An APP product model management method and system based on big data analysis

Abstract

This application relates to the field of financial technology, and in particular to a method and system for managing APP product models based on big data analytics. It includes: acquiring product information and business indicator information for multiple loan products; establishing a relational structure including nodes, impact paths, impact degree parameters, and time lag parameters; receiving product parameter adjustment information, determining the starting node and calculating the direct and indirect impacts level by level to obtain the impact path, impact result, and expected manifestation time, and generating a comprehensive impact view. This solves the technical problem in existing technologies where the lack of perception and management of complex relationships between products leads to the inability to assess the overall impact of automated parameter adjustments and the difficulty in tracing the impact path.

Description

Technical Field

[0001] This application relates to the field of financial technology, and in particular to a method and system for managing APP product models based on big data analysis. Background Technology

[0002] In the financial services sector, particularly in applications offering multiple loan products, product parameter management faces significant challenges. While existing intelligent management systems can automate parameter adjustments for individual loan products based on big data to optimize local performance, they overlook the widespread implicit or explicit business relationships and customer overlap between loan products. When the system adjusts the interest rate or entry threshold for a product, it may trigger customers to migrate to other related products, passively worsening the risk profile of unadjusted products or causing risk accumulation and return imbalance in the overall product portfolio. Due to the complex interaction mechanisms between products and the lag in the impact of automated adjustments, when overall business indicators become abnormal, managers struggle to quickly trace which adjustment sequence triggered the chain reaction through which path, let alone assess the comprehensive impact of adjustments on the overall product portfolio in advance. Existing technologies lack digital modeling of the inter-product relationship structure and a computable mechanism for the propagation of impact, making it impossible to achieve global optimization and transparent decision-making.

[0003] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0004] This application discloses an APP product model management method and system based on big data analysis, which aims to solve the technical problems in the prior art where the lack of perception and management of the complex relationships between products leads to the inability to assess the overall impact of automated parameter adjustments and the difficulty in tracing the impact path.

[0005] The technical solution of this application is as follows: Firstly, this application discloses an APP product model management method based on big data analysis, applied to the product model management of multiple loan products, including: Obtain product information for multiple loan products and business indicator information corresponding to the multiple loan products. Product information includes product identifiers and corresponding product parameters, and business indicator information includes business indicators used to evaluate a single loan product and / or the overall product portfolio. Based on product information and business indicator information, establish a relationship structure between products and business indicators. The relationship structure includes nodes representing loan products, product parameters and business indicators, as well as connection relationships representing the influence paths between nodes. The connection relationships have influence degree parameters and time lag parameters. Receive product parameter adjustment information, which includes the product identifier of the target loan product, the product parameters that have been adjusted, the parameter values ​​before adjustment, the parameter values ​​after adjustment, and the adjustment time. The target loan product is the loan product whose product parameters have been adjusted, as indicated by the product parameter adjustment information, among multiple loan products. Based on the product parameter adjustment information, the starting node in the associated structure is determined. Based on the influence degree parameter and time lag parameter of the connection relationship, the direct and indirect effects caused by the product parameter adjustment are calculated step by step from the starting node along the influence path in the associated structure. This yields other affected loan products and business indicators, the corresponding influence path, the influence results, and the expected manifestation time of the influence results. Presents the impact path, impact results, and expected manifestation time, and generates a comprehensive impact view of product parameter adjustments on the overall product portfolio based on other affected loan products and business metrics.

[0006] Secondly, this application also discloses an APP product model management system based on big data analysis, applied to the product model management of multiple loan products, including: The information acquisition module is used to acquire product information of multiple loan products and business indicator information corresponding to multiple loan products. The product information includes product identifiers and corresponding product parameters, and the business indicator information includes business indicators used to evaluate a single loan product and / or the overall product portfolio. The structure building module is used to establish the association structure between products and business indicators based on product information and business indicator information. The association structure includes nodes representing loan products, product parameters and business indicators, as well as the connection relationship representing the influence path between nodes. The connection relationship has an influence degree parameter and a time lag parameter. The receiving module is used to receive product parameter adjustment information. The product parameter adjustment information includes the product identifier of the target loan product, the product parameters that have been adjusted, the parameter values ​​before adjustment, the parameter values ​​after adjustment, and the adjustment time. The target loan product is the loan product whose product parameters have been adjusted, as indicated by the product parameter adjustment information, among multiple loan products. The impact calculation module is used to determine the starting node in the association structure based on the product parameter adjustment information, and based on the impact degree parameter and time lag parameter of the connection relationship, it calculates the direct and indirect impacts caused by the product parameter adjustment from the starting node along the impact path in the association structure, and obtains other affected loan products and business indicators, the corresponding impact path, the impact result, and the expected manifestation time of the impact result. The view-providing module presents the impact path, impact results, and expected display time, and generates a comprehensive impact view of product parameter adjustments on the overall product portfolio based on other affected loan products and business metrics.

[0007] Beneficial Effects: This application establishes a correlation structure including impact degree parameters and time lag parameters, enabling digital modeling of complex business relationships among multiple loan products. This allows the impact of product parameter adjustments to be quantitatively propagated and calculated step-by-step through the impact path, thus predicting the impact on other related loan products and the timing of their manifestation before the product parameter adjustments are implemented. This propagation calculation mechanism based on the correlation structure overcomes the limitations of isolated evaluation of individual products in existing technologies, solves the technical problems of difficulty in assessing the overall impact and tracing the impact path of automated parameter adjustments, effectively avoids overall suboptimal results or risk accumulation caused by local optimization, and achieves global optimization and decision-making transparency in financial product model management. Attached Figure Description

[0008] Figure 1 This application provides a flowchart illustrating an APP product model management method based on big data analysis.

[0009] Figure 2 A flowchart of an APP product model management system based on big data analysis is provided for this application.

[0010] In the diagram: 1. Information acquisition module; 2. Structure establishment module; 3. Adjustment and reception module; 4. Impact calculation module; 5. View provision module. Detailed Implementation

[0011] The technical solution of this application will be described below with reference to the accompanying drawings. The described embodiments are only some embodiments of this application and are not intended to limit the scope of protection of this application; other embodiments obtained by those skilled in the art without creative effort are all within the scope of protection of this application.

[0012] In the description of this application, similar reference numerals indicate similar items, and terms such as "first" and "second" are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0013] This application can be applied to the loan product management system of fintech platforms. Different loan products are related in terms of customer base, risk exposure, and funding costs. For example, when a platform offers both a 5% annual interest rate "Elite Loan" and a 7% annual interest rate "Business Working Capital Loan," some high-income entrepreneurs may meet the application requirements for both products. Lowering the credit score threshold for the Elite Loan from 700 to 650 might cause high-quality customers who would normally prefer the Business Working Capital Loan to switch to the Elite Loan, passively deteriorating the customer structure of the Business Working Capital Loan and increasing the average risk level.

[0014] In an automated parameter adjustment environment, the impact of different adjustments is often delayed. For example, the impact of credit limit adjustments on the final delinquency rate may take three to six months to materialize. When the platform's overall non-performing loan ratio rises abnormally, managers find it difficult to trace the specific adjustment path and assess its systemic impact on the overall product portfolio's risk-return characteristics before the adjustment is implemented.

[0015] To address the aforementioned issues, this application introduces a correlation structure. This structure describes the impact relationships between loan products, product parameters, and business indicators, including nodes representing these elements and connections representing the paths of influence between them. Each connection carries an impact degree parameter and a time lag parameter. The impact degree parameter quantifies the intensity of the impact of changes in upstream nodes on downstream nodes and can be represented by an elasticity coefficient, correlation coefficient, or model weight value. When using a correlation coefficient or normalized weight value, its value can be normalized to between -1 and +1. When using an elasticity coefficient, the actual elasticity coefficient value can be retained. The time lag parameter characterizes the time period required for the impact to manifest from its occurrence.

[0016] Reference Figure 1 The present application discloses a specific implementation method for APP product model management based on big data analysis as follows.

[0017] S1000: Obtain product information for multiple loan products and business indicator information corresponding to the multiple loan products. Product information includes product identifiers and corresponding product parameters. Business indicator information includes business indicators used to evaluate a single loan product and / or the overall product portfolio. S2000: Based on product information and business indicator information, establish the association structure between products and business indicators. The association structure includes nodes representing loan products, product parameters and business indicators, as well as connection relationships representing the influence paths between nodes. The connection relationships have influence degree parameters and time lag parameters. S3000: Receives product parameter adjustment information. The product parameter adjustment information includes the product identifier of the target loan product, the product parameters that have been adjusted, the parameter values ​​before adjustment, the parameter values ​​after adjustment, and the adjustment time. The target loan product is the loan product whose product parameters have been adjusted, as indicated by the product parameter adjustment information, among multiple loan products. S4000: Based on the product parameter adjustment information, determine the starting node in the associated structure, and based on the influence degree parameter and time lag parameter of the connection relationship, calculate the direct and indirect effects caused by the product parameter adjustment along the influence path in the associated structure from the starting node, and obtain other affected loan products and business indicators, the corresponding influence path, the influence result, and the expected manifestation time of the influence result; S5000: Presents the impact path, impact results, and expected manifestation time, and generates a comprehensive view of the impact of product parameter adjustments on the overall product portfolio based on other affected loan products and business metrics.

[0018] In the process of acquiring product information for multiple loan products and the corresponding business indicator information, product identifiers can be in the form of product codes, such as PRD2024001, to uniquely identify a specific loan product. The system can collect product codes, product names, product status, credit scoring thresholds, income-to-debt ratio requirements, occupational restriction lists, regional restriction codes, and historical business indicator data from the business database, risk control rule engine, and data warehouse through preset data extraction tasks. Product parameters cover business rule elements such as interest rates, terms, loan amounts, credit scoring thresholds, income-to-debt ratio requirements, occupational restrictions, and regional restrictions, and can be stored in key-value pairs, such as the interest rate parameter represented as key name "annual_rate" corresponding to key value "0.05". Business indicator information can include quantitative indicators such as application volume, conversion rate, loan amount, delinquency rate, default rate, risk-adjusted return, customer acquisition cost, and customer lifetime value. The above data can be cleaned and standardized, removing records with a missing rate exceeding 20%, unifying the date format, and performing primary key alignment to ensure consistency of product identifiers across different tables.

[0019] In establishing the relationship between products and business metrics, the system first creates product nodes, parameter nodes, and metric nodes. Product nodes have attributes including a unique product identifier (product_id), a product name (product_name), and a product type (product_type). Parameter nodes have attributes including parameter code (param_code), parameter type (param_type), current value (current_value), and value range (range). Metric nodes have attributes including metric code (metric_code), calculation cycle (calculation_cycle), and metric category (metric_category). Subsequently, the system creates directed edges based on preset association rules: if a parameter directly affects the admission decision of a product, a connection is created from the parameter node to the product node; if the performance of a product directly affects a business metric, a connection is created from the product node to the metric node; if two products have overlapping or competing customers, a connection is created connecting the two product nodes. This connection can be bidirectional or a set of opposite directed connections.

[0020] In this relational structure, product nodes store basic product attributes, parameter nodes store specific parameter configurations and their historical change records, and indicator nodes store the calculation logic and real-time values ​​of business indicators. Connections represent the influence paths between nodes; for example, a connection from a parameter node to a product node represents the impact of parameter configuration on product access, and a connection from a product node to an indicator node represents the contribution of product performance to business indicators. The degree of influence parameter can be determined through historical data regression analysis. In specific implementation, the system extracts historical monthly data, uses the interest rate parameter as the independent variable X_rate, and the application volume indicator as the dependent variable Y_apply, to establish a linear regression model Y_apply = α_reg + β_reg × X_rate + ε_reg, where X_rate is the interest rate parameter value, Y_apply is the application volume indicator value, α_reg is the regression intercept term, β_reg is the regression coefficient, and ε_reg is the regression error term; β_reg can be used as the degree of influence parameter to represent the expected change in application volume when the interest rate parameter changes by one unit. The time lag parameter can be determined through the cross-correlation function in time series analysis. In practice, the system calculates the cross-correlation coefficient between the credit score adjustment time series and the delinquency rate change series three months later, and identifies the lag order k_lag when the correlation coefficient reaches its peak. This k_lag value is the time lag parameter, where k_lag represents the number of periods in which the influence is delayed in the time series, which can be converted into the corresponding number of days according to the sampling period.

[0021] In a specific embodiment, for the two products, "Elite Career Loan" and "Business Working Capital Loan," the system's associated structure includes the following paths: The Elite Career Loan product node is connected to the Business Working Capital Loan product node via a customer overlap edge. The influence parameter of this edge is 0.3, indicating that for every percentage point decrease in the Elite Career Loan interest rate, the expected churn rate of high-quality customers for the Business Working Capital Loan is 3%. The time lag parameter is seven days, indicating that the migration of customer decisions and application behavior typically manifests within one week after the adjustment. Simultaneously, the Elite Career Loan interest rate parameter node is connected to its own application volume indicator node via a direct influence edge. The influence parameter is -2.5, and the time lag parameter is three days, indicating that a decrease in interest rates will rapidly stimulate application growth. The influence parameter of -2.5 uses an elasticity coefficient, which is not a normalized weight value and therefore does not conflict with the aforementioned normalization value range.

[0022] During the process of receiving product parameter adjustment information, the product parameter adjustment information includes the product identifier of the target loan product, the product parameter that has been adjusted, the parameter value before adjustment, the parameter value after adjustment, and the adjustment time. The target loan product is the loan product whose product parameter has been adjusted, as indicated by the product parameter adjustment information, among multiple loan products. This information can be entered through the front-end interface of the parameter management system or directly pushed by the automated decision engine. For example, the adjustment information received by the system indicates: the target loan product is the "Workplace Elite Loan" with identifier PRD2024001, the adjusted parameter is the annualized interest rate, the parameter value before adjustment is 5%, the parameter value after adjustment is 4.5%, and the adjustment time is 14:30 on March 15, 2024. Product parameter adjustment information can be transmitted via messages, and the message structure includes the product identifier of the target loan product, the product parameter that has been adjusted, the parameter value before adjustment, the parameter value after adjustment, and the adjustment time. Among them, the parameter value before adjustment and the parameter value after adjustment are used to calculate the change in the product parameter that has been adjusted.

[0023] In determining the starting node in the association structure based on product parameter adjustment information, and then propagating the influence step-by-step along the influence path from the starting node, the system locates the corresponding parameter node or product node in the association structure based on the product identifier of the target loan product and the adjusted product parameters, and designates this as the starting node. The system determines the initial change of the starting node based on the difference between the adjusted parameter value and the original parameter value; for example, when the annualized interest rate is adjusted from 0.05 to 0.045, the initial change is -0.005, which is a decrease of 0.5 percentage points. The propagation calculation uses a network propagation algorithm, starting from the starting node and passing the influence to adjacent nodes along the outgoing edges. For each hop of propagation, the influence is equal to the change of the upstream node multiplied by the influence degree parameter of the connection relationship. The specific calculation process is as follows: Let the change in the initial parameter node be ΔP_start. After propagating through the first edge to node A, the influence parameter of the edge is w_path,1. Then, the change in node A, ΔA_node = ΔP_start × w_path,1. If node A is connected to node B through edge e2, and the influence parameter of edge e2 is w_path,2, then the change in node B, ΔB_node = ΔA_node × w_path,2 = ΔP_start × w_path,1 × w_path,2. For time calculation, the time lag parameters of each path are accumulated, i.e., the total path lag time L_path = τ_1 + τ_2 + ... + τ_npath, where L_path is the total path lag time, τ_i is the time lag parameter of the i-th edge, and npath is the number of connections in the path. When the propagation path encounters a branch node, the influence of each branch path is calculated separately. When multiple propagation paths converge to the same node, the influence of each path can be superimposed or synthesized according to a preset synthesis rule.

[0024] For example, in the aforementioned interest rate adjustment scenario, the system first calculates the direct impact on the number of applications for the "Elite Worker Loan": a 0.5 percentage point decrease in the interest rate multiplied by a -2.5 elasticity coefficient yields an estimated 1.25 percentage point increase in application volume, with the expected effect appearing three days after the adjustment. Subsequently, the system propagates along the customer overlap edge to the "Corporate Working Capital Loan," calculating the indirect impact: the increase in "Elite Worker Loan" applications leads to a change in the quality of its customer base, affecting the customer structure of the "Corporate Working Capital Loan" through the customer overlap edge, and the expected delinquency rate of the "Corporate Working Capital Loan" will increase by 0.05 percentage points after seven days. Here, the estimated 1.25 percentage point increase in application volume is the predicted change calculated based on the elasticity coefficient and the interest rate change; the 0.05 percentage point increase in the delinquency rate of the "Corporate Working Capital Loan" is the indirect impact obtained after further propagation along the customer overlap edge.

[0025] Through hierarchical propagation calculations, the affected loan products and business indicators, corresponding impact paths, impact results, and the expected manifestation time of the impact results are obtained. The impact path is recorded in the form of a node sequence, such as from the interest rate parameter node to the workplace elite loan product node and then to the enterprise working capital loan product node. The impact results include the predicted changes in each affected node, such as the percentage change in application volume and the basis point change in delinquency rate. The expected manifestation time is determined by accumulating the time lag parameters of each connection relationship on the path, and can be converted into a specific expected manifestation date or time interval by combining the adjustment time.

[0026] In presenting the impact path, impact results, and expected manifestation time, and generating a comprehensive impact view of the overall product portfolio caused by product parameter adjustments, the comprehensive impact view can use a Sankey diagram to show the flow and intensity of the impact path, a timeline chart to show the pace of impact manifestation at different time points, and a dashboard to show the predicted changes in overall risk-return indicators. Managers can view the complete impact chain from micro-parameter adjustments to macro-portfolio indicators through an interactive interface, thereby assessing the systemic consequences before implementing adjustments.

[0027] Furthermore, before performing the stepwise propagation calculation, the impact parameters and time lag parameters related to the micro-customer groups can be updated through a pre-calibration process, and the stepwise propagation calculation can be performed based on the calibration results.

[0028] S4100: Adjust information and related structures based on product parameters to determine candidate customer groups; S4200: Acquire behavioral monitoring data of candidate customer groups after the adjustment period; S4300: Performs difference monitoring on behavioral monitoring data to identify valid abnormal differences; S4400: When valid outliers exist, identify the micro-customer groups corresponding to the valid outliers; S4500: Generates micro-experiment plans based on micro-customer groups; S4600: Collects experimental behavior data of the experimental group and the control group during the micro-experiment; S4700: Based on the differences in experimental behavior data between the experimental group and the control group, determine the micro-level influence parameters and micro-level time lag parameters corresponding to the micro-customer groups; S4800: Update the influence degree parameter and time lag parameter of the connection relationship related to the micro customer group in the association structure according to the micro influence degree parameter and the micro time lag parameter; S4900: Based on the influence degree parameter and time lag parameter of the connection relationship, the direct and indirect effects caused by the adjustment of product parameters are calculated step by step from the starting node along the influence path in the associated structure. This includes: when the influence path involves micro-customer groups, the updated influence degree parameter and time lag parameter are used for step-by-step propagation calculation; when the influence path does not involve micro-customer groups, the unupdated influence degree parameter and time lag parameter are used for step-by-step propagation calculation.

[0029] In determining the candidate customer group, the candidate customer group refers to the customer group associated with the target loan product and / or related loan products. Related loan products are those that have an influence path with the target loan product in the association structure. In specific implementation, the system can extract the historical borrower records of the target loan product and match them with the historical customer data of related loan products using the customer's unique identifier or the association identifier after de-identification, to identify the customer set that simultaneously holds or has applied for multiple products. For potential related customers who do not cross-hold but have similar risk characteristics or behavioral patterns, the system can use similarity calculation to select other product customer groups that have more than 80% similarity to the target product customer group in terms of age distribution, credit score distribution, and income level distribution as the potential related customer set. The union of the directly related customer set and the potential related customer set is then taken to form the candidate customer group.

[0030] In acquiring behavioral monitoring data of candidate customer groups after the adjustment period, the behavioral monitoring data includes at least one of the following: browsing behavior data, click behavior data, application behavior data, conversion behavior data, and repayment behavior data. The platform can collect data such as page visits, dwell time, clicked elements, application amount, application period, approval status, rejection reason code, repayment date, repayment status, and overdue days through preset data collection interfaces. This data can be aggregated, buffered, and stored in real time for querying to support the monitoring of behavioral changes in candidate customer groups after the adjustment period.

[0031] In identifying valid anomalies, a valid anomaly is defined as the deviation of the actual behavior of a candidate customer group from the predicted behavior baseline, provided that the deviation meets preset triggering conditions. The predicted behavior baseline can be generated by a preset prediction model. Training data includes time-series data such as historical application volume, conversion rate, and pageviews, as well as external macroeconomic features such as market interest rates, stock market indices, and holiday markers. The model can be a time-series prediction model or a neural network model, and the predicted behavior baseline model is obtained when the prediction error meets preset requirements. The prediction error can be evaluated using the root mean square error (RMSE) and the coefficient of determination (R²_score). RMSE represents the average error level between the predicted and actual values, while R²_score represents the model's explanatory power for changes in actual behavior. During actual monitoring, the system compares the actual conversion rate of the candidate customer group within the monitoring window with the conversion rate predicted by the baseline model, calculating the behavior deviation value. When the actual conversion rate drops by more than two percentage points from the baseline and persists for more than three days, it is initially marked as a preliminary anomaly, and further confirmed as a valid anomaly after meeting preset triggering conditions.

[0032] In determining the micro-customer groups corresponding to valid anomalies, the micro-customer group is a subset of customers from the candidate customer group who correspond to valid anomalies. The system can employ segmentation or clustering algorithms to identify the sub-customer groups that contribute significantly to the anomalies. In specific implementation, the system can use the anomaly index as the target variable and customer age, gender, credit score, income level, city tier, and device type as feature variables to identify customer characteristic combinations with conversion rates significantly below the average level. For example, customers aged 25 to 30, residing in a first-tier city, and with a credit score between 680 and 720. Customers meeting this characteristic combination are then selected from the candidate customer group and identified as the micro-customer group.

[0033] In generating micro-experiment plans based on micro-customer groups, the micro-experiment plan includes dividing the micro-customer group into an experimental group and a control group, and fine-tuning the parameters of the experimental group. The adjustment range of these parameters is less than a preset adjustment range threshold. In specific implementation, the system can use stratified random sampling combined with propensity score matching to stratify the micro-customer group based on key covariates such as age, income, and credit score. Within each stratum, customers are allocated to the experimental and control groups according to a preset proportion, ensuring that the distribution difference between the experimental and control groups on preset key customer characteristics is less than a preset distribution difference threshold. For example, the standardized difference between the two groups in age distribution and mean credit score should not exceed 0.1 standard deviation. The adjustment range is typically set to 10% to 20% of the original planned adjustment range. For example, if the original planned interest rate reduction is 0.5 percentage points, the micro-experiment will only reduce it by 0.05 to 0.1 percentage points to control potential risks.

[0034] During the collection of experimental behavioral data from the experimental and control groups, the experimental period is typically set at one to two weeks. During this period, browsing, clicking, application, conversion, and early repayment behaviors of both groups of customers are continuously tracked. The data collection method for experimental behavioral data can be consistent with that for behavioral monitoring data to ensure consistency in data definitions.

[0035] In determining the micro-level impact parameter and the micro-level time lag parameter, the micro-level impact parameter represents the strength of the influence of parameter fine-tuning on the behavioral data of the micro-customer group, while the micro-level time lag parameter represents the time required for the influence of parameter fine-tuning on the behavioral data of the micro-customer group to reach the preset manifestation conditions. The micro-level impact parameter can be calculated using the difference-in-differences method, with the formula I_micro=[(E_after-E_before)-(C_after-C_before)] / Δθ_micro, where I_micro is the micro-level impact parameter, E_after is the post-test mean of the experimental group, E_before is the pre-test mean of the experimental group, C_after is the post-test mean of the control group, C_before is the pre-test mean of the control group, and Δθ_micro is the non-zero adjustment magnitude of the parameter fine-tuning. The micro-level time lag parameter can be determined by the time series of behavioral changes in the experimental group, that is, identifying the number of days from the start of the micro-experiment to the time required for the rate of change of the behavioral indicator to reach its peak or stable value. For example, if the conversion rate of the experimental group reaches its highest improvement and remains stable on the third day after the start of the micro-experiment, then the micro time lag parameter is determined to be three days.

[0036] During the process of updating the influence parameters and time lag parameters of the connections related to micro-customer groups in the association structure, the system updates the connection parameters related to specific customer groups from general values ​​based on historical full data to segmented customer group-specific values ​​reflecting the current market state, based on the results of micro-experiments. If the difference between the micro-influence parameter and the original influence parameter exceeds 20%, the original influence parameter is replaced by the micro-influence parameter; if the difference is within 20%, a weighted average method is used for updating. The updated influence parameter w_path,new = 0.7 × w_path,old + 0.3 × I_micro, where w_path,new is the updated influence parameter, w_path,old is the original influence parameter, and I_micro is the micro-influence parameter. For time lag parameters, if the difference between the micro time lag parameter and the original time lag parameter exceeds a preset lag difference threshold, the original time lag parameter is replaced by the micro time lag parameter; if the difference does not exceed the preset lag difference threshold, a weighted average method is used to update the time lag parameter. The updated time lag parameter τ_new = 0.7 × τ_old + 0.3 × τ_micro, where τ_new is the updated time lag parameter, τ_old is the original time lag parameter, and τ_micro is the micro time lag parameter.

[0037] In the process of calculating the propagation of influence based on the connection relationship parameters and time lag parameters, when the influence path involves micro-customer groups, the updated influence parameters and time lag parameters are used for the propagation calculation; when the influence path does not involve micro-customer groups, the original influence parameters and time lag parameters are used. Whether a path involves micro-customer groups can be determined by checking whether the product nodes or parameter nodes on the propagation path match the characteristic tags of the micro-customer groups. For example, if the micro-customer group tag is "young white-collar workers," then the product node paths involving this customer group will use the updated parameters.

[0038] Furthermore, differential monitoring is performed on the behavioral monitoring data to identify valid anomalous differences, including: S4310: Determine the monitoring time window based on the target loan product, related loan products, candidate customer groups, and adjustment time; S4320: Based on historical business indicator information related to the target loan product and / or related loan products in the association structure, as well as current market environment information, generate a baseline for the predicted behavior of the candidate customer group within the monitoring time window; S4330: Calculate the behavioral deviation of the candidate customer group within the monitoring time window based on behavioral monitoring data and predicted behavioral baseline; S4340: Acquire information about external events related to the monitoring time window; S4350: Based on the historical behavioral data of the candidate customer group, determine the periodic patterns and trend patterns of the candidate customer group; S4360: Determine whether the behavioral deviation meets the preset event matching conditions with external event information; S4370: Determine whether the behavioral deviation conforms to a periodic or trend pattern; S4380: When the behavioral deviation and external event information meet the preset event matching conditions, the behavioral deviation does not conform to the periodic pattern and trend pattern, and the behavioral deviation exceeds the preset difference threshold within a preset number of consecutive monitoring time windows, the behavioral deviation will be identified as an effective abnormal difference driven by the external event. S4390: When the behavioral deviation does not meet the preset event matching conditions with the external event information, and the behavioral deviation does not conform to the periodic pattern and trend pattern, and the behavioral deviation exceeds the preset difference threshold within a continuous preset time window, the behavioral deviation will be identified as a valid abnormal difference related to the product parameter adjustment.

[0039] When determining the monitoring time window, it is necessary to consider the target loan product, related loan products, the behavioral cycles of the candidate customer group, and the adjustment period. For short-term consumer loan products, the window can be set to seven to thirty days after the adjustment; for medium- and long-term business operating loans, the window can be extended to ninety days; for those involving credit score adjustments, the window can be extended to one hundred and eighty days to observe risk exposure. The window starts at t_win,start and ends at t_win,end, and can be determined by querying a preset mapping table based on the parameter type. Here, t_win,start is the effective date of the adjustment, and t_win,end is the end date of the monitoring time window. The preset mapping table is used to record the observation period corresponding to different parameter types.

[0040] In generating the predictive behavior baseline, the baseline represents the expected behavior of the candidate customer group within the monitoring time window, assuming no abnormal disturbances. The system can generate the predictive behavior baseline based on historical business indicators related to the target loan product and / or related loan products in the correlation structure, as well as current market environment information. Model inputs may include historical weekly average application volume, conversion rate, current average market interest rate level, and the intensity of recent promotional activities by similar competitors. The model can employ time series forecasting models, nonlinear forecasting models, or regression models to output daily or weekly predictive behavior indicator values ​​and their confidence intervals within the monitoring time window.

[0041] In calculating behavioral bias, the behavioral bias is the deviation of the actual observed value from the predicted behavioral baseline, which can be expressed as a difference or ratio. For example, absolute behavioral bias can be expressed as D_abs = V_obs - V_base, and relative behavioral bias can be expressed as D_rel = V_obs / V_base - 1, where D_abs is the absolute behavioral bias, D_rel is the relative behavioral bias, V_obs is the actual observed value, and V_base is the predicted value corresponding to the predicted behavioral baseline. V_base is not zero when calculating relative behavioral bias. Specifically, the behavioral bias of the candidate customer group within the monitoring time window can be obtained by subtracting the predicted conversion rate from the actual conversion rate, or by dividing the actual application volume by the predicted application volume and then subtracting one.

[0042] In acquiring external event information, this information includes policy announcements and industry news, as well as data extracted from these sources, such as event occurrence time, affected region, affected customer groups, and affected product types. The system can access information streams from regulatory agency websites and financial news portals via pre-defined data interfaces and perform structured text processing. For example, it can extract time entities (e.g., "March 20, 2024"), location entities (e.g., "Beijing"), and institutional entities; identify event types such as "interest rate adjustment" and "purchase restriction policy announcement"; and structure the extracted results into event timestamps, affected region codes, affected customer group tags, and affected product categories.

[0043] In determining periodic and trend patterns, periodic patterns represent behavioral changes that repeat according to a preset time period, such as an increase in application volume every Friday evening or an increase in quota utilization at the beginning of each month. Trend patterns represent behavioral changes that continuously increase or decrease over multiple consecutive time windows, such as a month-on-month increase in overall application volume as the platform's brand awareness increases. The system can use time series decomposition to process the historical behavioral data of candidate customer groups. The decomposition formula can be expressed as X_beh,t = Trend_t + Season_t + Resid_t, where X_beh,t is the historical behavioral data at time t, Trend_t is the trend term at time t, Season_t is the seasonal term at time t, and Resid_t is the residual term at time t. The seasonal term is used to determine the periodic pattern, and the trend term is used to determine the trend pattern.

[0044] In determining whether behavioral deviations meet preset event matching conditions with external event information, the preset event matching conditions include at least one of the following: the occurrence time of the behavioral deviation matches the occurrence time of the event; the regional information of the candidate customer group matches the region affected by the event; the candidate customer group matches the customer group affected by the event; and the target loan product or related loan product matches the product type affected by the event. Specifically, if the absolute value of the difference between the timestamp of the behavioral deviation and the timestamp of the event is less than 72 hours, it is determined to be a time match; if the customer group's registration location or residence code intersects with the code of the region affected by the event, it is determined to be a regional match; if the customer group tag and the tag of the customer group affected by the event have an inclusion relationship, it is determined to be a customer group match; and if the product category code is consistent with the code of the product type affected by the event, it is determined to be a product match. When at least one matching condition is met, it is determined that the behavioral deviation and external event information meet the preset event matching conditions.

[0045] In determining whether behavioral deviations conform to cyclical or trend patterns, if the timing and magnitude of the behavioral deviation are consistent with historical cyclical fluctuations, or if the direction of the behavioral deviation aligns with the long-term trend, then it is determined to conform to a cyclical or trend pattern. Specifically, if the behavioral deviation value falls within one standard deviation of the cyclical pattern's predicted value, or if the direction of the behavioral deviation is consistent with the slope direction of the trend pattern and its duration exceeds two cycles, then the behavioral deviation is determined to conform to a cyclical or trend pattern.

[0046] In determining valid anomalies driven by external events, behavioral deviations are identified as valid anomalies driven by external events when they meet preset event matching conditions, do not exhibit periodic or trend patterns, and exceed a preset difference threshold within a preset number of consecutive monitoring time windows. These valid anomalies characterize the abnormal impact of external policies, industry news, or regional market events on the behavior of candidate customer groups and can be used as external event factors in impact explanations or risk warnings.

[0047] In determining valid abnormal differences related to product parameter adjustments, when behavioral deviations do not meet preset event matching conditions with external event information, and the behavioral deviations do not exhibit periodic or trend patterns, and the behavioral deviations exceed a preset difference threshold within a consecutive preset time window, the behavioral deviations are identified as valid abnormal differences related to product parameter adjustments. Here, the consecutive preset time window refers to a preset number of consecutive monitoring time windows, used to maintain consistency with the aforementioned judgment criteria. These valid abnormal differences characterize the correlation between product parameter adjustments and changes in the behavior of candidate customer groups, and can be used to subsequently identify micro-customer groups and trigger micro-experiments.

[0048] By identifying external event information and filtering periodic and trend patterns, it is possible to distinguish from behavioral deviations anomalies driven by external events, periodic or trend fluctuations, and anomalies related to product parameter adjustments, thereby reducing the possibility of misjudging external shocks as the impact of internal adjustments.

[0049] Furthermore, the following dynamic evaluation mechanism can be used to determine whether behavioral deviations meet preset trigger conditions: S4301: Determine the degree of market volatility based on current market environment information, determine the success rate of historical micro-experiments based on historical micro-experiment records, and determine potential risk exposure based on business indicator information corresponding to candidate customer groups; S4302: Adjust the initial difference threshold based on the degree of market volatility, the success rate of historical micro-experiments, and potential risk exposure to obtain the adjusted difference threshold; S4303: Determine the representativeness score of the candidate customer group based on the number of customers, behavioral activity, business contribution, and behavioral stability of the candidate customer group; S4304: When behavioral deviation exceeds the adjusted difference threshold within a consecutive preset number of monitoring time windows, and the customer group representativeness score reaches the preset representativeness threshold, the statistical significance verification process is initiated. S4305: In the statistical significance verification process, calculate the overall confidence score based on the duration of the behavioral deviation, the magnitude of the behavioral deviation, the number of customers in the candidate customer group, and the potential risk exposure. S4306: When the overall confidence score reaches the preset confidence threshold, it is determined that the behavioral deviation meets the preset triggering condition.

[0050] In determining the degree of market volatility, the success rate of historical micro-experiments, and potential risk exposure, the degree of market volatility can be quantified through market index volatility, interbank lending rate volatility, or the volatility of the platform's own business indicators. Specifically, the coefficient of variation can be calculated based on the standard deviation and mean of the platform's overall application volume or delinquency rate over the past 30 days, using the formula CV_vol=σ_biz / μ_biz, where CV_vol is the coefficient of variation corresponding to the degree of market volatility, σ_biz is the standard deviation of the platform's overall application volume or delinquency rate over the past 30 days, and μ_biz is the corresponding mean. When CV_vol is greater than 0.15, the degree of market volatility is considered high. The historical micro-experiment success rate refers to the proportion of micro-experiments that successfully identified significant impacts within the past six months; potential risk exposure refers to the proportion of the current outstanding loan balance of the candidate customer group to the total loan balance, or the scale of the delinquent amount of that group.

[0051] In obtaining the adjusted difference threshold, the difference threshold can be increased in a high-volatility market environment to reduce false alarms; when the historical success rate of micro-experiments is high or the potential risk exposure is large, the difference threshold can be decreased to improve sensitivity. The adjustment formula can be expressed as: θ_adj=θ_0×(1+α_M×M_vol-β_H×H_exp-γ_E×E_risk), where θ_adj is the adjusted difference threshold, θ_0 is the initial difference threshold, M_vol is the market volatility coefficient, H_exp is the historical micro-experiment success rate, E_risk is the risk exposure coefficient, and α_M, β_H, and γ_E are the weighting coefficients corresponding to the market volatility coefficient, the historical micro-experiment success rate, and the risk exposure coefficient, respectively.

[0052] In determining the representativeness score of candidate customer groups, the representativeness score can be calculated based on the number of customers, behavioral activity, business contribution, and behavioral stability. The number of customers reflects the sample size; behavioral activity can be measured by the login frequency over the past thirty days; business contribution can be calculated as the proportion of the group's historical loan amount; and behavioral stability can be assessed using the coefficient of variation of historical behavioral data. The representativeness score can be expressed as: S_rep=ω_rep,1×N_std+ω_rep,2×A_std+ω_rep,3×G_std+ω_rep,4×(1-CV_std), where S_rep is the representativeness score, N_std is the standardized number of customers, A_std is the standardized activity level, G_std is the standardized contribution, CV_std is the standardized coefficient of variation, and ω_rep,1 to ω_rep,4 are the representativeness score weights.

[0053] When behavioral deviations exceed the adjusted difference threshold within a preset number of monitoring time windows, and the customer group representativeness score reaches the preset representativeness threshold, the statistical significance verification process is initiated. The preset representativeness threshold is used to reduce the occurrence of micro-experiments triggered by low-representation customer groups.

[0054] The overall confidence score is calculated based on the duration and magnitude of the behavioral deviation, the number of customers in the candidate customer group, and the potential risk exposure. The overall confidence score can be expressed as: S_conf = (D_dev × √N_cust × L_dev × (1 + E_risk)) / SE_conf, where S_conf is the overall confidence score, D_dev is the magnitude of the behavioral deviation, N_cust is the number of customers in the candidate customer group, L_dev is the duration of the behavioral deviation, E_risk is the risk exposure coefficient, and SE_conf is the standard error. A longer duration, greater magnitude, larger number of customers, and higher potential risk exposure result in a higher overall confidence score.

[0055] When the overall confidence score reaches the preset confidence threshold, it is determined that the behavioral deviation meets the preset triggering conditions, and this judgment result can be used as the basis for starting the micro-experiment.

[0056] By dynamically adjusting thresholds and verifying statistical significance, we can combine market environment, historical micro-experiment success rate, potential risk exposure, and customer group representativeness to reduce the occurrence of micro-experiments triggered by random fluctuations or low representative samples.

[0057] Furthermore, based on the micro-customer groups, micro-experiment plans can be generated, and the following implementation methods, which include privacy protection and risk simulation mechanisms, can be adopted.

[0058] S4510: Anonymize the customer data corresponding to the micro-customer group to obtain anonymous customer data, and classify the customer characteristics in the anonymous customer data into privacy levels according to the preset privacy protection rules. Determine the set of customer characteristics that can be used for micro-experiment grouping based on the privacy levels obtained from the classification. S4520: Based on the customer characteristic set, select customers with data analysis authorization identifiers from the micro customer group, and divide the selected customers into experimental group and control group; S4530: Based on the preset parameter boundary rules and risk control rules, determine the compliant adjustment range corresponding to parameter fine-tuning, and determine the adjustment range of parameter fine-tuning within the compliant adjustment range; S4540: Perform a shadow mode preview of parameter fine-tuning before performing parameter fine-tuning; S4550: When the risk assessment results meet the preset risk control conditions, generate a micro-experiment plan that includes the experimental group, the control group, the adjustment range of parameters, and the execution time of the micro-experiment.

[0059] In the process of anonymizing customer data corresponding to micro-customer groups and determining the set of customer features that can be used for micro-experiment grouping, anonymization can employ K-anonymization or differential privacy techniques. Specifically, direct identifiers such as customer names, ID numbers, and mobile phone numbers are encrypted or deleted; quasi-identifiers such as age and income are generalized, for example, converting precise age into age ranges and precise income into income intervals; geolocation information is coarse-grained; and sensitive attributes are suppressed. Privacy levels can be divided into high, medium, and low: high-privacy-level features cannot be used for grouping, medium-privacy-level features can be used for grouping under encrypted transmission or access control, and low-privacy-level features can be used for grouping after permission verification. The system excludes customer features that cannot be used for grouping based on the privacy level and forms a customer feature set of customer features that are allowed for grouping.

[0060] In the process of screening customers with data analytics authorization identifiers and dividing them into experimental and control groups, the data analytics authorization identifier is used to indicate that the customer has consented to the use of their anonymized customer data for data analysis. The system only groups customers with data analytics authorization identifiers and can use propensity score matching methods to ensure that the distribution difference between the experimental and control groups on preset key customer characteristics is less than a preset distribution difference threshold. For example, the standardized difference between the two groups in age distribution and mean credit score does not exceed one standard deviation.

[0061] In determining the compliant adjustment range and adjustment magnitude for parameter fine-tuning, pre-defined parameter boundary rules are used to limit parameter adjustments from exceeding regulatory red lines, while risk control rules are used to limit the magnitude of a single adjustment and the number of consecutive adjustments in the same direction. Within the compliant adjustment range, the adjustment magnitude can be determined based on historical sensitivity analysis, calculated using the formula Δθ_micro=max(Δθ_min, Δθ_plan×10%), where Δθ_micro is the adjustment magnitude, Δθ_min is the minimum adjustment unit, Δθ_plan is the original planned magnitude, and max indicates taking the larger of the two values.

[0062] Before implementing parameter fine-tuning, a shadow model simulation is performed. This shadow model simulation involves simulating the risk assessment results corresponding to the parameter fine-tuning based on the experimental group's historical behavioral data, without actually adjusting the parameters to the experimental group's effective parameters. Based on similar historical adjustment cases, the system can predict the application conversion rate, delinquency probability, expected increase in delinquency amount, and expected increase in complaints for the experimental group's customers under the fine-tuned parameters, thus obtaining the risk assessment results.

[0063] When the risk assessment results meet the preset risk control conditions, such as the predicted increase in the delinquency rate not exceeding 0.5% and the expected loss amount being within an acceptable range, a micro-experiment plan is generated, including an experimental group, a control group, the adjustment range for parameter fine-tuning, and the execution time of the micro-experiment. The expected loss amount can be calculated based on the expected increase in delinquency, the average loss rate, and the loan balance.

[0064] In a preferred embodiment of this application, customer data corresponding to micro-customer groups is anonymized to obtain anonymous customer data. Then, based on preset privacy protection rules, customer characteristics within the anonymous customer data are classified into privacy levels, including: S4511: Receive customer information update notifications, which include customer data type update notifications and privacy protection requirement update notifications; S4512: Analyze customer information update notifications to determine the affected customer data fields and the corresponding compliance requirements; S4513: Based on the affected customer data fields and compliance requirements, select updated anonymization rules and updated customer feature classification standards from the preset policy library. The updated anonymization rules are used to determine at least one of the following processing methods for the affected customer data fields: deletion, replacement, encryption, generalization, and perturbation. The updated customer feature classification standards are used to determine the privacy level corresponding to the affected customer data fields. S4514: In a simulated environment, de-identified data is used to perform a pre-run of the updated anonymization rules and the updated customer characteristic classification standards to obtain strategy pre-run results. The strategy pre-run results include privacy protection strength assessment results and data availability assessment results. S4515: When the privacy protection strength assessment result meets the preset compliance index and the data availability assessment result meets the preset data availability index, the updated anonymization processing rule and the updated customer feature classification standard will be deployed to the micro-experiment solution generation process as the current execution rule. S4516: Anonymize the customer data corresponding to the micro-customer group according to the updated anonymization rules in the current execution rules to obtain anonymous customer data; S4517: Classify the privacy level of customer characteristics in anonymous customer data according to the updated customer characteristic classification standard in the current enforcement rules.

[0065] During the process of receiving customer information update notifications, these notifications can be pushed by the business system or the compliance management system. When a new data field is added to the business system, the customer data type update notification may include the field name, field data type, data source system identifier, and suggested sensitivity level. For example, when adding the field "face_feature_vector", it can be marked as a high-sensitivity field. When regulatory requirements change, the privacy protection requirement update notification may include the regulatory identifier, effective date, and binding clauses. The customer information update notification may carry one or both of the customer data type update notification and the privacy protection requirement update notification, so that the system can trigger subsequent processing based on changes in data fields or compliance requirements.

[0066] During the parsing of customer information update notifications, the system can extract affected customer data fields and their compliance requirements. Specifically, the system can identify field names, field types, regulatory identifiers, effective dates, sensitivity levels, and processing restrictions, and determine field categories through a dynamic field mapping table. The dynamic field mapping table can record mapping relationships such as "ID number" corresponding to "sensitive personal identification information" and "device fingerprint" corresponding to "indirect identifier," and expands as the parsing results of new notifications are obtained.

[0067] During the selection of updated anonymization rules and updated customer feature classification standards, a preset policy library can store processing rules corresponding to different field categories and business scenarios. For example, direct identifiers can be deleted, replaced, or encrypted; quasi-identifiers can be generalized; high-sensitivity fields can be processed using a combination of perturbation, encryption, or K-anonymization; medium-sensitivity fields can be used under encrypted transmission or access control; and low-sensitivity fields can be used for micro-experiment groups after completing permission verification. The use of low-sensitivity fields for micro-experiment groups does not imply bypassing data security management. When the newly added field "face_feature_vector" is parsed, the system can match it as a high-sensitivity field and select encrypted storage and differential privacy perturbation as the updated anonymization rules; the privacy budget parameter in differential privacy perturbation can be represented as ε_priv, which controls the perturbation strength.

[0068] In a preferred embodiment of this application, in a simulation environment, de-identified data is used to pre-run the updated anonymization rules and the updated customer characteristic classification criteria to obtain strategy pre-run results, including: S45141: Construct a simulation environment corresponding to the production environment topology. The simulation environment includes a data processing unit, a privacy protection policy execution unit, and a business indicator analysis unit. S45142: Generate anonymized data streams based on the statistical characteristics of historical customer data streams in the production environment. The statistical characteristics include data distribution, data volume, and data access patterns. S45143: Configure the data processing unit to receive and process the de-identified data stream; S45144: Deploy the updated anonymization rules and updated customer feature classification standards to the privacy protection policy enforcement unit; S45145: Run the updated anonymization rules and updated customer feature classification standards in a simulated environment, and monitor the simulated performance of data re-identification risk indicators and key business indicators through the business indicator analysis unit. Data re-identification risk indicators are used to represent the risk level of re-identifying customer identities in the processed de-identified data, and key business indicators are used to represent the usability of the processed de-identified data in micro-experimental analysis. S45146: Obtain historical baseline performance of the production environment, including baseline values ​​for data re-identification risk and key business metrics when the updated anonymization rules and updated customer feature classification standards have not been deployed. S45147: Compare the simulated performance of the data re-identification risk indicator with the baseline value of the data re-identification risk to obtain the privacy protection strength assessment result; S45148: Compare the simulated performance of key business indicators with the baseline values ​​of key business indicators to obtain the data availability assessment results; S45149: Generate policy simulation results based on the privacy protection strength assessment results and data availability assessment results.

[0069] During the construction of the simulation environment, it is used to verify the updated anonymization rules and updated customer characteristic classification standards without affecting the production environment. The simulation environment includes a data processing unit, a privacy protection policy enforcement unit, and a business indicator analysis unit. The data processing unit receives and processes the de-identified data stream, the privacy protection policy enforcement unit runs the updated anonymization rules and updated customer characteristic classification standards, and the business indicator analysis unit monitors the simulated performance of data re-identification risk indicators and key business indicators.

[0070] During the generation of the anonymized data stream, the system generates an anonymized data stream that does not contain real customer information, based on the statistical characteristics of historical customer data streams in the production environment. These statistical characteristics include data distribution, data volume, and data access patterns. For example, for the age field, the mean, standard deviation, and skewness of the age can be extracted, denoted as μ_age, σ_age, and γ_age, respectively; for daily active users, the mean daily active users μ_DAU can be extracted. Here, μ_age is the mean of the age field, σ_age is the standard deviation of the age field, γ_age is the skewness of the age field, and μ_DAU is the mean of daily active users. The generated anonymized data stream may contain fields such as synthetic customer identifiers, age, and income, while maintaining a data volume and access patterns similar to the production environment.

[0071] During the pre-execution simulation, the data processing unit receives the anonymized data stream and processes it according to normal business logic. The privacy protection policy execution unit applies the updated anonymization rules and updated customer characteristic classification standards in real time. The data re-identification risk indicator can be calculated using a record linking attack simulation method. This involves matching simulated external auxiliary information with the anonymized dataset and using the proportion of correctly matched records to the total number of records as the re-identification rate. The re-identification rate can be expressed as R_id = N_match / N_record,total, where R_id is the re-identification rate, N_match is the number of records successfully matched through the linking attack simulation, and N_record,total is the total number of data records participating in the simulation. Key business indicators can be reflected through the conversion rate calculation accuracy. The system uses anonymized data to calculate the simulated conversion rate and compares it with the conversion rate corresponding to the isolated and stored baseline statistical results to obtain the data availability indicator. The data availability metric can be expressed as E_use = |CR_mask - CR_base| / CR_base, where E_use is the absolute value of the relative error, CR_mask is the simulated conversion rate calculated based on the de-identified data, and CR_base is the conversion rate corresponding to the baseline statistical results, and CR_base is not zero.

[0072] In the process of obtaining historical baseline performance of the production environment, the system can extract baseline values ​​for data re-identification risk and baseline values ​​for key business indicators based on the historical production environment records. By comparing the simulated performance of the data re-identification risk indicators with the baseline values ​​for data re-identification risk, a privacy protection strength assessment result is obtained; by comparing the simulated performance of key business indicators with the baseline values ​​for key business indicators, a data availability assessment result is obtained. For example, if the simulated re-identification rate is lower than the historical baseline, it indicates that the privacy protection strength has improved; if the error between the simulated conversion rate and the baseline conversion rate remains within a preset range, it indicates that the data availability meets the requirements.

[0073] In a preferred embodiment of this application, a strategy simulation result is generated based on the privacy protection strength assessment result and the data availability assessment result, including: S451491: Based on the privacy protection strength assessment results and data availability assessment results, determine the overall strategy assessment results. The overall strategy assessment results are used to represent the privacy protection strength and data availability performance of the updated anonymization rules and the updated customer feature classification standards on the overall de-identified data stream. S451492: Acquire simulated performance data generated during the simulation environment when running the updated anonymization rules and updated customer feature classification standards. The simulated performance data includes data re-identification risk indicators, key business indicators and customer behavior characteristics segmented by micro-customer groups. S451493: Based on simulated performance data, identify unexpected changes in the strength of privacy protection or data availability of micro-customer groups. Unexpected changes are indicator changes that deviate from the historical baseline performance of the production environment and meet preset identification conditions. S451494: Identify the affected micro-customer groups based on unexpected changes; S451495: Conduct behavioral characteristic analysis on the affected micro-customer groups to obtain behavioral pattern difference values; S451496: Based on behavioral pattern differences, determine the micro-privacy protection strength assessment value and micro-data availability assessment value generated by the updated anonymization rules and updated customer characteristic classification standards for the affected micro-customer groups. S451497: Generate a micro-customer group impact report based on the micro-privacy protection strength assessment value and the micro-data availability assessment value; S451498: Integrate the micro-customer group impact report, overall strategy assessment results, and overall simulation performance data. The overall simulation performance data is the simulation performance data obtained from all de-identified data streams in the simulation environment when not divided according to micro-customer groups.

[0074] In determining the overall strategy evaluation result, the overall strategy evaluation result can be determined through a weighted scoring model. The calculation formula can be expressed as S_total=ω_priv×(1-R_id)×100+ω_use×A_use, where S_total is the overall strategy comprehensive score, ω_priv is the privacy protection strength weight, ω_use is the data availability weight, R_id is the re-identification rate, and A_use is the data availability accuracy rate. When S_total exceeds the preset overall qualification threshold, the overall strategy evaluation result is judged to be overall qualified.

[0075] In acquiring simulated performance data and identifying unexpected changes, the system can statistically analyze data re-identification risk indicators, key business indicators, and customer behavior characteristics for each micro-customer group based on preset customer segmentation dimensions such as age group, geographic level, and credit rating. For example, if the overall re-identification rate is 0.005, but the re-identification rate for the 25-30 age group increases to 0.075, and the conversion rate calculation error increases to 12.5%, then this local indicator change can be identified as an unexpected change. The increase in the re-identification rate can be expressed as G_rid = (R_id,current - R_id,base) / R_id,base × 100%, where G_rid is the increase in the re-identification rate, R_id,current is the current re-identification rate of the micro-customer group, and R_id,base is the corresponding historical baseline re-identification rate.

[0076] In identifying affected micro-customer groups and obtaining behavioral pattern difference values, the system can extract behavioral characteristics such as page browsing paths, click distribution, and application time distribution for these groups, and compare them with historical baseline behavioral characteristics to obtain behavioral pattern difference values. The behavioral pattern difference value can range from 0 to 1, with values ​​closer to 1 indicating a larger difference. For example, if the behavioral pattern difference value for a certain micro-customer group is 0.35, significantly higher than the 0.05-0.10 range for other groups, it indicates that this group is significantly affected by the updated anonymization rules and the updated customer characteristic classification standards.

[0077] In determining the micro-level privacy protection strength assessment value and the micro-level data availability assessment value, the system quantifies the micro-level privacy protection strength and micro-level data availability corresponding to the affected micro-level customer groups based on behavioral pattern differences, data re-identification risk indicators, and key business indicators. For example, a re-identification probability of 0.075 can be used as the basis for calculating the micro-level privacy protection strength assessment value, and a calculation error rate of 12.5% ​​can be used as the basis for calculating the micro-level data availability assessment value. A higher re-identification probability corresponds to lower micro-level privacy protection strength, and a higher calculation error rate corresponds to lower micro-level data availability.

[0078] During the generation and integration of micro-customer group impact reports, these reports can be output in a structured document format, including the affected group identifier, the magnitude of risk changes, and recommended adjustment measures. For example, the affected group identifier could be "25-30 years old_first-tier cities," and the recommended adjustment measure could be "adjust the K-anonymization parameter of this group from 3 to 5." The system integrates the micro-customer group impact reports, overall strategy evaluation results, and overall simulation performance data to form strategy pre-playback results, simultaneously reflecting the overall anonymized data stream performance and the differentiated impact on specific micro-customer groups.

[0079] In a preferred embodiment of this application, based on simulated performance data, identifying unexpected changes in the strength of privacy protection or data availability among micro-customer groups includes: S4514931: Perform quality checks on the simulated performance data to obtain quality-checked simulated performance data, wherein the quality check includes at least one of integrity check, consistency check, and value range check; S4514932: Group the simulated performance data after quality inspection according to micro-customer groups, and extract statistical features from the simulated performance data corresponding to each micro-customer group to obtain the corresponding statistical features. The statistical features include at least one of mean, variance, skewness, quantile and rate of change. S4514933: Obtain the baseline behavioral feature set corresponding to each micro-customer group. The baseline behavioral feature set is obtained by filtering historical customer behavior data that matches the micro-customer group and extracting features from historical customer behavior data in the production environment when the updated anonymization rules and updated customer feature classification standards have not been deployed. S4514934: Compare statistical features with the corresponding set of baseline behavioral features to obtain deviation information, which includes at least one of deviation direction, deviation magnitude, deviation duration, and the number of customers corresponding to the deviation. S4514935: Determine whether the deviation information meets the preset significance conditions. The preset significance conditions include the deviation magnitude exceeding the preset statistical significance threshold and / or the statistical test confidence level corresponding to the deviation information reaching the preset confidence threshold. S4514936: When the deviation information meets the preset significance condition, determine whether the deviation information meets the persistence standard, scale standard and magnitude standard. The persistence standard is used to indicate that the deviation information exists continuously within a preset number of time windows. The scale standard is used to indicate that the number of customers corresponding to the deviation information reaches a preset customer number threshold. The magnitude standard is used to indicate that the deviation magnitude reaches a preset business impact threshold. S4514937: When deviation information meets the persistence criterion, the size criterion, and the magnitude criterion, the deviation information is marked as an unexpected change.

[0080] During the quality check of simulated performance data, the system can perform integrity checks, consistency checks, and value range checks. Integrity checks are used to remove samples with a missing rate exceeding a preset missing threshold; consistency checks are used to compare the calculation results of the same indicator by different processing units; and value range checks are used to confirm that risk indicators are within the range of [0,1] and that percentage-based key business indicators are within the range of [0,100]%.

[0081] In the process of grouping and extracting statistical features from the simulated performance data after quality inspection, the system groups the data according to micro-customer group labels and extracts at least one statistical feature from the mean, variance, skewness, quantiles and rate of change. The statistical characteristics can be expressed as: μ_win=(x_1+x_2+...+x_7) / 7, σ²_win=Σ(x_i-μ_win)² / 6, γ_win=E[(X_win-μ_win)³] / σ³_win, r_day,t=(x_t-x_{t-1}) / x_{t-1}, where μ_win is the mean of the sliding window, x_i is the index value on day i within the sliding window, σ²_win is the sample variance of the sliding window, γ_win is the skewness of the sliding window, X_win is the random variable corresponding to the index value within the sliding window, σ_win is the standard deviation of the sliding window, r_day,t is the daily rate of change on day t, x_t is the index value on day t, and x_{t-1} is the index value on day t-1 and is not zero.

[0082] In acquiring the baseline behavioral feature set, the system can filter historical data from production environment historical customer behavior data that matches micro-customer group labels, and extract statistical features of the same type to form the corresponding baseline behavioral feature set. For example, it can filter historical data from the same period that match labels such as age and geographic level, for comparison with the current simulated performance data.

[0083] In the process of obtaining deviation information and determining whether the deviation information meets the preset significance condition, the system compares the current statistical characteristics with the baseline behavioral characteristic set to obtain the deviation direction, deviation magnitude, deviation duration, and the number of customers corresponding to the deviation. The deviation magnitude can be compared with a preset statistical significance threshold, and the statistical test result can be compared with a preset reliability threshold. In a specific embodiment, when the statistical significance probability p_sig is less than 0.05 and the deviation magnitude exceeds 20%, the deviation information is determined to meet the preset significance condition, where p_sig is used to represent the significance probability of the current deviation caused by random fluctuations in the statistical test.

[0084] In determining whether deviation information meets the persistence, scale, and magnitude standards, the persistence standard requires the deviation information to persist across multiple consecutive monitoring time windows; the scale standard requires the number of customers corresponding to the deviation information to reach a preset customer number threshold; and the magnitude standard requires the deviation magnitude to reach a preset business impact threshold, such as a conversion rate calculation error exceeding a preset percentage point, or a re-identification rate absolute value exceeding a preset risk threshold. The preset business impact threshold can be determined based on the sensitivity of historical business decisions to indicator errors.

[0085] When deviation information meets the persistence, scale, and magnitude criteria, the system marks the deviation information as an unexpected change and may trigger targeted policy adjustment suggestions, such as setting stricter privacy protection levels for affected micro-customer groups or excluding them from specific micro-experiments.

[0086] Reference Figure 2 This application also proposes an APP product model management system based on big data analysis, applicable to the product model management of multiple loan products, including: Information acquisition module 1 is used to acquire product information of multiple loan products and business indicator information corresponding to multiple loan products. The product information includes product identifiers and corresponding product parameters, and the business indicator information includes business indicators used to evaluate a single loan product and / or the overall product portfolio. Structure building module 2 is used to establish the association structure between products and business indicators based on product information and business indicator information. The association structure includes nodes representing loan products, product parameters and business indicators, as well as connection relationships representing the influence paths between nodes. The connection relationships have influence degree parameters and time lag parameters. The receiving module 3 is used to receive product parameter adjustment information. The product parameter adjustment information includes the product identifier of the target loan product, the product parameters that have been adjusted, the parameter values ​​before adjustment, the parameter values ​​after adjustment, and the adjustment time. The target loan product is the loan product whose product parameters have been adjusted, as indicated by the product parameter adjustment information, among multiple loan products. The impact calculation module 4 is used to determine the starting node in the association structure based on the product parameter adjustment information, and based on the impact degree parameter and time lag parameter of the connection relationship, to perform step-by-step propagation calculation of the direct and indirect impacts caused by the product parameter adjustment from the starting node along the impact path in the association structure, so as to obtain other affected loan products and business indicators, the corresponding impact path, the impact result, and the expected manifestation time of the impact result; The view provides module 5 to present the impact path, impact results, and expected display time, and to generate a comprehensive impact view of product parameter adjustments on the overall product portfolio based on other affected loan products and business metrics.

[0087] Through the above technical solutions, the management system automates and systematizes loan product model management by dividing and collaborating functional modules. It encapsulates complex data processing flows into independent modular units, improving the system's maintainability and scalability, while ensuring the efficiency and accuracy of product parameter adjustment impact analysis.

[0088] The above description is merely an embodiment of this application and is not intended to limit the scope of protection 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 scope of protection of this application.

Claims

1. A big data analysis-based APP product model management method applied to product model management of a plurality of loan products, characterized in that, include: Obtain product information of the multiple loan products and business indicator information corresponding to the multiple loan products. The product information includes product identifiers and corresponding product parameters. The business indicator information includes business indicators used to evaluate a single loan product and / or the overall product portfolio. Based on the product information and the business indicator information, an association structure between products and business indicators is established. The association structure includes nodes representing loan products, product parameters and business indicators, as well as connection relationships representing the influence paths between nodes. The connection relationships have influence degree parameters and time lag parameters. Receive product parameter adjustment information, which includes the product identifier of the target loan product, the product parameters that have been adjusted, the parameter values ​​before adjustment, the parameter values ​​after adjustment, and the adjustment time. The target loan product is the loan product whose product parameters have been adjusted, as indicated by the product parameter adjustment information, among the plurality of loan products. Based on the product parameter adjustment information, the starting node in the association structure is determined. Based on the influence degree parameter and time lag parameter of the connection relationship, the direct and indirect effects caused by the product parameter adjustment are calculated to propagate step by step from the starting node along the influence path in the association structure, so as to obtain other affected loan products and business indicators, the corresponding influence path, the influence result, and the expected manifestation time of the influence result. Present the impact path, the impact result, and the expected manifestation time, and generate a comprehensive impact view of the product parameter adjustment on the overall product portfolio based on other affected loan products and business indicators. 2.The APP product model management method based on big data analysis according to claim 1, wherein, Before performing a step-by-step propagation calculation of the direct and indirect effects caused by the product parameter adjustment from the starting node along the influence path in the associated structure based on the influence degree parameter and time lag parameter of the connection relationship, the method further includes: Based on the product parameter adjustment information and the association structure, a candidate customer group is determined. The candidate customer group is a customer group associated with the target loan product and / or related loan products. The related loan products are loan products that have an influence path with the target loan product in the association structure. Obtain behavioral monitoring data of the candidate customer group after the adjustment time, wherein the behavioral monitoring data includes at least one of browsing behavior data, click behavior data, application behavior data, conversion behavior data, and repayment behavior data; The behavior monitoring data is subjected to difference monitoring to identify valid abnormal differences, wherein the valid abnormal differences are the deviations of the actual behavior of the candidate customer group from the predicted behavior baseline, and the behavior deviations meet preset trigger conditions. When the effective abnormal difference exists, the micro customer group corresponding to the effective abnormal difference is determined, and the micro customer group is a subset of customers in the candidate customer group that corresponds to the effective abnormal difference; A micro-experiment plan is generated based on the micro-customer group. The micro-experiment plan includes dividing the micro-customer group into an experimental group and a control group, and performing parameter fine-tuning on the experimental group, wherein the adjustment range of the parameter fine-tuning is less than a preset fine-tuning range threshold. Collect experimental behavior data of the experimental group and the control group during the micro-experiment; Based on the differences in experimental behavior data between the experimental group and the control group, the micro-influence degree parameter and the micro-time lag parameter corresponding to the micro-customer group are determined. The micro-influence degree parameter is used to represent the strength of the influence of parameter fine-tuning on the behavior data of the micro-customer group, and the micro-time lag parameter is used to represent the time required for the influence of parameter fine-tuning on the behavior data of the micro-customer group to reach the preset manifestation conditions. Based on the micro-level impact parameter and the micro-level time lag parameter, update the impact parameter and time lag parameter of the connection relationship related to the micro-level customer group in the association structure; The influence degree parameter and time lag parameter based on the connection relationship are used to perform a step-by-step propagation calculation of the direct and indirect effects caused by the product parameter adjustment from the starting node along the influence path in the association structure, including: When the impact path involves the micro-customer group, the updated impact degree parameter and time lag parameter are used for step-by-step propagation calculation; when the impact path does not involve the micro-customer group, the unupdated impact degree parameter and time lag parameter are used for step-by-step propagation calculation. 3.The APP product model management method based on big data analysis according to claim 2, characterized in that, The step of performing difference monitoring on the behavior monitoring data and identifying valid abnormal differences includes: The monitoring time window is determined based on the target loan product, the related loan product, the candidate customer group, and the adjustment time. Based on historical business indicator information related to the target loan product and / or related loan products in the association structure, and current market environment information, a predictive behavior baseline for the candidate customer group within the monitoring time window is generated. Based on the behavior monitoring data and the predicted behavior baseline, the behavior deviation of the candidate customer group within the monitoring time window is calculated; Acquire external event information related to the monitoring time window, including policy release information and industry news information, as well as the event occurrence time, event-affected area, event-affected customer group, and event-affected product type extracted from the policy release information and industry news information; Based on the historical behavior data of the candidate customer group, the periodic pattern and trend pattern of the candidate customer group are determined. The periodic pattern is used to represent the behavioral change characteristics that repeat according to a preset time period, and the trend pattern is used to represent the behavioral change characteristics that continuously rise or fall within multiple consecutive time windows. Determine whether the behavioral deviation meets the preset event matching conditions with the external event information. The preset event matching conditions include at least one of the following: the occurrence time of the behavioral deviation matches the occurrence time of the event; the regional information corresponding to the candidate customer group matches the region affected by the event; the candidate customer group matches the customer group affected by the event; and the target loan product or the associated loan product matches the product type affected by the event. Determine whether the behavioral deviation conforms to the periodic pattern or the trend pattern; When the behavioral deviation and the external event information meet the preset event matching conditions, the behavioral deviation does not conform to the periodic pattern and the trend pattern, and the behavioral deviation exceeds the preset difference threshold within a preset number of consecutive monitoring time windows, the behavioral deviation is determined as an effective abnormal difference driven by an external event. When the behavioral deviation does not meet the preset event matching conditions with the external event information, and the behavioral deviation does not conform to the periodic pattern and the trend pattern, and the behavioral deviation exceeds the preset difference threshold within a preset number of consecutive monitoring time windows, the behavioral deviation is determined as a valid abnormal difference related to product parameter adjustment.

4. The APP product model management method based on big data analysis according to claim 3, characterized in that, Determining whether the behavioral deviation meets the preset triggering conditions includes: The degree of market volatility is determined based on current market environment information, the success rate of historical micro-experiments is determined based on historical micro-experiment records, and the potential risk exposure is determined based on the business indicator information corresponding to the candidate customer group. The initial difference threshold is adjusted based on the market volatility, the historical micro-experiment success rate, and the potential risk exposure to obtain the adjusted difference threshold. The representativeness score of the candidate customer group is determined based on the number of customers, behavioral activity, business contribution, and behavioral stability of the candidate customer group. When the behavioral deviation exceeds the adjusted difference threshold within a preset number of monitoring time windows, and the customer group representativeness score reaches the preset representativeness threshold, the statistical significance verification process is initiated. In the statistical significance verification process, a comprehensive confidence score is calculated based on the duration of the behavioral deviation, the magnitude of the behavioral deviation, the number of customers in the candidate customer group, and the potential risk exposure. When the overall confidence score reaches the preset confidence threshold, it is determined that the behavioral deviation meets the preset triggering condition.

5. The APP product model management method based on big data analysis according to claim 2, characterized in that, The process of generating micro-experiment plans based on the micro-customer groups includes: The customer data corresponding to the micro-customer group is anonymized to obtain anonymous customer data. Based on the preset privacy protection rules, the customer characteristics in the anonymous customer data are classified into privacy levels. Based on the privacy levels obtained, a set of customer characteristics that can be used for micro-experiment grouping is determined. Based on the customer feature set, customers with data analysis authorization identifiers are screened from the micro-customer group, and the screened customers are divided into an experimental group and a control group. The data analysis authorization identifier is used to indicate that the customer has agreed to use their anonymous customer data for data analysis, and the distribution difference between the experimental group and the control group on preset key customer features is less than a preset distribution difference threshold. Based on preset parameter boundary rules and risk control rules, determine the compliant adjustment range corresponding to the parameter fine-tuning, and determine the adjustment range of the parameter fine-tuning within the compliant adjustment range; Before performing the parameter fine-tuning, a shadow mode simulation is performed on the parameter fine-tuning. The shadow mode simulation refers to simulating the risk assessment result corresponding to the parameter fine-tuning based on the historical behavioral data of the experimental group without actually applying the parameter fine-tuning to the experimental group. When the risk assessment results meet the preset risk control conditions, a micro-experiment plan is generated, which includes the experimental group, the control group, the adjustment range of the parameter fine-tuning, and the micro-experiment execution time.

6. The APP product model management method based on big data analysis according to claim 5, characterized in that, The process involves anonymizing the customer data corresponding to the micro-customer group to obtain anonymous customer data, and then classifying the customer characteristics in the anonymous customer data into privacy levels according to preset privacy protection rules, including: Receive customer information update notifications, which include customer data type update notifications and privacy protection requirement update notifications; The customer information update notification is parsed to determine the affected customer data fields and the corresponding compliance requirements. Based on the affected customer data fields and the compliance requirements, an updated anonymization processing rule and an updated customer feature classification standard are selected from a preset policy library. The updated anonymization processing rule is used to determine at least one of the deletion, replacement, encryption, generalization, and perturbation processing methods for the affected customer data fields. The updated customer feature classification standard is used to determine the privacy level corresponding to the affected customer data fields. In a simulated environment, the updated anonymization rules and the updated customer characteristic classification standards are pre-run using de-identified data to obtain strategy pre-run results, which include privacy protection strength assessment results and data availability assessment results. When the privacy protection strength assessment result meets the preset compliance index and the data availability assessment result meets the preset data availability index, the updated anonymization processing rule and the updated customer feature classification standard are deployed as the current execution rules to the micro-experiment scheme generation process. According to the updated anonymization rules in the current execution rules, the customer data corresponding to the micro-customer group is anonymized to obtain the anonymous customer data; According to the updated customer characteristic classification standard in the current execution rules, the customer characteristics in the anonymous customer data are classified into privacy levels.

7. The APP product model management method based on big data analysis according to claim 6, characterized in that, In the simulated environment, the updated anonymization rules and the updated customer characteristic classification criteria are pre-run using de-identified data to obtain strategy pre-run results, including: Construct a simulation environment corresponding to the production environment topology, the simulation environment including a data processing unit, a privacy protection policy execution unit, and a business indicator analysis unit; Based on the statistical characteristics of historical customer data streams in the production environment, anonymized data streams are generated. The statistical characteristics include data distribution, data volume, and data access patterns. The data processing unit is configured to receive and process the de-identified data stream; The updated anonymization rules and the updated customer feature classification standards are deployed to the privacy protection policy enforcement unit; The updated anonymization rules and the updated customer characteristic classification standards are run in the simulation environment, and the simulated performance of data re-identification risk indicators and key business indicators is monitored through the business indicator analysis unit. The data re-identification risk indicators are used to represent the risk level of re-identifying customer identity after processing and de-identification data, and the key business indicators are used to represent the usability of processed and de-identified data in micro-experimental analysis. Obtain the historical baseline performance of the production environment, which includes the baseline values ​​of data re-identification risk and key business indicators when the updated anonymization rules and the updated customer feature classification standards were not deployed; The simulated performance of the data re-identification risk indicator is compared with the baseline value of the data re-identification risk to obtain the privacy protection strength assessment result; The simulated performance of the key business indicators is compared with the baseline value of the key business indicators to obtain the data availability assessment results; Based on the privacy protection strength assessment results and the data availability assessment results, the policy simulation results are generated.

8. The APP product model management method based on big data analysis according to claim 7, characterized in that, The step of generating the policy simulation results based on the privacy protection strength assessment results and the data availability assessment results includes: Based on the privacy protection strength assessment results and the data availability assessment results, an overall strategy assessment result is determined. The overall strategy assessment result is used to represent the privacy protection strength and data availability performance of the updated anonymization rules and the updated customer feature classification standards on the overall de-identified data stream. Acquire simulated performance data generated during the execution of the updated anonymization rules and the updated customer characteristic classification standards in the simulated environment. The simulated performance data includes data re-identification risk indicators, key business indicators, and customer behavior characteristics segmented according to micro-customer groups. Based on the simulated performance data, identify unexpected changes in the strength of privacy protection or data availability of micro-customer groups. The unexpected changes are indicator changes that deviate from the historical baseline performance of the production environment and meet preset identification conditions. Based on the aforementioned unexpected changes, identify the affected micro-customer groups; Behavioral characteristic analysis was performed on the affected micro-customer groups to obtain behavioral pattern difference values; Based on the behavioral pattern difference value, determine the micro-privacy protection strength assessment value and micro-data availability assessment value generated by the updated anonymization processing rule and the updated customer characteristic classification standard for the affected micro-customer group; Based on the micro-level privacy protection strength assessment value and the micro-level data availability assessment value, a micro-level customer group impact report is generated; The micro-customer group impact report, the overall strategy evaluation results, and the overall simulation performance data are integrated. The overall simulation performance data is the simulation performance data obtained from all de-identified data streams in the simulation environment when not divided according to micro-customer groups.

9. The APP product model management method based on big data analysis according to claim 8, characterized in that, The identification of unexpected changes in the strength of privacy protection or data availability of micro-customer groups based on the simulated performance data includes: The simulated performance data is subjected to a quality check to obtain the simulated performance data after quality check, wherein the quality check includes at least one of integrity check, consistency check and value range check; The simulated performance data after quality inspection is grouped according to micro-customer groups, and statistical features are extracted from the simulated performance data corresponding to each micro-customer group to obtain the corresponding statistical features. The statistical features include at least one of mean, variance, skewness, quantile and rate of change. Obtain a baseline behavioral feature set corresponding to each micro-customer group. The baseline behavioral feature set is obtained by filtering historical customer behavior data that matches the micro-customer group and extracting features from historical customer behavior data in the production environment when the updated anonymization processing rules and the updated customer feature classification standards were not deployed. The statistical features are compared with the corresponding set of baseline behavioral features to obtain deviation information, which includes at least one of deviation direction, deviation magnitude, deviation duration, and the number of customers corresponding to the deviation. Determine whether the deviation information meets the preset significance conditions. The preset significance conditions include the deviation magnitude exceeding the preset statistical significance threshold and / or the statistical test confidence level corresponding to the deviation information reaching the preset confidence threshold. When the deviation information meets the preset significance condition, it is determined whether the deviation information meets the persistence standard, the scale standard, and the magnitude standard. The persistence standard indicates that the deviation information exists continuously within a preset number of time windows. The scale standard indicates that the number of customers corresponding to the deviation information reaches a preset customer number threshold. The magnitude standard indicates that the deviation magnitude reaches a preset business impact threshold. When the deviation information meets the persistence criterion, the scale criterion, and the magnitude criterion, the deviation information is marked as the unexpected change.

10. An APP product model management system based on big data analysis, applied to the product model management of multiple loan products, characterized in that, include: The information acquisition module is used to acquire product information of the multiple loan products and business indicator information corresponding to the multiple loan products. The product information includes product identifiers and corresponding product parameters, and the business indicator information includes business indicators used to evaluate a single loan product and / or the overall product portfolio. The structure building module is used to establish a relationship structure between products and business indicators based on the product information and the business indicator information. The relationship structure includes nodes representing loan products, product parameters and business indicators, as well as connection relationships representing the influence paths between nodes. The connection relationships have influence degree parameters and time lag parameters. The receiving module is configured to receive product parameter adjustment information, which includes the product identifier of the target loan product, the product parameters that have been adjusted, the parameter values ​​before adjustment, the parameter values ​​after adjustment, and the adjustment time. The target loan product is the loan product whose product parameters have been adjusted, as indicated by the product parameter adjustment information, among the plurality of loan products. The impact calculation module is used to determine the starting node in the association structure based on the product parameter adjustment information, and based on the impact degree parameter and time lag parameter of the connection relationship, to perform a step-by-step propagation calculation of the direct and indirect impacts caused by the product parameter adjustment from the starting node along the impact path in the association structure, so as to obtain other affected loan products and business indicators, the corresponding impact path, the impact result, and the expected manifestation time of the impact result; The view-providing module is used to present the impact path, the impact result, and the expected manifestation time, and to generate a comprehensive impact view of the product parameter adjustment on the overall product portfolio based on other affected loan products and business indicators.

Images