Group insurance product recommendation algorithm and device based on big data and computer equipment

Abstract

The invention discloses a group insurance product recommendation algorithm and device based on big data and computer equipment, particularly relates to the field of insurance science and technology, and comprises feature fusion and standardization, space-time adaptation degree calculation, group preference propagation, dynamic demand prediction, multi-objective optimization sorting and feedback self-adaption. According to the method, deep integration of multi-source data is realized through multi-modal feature fusion and encryption standardization, service coverage radius change and policy periodic attenuation characteristics are dynamically captured by using a three-dimensional space-time quantification model, and recommendation timeliness is improved in combination with a group preference propagation mechanism; the innovatively designed double-target Pareto optimization strategy cooperates with enterprise profit and customer utility targets, and a space-time adaptation coefficient and a propagation weight are dynamically adjusted by collecting business indexes in real time, so that a closed-loop feedback system is formed. The technology effectively solves the problems of single data dimension, static rule lagging and benefit imbalance of a traditional method, and has the advantages of privacy protection, dynamic response and multi-target balance.

Description

Technical Field

[0001] The present invention relates to the field of insurance technology, and more specifically, to a group insurance product recommendation algorithm, device, and computer equipment based on big data. Background Art

[0002] Intelligent recommendation of group insurance products is an important research direction in the field of insurance technology. It aims to achieve precise product matching through data analysis. Existing technologies usually adopt a static matching mechanism between basic customer information and product terms. Its operation process is divided into three steps: first, structured data such as customer occupation type, company size and historical insurance policies are extracted from the business system; second, the matching degree between product and customer tags is calculated through a predefined rule library; finally, a recommendation list is generated based on single-dimensional sorting and the results are adjusted through manual underwriting.

[0003] The existing technology has three significant defects: first, the ability to process multi-source heterogeneous data is insufficient, and it is unable to effectively integrate unstructured data such as spatial location and temporal behavior, resulting in a single dimension of feature representation; second, there is a lack of dynamic spatiotemporal adaptation mechanism, and the impact of service outlet coverage radius, time-varying characteristics of occupational risks and policy cycle attenuation laws on recommendation effects are not considered; third, the use of a fixed weight sorting strategy makes it difficult to coordinate the balance between the profit goals of insurance institutions and the actual needs of customers, and there is a lack of online feedback and optimization capabilities. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a group insurance product recommendation algorithm, device and computer equipment based on big data, and solves the problems proposed in the above-mentioned background technology through the following scheme, namely, insufficient processing of multi-source heterogeneous data, lack of dynamic spatiotemporal adaptation mechanism, and difficulty in balancing the interests of multiple parties with fixed weight strategies and lack of online feedback tuning capabilities.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a group insurance product recommendation algorithm based on big data, comprising:

[0006] S1: Feature Fusion and Standardization: Integrate heterogeneous data sources such as product libraries, customer ledgers, and salesperson profiles to construct a cross-domain feature matrix. Convert geographic locations into raster coordinates through spatial encoding, use standardization methods to eliminate dimensional differences, encrypt sensitive fields, and output a set of feature vectors with a unified spatiotemporal granularity.

[0007] S2: Calculation of Spatiotemporal Compatibility: This comprehensively evaluates the degree of compatibility between the customer's geographic location and the product's service scope, the compatibility between occupational risk categories and insurance terms, and the fit between the policy period and time sensitivity, establishing a three-dimensional spatial-attribute-temporal quantitative model for customer product compatibility.

[0008] S3: Group Preference Propagation: Build a customer relationship graph based on customer feature similarity calculation and behavioral pattern correlation analysis. Design a preference propagation equation based on a graph neural network. Diffusion of explicit purchase behavior data along the network topology over multiple hops is performed to explore potential group preference and form an implicit demand correlation map between customer nodes.

[0009] S4: Dynamic Demand Forecasting: Integrating historical renewal cycle patterns and the influencing factors of real-time market events, we build a time series forecasting model to analyze the fluctuation characteristics of customers' potential insurance demand at different time points;

[0010] S5: Multi-objective optimization ranking: Construct a dual-objective optimization function that takes into account both corporate profit goals and customer utility value, set differentiated strategy parameters based on user role types, and generate a recommended product ranking sequence that satisfies the interests of multiple parties;

[0011] S6: Feedback Adaptation: By collecting click-through rate and conversion rate data of recommendation results in real time, an online learning mechanism is established to dynamically adjust model parameters to achieve continuous iterative optimization of recommendation strategies.

[0012] Preferably, the group insurance product recommendation device based on big data includes:

[0013] Feature Fusion and Standardization Module: This module integrates heterogeneous data sources such as product libraries, customer ledgers, and salesperson profiles to construct a cross-domain feature matrix. This module converts geographic locations into raster coordinates through spatial encoding, uses standardization methods to eliminate dimensional differences, encrypts sensitive fields, and outputs a set of feature vectors with a unified spatiotemporal granularity.

[0014] Spatiotemporal Compatibility Calculation Module: This module comprehensively evaluates the degree of compatibility between a customer's geographic location and the product's service scope, the compatibility between occupational risk categories and insurance terms, and the fit between the policy period and time sensitivity, establishing a three-dimensional spatial-attribute-temporal quantitative model for customer product compatibility.

[0015] Group Preference Propagation Module: This module constructs a customer relationship graph based on customer feature similarity calculation and behavioral pattern correlation analysis. It also designs a preference propagation equation based on a graph neural network. This module propagates explicit purchase behavior data across multiple hops along the network topology, mining potential group preference preferences and forming a mapping of implicit demand associations between customer nodes.

[0016] Dynamic Demand Forecasting Module: This module integrates historical renewal cycle patterns and the influencing factors of real-time market events to build a time series forecasting model and analyze the fluctuation characteristics of customers' potential insurance demand at different time points;

[0017] Multi-objective optimization ranking module: This module builds a dual-objective optimization function that takes into account both corporate profit goals and customer utility value, sets differentiated strategy parameters based on user role types, and generates a recommended product ranking sequence that satisfies the interests of multiple parties.

[0018] Feedback Adaptive Module: By collecting click-through rate and conversion rate data of recommendation results in real time, an online learning mechanism is established to dynamically adjust model parameters, thereby achieving continuous iterative optimization of recommendation strategies.

[0019] Preferably, a computer device stores a group insurance product recommendation device based on big data that can be loaded and executed by a processor as described above, and the computer device medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), magnetic disk or optical disk and other media that can store program code.

[0020] The technical effects and advantages of the present invention are as follows:

[0021] 1. This solution integrates heterogeneous data such as product libraries, customer ledgers, and salesperson profiles through multimodal feature fusion and encryption standardization. It uses spatial raster encoding to convert geographic locations into a computable latitude and longitude matrix. Combined with dynamic salt hash encryption, this approach protects privacy while improving data consistency. This addresses the incomplete feature representation problem of traditional methods due to the single data dimension.

[0022] 2. The spatiotemporal adaptation mechanism based on a three-dimensional quantitative model innovatively integrates the spatial compatibility index, occupational risk matching, and behavioral cycle fit. By constructing a customer relationship map, it propagates group preferences and effectively captures the dynamic changes in the coverage of service outlets, the time-varying patterns of occupational risk coefficients, and the decay characteristics of the policy validity period. This overcomes the shortcomings of static matching mechanisms that are often slow to respond to spatiotemporal factors.

[0023] 3. A strategy combining a dual-objective optimization function with online feedback tuning is adopted to balance corporate profits and customer utility goals within the Pareto optimal solution. By dynamically adjusting the spatiotemporal adaptation coefficient, propagation weight, and optimization parameters through real-time collection of business indicators, a closed-loop adaptive system is formed, breaking through the technical bottleneck of traditional fixed-weight strategies that are difficult to balance the interests of multiple parties and lack self-optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a schematic diagram of the overall process structure of the present invention.

[0025] Figure 2 It is a schematic structural diagram of a complete embodiment of the present invention.

[0026] Figure 3Schematic diagram of the data flow topology structure of the present invention. DETAILED DESCRIPTION

[0027] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0028] refer to Figure 1-Figure 3 The group insurance product recommendation algorithm based on big data shown includes:

[0029] S1: Feature fusion and standardization: Integrate heterogeneous data sources such as product libraries, customer ledgers, and salesperson portraits to build a cross-domain feature matrix. Convert geographic locations into raster coordinates through spatial encoding, use standardization methods to eliminate dimensional differences, encrypt sensitive fields, and output a set of feature vectors with a unified spatiotemporal granularity.

[0030] The S1 integrates three types of original inputs: product feature vectors, customer feature vectors, and behavioral time series data. It converts geographic locations into longitude and latitude coordinate systems through spatial rasterization encoding, uses Z-score standardization to eliminate numerical differences between features of different dimensions, implements dynamic salt value hash encryption for sensitive fields, and uses time window alignment technology to unify the time series granularity of multi-source data to generate a unified feature matrix containing spatial encoding, standardized values, and desensitized fields.

[0031] The product feature vector is extracted from the structured data storage of the product table. Full data synchronization is performed every morning through the scheduled task scheduling system. The complete status of product parameters is captured using database snapshot technology, and product information change events are monitored in real time through database triggers. The customer feature vector is constructed based on the historical transaction records of the ledger table. Data is incrementally extracted from the business database every hour through ETL tools, and customer status changes are captured in combination with transaction log parsing technology. The behavioral time series data extracts time series information from the effective date and renewal operation timestamp fields of the ledger table, as well as the group message records and tracking logs of the salesperson table.

[0032] The product feature vector is expressed as: P = (p1, ..., p n ), the customer feature vector is expressed as: C=(c1,...,c m ), the behavior time series data is expressed as: T=(t1,...,t k ); the standardization method is: x i Represents the original characteristic element, μ x represents the field mean, σx Represents the standard deviation; the dynamic salt value hash encryption is expressed as: Salt represents a dynamic salt value, which is updated every 24 hours according to the system time. Hash represents a composite hash function. s represents the plaintext of a sensitive field. Represents a bitwise exclusive OR operation.

[0033] S2: Calculation of spatiotemporal adaptability: Comprehensively evaluate the matching degree between the customer's geographical location and the coverage of the product service scope, the compatibility between the occupational risk category and the insurance terms, and the degree of fit between the policy period and time sensitivity, and establish a three-dimensional space-attribute-time customer product adaptability quantitative model.

[0034] The S2 uses the spatial compatibility index to evaluate the degree of matching between the customer's geographical location and the coverage radius of the product service outlets, uses the occupational risk matching metric to quantify the compatibility between the customer's occupational category and the underwriting terms of the insurance product, and combines the behavioral cycle fit to analyze the correlation strength between the customer's policy cycle and the current time node, constructing a multi-dimensional dynamic adaptation evaluation system to comprehensively judge the spatiotemporal matching characteristics of the customer and the insurance product.

[0035] The spatial compatibility index obtains the customer's geographic location coordinates in the ledger data and the service point coordinates marked in the product table, calculates the spherical distance between the two, normalizes the actual distance value with the maximum service radius of the product's historical underwriting cases, and finally generates a matching quantification value in the range of 0-1. The closer the value is to 1, the higher the match between the customer's geographic location and the service coverage of the product.

[0036] The spatial compatibility index is specifically expressed as: L c Indicates the client coordinates, L p Represents the coordinates of product service outlets, R max Indicates the maximum coverage radius of the product.

[0037] The occupational risk matching degree is determined by comparing the customer's occupational category with the major occupational category requirements specified in the insurance product underwriting terms item by item. A full matching value is assigned to occupational categories that fully meet the coverage scope, and a partial matching value is assigned to occupational categories that have overlapping risk levels but are not explicitly excluded. Finally, the weighted average of the matching results of all occupational categories is taken.

[0038] The occupational risk matching degree is specifically expressed as: o c Indicates the actual occupation category of the customer, o p represents the occupational categories allowed by the underwriting terms of the insurance product, j represents the jth occupational category, δ represents the occupational category matching function, δ = 1 for complete matching, δ = 0.5 for partial matching, and J represents the number of occupational categories required by the product.

[0039] The behavior cycle fit calculates the time sensitivity based on the date difference between the policy validity period and the current system time, and uses an exponential decay model to quantify the timeliness characteristics of customer renewal needs, where the decay coefficient is set according to the historical renewal cycle distribution law.

[0040] The behavior cycle fit is specifically expressed as: d expire Indicates the effective expiration date of the customer's current insurance policy, d current represents the current date when the system is calculating, and λ represents the decay coefficient.

[0041] S3: Group preference propagation: Build a customer relationship graph based on customer feature similarity calculation and behavioral pattern correlation analysis, design a preference propagation equation based on graph neural network, diffuse explicit purchase behavior data along the network topology structure in multiple hops, explore potential group selection preferences, and form an implicit demand association map between customer nodes.

[0042] The customer relationship graph calculates the basic correlation based on the cosine similarity of the customer's static feature vectors, introduces the cosine similarity of the temporal behavior vectors as a dynamic correction term, and weights and fuses the two types of similarities through the temporal weight coefficient optimized by grid search to construct a comprehensive correlation weight between customer nodes. The static features reflect the degree of matching of the customer's inherent attributes, the temporal behavior captures the similarity of dynamic interaction patterns, and the weight coefficient balances the contribution ratio of the two types of information.

[0043] The customer relationship map is specifically represented as follows: w ij represents the composite similarity weight in the customer relationship network, C i / C j represents the customer static feature vector, T i / T j represents the temporal behavior vector, α represents the temporal behavior weight coefficient, the value range of α is 0-1, and α = 0.35 is determined through grid search.

[0044] The preference propagation equation constructs a weighted relationship graph based on customer feature similarity and behavioral correlation, characterizes node connection strength through the Laplace matrix, performs inverse offset compensation on the initial purchase probability matrix and the identity matrix, controls the information diffusion range in combination with the empirically set propagation coefficient, and finally solves the linear equation group to obtain the steady-state preference distribution.

[0045] The preference propagation equation is specifically expressed as: new =(I-βL) -1 P init , P new represents the preference probability matrix, L represents the Laplace matrix, β represents the propagation coefficient, P initrepresents the initial purchase probability matrix.

[0046] The w ij As the edge weight matrix element of the customer relationship graph, it encodes the explicit similarity between customers and converts it into the structural constraint force of the topological space through the Laplace matrix L; new It is the steady-state solution under the action of the constraint, while retaining the initial purchase tendency P init On this basis, the preference values ​​of similar customer nodes are mutually infiltrated through weighted connection channels, and finally a new probability distribution is formed that reflects both individual characteristics and group influences.

[0047] S4: Dynamic demand forecasting: Integrate the historical renewal cycle patterns and the influencing factors of real-time market events to build a time series forecasting model to analyze the fluctuation characteristics of customers' potential insurance demand at different time points.

[0048] The S4 analyzes the time series patterns of customers' historical insurance behaviors through a baseline forecasting model to establish a basic demand forecast, introduces event impact factors to quantify the immediate disturbance effects of external events on insurance demand, combines time series trend forecasting with emergency response, and constructs a dynamic demand forecasting model with environmental perception capabilities, outputting the potential insurance demand intensity of customers in different scenarios in real time.

[0049] The baseline prediction model is based on historical policy effective dates and renewal cycle data, and uses autoregressive time series analysis methods to establish a basic prediction framework. The optimal lag period is determined through autocorrelation function analysis. The customer's historical insurance time point and renewal interval days are used as core input features. At the same time, real-time events are introduced as external intervention variables. The maximum likelihood estimation method is used to calculate the regression coefficients of each feature to construct a demand prediction benchmark model with strong interpretability.

[0050] The baseline prediction model is specifically expressed as: represents the forecast demand value at time t, y t-1 represents the actual demand observation value 1 day ago, y t-7 represents the actual demand observation value 7 days ago, φ1 and φ2 represent y t-1 and y t-7 The autoregressive coefficients at two historical points, ε t represents the random error term.

[0051] The time impact factor defines the types of market policy changes and marketing activities, quantifies the average impact of different events on insurance demand based on historical data regression analysis, uses a Boolean identification function within the event time window to capture the event status in real time, linearly weights the impact coefficient of each event type and its occurrence status, and dynamically corrects the baseline forecast value to reflect the impact effect of sudden market factors on customer demand.

[0052] The time impact factor is specifically expressed as: Δy represents the incremental impact of external events on insurance demand, γ k Indicates the intensity of the event impact. When the event is a policy adjustment, γ k =1.2, when the event is a promotion k =0.8, I(event k ) represents the event monitoring function.

[0053] The S4 passes Together with Δy, it constitutes a two-level correction mechanism of the prediction system, which is specifically expressed as: Y t Represents the final demand forecast value.

[0054] S5: Multi-objective optimization ranking: Construct a dual-objective optimization function that takes into account both corporate profit goals and customer utility value, set differentiated strategy parameters according to user role types, and generate a recommended product ranking sequence that satisfies the balance of interests of multiple parties.

[0055] The S5 constructs a dual-objective function that simultaneously optimizes the enterprise profit goal and customer utility value, designs constraints including maximum recommendation quantity restrictions and product selection variable constraints, and uses the Pareto front solution method to generate a recommended product ranking sequence that satisfies the balance between business strategy and customer needs, where the target weight is dynamically configured according to the user role type.

[0056] The dual-objective function constructs a profit maximization objective function based on the historical premium profit data of the enterprise side, and at the same time establishes a utility value evaluation function in combination with the client-side product adaptability index. The two types of objectives are integrated into a unified optimization framework through parameterized weighting, in which the weight parameters are dynamically configured according to the business scenario, and additional recommendation quantity restrictions and product mutual exclusivity constraints are added to form a complete mathematical programming model. Finally, a non-dominated sorting algorithm is used to generate a recommended solution set.

[0057] The dual objective function is specifically expressed as: Profit represents the profit target, Utility represents the utility target, and θ represents the adjustment parameter, which is determined according to the role type mapping table. Backstage staff = 0.7, salesperson = 0.3, r i represents the unit profit contribution value of the i-th insurance product, which is calculated by deducting the operating cost from the premium income of the product in the historical ledger data. j represents the customer utility score of the j-th insurance product, x i / x j represents the decision variable, when x i =1 indicates that the i-th product is recommended, x j =1 means recommending the jth product.

[0058] The u j Specifically expressed as: Δt represents the time interval between the current time and the customer demand forecast benchmark point.

[0059] The constraint design is based on actual business rules and system processing capabilities. The risk of information overload is controlled by setting a maximum number of recommendations. The threshold is determined based on the user interface carrying capacity and customer decision-making efficiency. The product selection binary variable constraint is used to ensure the discrete decision-making characteristics of the recommendation results. Each product has only two clear states: recommended or not recommended.

[0060] The constraint condition is specifically expressed as: ∑x i ≤N max ,x i ∈{0,1},N max Indicates the maximum recommended quantity.

[0061] S6: Feedback Adaptation: By collecting click-through rate and conversion rate data of recommendation results in real time, an online learning mechanism is established to dynamically adjust model parameters to achieve continuous iterative optimization of recommendation strategies.

[0062] The S6 uses real-time tracking of the click-through rate and conversion rate of recommended products as core effect evaluation indicators, establishes an online learning mechanism based on gradient descent, and dynamically adjusts the attenuation coefficient in the spatiotemporal adaptation calculation, the temporal behavior weight coefficient in the group preference propagation, and the adjustment parameters in the multi-objective optimization according to the changes in indicators, thereby realizing automatic iterative optimization of the recommendation strategy parameters. The parameter update process adopts a sliding window mechanism to retain the timeliness characteristics of recent behavior data, and corrects the calculation parameters of each process through backpropagation of the loss function to form a closed-loop feedback adjustment.

[0063] The effect evaluation indicators are constructed by collecting users' click behavior data on recommended products and insurance conversion data in real time to build a two-dimensional evaluation system. The click-through rate indicator calculates the ratio of the number of recommended product exposures to the actual number of user clicks to measure the attractiveness of the content. The conversion rate indicator counts the conversion ratio of users from viewing product details to completing insurance to evaluate the accuracy of demand matching. The data sources of the two types of indicators come from the front-end interaction log and the ledger system transaction records respectively.

[0064] The effect evaluation indicators specifically include: CTR stands for click-through rate, Conversion stands for conversion rate, Clicks stands for clicks, impressions stands for impressions, Details_Views stands for detail views, and Purchases stands for purchases.

[0065] The parameter update is based on the real-time collected click-through rate and conversion rate data of the recommendation results. The adjustment direction of each parameter is determined by calculating the partial derivatives of the effect indicators with respect to the model parameters. A dynamic learning rate is used to control the update step size. The learning rate decays exponentially over time to balance the convergence speed and stability. At the same time, parameter threshold boundary constraints are set to prevent over-adjustment, and a sliding window smoothing mechanism is activated for abnormal fluctuating data.

[0066] The update rule of the parameter update is specifically expressed as follows: α new represents the updated parameter value, α old Indicates the current parameter value. represents the partial derivative of the click rate with respect to the parameter, η represents the learning rate, with an initial value of 0.01 and decaying over time.

[0067] The present application also discloses a device for recommending group insurance products based on big data, including:

[0068] Feature Fusion and Standardization Module: This module integrates heterogeneous data sources such as product libraries, customer ledgers, and salesperson profiles to construct a cross-domain feature matrix. This module converts geographic locations into raster coordinates through spatial encoding, uses standardization methods to eliminate dimensional differences, encrypts sensitive fields, and outputs a set of feature vectors with a unified spatiotemporal granularity.

[0069] Spatiotemporal Compatibility Calculation Module: This module comprehensively evaluates the degree of compatibility between a customer's geographic location and the product's service scope, the compatibility between occupational risk categories and insurance terms, and the fit between the policy period and time sensitivity, establishing a three-dimensional spatial-attribute-temporal quantitative model for customer product compatibility.

[0070] Group Preference Propagation Module: This module constructs a customer relationship graph based on customer feature similarity calculation and behavioral pattern correlation analysis. It also designs a preference propagation equation based on a graph neural network. This module propagates explicit purchase behavior data across multiple hops along the network topology, mining potential group preference preferences and forming a mapping of implicit demand associations between customer nodes.

[0071] Dynamic Demand Forecasting Module: This module integrates historical renewal cycle patterns and the influencing factors of real-time market events to build a time series forecasting model and analyze the fluctuation characteristics of customers' potential insurance demand at different time points;

[0072] Multi-objective optimization ranking module: This module builds a dual-objective optimization function that takes into account both corporate profit goals and customer utility value, sets differentiated strategy parameters based on user role types, and generates a recommended product ranking sequence that satisfies the interests of multiple parties.

[0073] Feedback Adaptive Module: By collecting click-through rate and conversion rate data of recommendation results in real time, an online learning mechanism is established to dynamically adjust model parameters, thereby achieving continuous iterative optimization of recommendation strategies.

[0074] An embodiment of the present application also discloses a computer device that stores a group insurance product recommendation device based on big data that can be loaded and executed by a processor, and the computer device medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.

[0075] The present invention first integrates heterogeneous data sources such as product libraries, customer records, and salesperson portraits, converts geographic locations into longitude and latitude coordinates through spatial rasterization coding, adopts Z-score standardization to eliminate dimensional differences, and implements dynamic salt value hash encryption to generate a feature matrix with unified spatiotemporal granularity; then, a three-dimensional quantitative model is used to calculate the spherical distance matching between customer geographic locations and product service outlets, the matching weights between occupational categories and underwriting terms, and the exponential decay fit between the policy validity period and the current time; then, a customer relationship map based on feature similarity and temporal behavior association is constructed, and a preference propagation equation is designed using a weighted Laplace matrix to convert initial purchases into actual purchases. The probability diffuses along the network topology to form a group preference distribution; then, the autoregressive time series model and event influencing factors are integrated to perform dynamic demand forecasting, generating a demand intensity value that includes baseline trends and event disturbances; then, a dual-objective optimization function is constructed and a maximum recommendation quantity constraint is imposed, and a ranking sequence that takes into account both corporate profits and customer utility is generated through Pareto front solution; finally, click-through rate and conversion rate data are collected in real time, and the spatiotemporal adaptation attenuation coefficient, preference propagation weight and multi-objective parameters are adjusted online through gradient descent to form a closed-loop feedback optimization, completing the complete recommendation process from data preprocessing, adaptation evaluation, group propagation, demand forecasting, optimization sorting to parameter adaptation.

[0076] Secondly: The drawings of the embodiments disclosed in the present invention only involve structures related to the embodiments disclosed in the present invention. Other structures may refer to conventional designs. The same embodiment and different embodiments of the present invention may be combined with each other without conflict.

[0077] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A group insurance product recommendation algorithm based on big data, characterized by: include: S1: Feature Fusion and Standardization: Integrate heterogeneous data sources such as product libraries, customer ledgers, and salesperson profiles to construct a cross-domain feature matrix. Convert geographic locations into raster coordinates through spatial encoding, use standardization methods to eliminate dimensional differences, encrypt sensitive fields, and output a set of feature vectors with a unified spatiotemporal granularity. S2: Calculation of Spatiotemporal Compatibility: This comprehensively evaluates the degree of compatibility between the customer's geographic location and the product's service scope, the compatibility between occupational risk categories and insurance terms, and the fit between the policy period and time sensitivity, establishing a three-dimensional spatial-attribute-temporal quantitative model for customer product compatibility. S3: Group Preference Propagation: Build a customer relationship graph based on customer feature similarity calculation and behavioral pattern correlation analysis. Design a preference propagation equation based on a graph neural network. Diffusion of explicit purchase behavior data along the network topology over multiple hops is performed to explore potential group preference and form an implicit demand correlation map between customer nodes. S4: Dynamic Demand Forecasting: Integrating historical renewal cycle patterns and the influencing factors of real-time market events, we build a time series forecasting model to analyze the fluctuation characteristics of customers' potential insurance demand at different time points; S5: Multi-objective optimization ranking: Construct a dual-objective optimization function that takes into account both corporate profit goals and customer utility value, set differentiated strategy parameters based on user role types, and generate a recommended product ranking sequence that satisfies the interests of multiple parties; S6: Feedback Adaptation: By collecting click-through rate and conversion rate data of recommendation results in real time, an online learning mechanism is established to dynamically adjust model parameters to achieve continuous iterative optimization of recommendation strategies.

2. The big data-based group insurance product recommendation algorithm according to claim 1, characterized in that: The S1 integrates three types of original inputs: product feature vectors, customer feature vectors, and behavioral time series data. It converts geographic locations into longitude and latitude coordinate systems through spatial rasterization coding, uses Z-score standardization to eliminate numerical differences between features of different dimensions, implements dynamic salt value hash encryption for sensitive fields, and uses time window alignment technology to unify the time series granularity of multi-source data to generate a unified feature matrix containing spatial coding, standardized values, and desensitized fields. The S2 uses the spatial compatibility index to assess the degree of match between the customer's geographic location and the coverage radius of the product service outlets, uses the occupational risk matching metric to quantify the compatibility between the customer's occupational category and the insurance product's underwriting terms, and combines the behavioral cycle fit to analyze the correlation strength between the customer's policy cycle and the current time node, thereby constructing a multi-dimensional dynamic adaptation evaluation system to comprehensively judge the spatiotemporal matching characteristics between the customer and the insurance product; The S4 analyzes the time series patterns of customers' historical insurance behaviors through a baseline forecasting model to establish a basic demand forecast. It introduces event impact factors to quantify the immediate perturbation effect of external events on insurance demand. It combines time series trend forecasting with emergency response to build a dynamic demand forecasting model with environmental perception capabilities, and outputs the potential insurance demand intensity of customers in different scenarios in real time. The S5 constructs a dual-objective function that simultaneously optimizes the enterprise profit goal and customer utility value, designs constraints including the maximum number of recommendations and product selection variable constraints, and uses the Pareto frontier solution method to generate a recommended product ranking sequence that satisfies the balance between business strategy and customer needs. The objective weights are dynamically configured according to the user role type. The S6 uses real-time tracking of the click-through rate and conversion rate of recommended products as core effect evaluation indicators, establishes an online learning mechanism based on gradient descent, and dynamically adjusts the attenuation coefficient in the spatiotemporal adaptation calculation, the temporal behavior weight coefficient in the group preference propagation, and the adjustment parameters in the multi-objective optimization according to the changes in indicators, thereby realizing automatic iterative optimization of the recommendation strategy parameters. The parameter update process adopts a sliding window mechanism to retain the timeliness characteristics of recent behavior data, and corrects the calculation parameters of each process through backpropagation of the loss function to form a closed-loop feedback adjustment.

3. The big data-based group insurance product recommendation algorithm, device, and computer equipment according to claim 2, characterized in that: The product feature vector is extracted from the structured data storage of the product table. Full data synchronization is performed every morning through the scheduled task scheduling system. The complete status of product parameters is captured using database snapshot technology, and product information change events are monitored in real time through database triggers. The customer feature vector is constructed based on the historical transaction records of the ledger table. Data is incrementally extracted from the business database every hour through ETL tools, and customer status changes are captured in combination with transaction log parsing technology. The behavioral time series data extracts time series information from the effective date and renewal operation timestamp fields of the ledger table, as well as the group message records and tracking logs of the salesperson table.

4. The big data-based group insurance product recommendation algorithm according to claim 2, characterized in that: The spatial compatibility index is calculated by obtaining the customer's geographic location coordinates in the ledger data and the service point coordinates marked in the product table, calculating the spherical distance between the two, and normalizing the actual distance value with the maximum service radius of the product's historical underwriting cases to ultimately generate a matching quantified value between 0 and 1. The closer the value is to 1, the more closely the customer's geographic location matches the product's service coverage. The occupational risk matching degree is determined by comparing the customer's occupational category with the occupational category requirements specified in the insurance product underwriting terms and conditions. A full matching value is assigned to occupational categories that fully meet the coverage scope, and a partial matching value is assigned to occupational categories that have overlapping risk levels but are not explicitly excluded. The weighted average of the matching results of all occupational categories is finally taken. The behavior cycle fit calculates the time sensitivity based on the date difference between the policy validity period and the current system time, and uses an exponential decay model to quantify the timeliness characteristics of customer renewal needs, where the decay coefficient is set according to the historical renewal cycle distribution law.

5. The big data-based group insurance product recommendation algorithm according to claim 1, characterized in that: The customer relationship graph calculates the basic correlation based on the cosine similarity of the customer's static feature vector, introduces the cosine similarity of the time series behavior vector as a dynamic correction term, and uses the time series weight coefficient optimized by grid search to weight the two types of similarities to construct the comprehensive correlation weight between customer nodes, where Static features reflect the degree of matching of customers’ inherent attributes, temporal behaviors capture the similarity of dynamic interaction patterns, and the weight coefficient balances the contribution ratio of the two types of information; The preference propagation equation constructs a weighted relationship graph based on customer feature similarity and behavioral correlation, characterizes node connection strength through the Laplace matrix, performs inverse offset compensation on the initial purchase probability matrix and the identity matrix, controls the information diffusion range in combination with the empirically set propagation coefficient, and finally solves the linear equation group to obtain the steady-state preference distribution.

6. The big data-based group insurance product recommendation algorithm according to claim 2, characterized in that: The baseline forecasting model is based on historical policy effective date and renewal cycle data. It uses autoregressive time series analysis to establish a basic forecasting framework, determines the optimal lag period through autocorrelation function analysis, uses the customer's historical insurance time point and renewal interval as core input features, introduces real-time events as external intervention variables, and uses maximum likelihood estimation to calculate the regression coefficients of each feature to construct a highly interpretable demand forecasting benchmark model. The time impact factor defines the types of market policy changes and marketing activities, quantifies the average impact of different events on insurance demand based on historical data regression analysis, uses a Boolean identification function within the event time window to capture the event status in real time, linearly weights the impact coefficient of each event type and its occurrence status, and dynamically corrects the baseline forecast value to reflect the impact effect of sudden market factors on customer demand.

7. The big data-based group insurance product recommendation algorithm according to claim 2, characterized in that: The dual-objective function constructs a profit maximization objective function based on historical premium profit data on the enterprise side, and simultaneously establishes a utility value evaluation function in combination with the client-side product suitability index. These two objectives are integrated into a unified optimization framework through a parameterized weighting approach, where the weight parameters are dynamically configured according to the business scenario. A complete mathematical programming model is formed by adding a recommendation quantity limit and product mutual exclusivity constraints. Finally, a non-dominated sorting algorithm is used to generate a recommended solution set. The constraint design is based on actual business rules and system processing capabilities. The risk of information overload is controlled by setting a maximum number of recommendations. The threshold is determined based on the user interface carrying capacity and customer decision-making efficiency. The product selection binary variable constraint is used to ensure the discrete decision-making characteristics of the recommendation results. Each product has only two clear states: recommended or not recommended.

8. The big data-based group insurance product recommendation algorithm according to claim 2, characterized in that: The effectiveness evaluation index is constructed by collecting real-time user click behavior data on recommended products and insurance conversion data to build a two-dimensional evaluation system. The click-through rate index calculates the ratio of the number of recommended product exposures to the number of actual user clicks to measure the content's appeal. The conversion rate index calculates the conversion ratio of users from viewing product details to completing insurance purchases to evaluate the accuracy of demand matching. The data sources of these two types of indicators are respectively from the front-end interaction log and the ledger system transaction records; The parameter update is based on the real-time collected click-through rate and conversion rate data of the recommendation results. The adjustment direction of each parameter is determined by calculating the partial derivatives of the effect indicators with respect to the model parameters. A dynamic learning rate is used to control the update step size. The learning rate decays exponentially over time to balance the convergence speed and stability. At the same time, parameter threshold boundary constraints are set to prevent over-adjustment, and a sliding window smoothing mechanism is activated for abnormal fluctuating data.

9. A group insurance product recommendation device based on big data, used to implement the group insurance product recommendation algorithm based on big data as described in any one of claims 1 to 8, characterized in that: include: Feature Fusion and Standardization Module: This module integrates heterogeneous data sources such as product libraries, customer ledgers, and salesperson profiles to construct a cross-domain feature matrix. This module converts geographic locations into raster coordinates through spatial encoding, uses standardization methods to eliminate dimensional differences, encrypts sensitive fields, and outputs a set of feature vectors with a unified spatiotemporal granularity. Spatiotemporal Compatibility Calculation Module: This module comprehensively evaluates the degree of compatibility between a customer's geographic location and the product's service scope, the compatibility between occupational risk categories and insurance terms, and the fit between the policy period and time sensitivity, establishing a three-dimensional spatial-attribute-temporal quantitative model for customer product compatibility. Group Preference Propagation Module: This module constructs a customer relationship graph based on customer feature similarity calculation and behavioral pattern correlation analysis. It also designs a preference propagation equation based on a graph neural network. This module propagates explicit purchase behavior data across multiple hops along the network topology, mining potential group preference preferences and forming a mapping of implicit demand associations between customer nodes. Dynamic Demand Forecasting Module: This module integrates historical renewal cycle patterns and the influencing factors of real-time market events to build a time series forecasting model and analyze the fluctuation characteristics of customers' potential insurance demand at different time points; Multi-objective optimization ranking module: This module builds a dual-objective optimization function that takes into account both corporate profit goals and customer utility value, sets differentiated strategy parameters based on user role types, and generates a recommended product ranking sequence that satisfies the interests of multiple parties. Feedback Adaptive Module: By collecting click-through rate and conversion rate data of recommendation results in real time, an online learning mechanism is established to dynamically adjust model parameters, thereby achieving continuous iterative optimization of recommendation strategies.

10. A computer device, characterized in that: A computer program is stored which can be loaded by a processor and executes the method according to any one of claims 1 to 8.

Images