Credit portrait construction method and system based on dynamic weight and incremental learning

Abstract

The invention provides a credit portrait construction method and system based on dynamic weight and incremental learning, and belongs to the field of financial information science and technology. The method comprises the following steps: firstly, collecting multi-source heterogeneous user data and classifying the multi-source heterogeneous user data into feature data of multiple credit dimensions; then, weighting the feature data according to a dynamic weighting strategy, and fusing the real-time streaming data in a privacy protection environment to generate time sequence feature data; then, time sequence analysis and anomaly detection are carried out on the time sequence characteristic data to recognize abnormal behaviors, and response operation is triggered; then, through a credit evaluation model, a credit score is generated based on the time sequence characteristic data and the abnormal behavior, and an interpretability report is generated based on an attribution algorithm; and finally, monitoring a user behavior event, and performing incremental updating on the credit evaluation model when a preset condition is triggered. According to the method, the evaluation precision and dynamics can be improved, the risk response efficiency is enhanced, the black box problem of the model is solved, and the data privacy security is guaranteed.

Description

Technical Field

[0001] This application relates to the fields of financial information technology, software engineering, and data analysis, and in particular to a method and system for constructing credit profiles based on dynamic weights and incremental learning. Background Technology

[0002] In the field of financial risk control, personal credit assessment is a core component. Existing technologies mainly rely on static or semi-static credit scoring models, such as building scoring card models based on structured data like a user's historical credit records. While these methods are widely used, they have significant drawbacks.

[0003] First, its data dimensions are relatively singular, mainly relying on historical credit data, making it difficult to effectively utilize heterogeneous data such as social interactions and consumption behavior that can reflect a user's real-time credit status, resulting in incomplete and inaccurate assessment results. Second, the feature weights in the model are usually fixed or have long update cycles, such as monthly or quarterly batch updates. This makes the model slow to respond to sudden risk events that occur in the short term, such as frequent borrowing or drastic changes in consumption habits, resulting in a lag in risk identification. Although some technical solutions have attempted to introduce multi-source data and streaming computing technologies, their applications are mostly focused on simple abnormal behavior detection and have failed to build an adaptive credit scoring mechanism.

[0004] Furthermore, many advanced assessment models, especially complex deep learning models, suffer from a "black box" problem. Their internal decision-making logic is opaque, making it difficult to provide regulators and users with reasonable and clear explanations for their scores. This fails to meet the increasingly stringent compliance and explainability requirements of the financial sector. Therefore, existing technologies struggle to create a dynamic credit profiling system that is both responsive and accurate in its assessments while maintaining transparency in its decision-making. Summary of the Invention

[0005] The purpose of this application is to address the problems of static credit assessment models, single data dimensions, delayed risk response, and opaque decision-making processes in existing technologies by providing a method and system for constructing credit profiles based on dynamic weights and incremental learning. The aim is to solve the technical problem of how to integrate multi-source heterogeneous data within a unified framework to achieve dynamic, accurate, and interpretable assessment of user credit, and to enable rapid response to risk events and adaptive model adjustments.

[0006] To achieve the above objectives, this application provides a method for constructing credit profiles based on dynamic weights and incremental learning, the method comprising: The system receives credit profile construction requests from users and collects heterogeneous user data from multiple preset data sources, classifying the heterogeneous user data into feature data for multiple preset credit dimensions. The multiple preset data sources include: financial platforms, social networks, e-commerce platforms, and public credit databases. The multiple preset credit dimensions include: identity verification, performance capability, behavioral preferences, social connections, and environmental risks. The feature data is weighted according to a preset dynamic weighting strategy, and in a privacy-preserving environment, the weighted feature data is fused with real-time streaming data to generate time-series feature data. The time-series feature data is subjected to time-series analysis and anomaly detection to identify abnormal behavior, and a preset response operation is triggered when the abnormal behavior is identified; the response operation includes at least one of triggering manual review or storing the record of the abnormal behavior as evidence. A credit score is generated based on the time-series feature data and the identified abnormal behavior using a preset credit assessment model, and an interpretability report indicating the contribution of each feature data to the credit score is generated based on an attribution algorithm. Monitor user behavior events and incrementally update the credit assessment model when preset update conditions are triggered.

[0007] Furthermore, the collection of heterogeneous user data from multiple preset data sources includes: The system parses the user's credit profile construction request and captures data from various platforms through distributed task scheduling. The captured data from the various platforms is stream-cleaned, privacy-de-identified, and then categorized; among these processes, structured data is aggregated, semi-structured logs are standardized and parsed, and semantic features are extracted from unstructured text.

[0008] Furthermore, the process of classifying and processing the captured data from various platforms after streaming cleaning and privacy desensitization includes: Kafka is used to transmit the data from the various platforms to a preset cleaning layer. HTML tags and advertising noise are removed by a regular expression matching engine, and privacy is desensitized by AES-256 encryption and masking. The privacy-de-identified data is processed through a multimodal classification pipeline; structured data is aggregated using Flink SQL, semi-structured JSON logs are parsed into standardized key-value pairs using Apache NiFi, and unstructured text is extracted for semantic features using Tesseract OCR combined with NLP word segmentation.

[0009] Furthermore, the step of weighting the feature data according to a preset dynamic weighting strategy, and fusing the weighted feature data with real-time streaming data under a privacy-preserving environment to generate time-series feature data, includes: The weights of each feature data are determined by comprehensively considering the time decay effect and information entropy of the data. Based on the weights of each feature data, the weighted feature data and real-time streaming data are fused through federated learning to generate time-series feature data.

[0010] Furthermore, the step of performing time-series analysis and anomaly detection on the time-series feature data to identify abnormal behavior, and triggering a preset response operation when the abnormal behavior is identified, includes: The time-series feature data is decomposed into time series components to separate the residual components. Calculate the fluctuation entropy based on the residual components; The fluctuation entropy is provided as input to a preset anomaly detection model to identify the abnormal behavior, and a preset response operation is triggered when the abnormal behavior is identified.

[0011] Furthermore, the attribution algorithm is a Shapley value-based attribution algorithm.

[0012] Furthermore, the credit assessment model is incrementally updated, including: When the preset update condition is triggered, the underlying network parameters used for feature extraction in the credit assessment model are frozen, and only the top-level network parameters used for decision-making are fine-tuned.

[0013] Furthermore, the preset update condition is a composite triggering condition, which includes: When performing time series analysis and anomaly detection on the aforementioned time series feature data, the reconstruction error of the anomaly detection exceeds a dynamic threshold for a preset number of consecutive times; and Based on the attribution algorithm, an interpretability report is generated indicating the contribution of each feature data to the credit score. The contribution of the preset high-risk feature increases by more than a preset threshold within a preset time period.

[0014] To achieve the above objectives, this application also provides a credit profile construction system based on dynamic weights and incremental learning, applied to any of the above-described credit profile construction methods based on dynamic weights and incremental learning. The system includes: The data collection and classification module is used to receive credit profile construction requests from users and to collect heterogeneous user data from multiple preset data sources, and classify the heterogeneous user data into feature data of multiple preset credit dimensions; the multiple preset data sources include: financial platforms, social networks, e-commerce platforms and public credit databases; the multiple preset credit dimensions include: identity verification, performance ability, behavioral preferences, social connections and environmental risks. The dynamic weighted fusion module is used to weight the feature data according to a preset dynamic weighting strategy, and fuse the weighted feature data with real-time streaming data in a privacy-preserving environment to generate time-series feature data. The time series analysis and detection module is used to perform time series analysis and anomaly detection on the time series feature data to identify abnormal behavior, and to trigger a preset response operation when the abnormal behavior is identified. The response operation includes at least one of triggering manual review or storing the record of the abnormal behavior as evidence. The credit assessment generation module is used to generate a credit score based on the time-series feature data and the identified abnormal behavior through a credit assessment model, and to generate an interpretability report indicating the contribution of each feature data to the credit score based on an attribution algorithm. The model incremental update module is used to monitor user behavior events and perform incremental updates on the credit assessment model when preset update conditions are triggered.

[0015] Furthermore, the credit assessment generation module is also used for: The interpretability report is parsed to extract a preset number of features that contribute the most to the credit score; Based on the extracted features and their contribution, the data are fed back to the credit assessment model to adjust its internal weights.

[0016] Compared with the prior art, this application has the following beneficial effects: 1. Improved assessment accuracy and dynamism: By integrating multi-source heterogeneous data and adopting a time- and space-sensitive dynamic weighting strategy, it overcomes the shortcomings of traditional methods that rely on a single data source and fixed weights, making credit profiles more comprehensive and assessments more accurate. At the same time, in-depth analysis and anomaly detection of time-series data can capture dynamic changes in user behavior in a timely manner, improving the sensitivity of risk identification.

[0017] 2. Enhanced risk response efficiency: Through an incremental model update mechanism, when a high-risk event is detected, the system can quickly fine-tune the model and update the score. Compared with the traditional periodic batch update method, this greatly improves the response efficiency to sudden risks.

[0018] 3. It solves the "black box" problem of the model and meets compliance requirements. By using attribution algorithms to generate interpretable reports, it clearly shows the impact of various features on credit scores, making the assessment results transparent and traceable, and meeting the compliance requirements of financial supervision.

[0019] 4. It ensures data privacy and security. During the data fusion phase, privacy-preserving computing technologies such as federated learning are used to ensure that data from all parties can be jointly modeled without disclosing the original information, thus resolving the contradiction between the use of multi-source data and privacy protection. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating a credit profile construction method based on dynamic weights and incremental learning proposed in an embodiment of the present invention. Figure 2 This is a schematic diagram of the composition of a credit profile construction system based on dynamic weights and incremental learning proposed in an embodiment of the present invention. Detailed Implementation

[0023] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining this application and are not intended to limit the scope of protection of this application.

[0024] For specific implementation details, please refer to [link / reference]. Figure 1 , Figure 1 This is a flowchart illustrating a credit profile construction method based on dynamic weights and incremental learning. The method includes: S1. Receive a credit profile construction request from the user terminal, and collect heterogeneous user data from multiple preset data sources, and classify the heterogeneous user data into feature data of multiple preset credit dimensions; wherein, the multiple preset data sources include: financial platforms, social networks, e-commerce platforms and public credit databases; the multiple preset credit dimensions include: identity verification, performance capability, behavioral preferences, social connections and environmental risks.

[0025] S2. The feature data is weighted according to a preset dynamic weighting strategy, and under a privacy-preserving environment, the weighted feature data is fused with real-time streaming data to generate time-series feature data; S3. Perform time series analysis and anomaly detection on the time series feature data to identify abnormal behavior, and trigger a preset response operation when the abnormal behavior is identified; the response operation includes at least one of triggering manual review or storing the record of the abnormal behavior as evidence; S4. Using a preset credit assessment model, a credit score is generated based on the time-series feature data and the identified abnormal behavior, and an interpretability report indicating the contribution of each feature data to the credit score is generated based on an attribution algorithm. S5. Monitor user behavior events and incrementally update the credit assessment model when a preset update condition is triggered.

[0026] In specific implementation, as described in step S1, the process includes: receiving a credit profile construction request from the user, crawling heterogeneous data of the user from financial platforms, social networks, e-commerce behavior, and public credit databases, including structured transaction records, semi-structured JSON logs, and unstructured comment texts, through a distributed crawler framework, and classifying the heterogeneous data into historical feature data of five dimensions: identity verification, performance capability, behavioral preferences, social associations, and environmental risks. In step S1, a credit profile construction request is received from the user's client. A distributed crawler framework is used to crawl heterogeneous data about the user from financial platforms, social networks, e-commerce behavior, and public credit databases, including: After user authorization, the identity identifier in the profile request is parsed, and a crawling task queue is generated through a distributed task scheduler. The task is dynamically allocated according to the data source type. Financial data calls Headless Chrome to simulate logging into a bank page, social data is crawled using API, and public data is connected to the government OpenID interface. The tasks performed by the crawler nodes include: crawling transaction records and real-time liabilities from financial platforms over the past 5 years; extracting relationship topology and text sentiment tags from social networks; analyzing e-commerce behavior such as purchase frequency and return rate; verifying educational background and criminal records from public databases; cleaning the data through a streaming filter and marking low-trust sources; and entering the classification pipeline after initial screening.

[0027] In practice, the core of this step is to comprehensively acquire multi-dimensional user data through a distributed, multi-strategy crawling mechanism, under the premise of compliance (user authorization, official interface), and initially ensure data quality, laying the foundation for subsequent credit profile construction.

[0028] Furthermore, the data is cleaned by a streaming filter and low-confidence sources are marked. After initial screening, the data enters the classification pipeline, including: After the data enters the cleaning layer through the Kafka pipeline, the regular expression matching engine first removes HTML tags and advertising script noise, and then the privacy desensitization module performs AES-256 encryption on sensitive fields and adds a mask. The cleaned data enters the multimodal classification pipeline. Structured transaction records are aggregated and processed in real time using Flink SQL. Semi-structured JSON logs are parsed into standardized key-value pairs by Apache NiFi. Unstructured comment texts are extracted for semantic features using Tesseract OCR combined with NLP word segmentation technology.

[0029] In practice, the privacy desensitization module then performs AES-256 encryption on sensitive fields and adds a mask, including: The source credibility evaluator of the credit profile building device dynamically generates a confidence score based on the information website filing status, SSL certificate level and historical data accuracy. Data with a score lower than the preset value is marked with a low credibility label and automatically triggers a manual review process. After the cleaned data enters the multimodal classification pipeline, it is aligned and associated with the five credit dimensions by knowledge graph entities. Abnormal data is automatically transferred to the blockchain evidence queue to await subsequent completion, forming a fully automated data governance closed loop.

[0030] In specific implementation, as described in step S2, the feature data is weighted according to a preset dynamic weighting strategy, and under a privacy-preserving environment, the weighted feature data is fused with real-time streaming data to generate time-series feature data, including: The weights of each feature data are determined by comprehensively considering the time decay effect and information entropy of the data. Based on the weights of each feature data, the weighted feature data and real-time streaming data are fused through federated learning to generate time-series feature data.

[0031] In practice, a time decay function is used to reduce the weight of historical feature data, the entropy weight method is used to quantify historical feature data, an initial weight matrix is ​​generated by combining it with an expert rule base, and then a federated learning framework is introduced. In an encrypted state, the initial weight matrix is ​​fused with real-time streaming data of users obtained by the Internet of Things, and a three-dimensional tensor of user-feature-timestamp is output.

[0032] Specifically, the decay factor is dynamically set according to the user's industry attributes: λ=0.1 for high-volatility industries and λ=0.01 for stable industries. The weight of historical features is calculated using the exponential function e-λΔt, where Δt is the interval between the current time and the time when the feature was generated. Simultaneously, the entropy weight method is used to analyze the information entropy of the features and calculate the dispersion of each feature. The lower the information entropy, the higher the objective weight. The weights output by the entropy weight method are matched with the expert rule base. If the dispersion is lower than the preset value, the weights are forcibly increased for specific scenarios. Regulatory policy factors are dynamically injected to construct a three-dimensional weight matrix of user-feature-time. After aggregating data from multiple parties through federated learning, an encrypted weight tensor is output.

[0033] In practice, the impact of old data is reduced by time decay, and the importance of each feature is determined based on the degree of data change (entropy weight method) and expert experience to form initial weights. Then, real-time data is fused using encryption (federated learning) to finally obtain structured data containing user, feature and time information.

[0034] In specific implementation, as described in step S3, the time-series feature data is subjected to time-series analysis and anomaly detection to identify abnormal behavior, and a preset response operation is triggered when the abnormal behavior is identified, including: The time-series feature data is decomposed into time series components to separate the residual components. Calculate the fluctuation entropy based on the residual components; The fluctuation entropy is provided as input to a preset anomaly detection model to identify the abnormal behavior, and a preset response operation is triggered when the abnormal behavior is identified.

[0035] Specifically, the BERT-wwm pre-trained model is used to extract the user-feature-timestamp three-dimensional tensor. Then, the periodic term and trend term are separated by STL decomposition, the fluctuation entropy feature is extracted, and an adversarial autoencoder is generated to detect anomalies. When the error of the generated anomaly exceeds the dynamic threshold, manual review is triggered, and an anomaly report is generated and stored in the blockchain for evidence preservation.

[0036] More specifically, the periodic and trend terms are separated using STL (Seasonal and Trend Decomposition using LOESS), and fluctuation entropy features are extracted to generate an adversarial autoencoder for detecting anomalies, including: A sliding window is used to separate the periodic component, trend component, and residual term. The main period frequency is detected by Fourier transform, and the fluctuation entropy characteristic of the residual term is calculated, as shown in the following formula: (1) Where H represents the fluctuation entropy, p(rk) is the probability density of the residual amplitude in the interval k, and a high entropy value greater than 3.0 indicates potential behavioral anomalies; The fluctuation entropy feature is then input into the adversarial autoencoder. The generator reconstructs the input and calculates the Euclidean distance with the original data. When the reconstruction error ε>μt+2.5σt, an anomaly is marked, triggering the cross-platform verification process and completing the dynamic update of the threshold.

[0037] In specific implementation, as in step S4, a credit score is generated based on the time-series feature data and the identified abnormal behavior using a preset credit assessment model. An interpretability report indicating the contribution of each feature data point to the credit score is then generated based on an attribution algorithm, including: An incremental ensemble model is deployed to learn adversarial autoencoders to detect outlier features. MetaWeight-Net is used to dynamically adjust model weights, and model parameters are updated based on real-time prediction accuracy. A three-dimensional credit score is output, and the score results are used to generate a readable evaluation report with key influencing factors annotated.

[0038] In practice, the model parameters are updated based on the real-time prediction accuracy, and a three-dimensional credit score is output. The score results are used to generate a readable assessment report using Shapley values, highlighting key influencing factors, including: The base model weights are dynamically adjusted using MetaWeight-Net. If the AUC of the incremental ensemble model decreases by more than 5% in the new dataset, its weight is reduced by 0.15. The output three-dimensional credit score includes: a base score of 0-1000, reflecting historical performance; a volatility coefficient of 0-1, quantifying economic sensitivity; a fraud probability of 0-1, monitoring short-term anomalies. The score update delay is controlled within 90 seconds. The Shapley value attribution algorithm is used to calculate the marginal contribution of each feature to the score, as shown in the following formula: (2) Where: F is the set of all features, S is a subset of features, |S| is the number of features in subset S, f(S) is the output of the model trained with the features of subset S, and f(S∪{i}) is the output of the model after adding feature i. The Shapley value of feature i is calculated by the Shapley value attribution algorithm, and a visualization report is generated.

[0039] In specific implementation, as described in step S5, monitoring user behavior events and incrementally updating the credit assessment model when preset update conditions are triggered includes: When the preset update condition is triggered, the underlying network parameters used for feature extraction in the credit assessment model are frozen, and only the top-level network parameters used for decision-making are fine-tuned.

[0040] The preset update condition is a composite trigger condition, which includes: When performing time series analysis and anomaly detection on the aforementioned time series feature data, the reconstruction error of the anomaly detection exceeds a dynamic threshold for a preset number of consecutive times; and Based on the attribution algorithm, an interpretability report is generated indicating the contribution of each feature data to the credit score. The contribution of the preset high-risk feature increases by more than a preset threshold within a preset time period.

[0041] It should be noted that under what circumstances will an update be triggered (compound triggering conditions)? An update will only be initiated if two conditions are met simultaneously to ensure its necessity: Condition 1: During anomaly detection, the model's "reconstruction error" for anomalous behavior exceeds the dynamic threshold multiple times consecutively. (Simply put: if the model repeatedly fails to accurately identify anomalous behavior, it indicates that it is not adapted to new anomaly patterns.) Condition 2: In the interpretability report, the impact of pre-defined "high-risk characteristics" (such as "recent frequent borrowing" and "overdue payments on related accounts") on credit scores increases beyond the set value over a period of time. (Simply put: the importance of certain high-risk factors suddenly increases, and the original decision-making logic of the model needs to be adjusted to match the new risk landscape).

[0042] Overall, this mechanism makes model updates more "cautious" and "efficient," only initiating adjustments when necessary and modifying only key parts, thus ensuring that the model keeps pace with the times while maintaining the stability of the evaluation.

[0043] In practice, when a high-risk behavioral event is detected, the underlying feature extraction layer of the readability report module is immediately locked. Gradient masking technology is used to maintain the general pattern recognition capability obtained from pre-training, ensuring that the stability of the evaluation report is not affected by the event. With the feature layer locked, a small-step optimization algorithm is used to update the scoring weights based on new event samples, focusing on event-related feature dimensions, adding anti-overfitting constraints to complete the adjustment, and then updating the evaluation report synchronously to clearly label the event influencing factors.

[0044] More specifically, when a behavioral event risk occurs, the underlying feature extraction layer of the readability assessment report is frozen, and only the fully connected layer is fine-tuned, including: When a high-risk behavioral event is detected, such as more than 5 cross-platform loans in a single day, the underlying feature extraction layer of the readability report module is immediately locked, including text analysis, time series processing and graph neural network components. Gradient masking technology is used to maintain the general pattern recognition capability obtained from pre-training, ensuring that the stability of the evaluation report is not affected by the event. With the feature layer locked, only the top-level decision layer is fine-tuned. A small-step optimization algorithm is used to update the scoring weights based on new event samples, focusing on event-related feature dimensions, adding anti-overfitting constraints, and completing the adjustment within 15 seconds. After processing, the evaluation report is updated synchronously, and the influencing factors of the event are clearly marked.

[0045] This technical solution expands the boundaries of credit coverage, integrates similar user profiles into the cold start activation model to enhance the credibility of those without credit history, increases customer coverage for financial institutions, enables personalized privacy budgeting through differential privacy APIs, achieves a score correlation r>0.95 after noise injection, and utilizes authoritative data for compliant access through off-chain oracle verification in a trusted execution environment and zero-knowledge concise non-interactive knowledge argumentation, reducing the risk of violations to zero. Explainable knowledge graph attribution improves regulatory audit efficiency, shortens user-side appeal processing time, reduces the cost of cross-border credit migration, and comprehensively optimizes financial inclusion and security.

[0046] Compared with existing technologies, the dynamic user credit profile construction system and method provided by this invention have the following beneficial effects: 1. Improved assessment accuracy through time- and space-sensitive dynamic weight allocation; high accuracy in predicting bad debts of cyclical industry users in financial stress tests; STL fluctuation entropy + AAE adversarial coding for cross-chain collaborative anomaly detection enables cross-platform fraud confirmation within 90 seconds, greatly reducing the false alarm rate; incremental ensemble learning reduces model response latency from hours to 90 seconds, improving the efficiency of sudden risk identification; simultaneously resolving the contradiction between multi-source data utilization and privacy compliance; biometric features are homomorphically encrypted to meet GDPR requirements. 2. Complete the expansion of credit coverage boundaries, integrate similar user profiles into the cold start activation model to improve the credibility of those without credit history, increase the customer coverage of financial institutions, enable personalized privacy budgeting through differential privacy API, achieve a score correlation r>0.95 after noise injection, verify the oracle off-chain through a trusted execution environment and compliant access to authoritative data through zero-knowledge concise non-interactive knowledge argumentation, reduce the risk of violations to zero, improve the efficiency of regulatory audits through interpretable knowledge graph attribution, shorten the processing time of user-end appeals, promote the reduction of cross-border credit migration costs, and comprehensively optimize financial inclusion and security.

[0047] Please see Figure 2 , Figure 2 This is a schematic diagram illustrating the composition of a credit profile construction system based on dynamic weights and incremental learning, as proposed in an embodiment of the present invention. The system includes: The data collection and classification module 10 is used to receive credit profile construction requests from users and to collect heterogeneous user data from multiple preset data sources, and classify the heterogeneous user data into feature data of multiple preset credit dimensions; the multiple preset data sources include: financial platforms, social networks, e-commerce platforms and public credit databases; the multiple preset credit dimensions include: identity verification, performance capability, behavioral preferences, social connections and environmental risks. The dynamic weighted fusion module 20 is used to weight the feature data according to a preset dynamic weighting strategy, and fuse the weighted feature data with real-time streaming data in a privacy-preserving environment to generate time-series feature data. The time series analysis and detection module 30 is used to perform time series analysis and anomaly detection on the time series feature data to identify abnormal behavior, and when the abnormal behavior is identified, trigger a preset response operation, the response operation including at least one of triggering manual review or storing the record of the abnormal behavior as evidence. The credit assessment generation module 40 is used to generate a credit score based on the time-series feature data and the identified abnormal behavior through a credit assessment model, and to generate an interpretability report indicating the contribution of each feature data to the credit score based on an attribution algorithm. The model incremental update module 50 is used to monitor user behavior events and perform incremental updates on the credit assessment model when preset update conditions are triggered.

[0048] Furthermore, the credit assessment generation module is also used for: The interpretability report is parsed to extract a preset number of features that contribute the most to the credit score; Based on the extracted features and their contribution, the data are fed back to the credit assessment model to adjust its internal weights.

[0049] In this embodiment, the multi-source data acquisition and classification module includes: a request parsing and task scheduling unit, used to parse user authorized identity identifiers, generate a crawler task queue, and dynamically allocate tasks; a heterogeneous data real-time crawling unit, used to extract 5 years of transaction history and real-time liabilities from financial platforms, crawl relationship topology and sentiment tags from social networks, analyze consumption frequency and return rates from e-commerce behavior, verify educational background and criminal records from public databases, and initially screen and mark low-trust sources in the data stream; a knowledge graph entity alignment unit, used to map the raw data to five dimensions through a relational reasoning engine; and a cleaning and compliance unit, used to perform positive checks within the Kafka pipeline. The system filters, desensitizes privacy, and performs compliance checks, outputting structured data with confidence scores. A cold-start activation model integrates similar user profiles, enhancing the credibility of those without credit history and increasing customer coverage for financial institutions. A differential privacy API enables personalized privacy budgeting, and after noise injection, the score correlation r>0.95. Off-chain oracle verification, through a trusted execution environment and zero-knowledge concise non-interactive knowledge argumentation, ensures compliant access to authoritative data, reducing the risk of violations to zero. Explainable knowledge graph attribution improves regulatory audit efficiency, shortens user-side appeal processing time, reduces the cost of cross-border credit migration, and comprehensively optimizes financial inclusion and security.

[0050] This technical solution improves evaluation accuracy through time- and space-sensitive dynamic weight allocation, achieving high accuracy in predicting bad debts for cyclical industry users in financial stress tests. Cross-chain collaborative anomaly detection using STL volatility entropy + AAE adversarial coding enables cross-platform fraud confirmation within 90 seconds, simultaneously resolving the conflict between multi-source data utilization and privacy compliance. Biometric features are homomorphically encrypted to meet GDPR requirements. Differential privacy APIs enable personalized privacy budgeting. After noise injection, the score correlation r>0.95. Oracle off-chain verification utilizes a trusted execution environment and zero-knowledge concise non-interactive knowledge argumentation to compliantly access authoritative data, reducing the risk of violations to zero. Explainable knowledge graph attribution improves regulatory audit efficiency.

[0051] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for constructing credit profiles based on dynamic weights and incremental learning, characterized in that, include: Receive credit profile construction requests from users and collect heterogeneous user data from multiple preset data sources, and classify the heterogeneous user data into feature data of multiple preset credit dimensions; The multiple preset data sources include: financial platforms, social networks, e-commerce platforms, and public credit databases; the multiple preset credit dimensions include: identity verification, ability to fulfill obligations, behavioral preferences, social connections, and environmental risks. The feature data is weighted according to a preset dynamic weighting strategy, and in a privacy-preserving environment, the weighted feature data is fused with real-time streaming data to generate time-series feature data. The time-series feature data is subjected to time-series analysis and anomaly detection to identify abnormal behavior, and a preset response operation is triggered when the abnormal behavior is identified; the response operation includes at least one of triggering manual review or storing the record of the abnormal behavior as evidence. A credit score is generated based on the time-series feature data and the identified abnormal behavior using a preset credit assessment model, and an interpretability report indicating the contribution of each feature data to the credit score is generated based on an attribution algorithm. Monitor user behavior events and incrementally update the credit assessment model when preset update conditions are triggered.

2. The method according to claim 1, characterized in that, The collection of heterogeneous user data from multiple preset data sources includes: The system parses the user's credit profile construction request and captures data from various platforms through distributed task scheduling. The captured data from the various platforms is stream-cleaned, privacy-de-sensitized, and then categorized; among these processes, structured data is aggregated, semi-structured logs are standardized and parsed, and semantic features are extracted from unstructured text.

3. The method according to claim 2, characterized in that, The process of classifying and processing the captured data from various platforms after streaming cleaning and privacy desensitization includes: Kafka is used to transmit the data from the various platforms to a preset cleaning layer. HTML tags and advertising noise are removed by a regular expression matching engine, and privacy is desensitized by AES-256 encryption and masking. The privacy-de-identified data is processed through a multimodal classification pipeline; structured data is aggregated using Flink SQL, semi-structured JSON logs are parsed into standardized key-value pairs using Apache NiFi, and unstructured text is extracted for semantic features using Tesseract OCR combined with NLP word segmentation.

4. The method according to claim 1, characterized in that, The step of weighting the feature data according to a preset dynamic weighting strategy, and then fusing the weighted feature data with real-time streaming data under a privacy-preserving environment to generate time-series feature data includes: The weights of each feature data are determined by comprehensively considering the time decay effect and information entropy of the data. Based on the weights of each feature data, the weighted feature data and real-time streaming data are fused through federated learning to generate time-series feature data.

5. The method according to claim 1, characterized in that, The step of performing time-series analysis and anomaly detection on the time-series feature data to identify abnormal behavior, and triggering a preset response operation when the abnormal behavior is identified, includes: The time-series feature data is decomposed into time series components to separate the residual components. Calculate the fluctuation entropy based on the residual components; The fluctuation entropy is provided as input to a preset anomaly detection model to identify the abnormal behavior, and a preset response operation is triggered when the abnormal behavior is identified.

6. The method according to claim 1, characterized in that, The attribution algorithm is a Shapley value-based attribution algorithm.

7. The method according to claim 1, characterized in that, Incremental updates to the credit assessment model include: When the preset update condition is triggered, the underlying network parameters used for feature extraction in the credit assessment model are frozen, and only the top-level network parameters used for decision-making are fine-tuned.

8. The method according to claim 7, characterized in that, The preset update condition is a composite trigger condition, which includes: When performing time series analysis and anomaly detection on the aforementioned time series feature data, the reconstruction error of the anomaly detection exceeds a dynamic threshold for a preset number of consecutive times; and Based on the attribution algorithm, an interpretability report is generated indicating the contribution of each feature data to the credit score. The contribution of the preset high-risk feature increases by more than a preset threshold within a preset time period.

9. A credit profile construction system based on dynamic weights and incremental learning, applied to the credit profile construction method based on dynamic weights and incremental learning as described in any one of claims 1-8, characterized in that, include: The data acquisition and classification module is used to receive credit profile construction requests from users, collect heterogeneous user data from multiple preset data sources, and classify the heterogeneous user data into feature data of multiple preset credit dimensions. The multiple preset data sources include: financial platforms, social networks, e-commerce platforms, and public credit databases; the multiple preset credit dimensions include: identity verification, ability to fulfill obligations, behavioral preferences, social connections, and environmental risks. The dynamic weighted fusion module is used to weight the feature data according to a preset dynamic weighting strategy, and fuse the weighted feature data with real-time streaming data in a privacy-preserving environment to generate time-series feature data. The time series analysis and detection module is used to perform time series analysis and anomaly detection on the time series feature data to identify abnormal behavior, and to trigger a preset response operation when the abnormal behavior is identified. The response operation includes at least one of triggering manual review or storing the record of the abnormal behavior as evidence. The credit assessment generation module is used to generate a credit score based on the time-series feature data and the identified abnormal behavior through a credit assessment model, and to generate an interpretability report indicating the contribution of each feature data to the credit score based on an attribution algorithm. The model incremental update module is used to monitor user behavior events and perform incremental updates on the credit assessment model when preset update conditions are triggered.

10. The system according to claim 9, characterized in that, The credit assessment generation module is also used for: The interpretability report is parsed to extract a preset number of features that contribute the most to the credit score; Based on the extracted features and their contribution, the data are fed back to the credit assessment model to adjust its internal weights.

Images