Intelligent credit evaluation method and system based on multi-source data and knowledge graph

Abstract

The application discloses an intelligent credit evaluation method and system based on multi-source data and a knowledge graph, and belongs to the technical field of financial technology and artificial intelligence. The method collects and pre-processes multi-source heterogeneous data, constructs a financial knowledge graph containing entity association and risk transmission paths, adopts an attention mechanism to fuse multi-modal features and knowledge graph association features, realizes accurate credit evaluation through pre-training of a large language model, and dynamically optimizes the model through output of explainable results and continuous learning and emergency iteration mechanisms. The system is configured with data collection and preprocessing, knowledge graph construction, feature fusion, model reasoning, result explanation and continuous optimization modules, supports federated learning and risk transmission visualization, solves the problems of single data, weak explainability and insufficient dynamic adaptation of traditional credit evaluation, significantly improves the accuracy, timeliness and risk response capability of credit evaluation, and is suitable for various financial scenarios such as banks and consumer finance.

Description

Technical Field

[0001] This invention relates to the technical fields of financial technology and artificial intelligence, and more particularly to intelligent credit assessment methods and systems based on multi-source data and knowledge graphs. Background Technology

[0002] Credit assessment is a core component of financial risk prevention and control. Traditional credit assessment methods mainly rely on structured financial data and credit reports, which have the following significant drawbacks: Relying solely on static financial indicators, and lacking the integration of dynamic unstructured data such as transaction behavior, social connections, and public opinion, makes it difficult to comprehensively characterize the credit risk of an entity. Machine learning-based assessment models are often "black boxes," failing to clearly demonstrate risk transmission paths and key influencing factors, and thus failing to meet the interpretability requirements of financial regulators. Model updates rely on batches of historical data, making rapid iteration impossible in the face of sudden risk events (such as regional default waves or regulatory policy adjustments), resulting in assessment results lagging behind risk changes. The lack of a systematic knowledge graph makes it difficult to identify potential associated risks such as implicit guarantee chains and cross-entity risk transmission, resulting in weak capabilities to control complex network risks. Cross-institutional collaborative data modeling is prone to data privacy leaks, and traditional centralized data processing models are insufficient to meet compliance requirements.

[0003] In existing technologies, some solutions attempt to introduce knowledge graphs or multi-source data, but they suffer from problems such as insufficient feature fusion depth, poor model adaptability, and lack of emergency risk response mechanisms, and cannot effectively address the dynamic and complex needs of credit assessment in financial scenarios. Summary of the Invention

[0004] The present invention aims to overcome the shortcomings of the prior art and provide an intelligent credit assessment method and system based on multi-source data and knowledge graphs.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect of the present invention, an intelligent credit assessment method based on multi-source data and knowledge graphs is provided, comprising the following steps: Obtain multi-source heterogeneous data of the entity to be evaluated, including financial data, transaction behavior data, social connection data, public opinion data, structured business data, credit report data, tax data and supply chain transaction data, and clean, normalize, convert formats and de-identify the multi-source heterogeneous data; Based on the ontology rules of the financial field, the preprocessed multi-source data is mapped to entities and relationships in the knowledge graph, and a financial knowledge graph containing entity identity, association relationship, risk event and financial product interaction is constructed. The entity links and relationship weights in the graph are updated in real time through incremental data. Structured features, text features, and behavioral sequence features are extracted from the multi-source heterogeneous data. Combined with entity association features and risk transmission path features in the financial knowledge graph, the contribution of high-risk association features is enhanced through an attention mechanism to generate a fused feature vector. The pre-trained model is fine-tuned using a pre-trained large language model and a labeled dataset in the financial field to obtain a credit assessment model adapted to financial scenarios; the fused feature vector of the subject to be assessed is input into the model to output a credit score and risk level; The model interpretation technique generates interpretable explanations of the assessment results, and the financial knowledge graph is used to visualize the risk transmission path and key influencing factors. Based on real-time new data and evaluation feedback results, the financial knowledge graph and credit assessment model are incrementally updated and iteratively optimized, and the accuracy and timeliness of the assessment are balanced through reinforcement learning mechanism.

[0006] Preferably, the entities in the financial knowledge graph include the subject to be evaluated, related enterprises, financial institutions, and risk event nodes, and the relationships include guarantee relationships, transaction relationships, equity relationships, and risk transmission relationships; the update process includes entity disambiguation, dynamic weighting of relationship strength, and incremental addition of new entities to the database.

[0007] Preferably, the entities in the financial knowledge graph include the subject to be evaluated, related enterprises, financial institutions, and risk event nodes, and the relationships include guarantee relationships, transaction relationships, equity relationships, and risk transmission relationships; the update process includes entity disambiguation, dynamic weighting of relationship strength, and incremental addition of new entities to the database.

[0008] More preferably, the dynamic adjustment step includes: The structured features, text features, behavioral sequence features, and knowledge graph entity association features are assigned initial weights based on prior knowledge in the financial domain. A graph attention network (GAT) is used, combined with the type of association, time decay factor, and severity of risk events, to calculate the association strength weights between the entity to be evaluated and other entities in the knowledge graph. The contribution of each feature is dynamically adjusted based on the entity association attention weights. If the entity to be evaluated has a strong association with a high-risk entity, the weight of the corresponding feature is increased; if a specific feature has no high-risk association, its weight is appropriately reduced. Softmax normalization is performed on all adjusted feature weights to ensure the sum of the weights is 1. The attention weights and feature contributions are continuously optimized by combining real-time new data and evaluation feedback results.

[0009] Preferably, the fine-tuning process employs a combination of few-shot learning and Prompt Tuning, guiding the model to adapt to credit assessment scenarios through task instructions in the financial field; the inference stage achieves low-latency response through quantization compression and distributed deployment.

[0010] Preferably, the visualization includes knowledge graph subgraph rendering, risk factor heatmap, and assessment result tracing link, which are used to provide users with interactive exploration of risk transmission paths.

[0011] Preferably, the incremental update and iterative optimization includes feature distillation of incremental data, dynamic fine-tuning of model parameters, and incremental reasoning of the knowledge graph. An emergency iteration mechanism is triggered when model drift or new high-risk events are detected.

[0012] More preferably, the emergency iteration mechanism is triggered as follows: the AUC-ROC, KS value and prediction bias rate are tracked in real time by the model performance monitoring module. When the index drops below a preset threshold, or when a new systemic high-risk event is detected by the knowledge graph, emergency iteration is triggered. When the emergency iteration mechanism is triggered, incremental data related to the triggering event is collected first, and rapid cleaning, desensitization, and feature extraction are performed. Entities and relationships related to newly added high-risk events are incrementally added to the database, the weights of risk transmission paths are updated, and potential associated risks are mined through a semantic reasoning engine. The credit assessment model is rapidly fine-tuned based on the urgently collected dataset by combining few-shot learning and Prompt Tuning. The underlying weights of the pre-trained model are frozen, and only the parameters of the upper-level task adaptation layer are adjusted. The performance of the fine-tuned model is verified through online testing. If the preset indicators are met, it is quickly deployed while retaining the original model.

[0013] In a second aspect, the present invention also provides an intelligent credit assessment system based on multi-source data and knowledge graphs, applied to the intelligent credit assessment method based on multi-source data and knowledge graphs as described above, comprising: Data acquisition and preprocessing module: used to acquire multi-source heterogeneous data of the subject to be evaluated, and perform cleaning, normalization, privacy desensitization and cross-source alignment processing; Financial Knowledge Graph Construction Module: Used to construct and dynamically update a financial knowledge graph containing entities, relationships, and risk transmission paths based on the ontology rules of the financial field, and to configure ontology management, entity linking, and incremental update units; Multimodal feature fusion module: used to extract structured, textual and behavioral sequence features from multi-source data, combine them with the association features of the knowledge graph, and generate a fused feature vector through an attention mechanism; Credit assessment model module: used to load pre-trained large language models and fine-tune them for financial scenarios, perform credit assessment inference, and configure model fine-tuning, inference acceleration and drift monitoring units; Assessment Result Interpretation and Visualization Module: Used to generate interpretable descriptions, visualize risk transmission paths through knowledge graph rendering, heatmaps, and other methods, and support exporting reports in multiple formats; Continuous Learning and Optimization Module: Used to incrementally update and iteratively optimize the knowledge graph and model based on real-time data and feedback results, and configure incremental inference, parameter fine-tuning and drift triggering units.

[0014] Preferably, the financial knowledge graph construction module is equipped with a semantic reasoning engine, which mines potential related risks by combining rule-based reasoning and embedded reasoning; the credit assessment model module is equipped with a model caching unit to cache the feature vectors and results of high-frequency assessment requests, and improves the system's concurrent processing capability by combining a load balancing mechanism; the assessment result interpretation and visualization module is equipped with an interactive analysis unit, which allows users to explore the causes of risks and generate personalized risk improvement suggestions through graph drag-and-drop, feature filtering and other operations.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. By integrating heterogeneous data from multiple sources such as finance, transactions, social media, and public opinion, and combining the correlation characteristics of knowledge graphs, we break through the limitations of traditional single data dimensions, comprehensively depict the credit risk of entities, and improve the accuracy of credit scoring.

[0016] 2. By visually displaying the risk transmission path and key influencing factors, and combining explanation techniques such as SHAP and LIME, the interpretability problem of "black box" models is solved, meeting financial regulatory requirements, while supporting user interaction to explore the causes of risks.

[0017] 3. The attention mechanism dynamically adjusts feature weights, and the emergency iteration mechanism enables rapid iteration when the model drifts or high-risk events occur, significantly improving the model's response speed to sudden risks and reducing the risk of lag.

[0018] 4. The pre-trained large language model is fine-tuned to adapt to credit assessment tasks through financial scenarios. The continuous learning mechanism combines incremental data to dynamically optimize the model, so that the assessment results always keep pace with changes in market risk.

[0019] 5. Supports the fusion of encrypted features of cross-institutional data, enabling multi-source collaborative modeling without obtaining the original data, effectively avoiding the risk of data privacy leakage and meeting financial compliance requirements.

[0020] 6. The decoupled design of each module supports functional iteration and scenario expansion, and can be quickly adapted to various financial scenarios such as banking, consumer finance, and supply chain finance, and has strong commercial application value. Attached Figure Description

[0021] Figure 1 This is a flowchart of an intelligent credit assessment method based on multi-source data and knowledge graphs in a specific embodiment of the present invention; Figure 2This is a framework diagram of an intelligent credit assessment system based on multi-source data and knowledge graphs in a specific embodiment of the present invention; Detailed Implementation

[0022] The technical solutions of 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.

[0023] refer to Figure 1 As shown, this embodiment applies the intelligent credit assessment method based on multi-source data and knowledge graph proposed in this invention to upstream supplier B of core enterprise A in the supply chain, including: Step 1: Obtain multi-source heterogeneous data of the entity to be evaluated, including financial data, transaction behavior data, social connection data, public opinion data, structured business data, credit report data, tax data, and supply chain transaction data. Then, clean, normalize, convert formats, and anonymize the multi-source heterogeneous data, specifically including: Data from multiple sources is collected through three methods: API calls, web scraping, and cross-institutional federated collaboration. Structured data: Supplier B's financial statements for the past two years, tax return data, central bank credit report, and supply chain transaction data; The financial statements are from the supply chain management platform of core enterprise A. The tax declaration data is obtained from the tax platform through encrypted means via federated learning units. The central bank credit report is used to check for overdue records. The supply chain transaction data consists of transaction records with core enterprise A, including 120 purchase orders and 108 payment records, with amounts ranging from 50,000 to 800,000 yuan.

[0024] Unstructured data: 23 news articles and 156 user comments from Sina Finance and Qichacha about Supplier B and its related companies over the past 6 months, as well as the LinkedIn connections of the company's legal representative and core executives, and the interaction records of WeChat cooperation groups. Semi-structured data: Core enterprise A's performance evaluation report on supplier B, including quality score and delivery timeliness score.

[0025] Subsequently, the multi-source heterogeneous data is cleaned, normalized, format-converted, and de-identified for privacy. Data cleaning: Outliers in financial data are identified using the 3σ principle and retained after verification; meaningless text related to public opinion is filtered using regular expressions. Missing value imputation: Linear interpolation was used to fill in the missing two months of cash flow statement data, and blank public opinion comments were marked as "no valid information"; Normalization and Alignment: Normalize financial indicators such as debt-to-equity ratio and current ratio to the [0,1] range and align cross-source data using timestamps; Privacy desensitization: Hash algorithms encrypt sensitive corporate information, and tax data is homomorphically encrypted before participating in federated feature fusion to avoid the risk of privacy leakage.

[0026] Step 2: Based on the ontology rules of the financial domain, the preprocessed multi-source data is mapped to entities and relationships in a knowledge graph, constructing a financial knowledge graph that includes entity identities, relationships, risk events, and interactions with financial products. The entity links and relationship weights in the graph are updated in real time using incremental data. Specifically, this includes: The entities in the financial knowledge graph include entities to be evaluated, related enterprises, financial institutions, and risk event nodes. The relationships include guarantee relationships, transaction relationships, equity relationships, and risk transmission relationships. The update process includes entity disambiguation, dynamic weighting of relationship strength, and incremental addition of new entities to the database.

[0027] In this embodiment, a knowledge graph is generated based on ontology rule mapping data in the financial domain, containing 4 types of entities and 5 types of relationships: Entities: The entity to be evaluated is supplier B; related entities include core enterprise A, parent company C, cooperating bank D, and raw material supplier E; risk events include the policy of phasing out subsidies for new energy vehicle parts; financial products include working capital loans and factoring business; Relationships between entities: procurement transaction relationship (supplier B → core enterprise A, weight 0.8), equity relationship (parent company C → supplier B, weight 0.9, 65% shareholding), loan relationship (bank D → supplier B, weight 0.7, amount 5 million yuan), guarantee relationship (core enterprise A → supplier B, weight 0.85), policy relationship (industry policy → supplier B, weight 0.6).

[0028] New data is monitored via Kafka message queues, entity links are added in real time and relationship weights are adjusted (e.g., after 10 new orders, the transaction relationship weight increases to 0.83); a full calibration is performed every morning to complete entity disambiguation and weight normalization, ensuring the accuracy of the graph.

[0029] Step 3: Extract structured features, textual features, and behavioral sequence features from the multi-source heterogeneous data. Combine these with entity association features and risk transmission path features in the financial knowledge graph. Strengthen the contribution of high-risk association features through an attention mechanism to generate a fused feature vector, specifically including: The structured features include debt repayment capacity indicators, transaction frequency and amount features; the text features include public opinion sentiment polarity and social text semantic features; the behavioral sequence features include repayment cycle stability and fund flow concentration; the attention mechanism dynamically adjusts the degree of influence of different features on credit assessment by calculating the attention weights of entity associations in the knowledge graph.

[0030] The dynamic adjustment step includes: The structured features, text features, behavioral sequence features, and knowledge graph entity association features are assigned initial weights based on prior knowledge in the financial domain. A graph attention network (GAT) is used, combined with the type of association, time decay factor, and severity of risk events, to calculate the association strength weights between the entity to be evaluated and other entities in the knowledge graph. The contribution of each feature is dynamically adjusted based on the entity association attention weights. If the entity to be evaluated has a strong association with a high-risk entity, the weight of the corresponding feature is increased; if a specific feature has no high-risk association, its weight is appropriately reduced. Softmax normalization is performed on all adjusted feature weights to ensure the sum of the weights is 1. The attention weights and feature contributions are continuously optimized by combining real-time new data and evaluation feedback results.

[0031] Specifically, the extracted multimodal features, including structured features, text features, behavioral sequence features, and knowledge graph association features, are fused together: The structured features are derived by extracting indicators such as the debt-to-asset ratio of 42% and the current ratio of 1.3, forming a 28-dimensional structured feature vector; the text features are derived by extracting public opinion semantic features through the BERT-base model and combining them with sentiment polarity analysis, resulting in a 769-dimensional text feature vector; the behavioral sequence features are derived by processing the repayment flow sequence through the LSTM model, extracting features such as repayment stability, and generating a 64-dimensional sequence feature vector; and the knowledge graph association features are derived by querying the third-order related entity features, extracting information such as the length of the risk transmission path, and generating a 32-dimensional association feature vector.

[0032] Subsequently, dynamic adjustments were made through an attention mechanism: a two-layer graph attention network (GAT) was used to calculate the weights, and the association weights between supplier B and its parent company C were calculated by combining the relationship type, time decay factor, and risk severity. The association weights between supplier B and its parent company C were 0.52, and between supplier B and its core enterprise A were 0.78. Based on this, the feature contribution was adjusted, with the weight of the structured feature increasing from 0.3 to 0.38 and the weight of the knowledge graph association feature increasing from 0.2 to 0.25. After Softmax normalization, a 1093-dimensional fused feature vector was obtained.

[0033] Step 4: Fine-tune the pre-trained model using a pre-trained large language model and a labeled dataset in the financial field to obtain a credit assessment model adapted to financial scenarios; input the fused feature vector of the subject to be assessed into the model to output a credit score and risk level, specifically including: The fine-tuning process combines few-shot learning with Prompt Tuning, guiding the model to adapt to credit assessment scenarios through financial task instructions; the inference stage achieves low-latency response through quantization compression and distributed deployment.

[0034] Based on 100,000 pieces of labeled supply chain finance data, the BERT-base model was fine-tuned with a learning rate of 2e-5, a batch size of 32, and 15 training rounds. The weights of the first 8 layers were frozen, and only the upper adaptation layer and classification layer were fine-tuned. Through INT8 quantization and distributed deployment, inference latency is controlled within 450ms, and the output supplier B has a credit score of 78 (out of 100), a medium risk level, and a default probability of 3.2% within one year.

[0035] Step 5: Generate interpretability statements for the assessment results using model interpretation techniques, and visualize the risk transmission path and key influencing factors using financial knowledge graphs. Specifically, this includes: The top 3 key factors were determined by the SHAP value, including: "Transaction stability with core company A" (0.18), "Debt-to-equity ratio" (0.15), and "Overdue payment history of parent company C" (0.12). A partial report is generated using LIME to explain why the parent company's overdue payment resulted in a 5-point decrease in the score; Neo4j is used to visualize the related subplots, and Matplotlib is used to draw heatmaps, supporting interactive exploration by users.

[0036] The visualization includes knowledge graph subgraph rendering, risk factor heatmap, and assessment result tracing links, which are used to provide users with interactive exploration of risk transmission paths.

[0037] Step 6: Based on real-time new data and evaluation feedback results, incrementally update and iteratively optimize the financial knowledge graph and credit assessment model. Balance assessment accuracy and timeliness through reinforcement learning mechanisms, specifically including: Regular continuous learning: Collect new data daily, fine-tune the model incrementally every 3 days (3 rounds), and update the knowledge graph relationship weights synchronously.

[0038] Incremental updates and iterative optimizations include feature distillation of incremental data, dynamic fine-tuning of model parameters, and incremental inference of knowledge graphs. An emergency iteration mechanism is triggered when model drift or new high-risk events are detected.

[0039] The emergency iteration mechanism is triggered as follows: AUC-ROC, KS value and prediction bias rate are tracked in real time through the model performance monitoring module. When the index drops below the preset threshold, or when a new systemic high-risk event is detected through the knowledge graph, emergency iteration is triggered. When the emergency iteration mechanism is triggered, incremental data related to the triggering event is collected first, and rapid cleaning, desensitization, and feature extraction are performed. Entities and relationships related to newly added high-risk events are incrementally added to the database, the weights of risk transmission paths are updated, and potential associated risks are mined through a semantic reasoning engine. The credit assessment model is rapidly fine-tuned based on the urgently collected dataset by combining few-shot learning and Prompt Tuning. The underlying weights of the pre-trained model are frozen, and only the parameters of the upper-level task adaptation layer are adjusted. The performance of the fine-tuned model is verified through online testing. If the preset indicators are met, it is quickly deployed while retaining the original model.

[0040] Emergency Iteration Mechanism: On the 7th day after the assessment, core enterprise A experienced a large amount of overdue payment, and the model AUC dropped from 0.92 to 0.86, exceeding the threshold by 0.05, triggering an emergency iteration, the process of which is as follows: Step 601 (within 1 hour): Confirm AUC decrease of 0.06, mark as "systemic high-risk event", and generate trigger notification; Step 602 (within 6 hours): Prioritize collecting overdue details and unsettled order data from core enterprises, quickly preprocess and update federated shared features; Step 603 (within 4 hours): Add overdue event nodes, reduce the weight of guarantee and transaction relationships, infer the risk of implicit repayment delays and supplement the path; Step 604 (within 7 hours): Fine-tune the model based on 5000 related data points, freeze the weights of the first 10 layers, and optimize the attention mechanism through 5 rounds of training; Step 605 (within 2 hours): A / B testing verifies that the model's AUC has rebounded to 0.91 and KS value to 0.49. Deploy the new model and the original model as a backup at an 8:2 ratio. Step 606 (within 72 hours): Review the reasons, adjust the AUC warning threshold to 0.04, and increase the priority of core enterprise data monitoring.

[0041] Following an emergency iteration, Supplier B's credit score dropped to 70 points, its risk level was adjusted to "medium-high risk," and its default probability was revised to 6.8%, providing timely support for Bank D's credit strategy adjustment.

[0042] In a second aspect of the invention, an intelligent credit assessment system based on multi-source data and knowledge graphs is proposed, applied to the aforementioned intelligent credit assessment method based on multi-source data and knowledge graphs, comprising: Data acquisition and preprocessing module: used to acquire multi-source heterogeneous data of the subject to be evaluated, and perform cleaning, normalization, privacy desensitization and cross-source alignment processing; Financial Knowledge Graph Construction Module: Used to construct and dynamically update a financial knowledge graph containing entities, relationships, and risk transmission paths based on the ontology rules of the financial field, and to configure ontology management, entity linking, and incremental update units; Multimodal feature fusion module: used to extract structured, textual and behavioral sequence features from multi-source data, combine them with the association features and risk transmission path features of the knowledge graph, and generate fused feature vectors through an attention mechanism; Credit assessment model module: used to load pre-trained large language models and fine-tune them for financial scenarios, perform credit assessment inference, and configure model fine-tuning, inference acceleration and drift monitoring units; Assessment Result Interpretation and Visualization Module: Used to generate interpretable descriptions, visualize risk transmission paths through knowledge graph rendering, heatmaps, and other methods, and support exporting reports in multiple formats; Continuous Learning and Optimization Module: Used to incrementally update and iteratively optimize the knowledge graph and model based on real-time data and feedback results, and configure incremental inference, parameter fine-tuning and drift triggering units.

[0043] The financial knowledge graph construction module is equipped with a semantic reasoning engine, which mines potential related risks by combining rule-based reasoning and embedded reasoning; the credit assessment model module is equipped with a model caching unit to cache the feature vectors and results of high-frequency assessment requests, and improves the system's concurrent processing capabilities by combining a load balancing mechanism; the assessment result interpretation and visualization module is equipped with an interactive analysis unit, which allows users to explore the causes of risks and generate personalized risk improvement suggestions through graph drag-and-drop, feature filtering and other operations.

[0044] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. An intelligent credit assessment method based on multi-source data and knowledge graphs, characterized in that, Includes the following steps: Obtain multi-source heterogeneous data of the entity to be evaluated, including financial data, transaction behavior data, social connection data, public opinion data, structured business data, credit report data, tax data and supply chain transaction data, and clean, normalize, convert formats and de-identify the multi-source heterogeneous data; Based on the ontology rules of the financial field, the preprocessed multi-source data is mapped to entities and relationships in the knowledge graph, and a financial knowledge graph containing entity identity, association relationship, risk event and financial product interaction is constructed. The entity links and relationship weights in the graph are updated in real time through incremental data. Structured features, text features, and behavioral sequence features are extracted from the multi-source heterogeneous data. Combined with entity association features and risk transmission path features in the financial knowledge graph, the contribution of high-risk association features is enhanced through an attention mechanism to generate a fused feature vector. The pre-trained model is fine-tuned using a pre-trained large language model and a labeled dataset in the financial field to obtain a credit assessment model adapted to financial scenarios; the fused feature vector of the subject to be assessed is input into the model to output a credit score and risk level; The model interpretation technique generates interpretable explanations of the assessment results, and the financial knowledge graph is used to visualize the risk transmission path and key influencing factors. Based on real-time new data and evaluation feedback results, the financial knowledge graph and credit assessment model are incrementally updated and iteratively optimized, and the accuracy and timeliness of the assessment are balanced through reinforcement learning mechanism.

2. The intelligent credit assessment method based on multi-source data and knowledge graphs according to claim 1, characterized in that, The entities in the financial knowledge graph include entities to be evaluated, related enterprises, financial institutions, and risk event nodes. The relationships include guarantee relationships, transaction relationships, equity relationships, and risk transmission relationships. The update process includes entity disambiguation, dynamic weighting of relationship strength, and incremental addition of new entities to the database.

3. The intelligent credit assessment method based on multi-source data and knowledge graphs according to claim 1, characterized in that, The structured features include debt repayment capacity indicators, transaction frequency and amount features; the text features include public opinion sentiment polarity and social text semantic features; the behavioral sequence features include repayment cycle stability and fund flow concentration; the attention mechanism dynamically adjusts the degree of influence of different features on credit assessment by calculating the attention weights of entity associations in the knowledge graph.

4. The intelligent credit assessment method based on multi-source data and knowledge graphs according to claim 3, characterized in that, The dynamic adjustment steps include: The structured features, text features, behavioral sequence features, and knowledge graph entity association features are assigned initial weights based on prior knowledge in the financial domain. A graph attention network (GAT) is used, combined with the type of association, time decay factor, and severity of risk events, to calculate the association strength weights between the entity to be evaluated and other entities in the knowledge graph. The contribution of each feature is dynamically adjusted based on the entity association attention weights. If the entity to be evaluated has a strong association with a high-risk entity, the weight of the corresponding feature is increased; if a specific feature has no high-risk association, its weight is appropriately reduced. Softmax normalization is performed on all adjusted feature weights to ensure the sum of the weights is 1. The attention weights and feature contributions are continuously optimized by combining real-time new data and evaluation feedback results.

5. The intelligent credit assessment method based on multi-source data and knowledge graphs according to claim 1, characterized in that, The fine-tuning process employs a combination of few-shot learning and Prompt Tuning, guiding the model to adapt to credit assessment scenarios through task instructions in the financial field; the inference stage achieves low-latency response through quantization compression and distributed deployment.

6. The intelligent credit assessment method based on multi-source data and knowledge graphs according to claim 1, characterized in that, The visualization includes knowledge graph subgraph rendering, risk factor heatmap, and assessment result tracing links, which are used to provide users with interactive exploration of risk transmission paths.

7. The intelligent credit assessment method based on multi-source data and knowledge graphs according to claim 1, characterized in that, The incremental update and iterative optimization include feature distillation of incremental data, dynamic fine-tuning of model parameters, and incremental reasoning of knowledge graphs. An emergency iteration mechanism is triggered when model drift or new high-risk events are detected.

8. The intelligent credit assessment method based on multi-source data and knowledge graphs according to claim 7, characterized in that, The emergency iteration mechanism is triggered as follows: the AUC-ROC, KS value and prediction bias rate are tracked in real time by the model performance monitoring module. When the index drops below the preset threshold, or when a new systemic high-risk event is detected by the knowledge graph, emergency iteration is triggered. When the emergency iteration mechanism is triggered, incremental data related to the triggering event is collected first, and rapid cleaning, desensitization and feature extraction are performed. Incremental entry of entities and relationships related to newly added high-risk events into the database, updating the weights of risk transmission paths, and mining potential associated risks through a semantic reasoning engine; adopting a combination of few-shot learning and Prompt Tuning to quickly fine-tune the credit assessment model based on an urgently collected dataset; freezing the underlying weights of the pre-trained model and adjusting only the parameters of the upper-level task adaptation layer. The performance of the fine-tuned model is verified through online testing. If the preset indicators are met, it is quickly deployed while the original model is retained.

9. An intelligent credit assessment system based on multi-source data and knowledge graphs, applied to the intelligent credit assessment method based on multi-source data and knowledge graphs as described in any one of claims 1-8, characterized in that, include: Data acquisition and preprocessing module: used to acquire multi-source heterogeneous data of the subject to be evaluated, and perform cleaning, normalization, privacy desensitization and cross-source alignment processing; Financial Knowledge Graph Construction Module: Used to construct and dynamically update a financial knowledge graph containing entities, relationships, and risk transmission paths based on the ontology rules of the financial field, and to configure ontology management, entity linking, and incremental update units; Multimodal feature fusion module: used to extract structured, textual and behavioral sequence features from multi-source data, combine them with the association features and risk transmission path features of the knowledge graph, and generate fused feature vectors through an attention mechanism; Credit assessment model module: used to load pre-trained large language models and fine-tune them for financial scenarios, perform credit assessment inference, and configure model fine-tuning, inference acceleration and drift monitoring units; Assessment Result Interpretation and Visualization Module: Used to generate interpretable descriptions, visualize risk transmission paths through knowledge graph rendering, heatmaps, and other methods, and support exporting reports in multiple formats; Continuous Learning and Optimization Module: Used to incrementally update and iteratively optimize the knowledge graph and model based on real-time data and feedback results, and configure incremental inference, parameter fine-tuning and drift triggering units.

10. The intelligent credit assessment system based on multi-source data and knowledge graphs according to claim 9, characterized in that, The financial knowledge graph construction module is equipped with a semantic reasoning engine, which mines potential related risks by combining rule-based reasoning and embedded reasoning; the credit assessment model module is equipped with a model caching unit to cache the feature vectors and results of high-frequency assessment requests, and improves the system's concurrent processing capabilities by combining a load balancing mechanism; the assessment result interpretation and visualization module is equipped with an interactive analysis unit, which allows users to explore the causes of risks and generate personalized risk improvement suggestions through graph drag-and-drop, feature filtering and other operations.

Images