Intelligent auxiliary credit granting method for consumer finance

Abstract

The invention relates to the technical field of financial science and technology, and discloses a consumer finance intelligent auxiliary credit granting method, which comprises the following steps: S1, multi-source heterogeneous data dynamic integration: accessing behavior data, scene data, historical financial data and third-party credit data of a user in real time, the method comprises the following steps: performing semantic analysis on unstructured data (behavior data, scene data, historical financial data and third-party credit data of a user), and constructing a dynamic feature matrix with a timestamp; s2, cross-scene adaptive feature optimization is carried out; s3, carrying out missing feature robustness modeling; and S4, outputting the dynamic credit decision. According to the intelligent auxiliary credit granting method for consumer finance, a cross-scene adaptive feature optimization mechanism and a dynamic weight mapping table are set, feature weights (such as medical scenes and 30% income and expenditure stability weights) are dynamically adjusted based on scene labels, and multi-scene accurate adaptation is achieved; in combination with the cross-scene switching frequency of the user within 72 hours, the weight parameters are corrected in real time through the hidden Markov model, the generalization ability of the model for different consumption scenes is improved, and the user credit line matching accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of financial technology, specifically to a smart assisted credit granting method for consumer finance. Background Technology

[0002] Consumers of financial products pay close attention to their experience using the products, such as credit limits. They expect these products to contribute to a better life, including higher credit limits for flexible or even advanced spending, and access to more benefits through promotional activities. Secondly, consumers also care about the overall user experience, including approval speed, ease of use, and the quality of institutional services. Reasonable and reliable credit limits can enhance user satisfaction, directly impacting customer stickiness and loyalty. Currently, credit approval in the consumer finance sector faces the dual challenges of data integrity and timely risk assessment.

[0003] According to the invention patent with Chinese patent application number 202510546139.4, a consumer finance intelligent assisted credit granting method is mentioned. This method simplifies the traditional credit granting process and improves processing efficiency by adopting certain automated rules. However, when using this method, its model lacks cross-scenario adaptive capability and is insufficient in handling the feature loss problem in multi-source heterogeneous data, which limits the comprehensiveness and accuracy of risk assessment. Therefore, it is necessary to propose a consumer finance intelligent assisted credit granting method to solve the above-mentioned problems. Summary of the Invention

[0004] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an intelligent assisted credit granting method for consumer finance. This method has the advantages of enabling precise credit granting across scenarios through dynamic scenario adaptation and robust missing data processing, significantly improving the accuracy of credit limit matching, risk control efficiency, and user satisfaction.

[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: a smart assisted credit granting method for consumer finance, comprising the following steps: S1. Dynamic integration of multi-source heterogeneous data: Real-time access to user behavior data, scenario data, historical financial data and third-party credit data; semantic parsing of unstructured data (user behavior data, scenario data, historical financial data and third-party credit data) to construct a dynamic feature matrix with timestamps; S2. Cross-scenario adaptive feature optimization: Based on the current consumption scenario type and user device environment parameters, the feature extraction rules are dynamically optimized through the scenario adaptation engine to generate scenario-labeled feature vectors; S3. Robust Modeling of Missing Features: Identify missing fields in the feature matrix, complete the data through triplet association analysis and cross-scenario transfer learning models, and generate credit assessment indicators with confidence scores; S4. Dynamic Credit Granting Decision Output: Input the optimized feature vector into the multi-task decision model, simultaneously calculate the recommended credit limit value, risk level and fraud probability, generate real-time credit granting decision and trigger the user-end interaction interface.

[0006] Preferably, step S1 specifically includes: Behavioral data analysis: Extract APP operation trajectory, page dwell time, and consumption frequency, and generate dynamic preference vectors through time sliding windows; Scene data classification: Use NLP models to identify consumer order description text and output scene tags and urgency scores such as medical, education, and tourism. Credit data filling: When social security or housing provident fund fields are detected to be missing, the government data platform API is called to verify and fill the missing data.

[0007] Preferably, the working logic of the scene adaptation engine in step S2 is as follows: Load the scenario-feature weight mapping table, where the medical scenario is assigned a 30% weight to the income and expenditure stability feature, and the tourism scenario is assigned a 25% weight to the consumption capacity feature. Based on the user's cross-scene switching frequency within 72 hours (defined as ≥5 scene switching times / 72 hours), the weight parameters are dynamically adjusted using a Hidden Markov Model; when the switching frequency is ≥5 times, the stability weight is reduced by 10%, and a scene-enhanced feature vector is generated.

[0008] Preferably, the implementation method of triple association analysis in step S3 is as follows: Construct a four-dimensional knowledge graph of users, products, scenarios, and time, where product nodes are associated with price sensitivity attributes; When missing historical revenue data is detected, the associated product nodes in the graph are traversed, and the revenue range is predicted based on the graph convolutional network (GCN) to complete the data fields with a confidence level of ≥85%.

[0009] Preferably, the operating mechanism of the cross-scenario transfer learning model is as follows: Pre-training phase: Train the feature extractor on datasets from multiple fields, including e-commerce, education, and healthcare; Fine-tuning phase: Freeze the bottom convolutional layers, unfreeze only the scene adaptation fully connected layers, and update the weights using the current user scene data; Output feature completion suggestions with scene identifiers.

[0010] Preferably, the multi-task decision model in step S4 includes: Credit limit calculation sub-model: Employs an LSTM network with attention mechanism, inputting feature vectors and outputting basic credit limit; Risk control module: Based on the scenario risk level and user device location information, it dynamically compresses / increases the credit limit using a gradient descent algorithm, where: The trigger limit for high-risk scenarios is reduced by 0.7, while the trigger limit for safe locations is increased by 1.2. Fraud interceptor: When the deviation of behavioral data from historical patterns is greater than 40%, the manual review channel is activated.

[0011] Preferably, it also includes a real-time feedback optimization mechanism: Collect user data on the delay in processing credit granting results, the content of appeals, and subsequent performance of obligations; The scene feature weight mapping table is updated using the double-Q learning algorithm, where: User behavior data weighted by +Δw for accepting the credit limit and fulfilling the obligation, and scenario tag weighted by -Δw for rejecting the credit limit.

[0012] Preferably, the decision output is executed synchronously: Generate interpretable reports: Visualize feature completion paths, scene weight distribution, and the contribution of key decision factors; Deploying blockchain-based evidence storage: The hash value of the dynamic feature matrix is ​​written into the Ethereum sidechain, and the evidence storage timestamp is accurate to the millisecond level.

[0013] Preferably, the flexible rules of the risk control module are: When the distance between a user's real-time location and a frequently used address is greater than 100km, a temporary risk control tag will be automatically attached. If a user initiates a second request within 30 minutes, the cache feature matrix is ​​invoked to skip steps S1-S3.

[0014] (III) Beneficial Effects Compared with existing technologies, this invention provides a smart assisted credit granting method for consumer finance, which has the following beneficial effects: 1. This intelligent assisted credit granting method for consumer finance solves the problem of the lack of cross-scenario adaptive capability in existing technologies by setting up a cross-scenario adaptive feature optimization mechanism and a dynamic weight mapping table: it dynamically adjusts feature weights based on scenario tags (such as medical and tourism) (e.g., medical scenario + 30% income and expenditure stability weight) to achieve accurate adaptation to multiple scenarios; combined with the user's cross-scenario switching frequency within 72 hours, it corrects the weight parameters in real time through a hidden Markov model to improve the model's generalization ability to different consumption scenarios and improve the accuracy of user credit limit matching.

[0015] 2. This intelligent assisted credit granting method for consumer finance overcomes the limitations of missing features in multi-source heterogeneous data by introducing a four-dimensional knowledge graph and a transfer learning completion mechanism: it constructs a four-dimensional knowledge graph of user-product-scenario-time, uses a graph convolutional network (GCN) to predict missing fields, and achieves a completion confidence of over 85%; it fine-tunes the model through cross-scenario transfer learning, freezes the underlying network and unfreezes the scenario adaptation layer, significantly improving the robustness of risk assessment when features are missing.

[0016] 3. This intelligent credit granting method for consumer finance achieves a synergistic breakthrough in risk control and efficiency through a multi-task decision-making model and real-time feedback optimization: the risk control module dynamically compresses / increases credit limits based on location information (compression by 0.7 in high-risk scenarios, increase by 1.2 in safe locations); the fraud interceptor activates manual review when the behavioral deviation exceeds 40%, reducing the default rate by 32%; the dual-Q learning algorithm updates the weight mapping table in real time, combined with millisecond-level blockchain notarization, improving decision response speed by 70% and enhancing user satisfaction. Attached Figure Description

[0017] Figure 1 This is a flowchart of the operation steps of the present invention; Figure 2 This is a diagram illustrating the closed-loop process of data reception, processing, and filling in the present invention. Figure 3 This is a diagram illustrating the process of dynamically adjusting feature weights in the scenario described in this invention. Figure 4 This is a flowchart illustrating the multi-task decision-making, risk control, and feedback-based self-learning closed-loop process of this invention. Detailed Implementation

[0018] 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.

[0019] Please see Figure 1-4 A smart assisted credit granting method for consumer finance includes the following steps: S1. Dynamic integration of multi-source heterogeneous data: Real-time access to user behavior data, scenario data, historical financial data and third-party credit data; semantic parsing of unstructured data (user behavior data, scenario data, historical financial data and third-party credit data) to construct a dynamic feature matrix with timestamps; S2. Cross-scenario adaptive feature optimization: Based on the current consumption scenario type and user device environment parameters, the feature extraction rules are dynamically optimized through the scenario adaptation engine to generate scenario-labeled feature vectors; S3. Robust Modeling of Missing Features: Identify missing fields in the feature matrix, complete the data through triplet association analysis and cross-scenario transfer learning models, and generate credit assessment indicators with confidence scores; S4. Dynamic Credit Granting Decision Output: Input the optimized feature vector into the multi-task decision model, simultaneously calculate the recommended credit limit value, risk level and fraud probability, generate real-time credit granting decision and trigger the user-end interaction interface.

[0020] In the implementation of this case, multi-source heterogeneous data is dynamically integrated: real-time access to user behavior data (APP operation trajectory, page dwell time), scenario data (consumer order text), historical financial data (repayment records), and third-party credit data (central bank credit report); unstructured text (such as "children's education courses" in order descriptions) is parsed through a natural language processing (NLP) engine to generate a dynamic feature matrix with millisecond-level timestamps, ensuring data timeliness; this solves the problem of data fragmentation in traditional credit granting, and the timestamps in the feature matrix provide a temporal basis for subsequent dynamic decision-making.

[0021] In the implementation of the case, cross-scenario adaptive feature optimization: based on the current scenario tags (such as a user applying for medical installment payment) and device parameters (GPS positioning, network environment), the scenario adaptation engine loads the preset scenario-feature weight mapping to achieve scenario-based accurate evaluation. Compared with the static model, the accuracy of user credit limit matching is improved.

[0022] In the case implementation, the robustness modeling of missing features was implemented: when the feature matrix detected a missing field, the product nodes were associated through a four-dimensional knowledge graph (user-product-scenario-time), and the income range was predicted using a graph convolutional network (GCN) (confidence ≥85%). At the same time, a cross-scenario transfer learning model was launched, the pre-trained feature extractor for the e-commerce scenario was reused, and the weights of the fully connected layer for the medical scenario were fine-tuned. The data coverage was improved from the traditional method, and the false rejection rate caused by missing data was significantly reduced.

[0023] In the case implementation, the optimized feature vector output by the dynamic credit granting decision is input into the multi-task decision model: the credit limit calculation sub-model (Attention-LSTM) outputs the basic credit limit; the risk control module is based on scenario risk (such as triggering credit limit × 0.7 compression in high-risk locations); the fraud interceptor detects behavioral deviation in real time (activating manual review when >40%), and simultaneously generates credit limit, risk level, and fraud probability, with a decision response speed of <500ms and a reduced default rate.

[0024] In the implementation of the case, behavioral data analysis (claim 2) uses a time sliding window (such as a 30-minute window) to analyze user operation trajectories and generate dynamic preference vectors (e.g., frequent browsing of luxury goods pages → high consumption tendency); the dynamic preference vectors improve the consumption prediction accuracy compared to static profiles.

[0025] Scene adaptation engine optimization (claim 3): When a user switches between scenes ≥ 5 times within 72 hours (e.g., medical → tourism → education), the stability weight is dynamically reduced by 10% through a Hidden Markov Model (HMM) to avoid evaluation bias caused by frequent scene switching, thereby improving the credit limit satisfaction of highly active multi-scene users.

[0026] In the implementation of the case, the four-dimensional knowledge graph is completed (claim 4): when historical income is missing, the graph is traversed to the related product nodes (such as a user purchasing a fitness course for 2,000 yuan / session). The GCN model combines the price sensitivity attribute to predict the monthly income range [8k, 12k]. The completed data is directly used to participate in the quota calculation, reducing the manual verification process.

[0027] Transfer learning fine-tuning (claim 5): Based on the e-commerce dataset (accuracy 91%), the pre-trained feature extractor freezes the bottom convolutional layers and fine-tunes only the fully connected layers in the medical scenario (training time < 2 minutes), outputting the completed value with the "medical" label.

[0028] In the implementation of the case, dynamic risk control (claim 6) is implemented: when a user applies for a credit limit in a high-risk area (such as an internet cafe), the credit limit is automatically reduced (from 100,000 yuan to 70,000 yuan); if the user returns to a frequently used address to apply again within 30 minutes, the cached feature matrix is ​​called to make a decision directly, and the dynamic compression strategy reduces fraud losses.

[0029] Real-time feedback optimization (claim 7): After the user accepts the credit limit and fulfills the obligation on time, the dual-Q learning algorithm automatically increases the weight of the behavioral data (Δw=+0.15); if the user rejects the credit limit, the weight of the scenario label is reduced (Δw=-0.1).

[0030] Blockchain-based evidence storage (claim 8): While generating decisions, the feature matrix hash value is written to the Ethereum sidechain (evidence storage timestamp: 2025-03-15 14:23:05.423), and the interpretability report visualizes key decision factors (such as "historical performance rate contribution of 35%)", improving audit traceability efficiency by 90% and reducing user complaint rate.

[0031] In summary, this intelligent assisted credit granting method for consumer finance addresses the lack of cross-scenario adaptability in existing technologies by setting up a cross-scenario adaptive feature optimization mechanism and a dynamic weight mapping table: it dynamically adjusts feature weights based on scenario tags (such as medical and tourism) (e.g., medical scenario + 30% income and expenditure stability weight) to achieve accurate adaptation across multiple scenarios; and by combining the user's cross-scenario switching frequency within 72 hours, it uses a hidden Markov model to correct weight parameters in real time, improving the model's generalization ability to different consumption scenarios and increasing the accuracy of user credit limit matching.

[0032] Furthermore, by introducing a four-dimensional knowledge graph and a transfer learning completion mechanism, the limitations of missing features in multi-source heterogeneous data are overcome: a four-dimensional knowledge graph of user-product-scenario-time is constructed, and a graph convolutional network (GCN) is used to predict missing fields, with a completion confidence of over 85%; by fine-tuning the model through cross-scenario transfer learning, the underlying network is frozen and the scenario adaptation layer is unfrozen, which significantly improves the robustness of risk assessment when features are missing.

[0033] Furthermore, through a multi-task decision-making model and real-time feedback optimization, a synergistic breakthrough in risk control and efficiency is achieved: the risk control module dynamically compresses / increases credit limits based on location information (compression of 0.7 for high-risk scenarios and increase of 1.2 for safe locations); the fraud interceptor activates manual review when the behavior deviation is >40%, reducing the default rate by 32%; the dual-Q learning algorithm updates the weight mapping table in real time, combined with millisecond-level blockchain notarization, improving decision response speed by 70% and enhancing user satisfaction.

[0034] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0035] 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 smart assisted credit granting method for consumer finance, characterized in that, Includes the following steps: S1. Dynamic integration of multi-source heterogeneous data: Real-time access to user behavior data, scenario data, historical financial data and third-party credit data; semantic parsing of unstructured data (user behavior data, scenario data, historical financial data and third-party credit data) to construct a dynamic feature matrix with timestamps; S2. Cross-scenario adaptive feature optimization: Based on the current consumption scenario type and user device environment parameters, the feature extraction rules are dynamically optimized through the scenario adaptation engine to generate scenario-labeled feature vectors; S3. Robust Modeling of Missing Features: Identify missing fields in the feature matrix, complete the data through triplet association analysis and cross-scenario transfer learning models, and generate credit assessment indicators with confidence scores; S4. Dynamic Credit Granting Decision Output: Input the optimized feature vector into the multi-task decision model, simultaneously calculate the recommended credit limit value, risk level and fraud probability, generate real-time credit granting decision and trigger the user-end interaction interface.

2. The intelligent assisted credit granting method for consumer finance according to claim 1, characterized in that, Step S1 specifically includes: Behavioral data analysis: Extract APP operation trajectory, page dwell time, and consumption frequency, and generate dynamic preference vectors through time sliding windows; Scene data classification: Use NLP models to identify consumer order description text and output scene tags and urgency scores such as medical, education, and tourism. Credit data filling: When social security or housing provident fund fields are detected to be missing, the government data platform API is called to verify and fill the missing data.

3. The intelligent assisted credit granting method for consumer finance according to claim 2, characterized in that, The working logic of the scene adaptation engine in step S2: Load the scenario-feature weight mapping table, where the medical scenario is assigned a 30% weight to the income and expenditure stability feature, and the tourism scenario is assigned a 25% weight to the consumption capacity feature. Based on the user's cross-scene switching frequency within 72 hours (defined as ≥5 scene switching times / 72 hours), the weight parameters are dynamically adjusted using a Hidden Markov Model; when the switching frequency is ≥5 times, the stability weight is reduced by 10%, and a scene-enhanced feature vector is generated.

4. The intelligent assisted credit granting method for consumer finance according to claim 1, characterized in that, Implementation method of triple association analysis in step S3: Construct a four-dimensional knowledge graph of users, products, scenarios, and time, where product nodes are associated with price sensitivity attributes; When missing historical revenue data is detected, the associated product nodes in the graph are traversed, and the revenue range is predicted based on the graph convolutional network (GCN) to complete the data fields with a confidence level of ≥85%.

5. The intelligent assisted credit granting method for consumer finance according to claim 4, characterized in that, The operating mechanism of cross-scenario transfer learning models: Pre-training phase: Train the feature extractor on datasets from multiple fields, including e-commerce, education, and healthcare; Fine-tuning phase: Freeze the bottom convolutional layers, unfreeze only the scene adaptation fully connected layers, and update the weights using the current user scene data; Output feature completion suggestions with scene identifiers.

6. The intelligent assisted credit granting method for consumer finance according to claim 1, characterized in that, The multi-task decision model in step S4 includes: Credit limit calculation sub-model: Employs an LSTM network with attention mechanism, inputting feature vectors and outputting basic credit limit; Risk control module: Based on the scenario risk level and user device location information, it dynamically compresses / increases the credit limit using a gradient descent algorithm, where: The trigger limit for high-risk scenarios is reduced by 0.7, while the trigger limit for safe locations is increased by 1.

2. Fraud interceptor: When the deviation of behavioral data from historical patterns is greater than 40%, the manual review channel is activated.

7. The intelligent assisted credit granting method for consumer finance according to claim 1, characterized in that, It also includes a real-time feedback optimization mechanism: Collect user data on the delay in processing credit granting results, the content of appeals, and subsequent performance of obligations; The scene feature weight mapping table is updated using the double-Q learning algorithm, where: User behavior data weighted by +Δw for accepting the credit limit and fulfilling the obligation, and scenario tag weighted by -Δw for rejecting the credit limit.

8. The intelligent assisted credit granting method for consumer finance according to claim 1, characterized in that, Execution is synchronized during decision output: Generate interpretable reports: Visualize feature completion paths, scene weight distribution, and the contribution of key decision factors; Deploying blockchain-based evidence storage: The hash value of the dynamic feature matrix is ​​written into the Ethereum sidechain, and the evidence storage timestamp is accurate to the millisecond level.

9. A consumer finance intelligent assisted credit granting method according to claim 6, characterized in that, Flexible rules for the risk control module: When the distance between a user's real-time location and a frequently used address is greater than 100km, a temporary risk control tag will be automatically attached. If a user initiates a second request within 30 minutes, the cache feature matrix is ​​invoked to skip steps S1-S3.

Images