Method for realizing fast credit rating of financial customers by means of AI
By projecting multimodal credit data onto the credit feature manifold space and utilizing the Ricci flow evolution equation, the problems of multimodal data fusion and long-tail customer identification in credit rating are solved, achieving efficient and stable credit rating result output.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 孙治
- Filing Date
- 2026-01-29
- Publication Date
- 2026-05-29
AI Technical Summary
Existing credit rating technologies ignore the high-dimensional manifold topology of credit features when processing multimodal data fusion, leading to information collapse; they fail to effectively address the imbalance problem of long-tail customers and default samples, resulting in severe model robustness and decision bias, and lacking the ability to dynamically perceive risk boundaries.
By mapping multimodal credit data to the credit feature manifold space through embedding operators, and dynamically adjusting the manifold structure using the metric evolution equation of Ricci flow, a counterfactual reasoning closed-loop mechanism is constructed to achieve adaptive evolution and distribution alignment of credit features, and generate counterfactual prototypes for closed-loop correction.
It significantly improves the model's accuracy in identifying long-tail customers and the real-time performance of rating results, solves the information collapse problem in multimodal data fusion, ensures the stability and fairness of rating logic, and enhances the interpretability and rapid response capability of rating results.
Smart Images
Figure CN122115094A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of financial customer credit rating technology, specifically a method for achieving rapid credit rating of financial customers using AI. Background Technology
[0002] With the rapid development of financial technology, credit rating has shifted from traditional scoring cards based on expert experience to automated prediction systems based on artificial intelligence. In the current field of financial customer credit assessment, mainstream technical solutions mainly rely on ensemble learning algorithms represented by gradient boosting decision trees (GBDT) and preliminary graph neural networks (GNN). These methods have shown high execution efficiency when processing large-scale, highly feature-rich structured financial data. Meanwhile, in order to cope with the increasing amount of unstructured data (such as legal documents, social media sentiment, etc.), existing technologies usually introduce pre-trained natural language processing (NLP) models for feature extraction and simply concatenate them with financial features.
[0003] However, existing credit rating technologies still face severe challenges in practical applications. First, when processing multimodal data fusion, existing technologies often forcibly map features from different sources to the same Euclidean space. This linear processing method ignores the high-dimensional manifold topology behind credit features, which can easily lead to information collapse of key deep semantic information during feature dimensionality reduction or splicing, making it difficult for the model to capture the logical causal relationships hidden in unstructured text.
[0004] Secondly, for the long-tail customers (such as micro and small enterprises or emerging groups lacking credit records) and the serious imbalance of default samples that are common in the financial industry, most existing technologies are limited to data-level reinforcement, such as using synthetic minority oversampling technology (SMOTE) or simple loss function weighting. These methods are mathematically equivalent to local statistical corrections, which are not only prone to generating artificial noise, but also cannot solve the measurement offset problem between sparse long-tail distribution and core high-quality distribution from the perspective of underlying distribution evolution. This leads to the model often exhibiting extremely poor robustness and decision bias when facing credit boundary areas with extremely scarce samples.
[0005] Furthermore, most current credit rating models belong to a static mapping architecture, and their interpretability techniques (such as SHAP or LIME) usually exist as independent post-processing modules. This means that the model cannot feed back the ex-post interpretability analysis to the ex-ante feature representation learning, resulting in the rating system lacking the ability to dynamically perceive risk boundaries and reshape the feature space in real time. Under this disconnected architecture, when the market environment fluctuates rapidly or the customer's credit behavior pattern changes, the system is unable to make timely geometric corrections to the decision space through feedback loops, thus limiting the accuracy and real-time performance of the rating results. Therefore, this invention designs a method for rapid credit rating of financial customers using AI based on the problems mentioned above. Summary of the Invention
[0006] To address the shortcomings of existing technologies, this invention provides a method for rapid credit rating of financial customers using AI. This method solves the problems of feature fusion loss, distribution measure bias, and blurred decision boundaries caused by sample imbalance when processing long-tail data and multimodal unstructured data in existing financial credit rating methods.
[0007] To achieve the above objectives, this invention provides a method for rapid credit rating of financial clients using AI, comprising the following steps: S1. First, acquire the multimodal credit data of financial customers, and use the embedding operator to map it to the initialized credit feature manifold space to obtain the manifold representation of customer credit features. S2. Calculate the curvature features of the credit feature manifold space, and combine them with the counterfactual weight feedback term to dynamically adjust the metric tensor of the credit feature manifold space through the evolution equation, thereby realizing the adaptive evolution of the credit feature space. S3. Subsequently, the cost matrix of the evolved credit feature manifold space under geodesic distance is constructed. Using the optimal transmission plan with quality truncation, the long-tail sample distribution is mapped to the core sample distribution to complete the cross-distribution credit feature alignment. S4. Then, counterfactual reasoning is performed on the aligned samples to generate counterfactual prototypes. By analyzing the response entropy of the counterfactual prototypes in each dimension, the counterfactual weight feedback term is calculated and fed back to the evolution equation for closed-loop correction. S5. Finally, the feature representation of the customer to be rated is mapped to the aligned credit feature manifold space, and the credit rating result of the financial customer can be output through the classification operator.
[0008] Preferably, the step of mapping it to the initialized credit feature manifold space using the embedding operator specifically includes: using a pre-trained encoder to extract the structured financial features and unstructured text features of financial customers, and fusing them into an initial high-dimensional feature vector; The Laplace-Beltrami operator is applied as the embedding operator, and the initial coordinate points of the credit feature manifold space are generated by solving the eigenmap of the initial high-dimensional feature vector on the manifold.
[0009] Preferably, the evolution equation is a metric evolution equation based on Ricci curvature:
[0010] in, Let be the metric tensor of the credit feature manifold space. Let Ricci curvature tensor be the constant. For adjustment coefficients, This is the potential energy adjustment term driven by the fact weight feedback term.
[0011] Preferably, the adaptive evolution specifically includes: Identify risk fluctuation regions in the credit feature manifold space with curvature exceeding a preset threshold, and perform geometric smoothing on the regions using the evolution equation; Identify long-tailed sparse regions with sample distribution density below a preset threshold, and use the evolution equation to geometrically stretch these regions to increase the geometric spacing between samples of different credit levels.
[0012] Preferably, the cost matrix constructed based on geodesic distance specifically includes: Based on the evolved metric tensor Within the credit feature manifold space, the path with the minimum energy consumption between two points is searched as a geodesic. The integral length of the geodesic on the manifold is calculated accordingly, and it is used as the transportation cost to construct the cost matrix.
[0013] Preferably, the optimal transmission plan utilizing quality truncation specifically includes: A quality truncation operator is introduced to truncate and filter outlier samples in the source distribution that exceed the probability measure threshold; Under entropy regularization constraints, the unbalanced Wasserstein distance is solved based on the cost matrix, and the filtered long-tail sample measure is transferred to the probability distribution space of the core sample.
[0014] Preferably, the step of performing counterfactual reasoning to generate a counterfactual prototype specifically includes: For sample points with known credit rating results, find the minimum geometric perturbation vector required to flip their rating labels within the credit feature manifold space; Based on the minimum geometric perturbation vector, the counterfactual prototype is generated in the critical region of the rating decision boundary.
[0015] Preferably, the calculation of the counterfactual weight feedback term specifically includes: Extract the perturbation response of the counterfactual prototype in each dimension of its features; The information entropy of the disturbance response is calculated using the entropy weight method to generate a weight matrix that reflects the sensitivity of each dimension feature to rating changes, which serves as the counterfactual weight feedback term.
[0016] Preferably, the feedback to the evolution equation for closed-loop correction specifically includes: The weight matrix is injected in real time into the potential energy adjustment term in the evolution equation. middle; The geometric structure of the credit feature manifold space is driven to expand in a directional manner toward the risk-sensitive dimension, so that the manifold evolution process and credit risk judgment feedback can be dynamically decoupled and reconstructed.
[0017] Preferably, the output of the credit rating result for the financial customer specifically includes: The characteristic representation of the customer to be rated is linearly transformed by the mapping operator determined by the optimal transmission plan; Inference is performed by applying a pre-trained linear classification operator on the credit feature manifold space after evolution convergence, so as to realize the rapid calculation and output of rating results.
[0018] This invention provides a method for rapid credit rating of financial clients using AI. It offers the following advantages: 1. This invention dynamically adjusts the geometric structure of the credit feature manifold by introducing a metric evolution equation based on Ricci flow, thereby achieving curvature smoothing of high-risk volatility regions and geometric metric stretching of long-tail sparse regions. It solves the problem of ambiguous judgment boundaries caused by insufficient sample size of long-tail customers from the mathematical and geometric level, and significantly enhances the model's recognition accuracy and spatial separability for edge samples and credit status of new financial customers.
[0019] 2. This invention utilizes the Laplace-Beltrami operator to project multimodal credit data onto an initialized credit feature manifold space, achieving high-fidelity deep fusion of structured financial features and unstructured text semantic features. This effectively solves the information collapse problem that traditional feature splicing methods easily cause when processing unstructured data, and provides a topologically consistent underlying geometric representation for customer credit evaluation in complex financial contexts.
[0020] 3. This invention constructs an optimal transmission plan with quality truncation and uses manifold geodesic distance as the transportation cost for distribution alignment, realizing a nonlinear mapping between long-tail distribution and core high-quality distribution under a unified credit metric. This not only eliminates the systematic distribution bias between different customer groups, but also uses the truncation operator to accurately filter outlier noise interference in long-tail data, ensuring the stability and fairness of credit rating logic in scenarios with extremely imbalanced samples.
[0021] 4. This invention constructs a closed-loop correction mechanism of representation-decision-feedback through counterfactual prototype generation and response entropy weight calculation. It feeds back the risk sensitivity of the rating results to the potential energy term of the manifold evolution equation in real time, driving the credit space to undergo directional geometric reshaping towards key risk dimensions. This breaks the limitations of static training of traditional AI models, enabling the rating logic to evolve dynamically with risk characteristics, while significantly improving the interpretability of the model's decision-making process.
[0022] 5. This invention simplifies the online inference process by placing the complex nonlinear manifold evolution and distribution alignment process in the training convergence stage, thereby reducing the computational overhead and system response latency of a single rating task. While ensuring the depth of complex multimodal feature mining, it achieves the millisecond-level fast rating output required by financial business scenarios. Attached Figure Description
[0023] Figure 1 This is one of the schematic diagrams of the method flow of the present invention; Figure 2 This is a second schematic diagram of the method flow of the present invention; Figure 3 This is the third schematic diagram of the method flow of the present invention; Figure 4 This is the fourth schematic diagram of the method flow of the present invention; Figure 5 This is the fifth schematic diagram of the method flow of the present invention; Figure 6 This is the sixth schematic diagram of the method flow of the present invention; Figure 7 This is the seventh schematic diagram of the method flow of the present invention. Detailed Implementation
[0024] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0025] Please see the appendix Figure 1 -Appendix Figure 7 This invention provides a method for rapid credit rating of financial customers using AI, comprising the following steps: S1. First, acquire multimodal credit data of financial customers and use an embedding operator to map it to an initialized credit feature manifold space to obtain a manifold representation of customer credit features. Specifically, this includes: using a pre-trained encoder to extract structured financial features and unstructured text features of financial customers and fusing them into an initial high-dimensional feature vector; applying the Laplace-Beltrami operator as the embedding operator, and generating the initial coordinate points of the credit feature manifold space by solving the eigenmaps of the initial high-dimensional feature vector on the manifold. Specifically, in financial scenarios, customer credit data consists of structured financial data (such as income, liabilities, and cash flow) and unstructured text data (such as legal documents, public opinion, and social evaluations). This step first uses a pre-trained deep semantic encoder to transform unstructured text into high-dimensional semantic vectors, and then performs feature-level fusion with structured features to form the initial observation vector. In order to capture the complex nonlinear topological relationships between data, this implementation treats the feature space as a Riemannian manifold. ,in To measure tensors, the Laplace-Beltrami operator is used. The mathematical expression of this operator for embedding the observation vector is as follows:
[0026] By solving the eigenvalue mapping of this operator, high-dimensional heterogeneous data can be projected onto the manifold space. This process ensures that customer points that are similar in terms of credit risk remain close in terms of manifold geometry, thus providing an accurate initial topological layout for subsequent geometric evolution.
[0027] S2. Calculate the curvature features of the credit feature manifold space. Combined with the counterfactual weight feedback term, dynamically adjust the metric tensor of the credit feature manifold space through an evolutionary equation to achieve adaptive evolution of the credit feature space. Specifically, the evolutionary equation is a metric evolution equation based on Ricci curvature.
[0028] in, Let be the metric tensor of the credit feature manifold space. Let Ricci curvature tensor be the constant. For adjustment coefficients, The adaptive evolution, driven by the fact weight feedback term, specifically includes: identifying risk fluctuation regions in the credit feature manifold space with curvature higher than a preset threshold, and performing geometric smoothing on the regions through the evolution equation; identifying long-tailed sparse regions with sample distribution density lower than a preset threshold, and performing geometric measurement stretching on the regions through the evolution equation to expand the geometric spacing between samples of different credit ratings. Specifically, addressing the spatial sparse distribution problem of long-tail financial customers (such as micro and small enterprises or those with no credit history) due to small sample sizes, this step introduces a measure of Ricci flow to dynamically adjust the manifold. The Ricci flow equation is defined as:
[0029] in, The Ricci curvature tensor, which reflects the locality of the distribution of credit data, In this implementation, the evolution equations have a dual geometric adjustment effect, serving as feedback terms for subsequent steps: Smoothing high-risk areas: In areas where credit characteristics fluctuate drastically and curvature is too high, flow evolution is used to reduce curvature, making the decision boundary smoother and reducing the model's sensitivity to noise; Stretching sparse regions: In the sparse regions where long-tailed samples are located, the geometric space is artificially stretched by expanding the metric tensor, making the low-frequency samples that were originally crowded near the origin more distinguishable on a geometric scale, thereby improving the recognition efficiency of the classifier.
[0030] S3. Subsequently, the cost matrix of the evolved credit feature manifold space under geodesic distance is constructed. Using an optimal transfer plan with quality truncation, the long-tail sample distribution is mapped to the core sample distribution, completing cross-distribution credit feature alignment. The cost matrix under geodesic distance specifically includes: based on the evolved metric tensor... The process involves searching for the path with the minimum energy consumption between two points within the credit feature manifold space, which is then used as the geodesic. The integral length of the geodesic on the manifold is calculated and used as the transport cost to construct the cost matrix. The optimal transport plan with quality truncation specifically includes: introducing a quality truncation operator to truncate and filter outlier samples in the source distribution that exceed the probability measure threshold; and, under entropy regularization constraints, solving for the non-equilibrium Wasserstein distance based on the cost matrix to transport the filtered long-tailed sample measure to the probability distribution space of the core samples. Specifically, in order to eliminate the systematic distribution shift between long-tailed samples (source distribution) and core high-quality samples (target distribution), this step utilizes optimal transport theory to achieve cross-distribution alignment. First, based on the metric tensor evolved in step S2... Calculate the geodesic distances between points on the manifold as the cost matrix for transporting charge. Then, an unbalanced transmission plan with quality truncation is constructed:
[0031] During the solution process, a quality truncation operator is introduced to truncate outliers in the source distribution that exceed the probability measure threshold. This mechanism allows the model to only process the distribution portion with significant credit correlation, effectively preventing long-tailed noise data from interfering with the core evaluation system and ensuring that customers from different distribution backgrounds can be compared horizontally on the same credit scale.
[0032] S4. Then, counterfactual reasoning is performed on the aligned samples to generate counterfactual prototypes. By analyzing the response entropy of the counterfactual prototypes in each dimension, counterfactual weight feedback terms are calculated and fed back into the evolutionary equation for closed-loop correction. The calculation of counterfactual weight feedback terms specifically includes: extracting the perturbation response of the counterfactual prototypes in each dimension; calculating the information entropy of the perturbation response using the entropy weight method to generate a weight matrix reflecting the sensitivity of each dimension to rating changes, which serves as the counterfactual weight feedback term. Counterfactual reasoning is then performed to generate counterfactual prototypes, specifically including: for sample points with known credit rating results, finding the minimum geometric perturbation vector required to flip the rating label within the credit feature manifold space; based on the minimum geometric perturbation vector, generating counterfactual prototypes in the critical region of the rating decision boundary, and feeding them back into the evolutionary equation for closed-loop correction, specifically including: injecting the weight matrix into the potential energy adjustment term in the evolutionary equation in real time. In the middle; the geometric structure of the credit feature manifold space is driven to expand in a directional manner towards the risk-sensitive dimension, so that the manifold evolution process and credit risk judgment feedback can be dynamically decoupled and reconstructed; Specifically, by constructing a closed-loop feedback mechanism for model decision-making through counterfactual reasoning, the interpretability and robustness of the rating results are enhanced. For each sample point with a rating result, the system searches within the credit feature manifold space for the minimum geometric displacement that can reverse its rating result. This process generates a counterfactual prototype. Sensitivity analysis of the displacement process is performed, and the response entropy of each dimension's feature is calculated using the entropy weight method. Features with lower response entropy contribute more to the volatility of the rating result and have higher risk sensitivity. This sensitivity information is transformed into counterfactual weight feedback terms, which are fed back to the evolution equation in step S2 in real time. In this process, the closed-loop mechanism drives the manifold space to evolve in a direction toward higher-risk dimensions, enabling the representation space to adaptively reshape its shape according to the actual risk assessment logic.
[0033] S5. Finally, the feature representation of the customer to be rated is mapped to the aligned credit feature manifold space. The credit rating result of the financial customer can be output through the classification operator. Specifically, the feature representation of the customer to be rated is linearly transformed through the mapping operator determined by the optimal transfer plan; and the pre-trained linear classification operator is applied to the evolved converged credit feature manifold space for inference, so as to realize the rapid calculation and output of the rating result. Specifically, because the complex geometric evolution, distribution alignment, and feedback correction are all completed during the background training phase, the system exhibits extremely high efficiency once it enters the online inference phase: Spatial transformation: Newly entered customer multimodal data is directly projected onto the evolved credit feature manifold space through a preset mapping operator, and its position is corrected by the transmission mapping determined in step S3; Linear inference: On the manifold after evolutionary convergence, the credit rating task is transformed into a simple linear discrimination. Using a lightweight classification operator, the system can quickly calculate the credit rating result of the customer.
[0034] By transforming complex nonlinear feature mining into geometric location determination in manifold space, this method significantly reduces the computational overhead of real-time rating, enabling financial institutions to meet their business needs for rapid feedback on customer credit within minutes or even seconds.
[0035] In summary, this invention provides a method for rapid credit rating of financial customers using AI. By utilizing the Laplace-Beltrami operator to project multimodal credit data onto an initialized credit feature manifold space, it achieves high-fidelity deep fusion of structured financial features and unstructured textual semantic features. This effectively solves the information collapse problem that easily occurs when traditional feature splicing methods process unstructured data. It provides a topologically consistent underlying geometric representation for customer credit evaluation in complex financial contexts. Furthermore, by introducing a metric evolution equation based on Ricci flow to dynamically adjust the geometric structure of the credit feature manifold, it achieves curvature smoothing of high-risk volatility regions and geometric metric stretching of long-tailed sparse regions. From a mathematical and geometric perspective, it solves the problem of ambiguous judgment boundaries caused by insufficient long-tail customer sample size, significantly enhancing the model's recognition accuracy and spatial separability for marginal samples and the credit status of new financial customers.
[0036] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A method for rapid credit rating of financial customers using AI, characterized by: Includes the following steps: S1. First, acquire the multimodal credit data of financial customers, and use the embedding operator to map it to the initialized credit feature manifold space to obtain the manifold representation of customer credit features. S2. Calculate the curvature features of the credit feature manifold space, and combine them with the counterfactual weight feedback term to dynamically adjust the metric tensor of the credit feature manifold space through the evolution equation, thereby realizing the adaptive evolution of the credit feature space. S3. Subsequently, the cost matrix of the evolved credit feature manifold space under geodesic distance is constructed. Using the optimal transmission plan with quality truncation, the long-tail sample distribution is mapped to the core sample distribution to complete the cross-distribution credit feature alignment. S4. Then, counterfactual reasoning is performed on the aligned samples to generate counterfactual prototypes. By analyzing the response entropy of the counterfactual prototypes in each dimension, the counterfactual weight feedback term is calculated and fed back to the evolution equation for closed-loop correction. S5. Finally, the feature representation of the customer to be rated is mapped to the aligned credit feature manifold space, and the credit rating result of the financial customer can be output through the classification operator.
2. The method for rapid credit rating of financial customers using AI according to claim 1, characterized in that, The process of mapping the embedded features to the initialized credit feature manifold space using an embedding operator specifically includes: extracting structured financial features and unstructured text features of financial customers using a pre-trained encoder, and fusing them into an initial high-dimensional feature vector; The Laplace-Beltrami operator is applied as the embedding operator, and the initial coordinate points of the credit feature manifold space are generated by solving the eigenmap of the initial high-dimensional feature vector on the manifold.
3. The method for rapid credit rating of financial customers using AI according to claim 1, characterized in that, The evolution equation is specifically a metric evolution equation based on Ricci curvature: in, Let be the metric tensor of the credit feature manifold space. Let Ricci curvature tensor be the constant. For adjustment coefficients, This is the potential energy adjustment term driven by the fact weight feedback term.
4. The method for rapid credit rating of financial customers using AI according to claim 3, characterized in that, The adaptive evolution specifically includes: Identify risk fluctuation regions in the credit feature manifold space with curvature exceeding a preset threshold, and perform geometric smoothing on the regions using the evolution equation; Identify long-tailed sparse regions with sample distribution density below a preset threshold, and use the evolution equation to geometrically stretch these regions to increase the geometric spacing between samples of different credit levels.
5. The method for rapid credit rating of financial customers using AI according to claim 1, characterized in that, The cost matrix constructed based on geodesic distance specifically includes: Based on the evolved metric tensor Within the credit feature manifold space, the path with the minimum energy consumption between two points is searched as a geodesic. The integral length of the geodesic on the manifold is calculated accordingly, and it is used as the transportation cost to construct the cost matrix.
6. The method for rapid credit rating of financial customers using AI according to claim 1, characterized in that, The optimal transmission plan utilizing quality truncation specifically includes: A quality truncation operator is introduced to truncate and filter outlier samples in the source distribution that exceed the probability measure threshold; Under entropy regularization constraints, the unbalanced Wasserstein distance is solved based on the cost matrix, and the filtered long-tail sample measure is transferred to the probability distribution space of the core sample.
7. The method for rapid credit rating of financial customers using AI according to claim 1, characterized in that, The process of performing counterfactual reasoning to generate counterfactual prototypes specifically includes: For sample points with known credit rating results, find the minimum geometric perturbation vector required to flip their rating labels within the credit feature manifold space; Based on the minimum geometric perturbation vector, the counterfactual prototype is generated in the critical region of the rating decision boundary.
8. The method for rapid credit rating of financial customers using AI according to claim 1, characterized in that, The calculated counterfactual weight feedback term specifically includes: Extract the perturbation response of the counterfactual prototype in each dimension of its features; The information entropy of the disturbance response is calculated using the entropy weight method to generate a weight matrix that reflects the sensitivity of each dimension feature to rating changes, which serves as the counterfactual weight feedback term.
9. The method for rapid credit rating of financial customers using AI according to claim 3 or 8, characterized in that, The feedback to the evolution equation for closed-loop correction specifically includes: The weight matrix is injected in real time into the potential energy adjustment term in the evolution equation. middle; The geometric structure of the credit feature manifold space is driven to expand in a directional manner toward the risk-sensitive dimension, so that the manifold evolution process and credit risk judgment feedback can be dynamically decoupled and reconstructed.
10. The method for rapid credit rating of financial customers using AI according to claim 1, characterized in that, The credit rating results for the financial clients output specifically include: The characteristic representation of the customer to be rated is linearly transformed by the mapping operator determined by the optimal transmission plan; Inference is performed by applying a pre-trained linear classification operator on the credit feature manifold space after evolution convergence, so as to realize the rapid calculation and output of rating results.