correlation Meso-level correlation Micro-level correlation Input an industry-differentiated weight matrix library, and match corresponding learnable weight parameters based on the industry to which the target company belongs. , , ,satisfy Generate reconstruction factors : Based on the reconstruction factor The fused feature vector z obtained in step S3 is reconstructed using industry-adaptive dynamic weights to obtain the final state representation. and characterize the final state. As input for subsequent deep network classification prediction, where This represents the Hadamard product, which is the element-wise product.

5. The method according to claim 1, characterized in that, In step S6, the formula for calculating the weighted expected value of the marginal contribution of a feature using the SHAP algorithm based on cooperative game theory is as follows: in, The set of all input features. To remove the features under investigation A subset after that, This is the model's probability prediction output for this combination of features.

6. The method according to claim 1, characterized in that, In step S7, the specific steps for automatically generating risk management and audit recommendations based on problem semantic learning include: Step S71, Construct an audit suggestion library: Collect audit suggestion statements from existing audit reports, encode them using a pre-trained natural language model, and establish an audit suggestion vector library; Step S72, extract audit problem characteristics: based on the source tracing results, obtain the problem titles, problem descriptions, and qualitative basis of the main audit problems; Step S73, Hierarchical semantic encoding of the problem: The statement encoding is calculated using a Long Short-Term Memory (LSTM) network; a gated recurrent unit (GRU) is used to perform secondary weighted encoding of the description statement using the problem title as the query vector; qualitative basis encoding is generated by combining legal and regulatory codes; and the various parts are concatenated and integrated to obtain the embedding vector of the audit problem in the problem space. ; Step S74, Spatial Mapping and Recommendation Generation: A mapper constructed using a multilayer perceptron (MLP) is used to map the embedded vectors. Embedding vectors mapped to the audit recommendation space It generates relevant recommendations by combining them from the audit recommendation library through distance calculation.

7. The method according to claim 1, characterized in that: The macro level includes: Loan Prime Rate (LPR), Producer Price Index (PPI), and Gross Domestic Product (GDP); The meso-level includes: industry prosperity and profit margin; The micro level includes: industry average.

8. A device for dynamic enterprise credit risk assessment and audit recommendation generation based on multimodal machine learning, characterized in that, Includes: a multi-source data acquisition and compensation unit, used to acquire multimodal data of the target enterprise and use conditional generative adversarial networks for feature simulation compensation; A multi-head feature extraction unit is used to extract quantized features, text features, and graph topology embedding features in parallel. The cross-modal feature fusion unit is used to calculate the weights of each modality feature based on the Bayesian prior attention fusion layer and generate a fused feature vector; the grey relational adaptive reconstruction unit is used to calculate coefficients using grey relational analysis and perform industry dynamic weight reconstruction; the risk prediction and explainable tracing unit is used to predict credit risk through a deep classification network and call the SHAP algorithm engine to output an attribution path chart; the audit question mapping and suggestion generation unit is used to encode the semantic features of the extracted audit questions, map them through a multilayer perceptron, retrieve them from the audit suggestion library, and generate customized audit and risk control intervention suggestions.

Images