A risk assessment method and device, a computer device and a storage medium

Abstract

The application discloses a risk assessment method and device, computer equipment and a storage medium, belongs to the artificial intelligence technical field, and is applied to a financial loan risk assessment scene. The application obtains multi-source heterogeneous data from a preset data source and uniformly preprocesses the multi-source heterogeneous data to construct a structured data set. A multi-dimensional feature system is constructed by combining automatic feature engineering and expert experience, so that the model can simultaneously depict data statistical features and business semantic features. A structure in which an expert model and a machine learning model are fused is adopted to realize collaborative work of rule driving and data driving. A dynamic weight adjustment mechanism based on user behavior and time decay is introduced in the training process. Meanwhile, the output results of the expert model and the machine learning model are weighted and fused. The application enhances the expansibility and stability of the scoring system through modular and fused system architecture design, thereby effectively making up for the obvious defects of the existing personal customer scoring technology in the above aspects.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, specifically relating to a risk assessment method, apparatus, computer equipment, and storage medium. Background Technology

[0002] Currently, personal customer credit scoring models have become a core tool for risk management in financial institutions. However, existing scoring techniques have gradually revealed a series of limitations in practical applications, making them difficult to adapt to the increasingly complex financial risk environment. A thorough analysis of existing technologies reveals their main shortcomings in the following aspects: Insufficient data integration and feature engineering. Traditional credit scoring models primarily rely on structured data such as central bank credit reports, which are limited in scope and outdated, making it difficult to comprehensively reflect a customer's creditworthiness. The integration and utilization of multi-source heterogeneous data (such as e-commerce transactions, social media, and mobile device usage) faces technical bottlenecks, lacking effective means to handle issues like uneven data quality, missing values, and noisy data. While the industry has attempted to use rule-based cleaning algorithms combined with machine learning techniques for data preprocessing, feature engineering still heavily relies on expert experience, which is time-consuming, labor-intensive, and struggles to uncover complex nonlinear relationships within the data. Manual feature engineering is subject to subjective bias and cannot effectively handle unstructured data such as text and time-series behavior, resulting in an insufficiently comprehensive and multi-dimensional portrayal of customer risk profiles.

[0003] The intelligence and adaptability of existing scoring models are limited. Most existing scoring models are static and lack dynamic adjustment mechanisms. While traditional logistic regression or scorecard models offer strong interpretability, they struggle to capture complex nonlinear relationships between variables. Single machine learning models (such as random forests and support vector machines) improve prediction accuracy to some extent, but perform poorly in scenarios with imbalanced data (default samples are typically below 5%) or sparse features. More significantly, these models generally suffer from overfitting, exhibiting slow convergence, a tendency to get trapped in local optima, and poor generalization ability. Once deployed, the parameters and weights are often fixed, failing to adapt to external factors such as changes in user behavior or economic cycle fluctuations, leading to performance degradation over time.

[0004] The lack of a weight adjustment mechanism is a significant issue. Most existing scoring systems employ static weight allocation strategies, relying on expert experience or fixed rules, making them ill-suited to diverse scoring needs across various scenarios. This static weight allocation lacks the flexibility to respond to dynamic changes in user behavior, time decay effects, and adjustments to business objectives. When faced with frequent or high-value behavioral changes, the system cannot update the weights of each indicator in real time, resulting in scoring results that fail to accurately reflect the customer's latest risk status. Furthermore, static weight mechanisms are also prone to weight oscillations and overfitting, impacting the stability and fairness of the scoring results.

[0005] Insufficient interpretability and real-time performance. Many advanced machine learning models (such as deep neural networks), while possessing high predictive accuracy, operate like "black boxes," making their decision-making processes difficult to interpret. This severely limits their application in the financial sector, where compliance requirements are stringent. Simultaneously, traditional batch processing methods lead to delays in score updates, hindering real-time risk monitoring and early warning. Most systems lack the ability to trace intermediate results; when business personnel need to query details of the rating process, they often need to call different systems multiple times or manually compare large amounts of data, resulting in a poor query experience and time-consuming problem localization.

[0006] System architecture and query efficiency issues. Traditional customer rating processes suffer from low query efficiency, complex operations, and disorganized data structures. Especially when business or risk control personnel need to quickly query customer risk levels and rating process details, they often need to call different systems multiple times or manually compare large amounts of data, resulting in a poor query experience and impacting work efficiency and anti-money laundering management effectiveness. The system lacks an effective tracing mechanism, making it difficult to backtrack and verify rating results, thus reducing the reliability and credibility of the rating model.

[0007] Existing personal customer scoring technologies have significant shortcomings in data integration, model intelligence, dynamic weight adjustment, interpretability, and system architecture. There is an urgent need for a new scoring model that can integrate multi-source data, has an adaptive weight adjustment mechanism, and balances accuracy and interpretability. Summary of the Invention

[0008] The purpose of this application is to propose a risk assessment method, apparatus, computer equipment, and storage medium to address the shortcomings of existing personal customer scoring technologies in terms of data integration, model intelligence, dynamic weight adjustment, interpretability, and system architecture.

[0009] To address the aforementioned technical problems, this application provides a risk assessment method, employing the following technical solution: A risk assessment method, comprising: Obtain multi-source data related to customers from preset data sources, and preprocess the multi-source data to obtain a structured dataset; For structured datasets, a feature set is constructed from multiple dimensions by combining automated feature engineering with expert experience; Based on a feature set, a fusion architecture combining expert models and machine learning models is used to train the risk assessment model; During model training, a dynamic weight adjustment mechanism based on user behavior and time decay is introduced to dynamically adjust the feature weights in the feature system set. The output of the expert model and the output of the machine learning model are weighted and fused to obtain the risk assessment score; The risk level is determined by classifying risk levels based on risk assessment scores and preset ranges.

[0010] To address the aforementioned technical problems, this application also provides a risk assessment device, which employs the following technical solution: A risk assessment device, comprising: The data processing module is used to obtain multi-source data related to customers from preset data sources and preprocess the multi-source data to obtain a structured dataset. The feature building module is used to build a feature system set from multiple dimensions for structured datasets by combining automated feature engineering with expert experience. The model fusion module is used to train risk assessment models based on a feature set and employing a fusion architecture that combines expert models and machine learning models. The dynamic training module is used to introduce a dynamic weight adjustment mechanism based on user behavior and time decay to dynamically adjust the feature weights in the feature system set during model training. The weighted fusion module is used to weight and fuse the output of the expert model with the output of the machine learning model to obtain a risk assessment score. The risk assessment module is used to classify risk levels based on risk assessment scores and preset ranges, and to determine the risk assessment level.

[0011] To address the aforementioned technical problems, this application also provides a computer device that employs the following technical solution: A computer device includes a memory and a processor, the memory storing computer-readable instructions, the processor executing the computer-readable instructions to implement the steps of the risk assessment method as described in any of the preceding claims.

[0012] To address the aforementioned technical problems, this application also provides a computer-readable storage medium, employing the technical solution described below: A computer-readable storage medium storing computer-readable instructions that, when executed by a processor, implement the steps of the risk assessment method as described in any one of the preceding descriptions.

[0013] Compared with the prior art, the embodiments of this application have the following main advantages: This application discloses a risk assessment method, apparatus, computer equipment, and storage medium, belonging to the field of artificial intelligence technology, and applied to financial loan risk assessment scenarios. This application effectively solves the problems of scattered data sources, inconsistent formats, and difficulty in comprehensive utilization in existing technologies by acquiring multi-source heterogeneous data from preset data sources and performing unified preprocessing to construct a structured dataset. By combining automated feature engineering with expert experience to construct a multi-dimensional feature system, the model can simultaneously characterize statistical and business semantic features, improving the scoring model's ability to express complex risk features. At the model level, an architecture integrating expert models and machine learning models is adopted to achieve collaborative work between rule-driven and data-driven approaches, overcoming the problems of insufficient intelligence of single models or over-reliance on manual rules. During training, a dynamic weight adjustment mechanism based on user behavior and time decay is introduced, enabling feature weights to be dynamically updated according to changes in user behavior and time factors, improving the shortcomings of existing scoring models where weights are statically fixed and difficult to reflect real-time risk changes. Simultaneously, by weighted fusion of the outputs of expert models and machine learning models, the predictive ability of the model is guaranteed while retaining the rule-based reasoning path, improving the interpretability of the risk assessment process. This application enhances the scalability and stability of the scoring system through a modular and integrated system architecture design, thereby effectively making up for the obvious deficiencies of existing personal customer scoring technologies in the above aspects. Attached Figure Description

[0014] To more clearly illustrate the solutions in this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 An exemplary system architecture diagram is shown, in which this application can be applied; Figure 2 A flowchart of one embodiment of the risk assessment method according to this application is shown; Figure 3 It shows Figure 2 A flowchart of an embodiment of step S203; Figure 4 A schematic diagram of one embodiment of the risk assessment apparatus according to this application is shown; Figure 5 It shows Figure 4 A schematic diagram of an embodiment of the model fusion module 403; Figure 6 A schematic diagram of the structure of one embodiment of a computer device according to this application is shown. Detailed Implementation

[0016] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0017] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0018] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0019] like Figure 1 As shown, system architecture 100 may include terminal device 101, network 102, and server 103. Terminal device 101 may be a laptop 1011, tablet 1012, or mobile phone 1013. Network 102 is used as a medium to provide a communication link between terminal device 101 and server 103. Network 102 may include various connection types, such as wired, wireless communication links, or fiber optic cables.

[0020] Users can use terminal device 101 to interact with server 103 via network 102 to receive or send messages, etc. Various communication client applications can be installed on terminal device 101, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social media platform software, etc.

[0021] Terminal device 101 can be various electronic devices with a display screen and support web browsing. In addition to laptops 1011, tablets 1012, or mobile phones 1013, terminal device 101 can also be an e-book reader, an MP3 player (Moving Picture Experts Group Audio Layer III), an MP4 player (Moving Picture Experts Group Audio Layer IV), a laptop computer, and a desktop computer, etc.

[0022] Server 103 can be a server that provides various services, such as a backend server that provides support for the pages displayed on terminal device 101.

[0023] It should be noted that the risk assessment method provided in this application embodiment is generally executed by a server / terminal device, and correspondingly, the risk assessment device is generally set in the server / terminal device.

[0024] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative; the system can have any number of terminal devices, networks, and servers depending on implementation needs.

[0025] Continue to refer to Figure 2 A flowchart illustrating an embodiment of a risk assessment method according to this application is shown. The risk assessment method includes the following steps: S201: Obtain multi-source data related to customers from preset data sources, and preprocess the multi-source data to obtain a structured dataset; Specifically, the process of acquiring multi-source data related to customers from preset data sources includes unified access and management of multiple heterogeneous data sources. These data sources include at least business system data, transaction behavior data, device or terminal data, third-party credit or external risk data, and log data. Addressing the differences in data format, structure, and update frequency among different data sources, the raw data is first processed through data extraction, field mapping, and format conversion to eliminate structural differences. Subsequently, the extracted data undergoes cleaning, including missing value imputation, outlier identification and correction, duplicate data removal, and invalid record elimination. Based on this, time-related fields are standardized with unified timestamps, categorical fields are encoded, and numerical fields are normalized or standardized. Simultaneously, data from different data sources are linked and integrated based on the customer's unique identifier to form customer-centric data records. Finally, the cleaned, transformed, and integrated data is stored as a structured dataset.

[0026] In specific embodiments of this application, the system acquires multi-source heterogeneous data related to customers through various channels. This data includes: basic customer information (identity information, occupation information, and source of funds information), transaction behavior data (bank transaction records, POS transaction details), third-party data (People's Bank of China credit reports, blacklist information queried by third-party institutions, legal disputes and administrative penalty records, etc.), and alternative data (e-commerce transaction data, social media activities, mobile device usage behavior, etc.). To ensure data quality, this invention employs a complete data preprocessing workflow: using interpolation methods to handle outliers and missing values; identifying and correcting erroneous data through a combination of rule-based cleaning algorithms and machine learning algorithms; converting text data into vectors using Natural Language Processing (NLP) technology; and using normalization and standardization methods to unify the scale of numerical data. Furthermore, graph database technology is introduced to construct a knowledge graph containing various entities and relationships, integrating scattered data into structured information.

[0027] S202, for structured datasets, constructs a feature system set from multiple dimensions by combining automated feature engineering with expert experience; Specifically, the process of constructing a feature system set for structured datasets involves the collaborative design of automated feature engineering processes and human expert experience rules. First, based on a pre-defined automated feature engineering module, statistical feature extraction, time-series feature construction, cross-feature generation, and aggregate feature calculation are performed on the structured dataset. This includes sliding window statistics, frequency statistics, trend calculation, and volatility analysis of historical behavioral data. Simultaneously, feature selection algorithms are used to filter high-dimensional features, reducing redundancy. Second, domain experts define human-defined features based on business rules and risk experience. These include threshold features, interval features, and logical combination features, used to characterize business semantics that are difficult to express directly through automated feature engineering. Subsequently, automatically generated features and expert-defined features are uniformly named, grouped, and labeled, and classified at multiple levels according to data source, business attributes, and risk dimensions. Finally, all types of features are integrated to form a clearly structured and dimensionally scalable feature system set, which serves as the unified input feature space for the risk assessment model.

[0028] In the feature engineering phase, a combination of automated feature engineering and expert experience is employed. On one hand, deep learning algorithms (such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory Networks (LSTMs) are used to automatically extract deep features from the data. CNNs can automatically extract local features from data such as images and text, while LSTMs excel at processing time-series data, capturing key features such as repayment cycle patterns and consumption fluctuation trends. On the other hand, drawing on FICO criteria and other scoring card experience, and combining business knowledge, an expert feature system is constructed from four dimensions: repayment ability, repayment willingness, default risk, and credit history. For example, repayment ability features include merchants' average monthly transaction volume, household income and expenditure, and industry ranking; repayment willingness features are reflected through historical repayment records and credit account usage. This feature engineering approach, combining automation and expert experience, can both uncover complex nonlinear relationships in the data and ensure the business interpretability of the features.

[0029] S203, based on a feature set, uses a fusion architecture combining expert models and machine learning models to train risk assessment models; Specifically, when training a risk assessment model based on a feature set, a fusion architecture combining expert models and machine learning models is adopted. First, an expert model is constructed based on an expert rule base. Some key features from the feature set are input into the expert model, and rule matching, conditional judgment, and logical reasoning mechanisms are used to generate the expert model's risk assessment results. Simultaneously, the complete feature set is input into the machine learning model training module, and the model parameters are trained based on labeled sample data. The machine learning model may include one or more of logistic regression, tree models, ensemble models, or deep learning models. During training, feature adaptation and parameter initialization are performed for different models, and model learning is completed through cross-validation or sample partitioning. Subsequently, the outputs of the expert model and the machine learning model are aligned. Through this approach, the collaborative construction of rule-driven and data-driven models is achieved within the same training framework.

[0030] An architecture combining expert models and machine learning models is adopted to fully leverage the advantages of both. In the expert model component, initial weights are assigned to different features based on business experience to construct a highly interpretable scorecard model. In the machine learning model component, an LSTM model based on a sparse response neural network is selected for credit score estimation. This model improves convergence speed and reduces the risk of overfitting by introducing sparse terms into the loss function and automatically detecting and rejecting passive neurons.

[0031] S204 introduces a dynamic weight adjustment mechanism based on user behavior and time decay during model training to dynamically adjust the feature weights in the feature system set. Specifically, when introducing a dynamic weight adjustment mechanism based on user behavior and time decay during model training, the following steps are taken: First, user behavior data generated during the training period is continuously collected and updated, including operation frequency, behavior type changes, and behavior time distribution. Second, user behavior data is mapped to corresponding user behavior features in the feature system set, forming a behavior influence feature set. Based on this, a behavior influence factor is calculated according to behavior activity, behavior anomaly, or behavior change magnitude to reflect the degree of influence of behavior changes on feature importance. Simultaneously, a time decay function is introduced to decay the historical weights of each feature according to their update time, thereby weakening the impact of outdated information on the current model training. Subsequently, the behavior influence factor is combined with the time decay weights to generate dynamically updated feature weights. Finally, the updated feature weights are applied in real time during the model training iteration process to adjust the model parameter update process, enabling the model training process to dynamically respond to changes in user behavior and time factors.

[0032] A dynamic weight adjustment mechanism based on user behavior and time decay is introduced, enabling the rating system to adapt to changing environments. The weight adjustment formula is as follows:

[0033] in, Indicates the weight at the current moment. It is the weight from the previous moment. It is the smoothing coefficient. This is a new weighting method based on user behavior calculations. Specifically, the system dynamically adjusts the weights of relevant features by analyzing the frequency and value of user interaction behaviors (such as clicks, browsing, and purchases). Simultaneously, a time decay function (such as...) is introduced. ,in, To control the decay rate and reduce the impact of historical data on current weights, the scoring should better reflect the customer's latest situation.

[0034] In addition, the model incorporates Bayesian estimation methods to address the sparsity problem of new customer data, as shown in the formula:

[0035] in, For the final estimated score, For prior weights, The average rating on the platform. Rate users This mechanism assigns user weights. It effectively solves the cold start problem and ensures the reliability of new customer ratings.

[0036] S205, weighted fusion of the output of the expert model and the output of the machine learning model to obtain the risk assessment score; Specifically, the weighted fusion process of the expert model output and the machine learning model output includes unified processing and fusion calculation of the output results from different models. First, the risk judgment value output by the expert model and the prediction score output by the machine learning model are obtained, and the numerical range of the two types of outputs is normalized to eliminate differences in scale between the outputs of different models. Second, according to the model weights or dynamic weight allocation strategy determined during the model training phase, fusion weights are assigned to the expert model output and the machine learning model output respectively. Subsequently, in the fusion layer, the two types of output results are linearly or non-linearly combined according to preset weighting rules to generate a fused risk score. During the fusion process, smoothing or calibration mechanisms can be introduced to constrain extreme outputs. Finally, a risk assessment score in a unified format is output.

[0037] S206. Based on the risk assessment score and the preset range, the risk level is divided and the risk assessment level is determined.

[0038] Specifically, the process of classifying risk levels based on risk assessment scores and preset intervals includes two stages: risk score mapping and level determination. First, multiple risk score intervals are pre-defined based on business needs or historical data analysis, each interval corresponding to a risk level identifier. These risk score intervals can be set using a fixed threshold or dynamically based on sample distribution. Then, the risk assessment score obtained in step S205 is compared with the preset intervals to determine the range within which the score falls. During the comparison process, boundary condition processing rules can be incorporated to avoid ambiguity when the score falls within the interval's critical value. Finally, the corresponding risk assessment level is determined based on the matched interval, and this risk assessment level is stored as the model output or output to the upper-level business system.

[0039] In the scoring fusion stage, this invention weights and combines the output of the expert model and the prediction results of the machine learning model. The expert model outputs a credit score (e.g., 0-850), and the machine learning model outputs the probability of default. For example, through the formula:

[0040] Calculate the final score ,in, Output scores for the expert model. This represents the default probability output by the machine learning model. The weighting can be dynamically adjusted based on model performance. Initially, expert models have higher weights (e.g., 0.8), which can be gradually increased as the machine learning model is optimized. Ultimately, customers are categorized into different risk levels based on preset risk level ranges (e.g., blacklisted users: 100 points and below; low-risk users: below 350 points; eligible users: 350-500 points; eligible users: 500-650 points; high-quality users: 650-850 points).

[0041] Further, please refer to Figure 3 The steps for training a risk assessment model based on a feature set and using a fusion architecture combining expert models and machine learning models include: S301: Based on the feature system set, an expert rule base is constructed, and an expert model is generated using the expert rule base. Based on the expert model, rule reasoning is performed on the input features to obtain the first risk assessment result. S302, input the feature system set into the preset machine learning model, and perform supervised learning training on the machine learning model, wherein the machine learning model includes at least one or more of the following: logistic regression model, decision tree model, ensemble learning model or deep learning model, in order to obtain the second risk assessment result; S303, performs independent parameter optimization and performance evaluation on expert models and machine learning models respectively, and determines the initial weight coefficients of each model based on historical sample data; S304, Construct a fusion layer based on the initial weighting coefficients to form a fusion risk assessment model; S305, the first risk assessment result and the second risk assessment result are merged based on the initial weighting coefficients to obtain the risk fusion assessment result; S306. Based on the feature system set and the risk fusion assessment results, the fusion risk assessment model is jointly trained. The parameters of the fusion model are iteratively updated by minimizing the preset loss function to complete the training of the risk assessment model.

[0042] In this embodiment, when training a risk assessment model using a fusion architecture combining expert models and machine learning models based on a feature set, the feature set is first encapsulated with a unified data interface to ensure it can be called by both the expert model and the machine learning model. By filtering and mapping key risk features in the feature set, an expert rule base is constructed and organized into an executable set of rules for generating the expert model. During the expert model's operation, rule matching and logical reasoning mechanisms are used to judge each rule of the input features, outputting the corresponding first risk assessment result. Simultaneously, the same feature set is input into a preset machine learning model, and supervised learning training is performed using historical sample data with risk labels. Model learning is completed through parameter iteration and error backpropagation to obtain a second risk assessment result. Subsequently, independent parameter optimization and performance evaluation are performed on the expert model and the machine learning model, and the initial weight coefficients for each model are determined based on the prediction accuracy, stability, or other evaluation indicators in historical samples. Based on this, a fusion layer is constructed according to the initial weight coefficients. The outputs of the expert model and the machine learning model are used as input parameters for the fusion layer to form a fusion risk assessment model. Further, the first risk assessment result and the second risk assessment result are weighted and fused according to the initial weight coefficients to generate a risk fusion assessment result. Finally, the feature set and the risk fusion assessment result are used as training inputs to jointly train the fusion risk assessment model. The parameters of the fusion model are iteratively updated by minimizing a preset loss function, allowing the weights of the fusion layer and the model parameters to gradually converge during training, thus completing the overall training process of the fusion risk assessment model.

[0043] Through the above steps, the collaborative training of expert rule reasoning and data-driven models is realized, enabling the risk assessment model to improve its ability to identify complex risk patterns while maintaining the controllability of rules.

[0044] Furthermore, based on the feature system set, an expert rule base is constructed, and an expert model is generated using the expert rule base. The first risk assessment result is obtained by performing rule-based reasoning on the input features based on the expert model. Specifically, this includes: Based on historical risk case data and domain expert experience, key features in the feature system set are screened, and corresponding risk judgment rules and threshold conditions are defined to form an initial expert rule set. The initial expert rule set is standardized and structured to convert the risk assessment rules into an executable rule expression and store them in the expert rule base; An expert model is built based on an expert rule base. After receiving the input features of the object to be evaluated, the model performs rule reasoning on the input features according to a preset rule matching and reasoning strategy, and generates intermediate risk judgment results corresponding to each rule. The intermediate risk assessment results are conflict-resolved and consistency-verified, and then aggregated and calculated according to preset rule weights or priorities to output the first risk assessment result.

[0045] In this embodiment, the process of constructing an expert rule base and generating an expert model based on a feature system set first uses historical risk case data and domain expert experience as the rule source. Risk correlation analysis is performed on the features in the feature system set to screen out key features that have a significant impact on risk assessment. For these key features, corresponding risk assessment rules and threshold conditions are set according to different risk scenarios to describe the mapping relationship between feature values ​​and risk states, thus forming an initial expert rule set. Subsequently, the initial expert rule set undergoes unified standardization and structuring, converting rule conditions, judgment logic, and output results into executable rule expressions, and storing them in the expert rule base according to rule type, scope of application, or priority. Based on this, an expert model is constructed using an expert rule base. When the input features of the object to be evaluated are received, the expert model compares the input features with the rule conditions one by one according to the preset rule matching strategy, and generates intermediate risk judgment results corresponding to each rule through the rule reasoning mechanism. After the rule reasoning is completed, the intermediate risk judgment results are conflict-resolved and consistency-verified. The conflict results are processed by comparing rule weights, priorities or applicable conditions. Finally, the effective rule results are integrated according to the preset aggregation calculation method, and the first risk assessment result is output.

[0046] Through the above steps, rule-based risk reasoning based on expert knowledge is realized, enabling the risk assessment process to have clear logical basis and good interpretability.

[0047] Taking customer risk assessment in financial loan scenarios as an example, the expert rule base can set rules for the key feature of "number of overdue payments in the past 6 months": if the number of overdue payments is ≥3 times, a high-risk warning is triggered (the intermediate risk assessment result is "high risk"); for the feature of "monthly income-to-debt ratio", a rule can be set: when this ratio is >60%, it is judged as high risk (the intermediate risk assessment result is "medium-high risk"). In the rule reasoning process, if a customer meets both of the above rules, the expert model will first resolve the conflict and make a comprehensive judgment based on the preset priority (e.g., the "number of overdue payments" rule has a higher priority than the "monthly income-to-debt ratio" rule) or the rule weight (e.g., the weight of the overdue payment rule is 0.6, and the weight of the monthly income-to-debt ratio rule is 0.4). Assuming the customer has 4 overdue payments (triggering high risk) and a monthly income-to-debt ratio of 65% (triggering medium-high risk), if a priority strategy is used, "high risk" will be directly output as the first risk assessment result. If a weighted aggregation strategy is used, combined with intermediate results from other rules (such as the "credit inquiry number" rule outputting "low risk" with a weight of 0.3), a weighted calculation (0.6 × high risk score + 0.4 × medium-high risk score + 0.3 × low risk score) will be used to finally output a quantitative first risk assessment result (such as a risk score of 750 points, corresponding to the "high risk" level).

[0048] Simultaneously, the machine learning model inputs the customer's characteristics, such as "number of overdue payments in the past 6 months," "monthly income-to-debt ratio," and "number of credit inquiries," along with their historical loan repayment records and consumption behavior data, into the LSTM model. The model analyzes the time-series changes in these data (e.g., whether the frequency of overdue payments is increasing or whether income is declining), and outputs a second risk assessment result (e.g., a default probability of 0.35). Subsequently, the system performs a fusion calculation on the first risk assessment result (750 points) and the second risk assessment result (0.35 default probability converted to a corresponding score, such as 600 points) according to preset weights (e.g., 0.7 for the expert model and 0.3 for the machine learning model) (750×0.7+600×0.3=705 points), ultimately determining the customer's risk level as "high risk," thus assisting financial institutions in making decisions regarding whether to grant a loan.

[0049] Furthermore, the step of fusing the first risk assessment result and the second risk assessment result based on the initial weighting coefficients to obtain the risk fusion assessment result specifically includes: Based on the initial weight coefficients, corresponding model weights are assigned to the first risk assessment result and the second risk assessment result, and the model weights are used as input parameters of the fusion layer. In the fusion layer, the first risk assessment result and the second risk assessment result are weighted and combined based on the model weights to generate the intermediate result of the fusion risk assessment. The intermediate results of the fusion risk assessment are smoothed or calibrated to eliminate model output bias and output the risk fusion assessment result of the fusion risk assessment model.

[0050] In this embodiment, when fusing the first risk assessment result and the second risk assessment result based on the initial weight coefficients, firstly, according to the initial weight coefficients determined during the model training phase, corresponding model weights are assigned to the first risk assessment result output by the expert model and the second risk assessment result output by the machine learning model, respectively, and these model weights are managed as the core input parameters of the fusion layer. To ensure the comparability of different model outputs during the fusion process, the first and second risk assessment results are processed with a unified data format and numerical scale calibration before entering the fusion layer. Subsequently, in the fusion layer, the two types of risk assessment results are weighted and combined according to the model weights, and intermediate fusion risk assessment results are generated through linear weighting or preset combination rules. The fusion layer can be embedded as an independent computing module in the overall risk assessment model architecture for centralized processing of multiple model outputs. After obtaining the intermediate fusion risk assessment results, they are smoothed or calibrated, for example, by using moving averages, boundary constraints, or distribution adjustments to correct abnormal fluctuations or extreme values, so as to reduce the impact of model output bias on the final assessment result. Finally, the processed result is output as the risk fusion assessment result of the fusion risk assessment model.

[0051] Through the above steps, the risk assessment results of different models are unified and integrated, improving the stability and consistency of risk score output.

[0052] Furthermore, during model training, a dynamic weight adjustment mechanism based on user behavior and time decay is introduced to dynamically adjust the feature weights in the feature set. This includes the following steps: Collect user behavior data within a preset time window, and associate the user behavior data with the corresponding features in the feature system set to form a user behavior impact feature set; Based on the user behavior impact feature set, calculate the behavior impact factor corresponding to each feature. The behavior impact factor is used to characterize the degree of influence of changes in user behavior on the importance of the feature. Based on the preset time decay function, the historical weights of each feature in the feature system set are decayed to obtain the time decay weights corresponding to each feature. The behavioral influence factor and the time decay weight are combined for calculation, and the feature weights of each feature in the feature system set are dynamically updated.

[0053] In this embodiment, when introducing a dynamic weight adjustment mechanism based on user behavior and time decay during model training, user behavior data is first continuously collected within a preset time window. This user behavior data may include information such as operation frequency, operation type, behavior sequence, and behavior changes. The user behavior data is then mapped and associated with corresponding user behavior-related features in the feature set using user identifiers, thereby forming a user behavior impact feature set. Subsequently, based on this user behavior impact feature set, the changes in user behavior corresponding to different features are quantitatively analyzed. By statistically analyzing indicators such as behavioral activity, change magnitude, or anomaly degree, the behavioral impact factor corresponding to each feature is calculated to characterize the degree of influence of user behavior changes on feature importance. On this basis, a preset time decay function is introduced to decay the historical weights of each feature in the feature set. The time decay function reduces the historical weights based on the time interval between the feature weight update time and the current time to reflect the effectiveness of feature information changing over time. The time decay weights corresponding to each feature are obtained through the above calculations. Finally, the behavioral influence factor and the time decay weight are combined to generate dynamically updated feature weights, and the updated feature weights are applied to the parameter update stage in the model training process, thereby realizing dynamic adjustment of feature weights during training iterations.

[0054] Through the above steps, the feature weights can be dynamically updated according to changes in user behavior and time factors, thereby enhancing the model training process's adaptability to real-time information.

[0055] Furthermore, the steps for calculating the behavioral influence factor corresponding to each feature based on the user behavior influence feature set specifically include: The user behavior features in the user behavior impact feature set are classified and processed, and the user behavior is quantitatively represented according to behavior type, frequency of occurrence and behavior intensity; Based on the quantified user behavior characteristics, the activity index of each characteristic within a preset time window is calculated. The activity index is used to reflect the degree of change in user behavior. The correlation analysis between behavioral activity indicators and the historical risk contribution of corresponding user behavioral characteristics was conducted to calculate the behavioral correlation coefficient corresponding to each user behavioral characteristic. Based on the behavioral correlation coefficient, the behavioral activity index is weighted and calculated to obtain the behavioral influence factor corresponding to each feature.

[0056] In this embodiment, when calculating the behavior impact factor corresponding to each feature based on the user behavior impact feature set, the user behavior features in the user behavior impact feature set are first classified. User behavior is divided into different types according to preset behavior classification rules, and the frequency and intensity of each type of user behavior are quantified. This quantification process can be achieved through methods such as counting frequency, duration, or magnitude of change. Subsequently, based on the quantified user behavior features, statistical analysis is performed on the user behavior corresponding to each feature within a preset time window to calculate a behavior activity index reflecting the degree of change in user behavior, used to describe the fluctuation of user behavior over time. On this basis, the behavior activity index is correlated with the risk contribution of the corresponding user behavior feature in historical risk assessment samples. The correlation between user behavior features and risk outcomes is calculated through correlation analysis or regression analysis, thereby obtaining the behavior correlation coefficient corresponding to each user behavior feature. Finally, the behavior activity index is weighted based on the behavior correlation coefficient, comprehensively quantifying the magnitude of behavior change and its risk correlation to generate the behavior impact factor corresponding to each feature.

[0057] Through the above steps, a quantitative characterization of the correlation between changes in user behavior and risk is achieved, making the adjustment of feature weights more precise and based on evidence.

[0058] Furthermore, the steps of performing correlation analysis between behavioral activity indicators and the historical risk contribution of corresponding user behavioral characteristics, and calculating the behavioral correlation coefficient corresponding to each user behavioral characteristic, specifically include: Obtain the risk contribution data of each user behavior feature in the historical risk assessment sample, and standardize the risk contribution data. The standardized risk contribution data is time-aligned with the corresponding user behavior activity indicators to construct risk behavior correlation data pairs. Based on risk behavior-related data pairs, correlation analysis or regression analysis methods are used to calculate the degree of correlation between user behavior characteristics and risk outcomes, and to obtain the initial behavior correlation coefficient. The initial behavioral correlation coefficients are smoothed or confidence-corrected to obtain the behavioral correlation coefficients corresponding to each user's behavioral characteristics.

[0059] In this embodiment, when performing correlation analysis between behavioral activity indicators and the historical risk contribution of corresponding user behavioral features, the risk contribution data corresponding to each user behavioral feature is first extracted from the historical risk assessment samples. This risk contribution data describes the degree of influence of different user behavioral features in the historical risk assessment results. To ensure comparability between different features, the risk contribution data is standardized and mapped to a uniform numerical range. Subsequently, the standardized risk contribution data is time-aligned with the behavioral activity indicators of the corresponding user behavioral features, i.e., the two types of data are matched according to the same time window or timestamp, constructing risk behavior correlation data pairs containing behavioral change information and risk outcome information. Based on this, correlation analysis or regression analysis is used to model the relationship between user behavioral features and risk outcomes, calculating an initial behavioral correlation coefficient reflecting the degree of correlation between the two. Finally, the initial behavioral correlation coefficient is smoothed or confidence-corrected to reduce the impact of sample fluctuations or extreme values, obtaining stable and reliable behavioral correlation coefficients corresponding to each user behavioral feature.

[0060] The above steps improve the stability and reliability of the correlation calculation between user behavior characteristics and risk outcomes.

[0061] Furthermore, the step of calculating the decay of the historical weights of each feature in the feature system set based on a preset time decay function to obtain the time decay weights corresponding to each feature specifically includes: Obtain the historical weights and update time information of each feature in the feature system set, and calculate the time interval between the current time and the update time; Based on the time interval, each feature is divided into different time decay intervals, and a corresponding decay parameter is configured for each time decay interval; Substitute the time interval and the corresponding decay parameter into the preset time decay function to calculate the decay of the historical weights of each feature, and obtain the initial time decay weights of each feature. The initial time decay weights are normalized or subjected to threshold constraints to obtain the final time decay weights corresponding to each feature.

[0062] In this embodiment, when calculating the decay of historical weights for each feature in the feature set based on a preset time decay function, the historical weights of each feature in the feature set during historical model training or evaluation are first obtained, and the most recent update time information of the historical weights is simultaneously obtained. Then, by calculating the difference between the current time and the update time, the time interval corresponding to each feature weight is determined to reflect the time span of the feature information from the current moment. Based on this, according to the size of the time interval, each feature is divided into different time decay intervals, such as short-term, medium-term, or long-term intervals, and corresponding decay parameters are pre-configured for different time decay intervals to reflect the degree of decay of feature weights under different time spans. Subsequently, the time intervals and corresponding decay parameters are substituted into the preset time decay function to calculate the historical weights of each feature, thereby obtaining the initial time decay weights corresponding to each feature. Finally, the initial time decay weights are normalized or subjected to threshold constraints to limit the range of weight values, avoiding excessively large or small weights from interfering with the model training process, ultimately obtaining the time decay weights corresponding to each feature.

[0063] By following the steps above, the feature weights can automatically decay over time, reducing the interference of historical information on the current model training.

[0064] For example, suppose a customer's "credit inquiry frequency" feature had a historical weight of 0.25 three months ago, its most recent update was January 1, 2023, and the current time is May 1, 2023, resulting in a time interval of 4 months. If the preset time decay interval is divided into: short-term (≤2 months) decay parameter 0.9, medium-term (3-6 months) decay parameter 0.7, and long-term (>6 months) decay parameter 0.4, then the 4-month time interval for this feature belongs to the medium-term interval, corresponding to a decay parameter of 0.7. Substituting the historical weight 0.25, the 4-month time interval, and the decay parameter 0.7 into the time decay function (e.g., initial time decay weight = historical weight × e^(-decay parameter × time interval / 30), where 30 is the monthly time base), we can calculate the initial time decay weight = 0.25 × e^(-0.7 × 4 / 30) = 0.25 × e^(-0.093) ≈ 0.25 × 0.911 ≈ 0.228. The initial time decay weight of 0.228 was then normalized. Assuming the sum of the initial time decay weights of all features is 1.8, the final time decay weight of this feature after normalization is approximately 0.228 / 1.8 ≈ 0.127. The weight threshold range was set to [0.05, 0.4]. Since 0.127 falls within this range, the final time decay weight for this feature was determined to be 0.127. Through this calculation, the weight of the "number of credit inquiries" feature for this customer dynamically decays from a historical 0.25 to 0.127, effectively reducing the impact of historical weights from three months ago on the current model training, and making the model focus more on recent credit inquiry behavior.

[0065] In this embodiment, the risk assessment method operates on electronic devices (e.g., Figure 1 The server shown can receive instructions or acquire data via wired or wireless connection. It should be noted that the aforementioned wireless connection methods may include, but are not limited to, 3G / 4G connections, WiFi connections, Bluetooth connections, WiMAX connections, Zigbee connections, UWB (ultra-wideband) connections, and other currently known or future wireless connection methods.

[0066] It should be emphasized that, in order to further ensure the privacy and security of the aforementioned multi-source data, the aforementioned multi-source data can also be stored in a node of a blockchain.

[0067] The blockchain referred to in this application is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Essentially, a blockchain is a decentralized database, a chain of data blocks linked together using cryptographic methods. Each data block contains information about a batch of network transactions, used to verify the validity of the information (anti-counterfeiting) and generate the next block. A blockchain can include an underlying blockchain platform, a platform product service layer, and an application service layer.

[0068] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0069] Foundational artificial intelligence technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, risk assessment, natural language processing, and machine learning / deep learning.

[0070] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing related hardware through computer-readable instructions. These computer-readable instructions can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, optical disk, or read-only memory (ROM), or random access memory (RAM).

[0071] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0072] Further reference Figure 4 As a response to the above Figure 2 To implement the method shown, this application provides an embodiment of a risk assessment device, which is similar to... Figure 2 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0073] like Figure 4 As shown, the risk assessment device 400 described in this embodiment includes: The data processing module 401 is used to obtain multi-source data related to customers from a preset data source and preprocess the multi-source data to obtain a structured dataset. Feature construction module 402 is used to construct a feature system set from multiple dimensions for structured datasets by combining automated feature engineering with expert experience; Model fusion module 403 is used to train a risk assessment model based on a feature system set and using a fusion architecture that combines expert models and machine learning models. The dynamic training module 404 is used to introduce a dynamic weight adjustment mechanism based on user behavior and time decay to dynamically adjust the feature weights in the feature system set during the model training process. The weighted fusion module 405 is used to weight and fuse the output of the expert model with the output of the machine learning model to obtain a risk assessment score. Risk assessment module 406 is used to classify risk levels based on risk assessment scores and preset ranges, and to determine the risk assessment level.

[0074] Further, please refer to Figure 5 Model fusion module 403 specifically includes: The expert reasoning submodule 501 is used to construct an expert rule base based on the feature system set, generate an expert model using the expert rule base, and perform rule reasoning on the input features based on the expert model to obtain the first risk assessment result; The machine learning submodule 502 is used to input the feature system set into a preset machine learning model and perform supervised learning training on the machine learning model. The machine learning model includes at least one or more of the following: logistic regression model, decision tree model, ensemble learning model or deep learning model, in order to obtain a second risk assessment result. The parameter optimization submodule 503 is used to perform independent parameter optimization and performance evaluation on the expert model and the machine learning model respectively, and to determine the initial weight coefficients of each model based on historical sample data. Model fusion submodule 504 is used to construct a fusion layer based on the initial weight coefficients to form a fusion risk assessment model; The result fusion submodule 505 is used to fuse the first risk assessment result and the second risk assessment result based on the initial weight coefficient to obtain the risk fusion assessment result; The joint training submodule 506 is used to jointly train the fusion risk assessment model based on the feature system set and the risk fusion assessment results. It iteratively updates the parameters of the fusion model by minimizing the preset loss function to complete the training of the risk assessment model.

[0075] Furthermore, the expert reasoning submodule 501 specifically includes: The feature filtering unit is used to filter key features in the feature system set based on historical risk case data and domain expert experience, and to define corresponding risk judgment rules and threshold conditions to form an initial expert rule set. The structured processing unit is used to standardize and structure the initial expert rule set, converting the risk assessment rules into an executable rule expression and storing them in the expert rule base; The rule reasoning unit is used to build an expert model based on the expert rule base. After receiving the input features of the object to be evaluated, it performs rule reasoning on the input features according to the preset rule matching and reasoning strategy to generate intermediate risk judgment results corresponding to each rule. The consistency verification unit is used to resolve conflicts and verify the consistency of intermediate risk assessment results, and to summarize and calculate according to preset rule weights or priorities, and output the first risk assessment result.

[0076] Furthermore, the results fusion submodule 505 specifically includes: The weight configuration unit is used to assign corresponding model weights to the first risk assessment result and the second risk assessment result according to the initial weight coefficients, and to use the model weights as input parameters of the fusion layer. The weighted combination unit is used in the fusion layer to weight and combine the first risk assessment result and the second risk assessment result based on the model weights to generate an intermediate fusion risk assessment result. The smoothing and calibration unit is used to smooth or calibrate the intermediate results of the fusion risk assessment, eliminate model output bias, and output the risk fusion assessment result of the fusion risk assessment model.

[0077] Furthermore, the dynamic training module 404 specifically includes: The behavior association submodule is used to collect user behavior data within a preset time window and associate the user behavior data with the corresponding features in the feature system set to form a user behavior impact feature set. The Influence Factor submodule is used to calculate the behavior influence factor corresponding to each feature based on the user behavior influence feature set. The behavior influence factor is used to characterize the degree of influence of changes in user behavior on the importance of features. The decay calculation submodule is used to calculate the decay of the historical weights of each feature in the feature system set based on a preset time decay function, so as to obtain the time decay weights corresponding to each feature. The weight combination submodule is used to combine behavioral influence factors with time decay weights to calculate and dynamically update the feature weights of each feature in the feature system set.

[0078] Furthermore, the impact factor sub-module specifically includes: The classification processing unit is used to classify user behavior features in the user behavior impact feature set and to quantify user behavior according to behavior type, frequency of occurrence and behavior intensity. The activity index unit is used to calculate the activity index of each feature within a preset time window based on the quantified user behavior characteristics. The activity index is used to reflect the degree of change in user behavior. The correlation analysis unit is used to perform correlation analysis between behavioral activity indicators and the historical risk contribution of corresponding user behavioral characteristics, and to calculate the behavioral correlation coefficient corresponding to each user behavioral characteristic. The weighted calculation unit is used to perform weighted calculations on behavioral activity indicators based on behavioral correlation coefficients to obtain the behavioral influence factors corresponding to each feature.

[0079] Furthermore, the correlation analysis unit specifically includes: The risk contribution subunit is used to obtain the risk contribution data of each user's behavioral characteristics in the historical risk assessment sample, and to standardize the risk contribution data. The time alignment subunit is used to align the standardized risk contribution data with the corresponding user behavior activity index over time to construct risk behavior related data pairs. The correlation calculation subunit is used to calculate the degree of correlation between user behavior characteristics and risk outcomes based on risk behavior correlation data pairs, using correlation analysis or regression analysis methods, and obtain the initial behavior correlation coefficient. The smoothing correction subunit is used to smooth or correct the confidence level of the initial behavior correlation coefficients to obtain the behavior correlation coefficients corresponding to each user behavior feature.

[0080] To address the aforementioned technical problems, embodiments of this application also provide a computer device. Please refer to [link / reference needed]. Figure 6 , Figure 6 This is a basic structural block diagram of the computer device in this embodiment.

[0081] The computer device 6 includes a memory 61, a processor 62, and a network interface 63 that are interconnected via a system bus. It should be noted that only the computer device 6 with memory 61, processor 62, and network interface 63 is shown in the figure; however, it should be understood that it is not required to implement all the components shown, and more or fewer components can be implemented alternatively. Those skilled in the art will understand that the computer device described here is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0082] The computer device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device can interact with the user via a keyboard, mouse, remote control, touchpad, or voice control.

[0083] The memory 61 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 61 may be an internal storage unit of the computer device 6, such as the hard disk or memory of the computer device 6. In other embodiments, the memory 61 may also be an external storage device of the computer device 6, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 6. Of course, the memory 61 may also include both the internal storage unit and its external storage device of the computer device 6. In this embodiment, the memory 61 is typically used to store the operating system and various application software installed on the computer device 6, such as computer-readable instructions for risk assessment methods. In addition, the memory 61 can also be used to temporarily store various types of data that have been output or will be output.

[0084] In some embodiments, the processor 62 may be a central processing unit (CPU), controller, microcontroller, microprocessor, or other data processing chip. The processor 62 is typically used to control the overall operation of the computer device 6. In this embodiment, the processor 62 is used to execute computer-readable instructions stored in the memory 61 or to process data, such as executing computer-readable instructions for the risk assessment method.

[0085] The network interface 63 may include a wireless network interface or a wired network interface, which is typically used to establish communication connections between the computer device 6 and other electronic devices.

[0086] This application also provides an embodiment, namely, a computer device including a memory and a processor, wherein the memory stores computer-readable instructions, and the processor executes the computer-readable instructions to implement the steps of the risk assessment method described above.

[0087] This application also provides another embodiment, namely, providing a computer-readable storage medium storing computer-readable instructions that can be executed by at least one processor to cause the at least one processor to perform the steps of the risk assessment method described above.

[0088] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0089] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0090] It should be noted that the software tools or components not belonging to this company that appear in the various embodiments of this application are merely illustrative examples and do not represent actual use.

[0091] Obviously, the embodiments described above are only some embodiments of this application, not all embodiments. The accompanying drawings show preferred embodiments of this application, but do not limit the patent scope of this application. This application can be implemented in many different forms; rather, the purpose of providing these embodiments is to provide a more thorough and comprehensive understanding of the disclosure of this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this application's specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this application.

Claims

1. A risk assessment method, characterized in that, include: Obtain multi-source data related to customers from a preset data source, and preprocess the multi-source data to obtain a structured dataset; For the structured dataset, a feature system set is constructed from multiple dimensions by combining automated feature engineering with expert experience; Based on the aforementioned feature set, a fusion architecture combining expert models and machine learning models is used to train the risk assessment model; During model training, a dynamic weight adjustment mechanism based on user behavior and time decay is introduced to dynamically adjust the feature weights in the feature system set. The output of the expert model and the output of the machine learning model are weighted and fused to obtain a risk assessment score; The risk assessment level is determined by classifying risk levels based on the risk assessment score and preset range.

2. The risk assessment method as described in claim 1, characterized in that, The step of training the risk assessment model based on the aforementioned feature set using a fusion architecture combining expert models and machine learning models specifically includes: Based on the aforementioned feature system set, an expert rule base is constructed, and an expert model is generated using the expert rule base. Based on the expert model, rule reasoning is performed on the input features to obtain the first risk assessment result. The feature set is input into a preset machine learning model, and the machine learning model is trained by supervised learning. The machine learning model includes at least one or more of the following: logistic regression model, decision tree model, ensemble learning model or deep learning model, in order to obtain a second risk assessment result. Independent parameter optimization and performance evaluation are performed on the expert model and the machine learning model respectively, and the initial weight coefficients of each model are determined based on historical sample data; A fusion layer is constructed based on the initial weighting coefficients to form a fusion risk assessment model; The first risk assessment result and the second risk assessment result are fused based on the initial weighting coefficients to obtain the risk fusion assessment result; The fusion risk assessment model is jointly trained based on the feature set and the risk fusion assessment results. The parameters of the fusion model are iteratively updated by minimizing the preset loss function to complete the training of the risk assessment model.

3. The risk assessment method as described in claim 2, characterized in that, The steps of constructing an expert rule base based on the feature system set, generating an expert model using the expert rule base, and performing rule reasoning on the input features based on the expert model to obtain the first risk assessment result specifically include: Based on historical risk case data and domain expert experience, key features in the feature system set are screened, and corresponding risk judgment rules and threshold conditions are defined to form an initial expert rule set. The initial expert rule set is standardized and structured to convert the risk assessment rules into an executable rule expression and store them in the expert rule base. Based on the expert rule base, an expert model is constructed. After receiving the input features of the object to be evaluated, the model performs rule reasoning on the input features according to the preset rule matching and reasoning strategy to generate intermediate risk judgment results corresponding to each rule. The intermediate risk assessment results are subjected to conflict resolution and consistency verification, and are summarized and calculated according to preset rule weights or priorities to output the first risk assessment result.

4. The risk assessment method as described in claim 2, characterized in that, The step of fusing the first risk assessment result and the second risk assessment result based on the initial weighting coefficient to obtain the risk fusion assessment result specifically includes: Based on the initial weight coefficients, corresponding model weights are assigned to the first risk assessment result and the second risk assessment result, and the model weights are used as input parameters of the fusion layer. In the fusion layer, the first risk assessment result and the second risk assessment result are weighted and combined based on the model weights to generate an intermediate fusion risk assessment result; The intermediate results of the fusion risk assessment are smoothed or calibrated to eliminate model output bias and output the risk fusion assessment result of the fusion risk assessment model.

5. The risk assessment method as described in claim 1, characterized in that, The step of dynamically adjusting the feature weights in the feature system set by introducing a dynamic weight adjustment mechanism based on user behavior and time decay during model training specifically includes: Collect user behavior data within a preset time window, and associate the user behavior data with the corresponding features in the feature system set to form a user behavior influence feature set; Based on the user behavior impact feature set, calculate the behavior impact factor corresponding to each feature, wherein the behavior impact factor is used to characterize the degree of influence of user behavior changes on the importance of the feature; Based on a preset time decay function, the historical weights of each feature in the feature system set are decayed to obtain the time decay weights corresponding to each feature. The behavioral influence factor and the time decay weight are combined to calculate and dynamically update the feature weights of each feature in the feature system set.

6. The risk assessment method as described in claim 5, characterized in that, The step of calculating the behavior influence factor corresponding to each feature based on the user behavior influence feature set specifically includes: The user behavior features in the user behavior impact feature set are classified and processed, and the user behavior is quantitatively represented according to behavior type, frequency of occurrence and behavior intensity. Based on the quantified user behavior characteristics, the activity index of each characteristic within a preset time window is calculated, wherein the activity index is used to reflect the degree of change in user behavior. The correlation analysis between the behavioral activity index and the historical risk contribution of the corresponding user behavioral characteristics is performed to calculate the behavioral correlation coefficient corresponding to each user behavioral characteristic. Based on the behavioral correlation coefficient, the behavioral activity index is weighted and calculated to obtain the behavioral influence factor corresponding to each feature.

7. The risk assessment method as described in claim 6, characterized in that, The step of performing correlation analysis between the behavioral activity index and the historical risk contribution of the corresponding user behavioral characteristics, and calculating the behavioral correlation coefficient corresponding to each user behavioral characteristic, specifically includes: Obtain the risk contribution data of each user behavior feature in the historical risk assessment sample, and standardize the risk contribution data. The standardized risk contribution data is time-aligned with the corresponding user behavior activity index to construct risk behavior association data pairs; Based on the risk behavior-related data pairs, correlation analysis or regression analysis methods are used to calculate the degree of correlation between the user behavior characteristics and the risk results, and to obtain the initial behavior correlation coefficient. The initial behavior correlation coefficients are smoothed or confidence-corrected to obtain the behavior correlation coefficients corresponding to each user behavior feature.

8. A risk assessment device, characterized in that, include: The data processing module is used to obtain multi-source data related to customers from a preset data source, and to preprocess the multi-source data to obtain a structured dataset; The feature construction module is used to construct a feature system set from multiple dimensions for the structured dataset by combining automated feature engineering with expert experience. The model fusion module is used to train a risk assessment model based on the aforementioned feature system set, using a fusion architecture that combines expert models and machine learning models. The dynamic training module is used to dynamically adjust the feature weights in the feature system set by introducing a dynamic weight adjustment mechanism based on user behavior and time decay during the model training process. The weighted fusion module is used to weight and fuse the output of the expert model with the output of the machine learning model to obtain a risk assessment score. The risk assessment module is used to classify risk levels based on the risk assessment score and preset range, and to determine the risk assessment level.

9. A computer device, characterized in that, The system includes a memory and a processor, wherein the memory stores computer-readable instructions, and the processor executes the computer-readable instructions to implement the steps of the risk assessment method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the risk assessment method as described in any one of claims 1 to 7.

Images