Insurance product recommendation method and device, storage medium and computer device

Abstract

The application relates to the technical field of product recommendation, and discloses an insurance product recommendation method and device, a storage medium and computer equipment, the method comprising the following steps: inputting customer features of each customer into a gradient boosting decision tree model which has been trained to obtain a satisfaction prediction value of the customer; analyzing contribution values and contribution directions of each customer feature type to the satisfaction prediction value of each customer to generate a feature contribution degree vector; dividing the customers into multiple levels according to the satisfaction prediction values of the customers; performing aggregation processing on the feature contribution degree vectors of the customers for each level; calculating average contribution values and average contribution directions of each customer feature type in the level; determining key constraint bottleneck feature types of a customer group in the level; matching corresponding insurance product labels; and screening a candidate insurance product set matched with the insurance product labels to determine an insurance product recommendation result in combination with historical insurance application records of the customers. The application can be applied to a financial technology scene.

Description

Technical Field

[0001] This application relates to the field of product recommendation technology, and in particular to a method, apparatus, storage medium, and computer equipment for recommending insurance products. Background Technology

[0002] In the context of digital transformation in the insurance industry, accurately identifying customer needs and recommending suitable products has become crucial for improving customer satisfaction. Due to the low-frequency transaction and long decision-making cycle of insurance products, traditional general recommendation methods are ill-suited to increasingly segmented customer groups. The industry urgently needs an intelligent recommendation solution that can explain the reasons behind customer preferences and dynamically adapt to different customer segments to support refined operations.

[0003] Currently, insurance product recommendations mainly rely on rule matching (such as based on static labels like age and region), collaborative filtering, or simple machine learning models (such as logistic regression and random forests) to score and rank customers' purchase intentions. These methods typically output a comprehensive prediction score, but fail to deeply analyze the specific impact and direction of each customer characteristic (such as premium affordability, protection preferences, and historical interaction behavior) on the prediction result. Furthermore, the recommendation logic lacks transparency and is difficult to explain clearly to sales personnel or customers.

[0004] Therefore, the existing technology has the following shortcomings: First, it cannot identify the key constraints that lead to low customer satisfaction, making optimization measures lack a basis; second, it uses a uniform recommendation strategy to process all customers, ignoring differences within the group, resulting in inconsistent recommendation effects; third, it lacks an effective conversion mechanism from feature analysis to product selection, making it difficult to quickly translate analysis conclusions into actionable recommendation rules, and the overall accuracy, interpretability, and adaptability of the recommendation need to be improved. Summary of the Invention

[0005] In view of the above, embodiments of this application provide an insurance product recommendation method, apparatus, storage medium, and computer equipment.

[0006] According to one aspect of this application, an insurance product recommendation method is provided, the method comprising: The system acquires customer features from multiple customers, inputs each customer's features into the trained gradient boosting decision tree model, and infers and outputs the predicted satisfaction value of each customer within a predetermined time period. Attribution analysis is then used to calculate the contribution value and direction of each customer feature type to the predicted satisfaction value of each customer, generating a feature contribution vector for each customer. Based on the predicted satisfaction values ​​of each customer, each customer is divided into multiple levels. For each level, the feature contribution vectors of customers belonging to that level are aggregated. The average contribution value and average contribution direction of each customer feature type within that level are calculated. Based on the average contribution value with a negative average contribution direction, the key bottleneck feature type of the customer group in that level is determined. Based on the key constraint bottleneck feature types corresponding to each level, the corresponding insurance product tags are matched from the feature-policy mapping rule base, and a set of candidate insurance products matching the insurance product tags is selected from the insurance product database; Based on each customer's historical insurance records and the set of candidate insurance products corresponding to each customer's level, the recommended insurance products for each customer are determined.

[0007] According to another aspect of this application, an insurance product recommendation device is provided, the device comprising: The prediction module is used to acquire customer features of multiple customers, input each customer's customer features into the trained gradient boosting decision tree model, infer and output the predicted satisfaction value of each customer in the future within a predetermined time period, and use the attribution analysis method to calculate the contribution value and direction of each customer feature type to the predicted satisfaction value of each customer, generating the feature contribution vector of each customer. The segmentation module is used to divide each customer into multiple levels based on the predicted satisfaction value of each customer, and for each level, aggregate the feature contribution vectors of customers belonging to the level, calculate the average contribution value and average contribution direction of each customer feature type within the level, and determine the key constraint bottleneck feature type of the customer group in the level based on the average contribution value with a negative average contribution direction. The filtering module is used to match the corresponding insurance product tags from the feature-policy mapping rule base according to the key constraint bottleneck feature types corresponding to each level, and to filter the set of candidate insurance products that match the insurance product tags from the insurance product database. The recommendation module is used to determine the recommended insurance products for each customer based on their historical insurance records and the set of candidate insurance products corresponding to their respective insurance tier.

[0008] According to another aspect of this application, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described insurance product recommendation method.

[0009] According to another aspect of this application, a computer device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the program to implement the above-described insurance product recommendation method.

[0010] Using the above technical solutions, the insurance product recommendation method, apparatus, storage medium, and computer equipment provided in this application embodiment acquire customer characteristics of multiple customers and input them into a trained gradient boosting decision tree model to infer and output the predicted satisfaction value of each customer within a predetermined future time period. Simultaneously, attribution analysis is used to calculate the contribution value and direction of each customer characteristic type to the predicted satisfaction value of each customer, generating a feature contribution vector for each customer. Based on this, all customers are divided into multiple levels according to their predicted satisfaction values, and the feature contribution vectors of all customers within each level are aggregated to calculate the average contribution value and average contribution direction of each customer characteristic type within that level. Then, based on the average contribution value with a negative average contribution direction, the key bottleneck characteristic type of the customer group at that level is determined. Subsequently, based on the key bottleneck characteristic type corresponding to each level, corresponding insurance product tags are matched from the feature-policy mapping rule base, and a set of matching candidate insurance products is selected from the insurance product database. Finally, based on each customer's historical insurance records and the set of candidate insurance products corresponding to each customer's level, the insurance product recommendation result for each customer is determined. This application achieves interpretability of satisfaction prediction results by combining gradient boosting decision trees with attribution analysis, solving the problem that black-box models have difficulty in explaining the basis of predictions. Through customer segmentation and group feature aggregation, it achieves quantitative identification and statistically stable output of the core shortcomings of customers at each level. Through the matching mechanism of feature-policy mapping rule base, bottleneck features are accurately transformed into product recommendation directions. Through gap-driven scoring and ranking, it ensures that the recommendation results accurately match the customer's existing protection status and avoids homogeneous and duplicate recommendations.

[0011] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0012] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating an insurance product recommendation method provided in an embodiment of this application is shown. Figure 2 A flowchart illustrating another insurance product recommendation method provided in an embodiment of this application is shown; Figure 3 This paper shows a schematic diagram of the structure of an insurance product recommendation device provided in an embodiment of this application; Figure 4 A schematic diagram of the device structure of a computer device provided in an embodiment of this application is shown. Detailed Implementation

[0013] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0014] This embodiment provides a method for recommending insurance products, such as... Figure 1 As shown, the method includes: Step 101: Obtain customer features of multiple customers, input each customer's customer features into the trained gradient boosting decision tree model, infer and output the predicted satisfaction value of each customer in the future predetermined time period, and use the attribution analysis method to calculate the contribution value and contribution direction of each customer feature type to the predicted satisfaction value of each customer, and generate the feature contribution vector of each customer.

[0015] Specifically, obtaining customer characteristics from multiple customers includes: acquiring raw data, including: collecting raw data from multiple data sources for each customer, including a core business system, a customer service system, and a marketing system; the data collected from the core business system includes at least basic customer information data, policy information data, premium payment data, and claims information data; the data collected from the customer service system includes at least consultation record data, complaint record data, and service acceptance record data; the data collected from the marketing system includes at least marketing outreach record data, conversion record data, and activity participation record data; and anonymizing sensitive personal information in the collected raw data.

[0016] Furthermore, the acquisition of customer features from multiple customers also includes: preprocessing the collected raw data to generate the customer features. The preprocessing includes: filling missing values ​​of categorical features with the mode of the feature, filling missing values ​​of numerical features with the median of the feature, and removing features with a missing value rate exceeding a predetermined threshold; performing outlier detection on all numerical features using statistical distribution criteria, calculating the mean and standard deviation of each feature, and replacing values ​​exceeding the mean plus or minus a predetermined multiple of the standard deviation with the corresponding boundary values; using weighted encoding for unordered categorical features and label encoding for ordered categorical features; and standardizing all numerical features by scaling each feature to a standard distribution with a mean of 0 and a variance of 1.

[0017] For example, the customer characteristics include multiple features from the following six dimensions: Basic customer attributes, including multiple features such as age, gender, occupation, education level, marital status, family structure, city tier, risk preference, and social security payment status; Policy lifecycle, including multiple features such as duration of insurance coverage, number of valid policies, type of insurance, cumulative coverage amount, annual premium amount, payment method, payment period, coverage period, cumulative renewal frequency, historical renewal rate, number of policy lapses, grace period payment status, and number of policy reinstatements; and Consumer behavior, including the most recent payment date, cumulative payment frequency, payment amount in the most recent time period, average annual payment amount, annual premium growth rate, cumulative number of policy additions, and cumulative number of policy additions. The data includes multiple factors such as amount, number of supplementary insurance purchases, and preferred insurance channels; claims behavior dimensions, including multiple factors such as cumulative number of claims, cumulative claim amount, number of claims within the recent scheduled time period, claim type, claim processing time, claim settlement rate, number of claim complaints, number of illness-related claims, and number of accident-related claims; service interaction dimensions, including multiple factors such as the time of the most recent service interaction, cumulative interaction frequency, interaction channel preference, type of consultation content, cumulative number of complaints, complaint resolution rate, participation rate in activities within the recent scheduled time period, marketing outreach response rate, and customer service connection rate; and macroeconomic dimensions, including multiple factors such as per capita disposable income of residents in the customer's city, insurance industry depth in the province, insurance industry growth rate, guaranteed interest rate, and consumer price index.

[0018] Among them, the gradient boosting decision tree model refers to a machine learning model that uses a decision tree as a base learner and performs ensemble learning by iteratively fitting the negative gradient direction. Attribution analysis methods refer to algorithmic mechanisms based on game theory or gradient propagation principles to calculate the degree and direction of influence of each input feature on the model's output. Its output includes the contribution value (i.e., the magnitude of influence) and the contribution direction (i.e., positive promotion or negative constraint) of each feature. The feature contribution vector is a vectorized data structure composed of the contribution values ​​and directions corresponding to each customer feature type, used to represent the reasons for the formation of a single customer satisfaction prediction result.

[0019] Specifically, the process begins by extracting raw data from multiple customers from the customer relationship management system, historical policy database, and third-party credit data sources. After data cleaning and standardization, a customer feature set is formed. This set of customer features includes, but is not limited to, age range, annual income level, number of previously purchased insurance products, number of historical claims, average claim amount, policy holding duration, number of days since the last interaction, and customer service evaluation score. The customer features of each customer are then assembled into feature vectors according to the model input format and input into a gradient boosting decision tree model that has been trained offline. This model performs a layer-by-layer nonlinear transformation on the input features based on the splitting logic of each decision tree node, ultimately outputting a predicted satisfaction value for each customer within a predetermined future time period (e.g., the next three or six months). This predicted satisfaction value is a continuous value between 0 and 1, representing the customer's overall expectation level regarding the current insurance service experience. Subsequently, for each customer, attribution analysis is used to reverse-interpret the model's output satisfaction prediction. Specifically, the SHAP method based on Shapley values ​​or the integral gradient method based on gradient integrals can be used to calculate the contribution of each customer feature type to the customer's satisfaction prediction (e.g., a feature's contribution is +0.12 or -0.08) and the direction of contribution (positive indicates that the current value of the feature positively promotes satisfaction improvement, negative indicates that the current value of the feature has an inhibitory effect on satisfaction). The contribution values ​​and directions of all customer feature types are then combined to form the customer's feature contribution vector. It should be noted that in actual execution, if some customer features are missing for a certain customer, the missing features are filled with preset default values ​​or median values ​​before being used in inference. If the computational load exceeds a preset threshold due to the complexity of the model structure during attribution analysis, an approximate attribution algorithm (such as Kernel SHAP) is used instead of precise calculation to ensure processing efficiency.

[0020] By combining model reasoning and attribution analysis in this step, the satisfaction prediction results not only have numerical accuracy, but also causal interpretability. The attribution analysis mechanism enables the system to clearly identify the specific direction and degree of influence of each customer characteristic on the prediction results, providing a data foundation for subsequent development of differentiated optimization strategies for different customer groups.

[0021] In one optional embodiment, the training method of the gradient boosting decision tree model includes: Acquire historical customer sample data, and divide the historical customer sample data into a training sample set and a validation sample set. The historical customer sample data includes historical customer characteristics and corresponding satisfaction tags. A gradient boosting decision tree model is trained using a gradient boosting decision algorithm. Before training begins, each continuous feature in the historical customer features is binned, and multiple bin boundaries for each continuous feature are determined. The values ​​of each continuous feature are mapped to their corresponding discrete bin indices. At the start of each iteration, a new decision tree is initialized as the current tree, and the maximum number of leaf nodes and / or maximum depth parameters of the current tree are set. During the growth of the current tree, the following leaf node splitting steps are repeated until the current tree reaches the maximum number of leaf nodes or the maximum depth: all leaf nodes in the current tree are obtained; for each leaf node, a histogram is constructed based on the bin indices of the feature values ​​of the training samples falling into that leaf node; the histogram is traversed to search for candidate split points that maximize the splitting gain, and the candidate split points and their corresponding splitting gains are recorded; after traversal, the target leaf node with the maximum splitting gain is selected from all leaf nodes, and the splitting operation is performed according to the candidate split points corresponding to the target leaf node to generate two new child leaf nodes. After each iteration, the current model is used to predict the validation sample set and calculate the preset evaluation index value. When the improvement of the current evaluation index value relative to the historical best evaluation index value is less than the predetermined minimum improvement threshold, the counter for consecutive unimproved rounds is incremented by 1. When the value of the counter reaches the predetermined maximum tolerable rounds, the iteration stops and the model parameters are rolled back to the iteration round corresponding to the historical best evaluation index value. The model parameters corresponding to the iteration round with the best performance on the validation sample set are used as the trained gradient boosting decision tree model.

[0022] The gradient boosting decision tree algorithm is an ensemble learning algorithm that iteratively combines multiple weak learners (decision trees) into a strong learner. Each iteration generates a new decision tree to fit the prediction residual of the previous tree, allowing the overall model to gradually approach the true value in each iteration. Gradient boosting decision tree models include XGBoost (eXtreme Gradient Boosting), CatBoost (Categorical Boosting), and LightGBM (Lightweight Gradient Boosting Machine). This application uses the LightGBM model as an example. LightGBM (Light Gradient Boosting Machine, an open-source distributed gradient boosting framework) is an efficient implementation based on the gradient boosting decision tree framework. It significantly reduces training time and memory consumption while maintaining prediction accuracy through histogram binning and leaf node growth strategies. Histogram binning involves discretizing continuous feature values ​​into a finite number of discrete intervals before model training begins. Each interval is called a bin, and the values ​​of each continuous feature are mapped to the corresponding discrete bin index. The model then performs calculations based on these discrete values ​​during subsequent training. Binning boundaries refer to the critical values ​​between consecutive intervals when dividing the value range of a continuous feature into multiple consecutive intervals. Determining the binning boundaries determines the fineness of feature discretization. Leaf node split gain is a quantitative indicator used in decision tree training to measure the effect of a candidate split point on reducing model prediction error. A larger split gain indicates that the split point has a more significant effect on improving model performance.

[0023] In this optional embodiment, the gradient boosting decision tree model is trained in the following way: During the data preparation phase, historical customer sample data is acquired. This data originates from the insurance company's accumulated historical business data. Each sample contains customer characteristic data for a past period and the corresponding satisfaction tag. The customer characteristic data covers multiple dimensions, including basic customer information, policy holding status, payment behavior, claims records, and service interaction records. The satisfaction tag is a pre-determined quantitative value of the customer's satisfaction during the corresponding historical period. After acquiring the historical customer sample data, it is divided into a training sample set and a validation sample set.

[0024] During the binning phase, before training begins, each continuous feature in the historical customer features is binned. Specifically, for each continuous feature, its value distribution across all training samples is extracted. Based on this distribution, multiple bin boundaries are determined, and the continuous values ​​of each sample for that feature are mapped to their corresponding discrete bin indices. After binning, the value of each continuous feature is converted into a discrete bin number, pointing to the value range to which the feature value belongs. Several implementation methods can be used to determine the bin boundaries. One method is equal-width binning, which divides the feature's value range into a predetermined number of equal intervals, ensuring each interval has the same width. Another method is equal-frequency binning, which adjusts the bin boundaries to ensure that the number of samples falling into each bin is approximately equal. Yet another method is histogram-based binning, which involves sampling the feature values ​​and constructing a histogram. The bin boundaries are then determined based on the histogram's distribution pattern, ensuring that the bin boundaries fall at the "valley" of the data distribution, thus grouping areas with concentrated feature value distributions into the same bin. By using histogram binning, continuous features are discretized into a finite number of bin indices, which significantly reduces memory usage and computational complexity during model training. When calculating split gain, the model only needs to traverse each bin instead of traversing the original feature values ​​of each sample, thus significantly improving training speed.

[0025] During the iterative training phase, the gradient boosting decision tree algorithm is used to train the initial gradient boosting decision tree model, i.e., the decision tree sequence is gradually constructed through multiple iterations. The specific process of each iteration is as follows: At the beginning of each iteration, a new decision tree is initialized as the current tree, and the maximum number of leaf nodes and the maximum depth of the current tree are set. The maximum number of leaf nodes parameter limits the maximum number of leaf nodes that the tree can generate, and the maximum depth parameter limits the longest path length from the root node to any leaf node. In practice, several implementation methods can be adopted: one method is to set both the maximum number of leaf nodes and the maximum depth simultaneously, and stop growing when the tree reaches either of these limits; another method is to set only the maximum number of leaf nodes without limiting the depth, and let the number of leaf nodes control the complexity of the tree; yet another method is to set only the maximum depth without limiting the number of leaf nodes, and let the depth control the complexity of the tree. The specific values ​​of the maximum number of leaf nodes and the maximum depth are determined by those skilled in the art based on practical factors such as the training sample size and feature dimensions. During the growth of the current tree, the leaf node splitting step is repeated until the current tree reaches the maximum number of leaf nodes or the maximum depth.

[0026] The leaf node splitting step specifically includes: obtaining all leaf nodes in the current tree. For each leaf node in the current tree, obtaining a subset of training samples that have fallen into that leaf node; these samples have reached the node along the path in the current decision tree but have not yet been split. For that leaf node, constructing a histogram based on the bin indices of the feature values ​​of the training samples falling into that node. The histogram is constructed as follows: for each feature type and its corresponding bins, the accumulated gradient value and the accumulated Hessian matrix value of the training samples falling into that bin are calculated. After the histogram is constructed, the histogram is traversed through each bin index to search for candidate split points that maximize the splitting gain. That is, for each bin, the splitting gain when the sample is divided into left and right parts using that bin as the splitting point is calculated, and the position of the bin that maximizes the splitting gain and its corresponding splitting gain value are recorded. For each leaf node, the optimal candidate split point and its corresponding splitting gain are recorded. After all candidate split points for all leaf nodes have been recorded, the target leaf node with the largest splitting gain is selected from all leaf nodes. This selection operation ensures that each split prioritizes the node most beneficial for reducing model prediction error, thereby maximizing the model's fitting ability within a limited tree size. After selection, the split operation is performed according to the candidate split point corresponding to the target leaf node, splitting the leaf node into two new child leaf nodes. After splitting, the new child leaf nodes are added to the current tree's leaf node list, and the next round of splitting continues. For example, in the specific calculation method of split gain, one implementation is based on variance reduction gain calculation, that is, the split gain equals the mean squared error of the parent node minus the sum of the mean squared errors of the two child nodes; another implementation is based on log loss gain calculation, that is, the split gain equals the log loss of the parent node minus the sum of the log losses of the two child nodes. Through the leaf node growth strategy, each split selects the globally optimal leaf node for expansion, achieving a better fitting effect than layer-by-layer growth within the same tree size. At the same time, the setting of the maximum number of leaf nodes and the maximum depth parameters effectively prevents the risk of overfitting caused by the overgrowth of a single tree.

[0027] During the validation phase, after each iteration, the current model is used to predict the validation sample set, and a preset evaluation metric value is calculated. Evaluation metrics such as mean squared error and mean absolute error can be used. After calculating the evaluation metric value for this round, it is compared with the historical best evaluation metric value. Specifically, the current evaluation metric value is compared with the best evaluation metric value recorded in all rounds up to now: if the current evaluation metric value is better than the historical best value (i.e., the improvement reaches or exceeds the predetermined minimum improvement threshold), the historical best evaluation metric value is updated with the current evaluation metric value, and the counter for consecutive rounds without improvement is reset to its initial value; if the improvement of the current evaluation metric value relative to the historical best evaluation metric value is less than the predetermined minimum improvement threshold, the counter for consecutive rounds without improvement is incremented by 1. When the value of the counter for consecutive rounds without improvement reaches the predetermined maximum tolerable number of rounds, the iterative training process stops. After stopping the iteration, the model parameters are rolled back to the model parameters of the iteration round corresponding to the historical best evaluation metric value, and the model parameters corresponding to the best-performing iteration round on the validation set are output as the trained gradient boosting decision tree model. By controlling the early stopping strategy, the model training is actively terminated when the performance on the validation set no longer improves, avoiding meaningless extra iterations. At the same time, rolling back to the historical best model parameters ensures that the final model has the best generalization performance.

[0028] For example, in a specific implementation, the gradient boosting decision tree model (such as LightGBM) mainly consists of a data input and preprocessing module, a feature engineering and binning module, a tree model integration module, a model parameter configuration module, and an output module. The tree model integration module, as the core processing unit, is composed of multiple CART regression trees cascaded together using a gradient boosting framework. The trees are connected in series, with the prediction residual of the previous tree serving as the fitting target for each subsequent tree. Each decision tree internally adopts a binary tree structure, recursively composed of a root node, intermediate nodes, and leaf nodes. Nodes are connected by conditional branches. Sample data flows in from the root node, routes downwards layer by layer according to the feature splitting conditions at each node, and finally falls into the leaf node and is output. The tree's predicted value is output. During the tree's growth, a depth-constrained leaf-based growth strategy is adopted. That is, in each round of splitting, all current leaf nodes are traversed, the gain of each candidate split point is calculated, and the leaf node with the largest gain is selected for splitting until the preset maximum number of leaf nodes or maximum depth constraint is reached. The prediction results of each decision tree are weighted and accumulated in the output module, and the final satisfaction prediction value is generated after activation function or identity mapping. During the training phase, the prediction residuals are propagated back along the tree structure as the fitting target for the new decision trees in the next iteration, forming a series of error correction closed loops. At the same time, the validation set performance is monitored through an early stopping strategy. When the validation set evaluation index no longer improves in multiple consecutive iterations, the training is automatically terminated to ensure that the model has good generalization ability.

[0029] In an optional embodiment, the satisfaction label is determined as follows: The first satisfaction score is determined based on the recommendation ratings given by customers after historical service interactions, wherein the recommendation rating is a quantitative rating given by customers within a predetermined rating range; A second satisfaction score is determined based on the historical complaint records and service handling records of historical customers, and the number of complaints and / or complaint resolution rate are mapped to the satisfaction score. The third satisfaction score is determined based on the customer's historical renewal behavior data. The third satisfaction score is the first value when the historical customer completes the renewal within the predetermined time after the policy expires, and the second value otherwise. The fourth satisfaction rating is determined based on the number of times insurance was added and the premium growth rate in the historical consumption behavior data of customers. The satisfaction label is determined based on the first satisfaction score, the second satisfaction score, the third satisfaction score, and the fourth satisfaction score.

[0030] In this optional embodiment, the satisfaction label is determined by a combination of the first satisfaction score, the second satisfaction score, the third satisfaction score, and the fourth satisfaction score.

[0031] The primary method for determining customer satisfaction ratings is by obtaining recommendation ratings from past customer interactions. Specifically, after a customer receives a service from an insurance company (such as claims processing, policy changes, or consultation), the insurance company's customer service system sends a satisfaction rating invitation. The customer provides a quantitative rating within a predetermined range, and this rating becomes the recommendation rating. For example, after a customer completes a claims process, the system automatically sends an evaluation SMS or in-app push notification, inviting the customer to rate their satisfaction with the claims service. The rating range is 0 to 10 points, and the score selected by the customer becomes the recommendation rating for that service interaction.

[0032] The second satisfaction score is determined by obtaining historical customer complaint records and service handling records, and then determining the second satisfaction score based on the number of complaints and the complaint resolution rate. Specifically, the customer's complaint event records and service handling records within a predetermined historical time period are extracted from the customer service system. The total number of complaint events and the number of resolved and customer-approved complaints are counted. The complaint resolution rate is obtained by dividing the number of resolved complaints by the total number of complaints. Several implementation methods can be used to map the number of complaints to the satisfaction score: One method is to use the number of complaints within a predetermined historical time period as input; the fewer the number of complaints, the higher the satisfaction score. When the number of complaints is zero, this dimension of the second satisfaction score takes its highest value; when the number of complaints reaches or exceeds a predetermined threshold, this dimension takes its lowest value. Another method is to divide the number of complaints into preset intervals, with different intervals corresponding to different satisfaction score values. Mapping complaint resolution rate to satisfaction score can be implemented in several ways: one approach is to directly map the percentage of complaint resolution rate to the corresponding satisfaction score value, with a higher resolution rate corresponding to a higher satisfaction score; another approach is to take the highest value for that dimension of the second satisfaction score when the complaint resolution rate reaches or exceeds a predetermined threshold, and a lower value when the resolution rate is below the threshold. When considering both the number of complaints and the resolution rate, the score values ​​for each factor can be determined separately first, and then the two scores can be weighted and summed to obtain the second satisfaction score.

[0033] The third satisfaction rating is determined by obtaining historical renewal behavior data from existing customers. For each expiring policy, it is recorded whether the customer completed the renewal payment within the scheduled time after the policy's expiration. When a historical customer completes the renewal within the scheduled time after the policy's expiration, the third satisfaction rating is set to the highest value (indicating that the renewal behavior reflects high customer satisfaction); when a historical customer fails to complete the renewal within the scheduled time after the policy's expiration, the third satisfaction rating is set to the lowest value (indicating that the non-renewal behavior reflects low customer satisfaction).

[0034] The fourth satisfaction score is determined by obtaining historical customer consumption behavior data, including the number of times the customer added insurance coverage and the premium growth rate. The number of times the customer added insurance coverage refers to the number of times they purchased new insurance products within a predetermined historical period. Each new purchase is considered a single insurance addition; a higher number of additions indicates a stronger willingness to increase insurance coverage. The premium growth rate refers to the increase in the customer's premium payment amount within adjacent time intervals. Specifically, it can be calculated by comparing the premium payment amount in the current time interval with that in the previous time interval and calculating the ratio of the difference to the premium payment amount in the previous time interval. After mapping the number of additions and the premium growth rate to satisfaction scores, the fourth satisfaction score is determined by combining both. One implementation method is to divide the number of additions according to predetermined intervals, with different score values ​​corresponding to different intervals, and higher score values ​​for higher additions. Similarly, the premium growth rate is divided according to predetermined growth rate intervals, with different score values ​​corresponding to different growth rate intervals, and higher score values ​​for positive and higher growth rates, and lower score values ​​for zero or negative growth rates. The fourth satisfaction score is obtained by weighted summation of the scores corresponding to the two dimensions.

[0035] The method for determining the fusion of satisfaction labels: After obtaining the first, second, third, and fourth satisfaction scores respectively, the final satisfaction label is determined by combining these four scores. For example, the four scores can be weighted and summed, with the weight coefficients for each score determined by those skilled in the art based on business experience or historical data fitting. Providing multiple scoring fusion methods enables this technical solution to have flexible adaptability in different business scenarios.

[0036] Step 102: Based on the predicted satisfaction values ​​of each customer, divide each customer into multiple levels. For each level, aggregate the feature contribution vectors of customers belonging to that level, calculate the average contribution value and average contribution direction of each customer feature type within that level, and determine the key bottleneck feature type of the customer group in that level based on the average contribution value with a negative average contribution direction.

[0037] Here, "tier" refers to the subgroups obtained by dividing the customer group according to the numerical distribution range of the predicted satisfaction value. Customers within the same tier have similar expected satisfaction levels. Aggregation processing involves statistically summarizing the feature contribution vectors of multiple customers within the same tier according to the feature type dimension to extract common contribution patterns for that tier. The average contribution value is the arithmetic mean of the contribution values ​​of all customers within the same tier on the same feature type, used to characterize the strength of the influence of that feature type on the overall satisfaction of the tier. The average contribution direction is the comprehensive direction determined by majority voting or the mean sign based on the contribution directions of all customers within the same tier on the same feature type. Key bottleneck feature types refer to customer feature types within a certain tier where the average contribution direction is negative and the absolute value of the average contribution value meets preset screening criteria. These features are considered the core bottleneck factors limiting the improvement of customer satisfaction at that tier.

[0038] Specifically, firstly, based on the predicted customer satisfaction values, all customers are divided into multiple tiers. The tiering method can be equal-width binning (dividing the numerical range of 0 to 1 evenly into a predetermined number of sub-intervals, e.g., 3 to 5), equal-frequency binning (ensuring the number of customers within each tier is roughly equal), or clustering algorithms (such as K-means) to naturally cluster the predicted satisfaction values ​​into multiple groups, with each group corresponding to a tier. After tiering, each customer is labeled with its tier identifier. For each tier, the feature contribution vectors of all customers belonging to that tier are obtained, and these vectors are aggregated: for each customer feature type, the arithmetic mean of the contribution values ​​of all customers in that tier for that feature type is calculated to obtain the average contribution value of that feature type in that tier; simultaneously, based on the positive and negative distribution of the contribution direction of each customer in that tier for that feature type, the average contribution direction of that feature type in that tier is determined by majority voting (i.e., counting the number of positive and negative contributions and taking the one with the larger number as the overall direction) or by weighted summation of contribution values ​​followed by sign. After obtaining the average contribution value and average contribution direction of each feature type, the feature types with a negative average contribution direction are first screened out. Then, from these negative feature types, key bottleneck feature types are further identified. Specifically, they can be sorted by average contribution value from smallest to largest (i.e., negative degree from largest to smallest), and a preset number of feature types at the top of the sorting are selected as key bottleneck feature types. Alternatively, a negative threshold is set, and all feature types with an average contribution value below the negative threshold are identified as key bottleneck feature types. For example, if the average contribution value of the "premium sensitivity" feature in a certain level is -0.15 and its direction is negative, while the average contribution value of other negative features is greater than -0.05, then the "premium sensitivity" feature is identified as the key bottleneck feature type of that level. Optionally, before aggregation, outlier detection can be performed on the feature contribution vectors within the same level to remove customer vectors whose contribution values ​​deviate from the mean of the feature type in that level by more than a preset multiple of the standard deviation, in order to prevent extreme values ​​from having an excessive impact on the average result. In addition, if the number of customers in a certain level is less than the preset minimum sample size, the level will be merged with the adjacent level before the aggregation process is performed to ensure the stability of the statistical results.

[0039] Through customer segmentation and bottleneck feature identification in this step, the recommendation system can locate the core shortcomings that lead to low customer satisfaction at each level from the group level. The segmentation process ensures that customer groups with different satisfaction levels can obtain differentiated analysis conclusions, avoiding uniform analysis from obscuring the characteristics of subgroups. The aggregation process eliminates individual noise interference, so that the bottleneck feature identification results reflect the common patterns of the group rather than individual anomalies.

[0040] In one optional implementation, the step of dividing customers into multiple tiers based on their predicted satisfaction values ​​includes: Obtain the percentile ranking of each customer's predicted satisfaction value among all customers, and obtain the current life cycle stage label of each customer, which includes the new customer stage, growth stage, maturity stage, decline stage, and churn stage. The hierarchical division will be executed in the following priority order: Those whose percentile ranking falls within the top 100th percentile are classified into the first tier; Customers whose percentile ranking falls between the first percentile and the second percentile, or whose lifecycle stage tag belongs to the first pre-booking stage set, are classified into the second tier. The first pre-booking stage set includes the growth stage and the new customer stage. Customers whose percentile ranking falls between the second percentile and the third percentile, or whose lifecycle stage label belongs to the second pre-sales stage set, are classified into the third tier. The second pre-sales stage set includes the new customer period and the decline period. The remaining customers who do not meet any of the above classification criteria will be classified into the fourth tier.

[0041] Percentile ranking refers to the relative position of each customer after sorting all customers according to their predicted satisfaction values ​​from highest to lowest, expressed as a percentage. For example, ranking in the top 10% means that the customer's predicted satisfaction value is higher than that of 90% of the other customers in the total customer base. Lifecycle stage labels refer to classification markers that characterize the stage of the business relationship between a customer and an insurance company. Specifically, the new customer stage refers to the initial exploratory stage after a customer establishes a business relationship with an insurance company. At this stage, the customer's insurance needs have not yet been fully developed, and consumption behavior is mainly trial purchases. The growth stage refers to the period when the business relationship between the customer and the insurance company gradually deepens, and the frequency and amount of consumption increase. At this stage, the customer's understanding of insurance products gradually deepens, and the insurance behavior becomes more active. The maturity stage refers to the period when the business relationship between the customer and the insurance company reaches a stable state. At this stage, the customer holds a large number of policies, premium payments are stable, and there is a high degree of recognition of the insurance company's services. The decline stage refers to the period when the business relationship between the customer and the insurance company shows a downward trend. At this stage, the customer may gradually reduce business dealings with the insurance company due to policy expiration, changes in economic conditions, or other reasons. The churn stage refers to the period when the customer has basically or completely terminated business dealings with the insurance company. At this stage, the customer has stopped purchasing or renewing insurance, and the marketing outreach effect of the insurance company is significantly reduced.

[0042] In this optional embodiment, firstly, the predicted satisfaction value for each customer in the entire customer base is obtained. This predicted satisfaction value is inferred and output by a trained gradient boosting decision tree model. The predicted satisfaction values ​​of all customers are sorted in descending order, and the percentile ranking of each customer is calculated based on their position in the sorted sequence. For example, when a customer's predicted satisfaction value is higher than 90% of the other customers in the entire customer base, that customer's percentile ranking is in the top 10%. Simultaneously, the current lifecycle stage label for each customer is obtained. The lifecycle stage label is maintained by the insurance company's customer relationship management system or business system. This system dynamically determines the lifecycle stage of each customer based on multiple dimensions of information, such as the length of business dealings between the customer and the insurance company, the number of policies held, the time of the most recent business interaction, and the continuity of premium payments. The lifecycle stage labels include new customer stage, growth stage, maturity stage, decline stage, and churn stage. For example, in determining the specific method of lifecycle stage labels, the interval between the customer's first insurance purchase time and the current time can be used as the main dividing criterion. When the interval is short (such as less than a predetermined time threshold), it is marked as the new customer period; when the interval is in the middle range and the number of policies is on the rise, it is marked as the growth period; when the interval is long and the number of policies is stable, it is marked as the maturity period; when the interval is long and the number of policies is on the decline period, it is marked as the decline period; and when the interval is extremely long and there are no valid policies, it is marked as the churn period.

[0043] After obtaining percentile rankings and lifecycle stage labels, the tiering is performed sequentially in descending order of priority. The purpose of this priority order is to prioritize customers who meet the criteria for multiple tiers, placing them in the higher-ranked tier to avoid duplicate tier assignments. Specifically, the first tier is defined as follows: customers whose percentile ranking falls within the top 100th percentile are assigned to the first tier. This first tier is based solely on percentile ranking. The specific value of the top 100th percentile can be pre-defined by those skilled in the art based on the insurance company's customer distribution characteristics and business strategy. For example, in one implementation, the first percentile is set to 10%, meaning customers whose predicted satisfaction value ranks in the top 10% are assigned to the first tier. The second tier is defined as follows: customers who do not meet the criteria for the first tier but whose percentile ranking falls between the first and second percentiles, or whose lifecycle stage labels belong to a first predetermined stage set, are assigned to the second tier. The first predetermined stage set includes the growth stage and the new customer stage. The second percentile is greater than the first percentile. For example, if the first percentile is 10%, the second percentile could be 40%. Specifically, customers meeting either of the following conditions are classified into the second tier: Condition 1: their percentile ranking falls between the first and second percentiles (e.g., between 10% and 40%); Condition 2: their lifecycle stage label belongs to the first predetermined stage set (i.e., growth stage or new customer stage). If a customer does not meet the first tier conditions but meets at least one of conditions 1 and 2, they are classified into the second tier. If a customer meets both the second and third tier conditions, since the second tier has higher priority than the third tier, the customer is classified into the second tier and will not be considered for the third tier. The third tier is defined as follows: customers who do not meet the above tier classification conditions but whose percentile ranking falls after the second percentile and before the third percentile, or whose lifecycle stage label belongs to the second predetermined stage set, are classified into the third tier. The second predetermined stage set includes the new customer stage and the decline stage. The third percentile is greater than the second percentile; for example, if the second percentile is 40%, the third percentile could be 80%. Specifically, customers meeting either of the following conditions are classified into the third tier: Condition 1: their percentile ranking falls between the second and third percentiles (e.g., between 40% and 80%); Condition 2: their lifecycle stage label belongs to the second predetermined stage set (i.e., the new customer phase or the decline phase). When a customer does not meet the conditions for the first and second tiers, but meets at least one of conditions 1 and 2, that customer is classified into the third tier. The fourth tier is defined as follows: the remaining customers who do not meet any of the above classification conditions are classified into the fourth tier. That is, after sequential screening through the first, second, and third tiers, customers who do not meet the conditions for the first, second, or third tier are all classified into the fourth tier.In specific application scenarios, different insurance companies can flexibly set the specific values ​​for each percentile based on their actual customer satisfaction distribution and business strategy needs. Through the hierarchical division of the above priority order, each customer in the total customer base is ultimately uniquely assigned to one of the first, second, third, or fourth tiers. There are no customers who are not assigned, nor are there customers who are repeatedly assigned to multiple tiers.

[0044] In one alternative implementation, after obtaining the customer satisfaction forecast value and before dividing each customer into multiple tiers, the customer value forecast value for each customer is first calculated based on the satisfaction forecast value, and then the tiers are divided based on the customer value forecast value. This process is implemented in the following way: For each customer, after obtaining their predicted satisfaction value, the renewal probability is first determined based on this value. Specifically, the predicted satisfaction value is used as input, and a predetermined satisfaction-renewal probability mapping function is used to calculate the corresponding renewal probability. This mapping function is a monotonically increasing function, and its basic property is: the higher the predicted satisfaction value, the higher the mapped renewal probability; the lower the predicted satisfaction value, the lower the mapped renewal probability. The specific form of this mapping relationship can be predetermined based on the statistical correlation between satisfaction scores and subsequent renewal behavior in the insurance company's historical data.

[0045] After obtaining the renewal probability value, substitute it into the projected value calculation formula to calculate the customer's projected customer value. The projected customer value calculation formula is: Projected Customer Value = Expected Premium Income over the Future Projected Time Period × Renewal Probability Value × Average Profit Margin of the Insurance Product Expected Service and Claims Costs. The expected premium income for a predetermined future period is determined as follows: All currently valid policies held by the customer are retrieved, and the total premium payable for each policy within the predetermined future period is calculated. Based on this basic method, the expected premium income can be further adjusted. One adjustment is to multiply the total premium payable by an expected increase in coverage growth rate coefficient. This coefficient is derived from time-series data of the customer's historical coverage increase behavior and reflects the potential premium increase from new insurance purchases by the customer within the predetermined future period. The average profit margin for each insurance type is determined as follows: The insurance types corresponding to all currently valid policies held by the customer are identified. The historical average profit margin for each insurance type in the insurance company's system is obtained. A weighted average is calculated based on the proportion of each insurance type's premium amount to the customer's total premium amount, yielding the customer's specific average profit margin for each insurance type. When the customer currently holds no valid policies, the average profit margin of all customers of the insurance company is used as the customer's average profit margin for each insurance type, or the typical profit margin corresponding to the customer's tier is used. The expected service and claims costs are determined as follows: Historical claims data and service data for the customer are obtained. Historical claims data includes the number of claims, claim amounts, and claim types made by the customer in the past period. Service data includes the number of inquiries, complaints, and service requests made by the customer. Based on this historical data, a pre-defined cost fitting model is used to estimate the sum of the customer's potential service costs and claims payment costs in the future within a predetermined time period. When a customer has no historical claims record, the average cost of customers at the same level is used as the customer's expected cost.

[0046] Furthermore, after calculating the predicted customer value for all customers, the predicted customer satisfaction value in the above-mentioned tiered classification method can be replaced with the predicted customer value value. Based on the predicted customer value value for each customer, the entire customer base can be divided into multiple tiers. Specifically, the percentile ranking of each customer's predicted customer value among all customers is obtained, as well as the current lifecycle stage label for each customer, which includes the new customer stage, growth stage, maturity stage, decline stage, and churn stage. The tiered classification is performed in the following priority order: customers whose percentile ranking falls within the top 1 percentile are classified as tier 1; customers whose percentile ranking falls between the first and second percentiles, or whose lifecycle stage label belongs to the first predetermined stage set, are classified as tier 2, which includes the growth stage and the new customer stage; customers whose percentile ranking falls between the second and third percentiles, or whose lifecycle stage label belongs to the second predetermined stage set, are classified as tier 3, which includes the new customer stage and the decline stage; the remaining customers who do not meet any of the above classification conditions are classified as tier 4.

[0047] In one optional implementation, determining the key bottleneck characteristic type of the customer group at the level based on the average contribution value with a negative average contribution direction in each customer characteristic type includes: When the level is not a preset reference level: The group feature vector of the level is compared with the group feature vector of the preset reference level using a feature-by-feature difference calculation to obtain a difference feature sequence, wherein the group feature vector is composed of the average contribution value of each customer feature type within the level; based on the difference feature sequence, customer feature types whose absolute difference value exceeds a predetermined threshold are marked as significant difference feature types of the level; customer feature types among the significant difference feature types whose average contribution direction is negative are identified as key bottleneck feature types of the customer group at the level. When the level is a preset reference level, the customer feature types whose average contribution direction is positive and whose average contribution value ranks in the top K positions in the group feature vector of the level are obtained as positive advantage feature types.

[0048] The group feature vector refers to the vector obtained by aggregating the feature contribution vectors of all customers belonging to the same level. Each element of this vector corresponds to the average contribution value of a feature type within that level, representing the overall performance of the customer group at that level in each feature dimension. The preset reference level refers to a level pre-specified among multiple levels as a comparison benchmark. The group feature vector of this level serves as a reference standard for measuring the performance of other levels. In this embodiment, the first level is specifically selected. The difference feature sequence refers to the difference sequence obtained by calculating the feature difference between the group feature vector of the level to be analyzed and the group feature vector of the reference level. Each difference in this sequence corresponds to the performance gap of a feature type between the two levels. A significantly different feature type refers to a feature type in the difference feature sequence whose absolute value of the difference exceeds a predetermined threshold. This type of feature represents a dimension where there is a substantial gap between the level to be analyzed and the reference level. A key bottleneck feature type refers to a feature type in the significantly different feature type whose average contribution direction is negative. This type of feature represents the core shortcomings of the customer group at the level to be analyzed compared to the customer group at the reference level in terms of satisfaction. Positive advantage feature type refers to the feature type in the group feature vector where the average contribution direction is positive and the average contribution value ranks high. This type of feature represents the core advantage dimension of the customer group at this level in terms of satisfaction.

[0049] In this optional embodiment, the process of determining the key bottleneck characteristic types of each customer group at each level based on the average contribution value with a negative average contribution direction is implemented in the following ways, taking the first level as a preset reference level as an example: For each level other than the preset reference level (i.e., the second, third, and fourth levels), the group feature vector of the level to be analyzed is compared with the group feature vector of the preset reference level using feature-by-feature difference calculation to obtain a difference feature sequence. The group feature vector is composed of the average contribution value of each customer feature type within that level. That is, for each feature type, the arithmetic mean of the contribution values ​​of that feature type among all customers within that level is calculated, and the average contribution values ​​of all feature types together form the group feature vector of that level. Feature-by-feature difference calculation refers to calculating the difference between the average contribution value of the level to be analyzed and the average contribution value of the reference level for that feature (the value of the level to be analyzed minus the value of the reference level) for each feature type. The differences corresponding to each feature are arranged according to the feature type to form a difference feature sequence. After obtaining the difference feature sequence, based on the absolute value of the difference corresponding to each feature in the difference feature sequence, customer feature types whose absolute value of the difference exceeds a predetermined threshold are marked as significant difference feature types for that level. Significant difference feature types represent a substantial difference between the level to be analyzed and the reference level in the corresponding feature dimension. After identifying the significantly different feature types, customer feature types with a negative average contribution direction are identified as the key bottleneck feature types for that customer level. Specifically, for each feature type marked as significantly different, the average contribution direction of that feature type within the analyzed level is examined. When the average contribution value of that feature type within the analyzed level is negative, it indicates that the feature type generally contributes negatively to the customer group at that level, meaning that this feature dimension lowers the predicted satisfaction value for customers at that level. These significantly different feature types with a negative average contribution direction are identified as the key bottleneck feature types for that customer level. Alternatively, all feature types that meet the condition of "significant difference and negative average contribution direction" can be output as key bottleneck feature types, or only a predetermined number (such as the top 3 or top 5) of feature types with the most significant negative degree (i.e., the smallest average contribution value, or the largest absolute value of the negative difference with the reference level) can be selected as key bottleneck feature types; this is not limited here. By identifying key bottleneck characteristics, the core shortcomings of each non-first-tier customer group can be quantitatively identified. The feature difference calculation reveals the specific gaps between the analyzed tier and the reference tier in each dimension. The negative screening of contribution direction ensures that the identified bottleneck characteristics have indeed lowered the satisfaction of customers at that tier. The combination of the two ensures the accuracy of bottleneck characteristic identification and its business guidance value.

[0050] When the level to be analyzed is the first level, since it has the highest predicted satisfaction value, it inherently lacks any negative bottleneck features requiring improvement. For the first level, the customer feature types with a positive average contribution direction and ranking in the top K of the average contribution values ​​in the group feature vector of this level are identified as positive advantage feature types. Specifically, all feature types with positive average contribution values ​​are selected from the group feature vector of the first level, and these feature types are sorted from largest to smallest according to their average contribution value. The top K feature types (e.g., K=3 or K=5) are selected as the positive advantage feature types of the first level. The identification and utilization of positive advantage feature types allows the core advantages of this level to be recognized, providing a differentiated direction for subsequent product recommendations for this level.

[0051] In one optional implementation, the step of matching the corresponding insurance product tag from the feature-policy mapping rule base according to the key constraint bottleneck feature type corresponding to each level includes: For each level that is not a preset reference level, based on the key constraint bottleneck feature type of the level, the insurance product tag corresponding to the key constraint bottleneck feature type is matched from the feature-policy mapping rule base; For the preset reference level, based on the positive advantage feature type, the insurance product label corresponding to the positive advantage feature type is matched from the feature-policy mapping rule base.

[0052] In this optional embodiment, the feature-strategy mapping rule base needs to be pre-built and dynamically maintained before matching insurance product tags. Specifically, historical marketing activity data from the insurance company is obtained. This data includes the characteristic data of each customer at the time of historical marketing outreach, the historical marketing strategies executed, and the corresponding conversion performance metrics. An association rule mining algorithm is used to perform frequent pattern mining on the historical marketing activity data, extracting association rules between multiple feature combinations and the optimal marketing strategy. An initial feature-strategy mapping rule base is built based on these association rules. Each rule in the feature-strategy mapping rule base includes a feature condition field and an insurance product tag field. The feature condition field stores a combination of one or more feature types and their corresponding value conditions, while the insurance product tag field stores the insurance product classification label associated with that feature condition. For example, one rule can be expressed as: when the feature condition is "number of supplementary insurance purchases = 0 and insurance type = single medical insurance", the associated insurance product tags are "supplementary medical insurance" and "supplementary accident insurance"; another rule can be expressed as: when the feature condition is "cumulative number of additional insurance purchases ≥ 3 and insurance type = annuity insurance", the associated insurance product tags are "increasing whole life insurance" and "pension insurance". For example, several implementation methods can be adopted for the dynamic maintenance of the feature-strategy mapping rule base. One implementation method is to acquire newly added historical marketing activity data at a predetermined period (e.g., monthly or quarterly), recalculate the support and confidence of each associated rule in the rule base, and remove or downgrade associated rules with support or confidence below a predetermined threshold to dynamically maintain the rule base. Another implementation method is to adopt an incremental update strategy, triggering rule updates when the amount of newly added historical data reaches a predetermined scale, rather than updating according to a fixed period, to balance the timeliness of the rule base and the consumption of computing resources.

[0053] For levels other than the first level (i.e., the second, third, and fourth levels), based on the key constraint bottleneck feature type of that level, insurance product tags corresponding to that key constraint bottleneck feature type are matched from the feature-policy mapping rule base. Rules containing that key constraint bottleneck feature type as a feature condition are searched from the feature-policy mapping rule base, and the insurance product tags associated with the searched rules are output as matching results. For example, when the key constraint bottleneck feature type of a certain level is "quantity of supplementary insurance purchased," all rules with "quantity of supplementary insurance purchased" as a feature condition are searched from the rule base, and the corresponding insurance product tags (such as "supplementary medical insurance," "supplementary accident insurance," and "supplementary critical illness insurance") are extracted and used as matching results for that level. When there are multiple key constraint bottleneck feature types, multiple bottleneck feature types are combined as query conditions, and rules containing multiple bottleneck feature types as feature conditions are matched from the rule base. The insurance product tags associated with the matched rules are used as matching results. This combined matching method improves matching accuracy and avoids recommendation bias caused by matching only a single feature type. For example, when the key constraint bottleneck feature type at a certain level is both "number of supplementary insurance purchases" and "single type of insurance", searching the rule base for rules that contain both of these features will result in matching insurance product tags such as "critical illness insurance upgrade plan" and "medical insurance supplementary package", rather than "supplementary medical insurance" generated when matching a single feature.

[0054] For the first tier, since it lacks negative bottleneck features, bottleneck feature-based matching is not performed. Instead, based on the positive advantage feature types of this tier, insurance product tags corresponding to the positive advantage feature types are matched from the feature-policy mapping rule base. Rules containing the positive advantage feature type as a feature condition are searched from the feature-policy mapping rule base, and the insurance product tags associated with the searched rules are used as the matching results. Since the feature types with a positive average contribution direction in the feature contribution vector of first-tier customers represent the core advantage dimensions of this customer group, the insurance product tags matched based on these advantage features are used to further meet the core needs of this customer level. For example, when the positive advantage feature type of the first tier is "number of annuity insurance holdings," the matched insurance product tags might be "increasing whole life insurance," "pension insurance," or "wealth transfer plan," etc. These products aim to further explore the in-depth insurance needs of highly satisfied customers.

[0055] Step 103: Based on the key constraint bottleneck feature types corresponding to each level, match the corresponding insurance product tags from the feature-policy mapping rule base, and filter the set of candidate insurance products that match the insurance product tags from the insurance product database.

[0056] The feature-strategy mapping rule base refers to a pre-built set of association rules. Each rule in this set defines the correspondence between key bottleneck feature types and one or more insurance product tags. This correspondence is formed based on the experience of insurance business experts or historical successful cases. Insurance product tags are standardized labels used to describe the core attributes of insurance products, including but not limited to product type tags (such as critical illness insurance, medical insurance, accident insurance, annuity insurance), coverage period tags (such as one-year, long-term, lifetime), payment method tags (such as lump sum payment, installment payment), additional service tags (such as green channel, overseas medical treatment), and applicable population tags (such as exclusive for the elderly, exclusive for children, family-shared). The candidate insurance product set refers to the set of all products initially selected from the insurance product database that match the target product tags, serving as the final screening range for subsequent recommendations.

[0057] Specifically, firstly, based on the key constraint bottleneck feature types corresponding to each level, a pre-built feature-strategy mapping rule base is accessed. This rule base is stored in key-value pair format, where the key is the feature type or combination of feature types, and the value is the corresponding set of product tags. The key constraint bottleneck feature type of the current level is input into the rule base as the query condition, and a matching operation is performed. The matching method can be exact matching (i.e., the key constraint bottleneck feature type is completely consistent with the feature type in the rule base) or fuzzy matching (i.e., matching is based on the semantic similarity or business relevance of the feature types, for example, "premium sensitivity" and "rate high sensitivity" are considered semantically similar feature types). If the match is successful, one or more insurance product tags corresponding to the key constraint bottleneck feature type are obtained; if the match fails (i.e., the mapping relationship of the feature type is not defined in the rule base), a fallback strategy is adopted, and at least one secondary negative feature type of the level is combined and queried again, or a preset general product tag is directly selected as a candidate. After obtaining the insurance product tag, all insurance products with the tag are filtered from the insurance product database using the tag as the search condition, forming a candidate insurance product set. It should be noted that each product in the insurance product database is pre-labeled with one or more insurance product tags, and the screening process is the tag matching process; if a product has multiple matching tags, it still belongs to the candidate set.

[0058] Through the mapping rule base matching and product screening in this step, the transformation process from bottleneck feature analysis to specific product pools has a structured and configurable execution path. The introduction of the rule base enables business experts' experience and knowledge to be solidified in the form of explicit rules and reused in the recommendation process, avoiding manual intervention each time re-analysis is performed.

[0059] Step 104: Based on each customer's historical insurance records and the set of candidate insurance products corresponding to each customer's level, determine the insurance product recommendation results for each customer.

[0060] Historical insurance records refer to a list of insurance products that a customer has purchased or held at a past point in time, along with related information. This list includes at least the product identifier, application date, policy status (active, expired, or cancelled), and insured information. Insurance product recommendation results refer to a final list of recommendations containing one or more insurance products for a single customer. These products are displayed or pushed to the customer for selection.

[0061] Specifically, the process begins by obtaining each customer's historical insurance records, which can be extracted from a customer profile database or policy management system. Simultaneously, a set of candidate insurance products corresponding to the customer's current insurance tier is obtained. These candidate product sets are then compared and integrated with the customer's historical insurance records to determine the final insurance product recommendation. The specific determination methods include at least the following implementation approaches: One approach is to remove products already purchased by the customer and currently in effect from the candidate product set to avoid duplicate recommendations of the same product; the remaining products are then sorted according to their matching degree with key constraint bottleneck characteristics and output as the recommendation result. Another approach is to prioritize products from the candidate product set that complement (rather than duplicate) the coverage of the customer's historical insurance products. For example, if the customer already has critical illness insurance, medical insurance or accident insurance is prioritized to form a protection combination recommendation. A further approach is to select similar products from the candidate product set whose coverage amount or scope differs from the customer's existing products as the recommendation result, or to recommend product upgrade / renewal plans to the customer. In determining the recommendation results, a limit on the number of recommendations can be introduced, for example, pre-setting a maximum of 3 to 5 products per customer. Optionally, a conflict detection step can be set before outputting the recommendation results to check whether there is overlap in coverage or mutual exclusion between multiple products in the recommendation results. If a conflict exists, the order is adjusted or conflicting products are removed to ensure the logical consistency within the recommendation results. This ensures that the final recommendation results not only consider the common bottlenecks of the customer's group but also take into account the customer's individual insurance history, avoiding the recommendation of products that the customer already owns or that are not suitable, thereby improving the personalization, practicality, and accuracy of the recommendations.

[0062] By applying the technical solution of this embodiment, customer characteristics of multiple customers are acquired and input into a trained gradient boosting decision tree model. The model then infers and outputs the predicted satisfaction value of each customer within a predetermined future time period. Simultaneously, attribution analysis is used to calculate the contribution value and direction of each customer characteristic type to the predicted satisfaction value of each customer, generating a feature contribution vector for each customer. Based on this, all customers are divided into multiple levels according to their predicted satisfaction values. For each level, the feature contribution vectors of all customers within that level are aggregated to calculate the average contribution value and average contribution direction of each customer characteristic type within that level. Then, based on the average contribution value with a negative average contribution direction, the key bottleneck characteristic type of the customer group at that level is determined. Subsequently, based on the key bottleneck characteristic type corresponding to each level, corresponding insurance product tags are matched from the feature-policy mapping rule base, and a set of matching candidate insurance products is selected from the insurance product database. Finally, based on each customer's historical insurance records and the set of candidate insurance products corresponding to each customer's level, the insurance product recommendation result for each customer is determined. This application achieves interpretability of satisfaction prediction results by combining gradient boosting decision trees with attribution analysis, solving the problem that black-box models have difficulty in explaining the basis of predictions. Through customer segmentation and group feature aggregation, it achieves quantitative identification and statistically stable output of the core shortcomings of customers at each level. Through the matching mechanism of feature-policy mapping rule base, bottleneck features are accurately transformed into product recommendation directions. Through gap-driven scoring and ranking, it ensures that the recommendation results accurately match the customer's existing protection status and avoids homogeneous and duplicate recommendations.

[0063] Furthermore, as a refinement and extension of the specific implementation of the above embodiments, and to fully illustrate the specific implementation process of this embodiment, another method for recommending insurance products is provided, such as... Figure 2 As shown, the method includes: Step 201: Obtain the first insurance category and first coverage amount of the customer's currently valid insurance policy, and obtain the second insurance category and second coverage amount of each candidate insurance product in the candidate insurance product set corresponding to the customer's level; Step 202: Compare the insurance category of each candidate insurance product with the insurance category of the valid policies. If no valid policy has the same insurance category as the candidate insurance product, mark the candidate insurance product with the highest recommended score. If a valid policy has the same insurance category as the candidate insurance product, obtain the existing coverage amount of that policy, calculate the difference between the coverage amount of the candidate insurance product and the existing coverage amount. If the difference is greater than zero, determine the recommended score of the candidate insurance product based on the magnitude of the difference; if the difference is less than or equal to zero, remove the candidate insurance product. Step 203: Based on the recommendation scores of each candidate insurance product, rank the candidate insurance products and select at least one candidate insurance product with the highest ranking as the insurance product recommendation result for the customer.

[0064] In this embodiment, firstly, the insurance types and coverage amounts of all currently valid insurance policies held by the customer are obtained. Valid policies are defined as follows: the policy's coverage period has not expired, the customer has paid the premiums due as stipulated in the contract, the customer has not voluntarily requested to cancel the policy, and the insurance company has not deemed the policy invalid for specific reasons (such as failure to disclose information truthfully, failure to pay premiums on time exceeding the grace period, etc.). Policies meeting all of the above conditions are included in the customer's set of valid policies. For each valid policy in this set, its corresponding insurance type and coverage amount are extracted. The insurance type can be extracted from the policy's insurance code or product classification field, and the coverage amount can be directly read from the policy's sum insured field. Simultaneously, the insurance types and coverage amounts of each candidate insurance product in the candidate insurance product set corresponding to the customer's level are obtained. The candidate insurance product set is generated by the aforementioned steps, and each candidate insurance product in this set has been screened through level division and bottleneck feature matching. For each product in the candidate insurance product set, its corresponding insurance category tag and coverage amount field are extracted from the insurance product database. The insurance category tag is pre-labeled according to the insurance company's product classification system when the product is entered into the database, and the coverage amount field can be directly read from the product information.

[0065] Furthermore, the insurance category of each candidate insurance product is compared with the insurance category of the customer's currently held valid policies. If none of the customer's currently held valid policies have the same insurance category as the candidate insurance product, it is determined that the customer has a complete gap in that insurance category. For example, if the candidate insurance product is an accident insurance product, and none of the customer's currently held valid policies belong to the accident insurance category, then the customer is completely uninsured in the accident insurance dimension. In this case, the candidate insurance product is marked with the highest recommendation score. Because this type of product can fill the customer's complete gap in a certain protection dimension, its recommendation priority should be higher than other types of candidate products. When the customer's currently held valid policies have the same insurance category as the candidate insurance product, the existing coverage amount of that policy is obtained, and the difference between the coverage amount of the candidate insurance product and the existing coverage amount is calculated (i.e., the coverage amount of the candidate product minus the coverage amount of the customer's existing policies). If the difference is greater than zero, it is determined that the customer has a coverage gap in that insurance category, meaning that the customer's existing coverage is insufficient to cover the coverage level provided by the candidate product. A positive difference exists between the two, and this difference is the coverage gap. The recommendation score for the candidate insurance product is determined based on the magnitude of this difference. For example, the difference is input as an independent variable into a predefined linear or nonlinear function to calculate a continuous recommendation score; the larger the difference, the higher the output score. When the customer currently holds a valid policy in the same insurance category as the candidate insurance product, and the existing coverage of that policy is greater than or equal to the coverage of the candidate insurance product, it is determined that the customer has no coverage gap in that insurance category. That is, the customer's existing coverage level in that category has reached or exceeded the coverage level provided by the candidate product, and the customer does not need to supplement their coverage through the candidate product. In this case, the candidate insurance product is removed from the recommendation results and will no longer participate in subsequent ranking and recommendation.

[0066] After comparing and scoring all candidate insurance products, a recommendation score is obtained for each candidate product (products that are eliminated are not included in the subsequent ranking). Based on the recommendation scores of each candidate insurance product, the candidate insurance products are ranked in descending order of recommendation score, and at least one candidate insurance product with the highest ranking is output as the insurance product recommendation result for that customer.

[0067] By applying the technical solution of this embodiment, insurance product recommendations are made capable of personalized matching based on the customer's existing coverage status. Customers with complete coverage gaps receive the highest score to prioritize coverage of the missing coverage dimensions. The larger the coverage gap, the higher the recommendation score to achieve on-demand configuration. Similar products with sufficient coverage are automatically removed to avoid duplicate recommendations.

[0068] Furthermore, as Figure 1 In terms of specific implementation of the method, this application provides an insurance product recommendation device, such as... Figure 3 As shown, the device includes: The prediction module is used to acquire customer features of multiple customers, input each customer's customer features into the trained gradient boosting decision tree model, infer and output the predicted satisfaction value of each customer in the future within a predetermined time period, and use the attribution analysis method to calculate the contribution value and direction of each customer feature type to the predicted satisfaction value of each customer, generating the feature contribution vector of each customer. The segmentation module is used to divide each customer into multiple levels based on the predicted satisfaction value of each customer, and for each level, aggregate the feature contribution vectors of customers belonging to the level, calculate the average contribution value and average contribution direction of each customer feature type within the level, and determine the key constraint bottleneck feature type of the customer group in the level based on the average contribution value with a negative average contribution direction. The filtering module is used to match the corresponding insurance product tags from the feature-policy mapping rule base according to the key constraint bottleneck feature types corresponding to each level, and to filter the set of candidate insurance products that match the insurance product tags from the insurance product database. The recommendation module is used to determine the recommended insurance products for each customer based on their historical insurance records and the set of candidate insurance products corresponding to their respective insurance tier.

[0069] In an optional embodiment, the prediction module is configured to train a gradient boosting decision tree model in the following manner: Acquire historical customer sample data, and divide the historical customer sample data into a training sample set and a validation sample set. The historical customer sample data includes historical customer characteristics and corresponding satisfaction tags. A gradient boosting decision tree model is trained using a gradient boosting decision algorithm. Before training begins, each continuous feature in the historical customer features is binned, and multiple bin boundaries for each continuous feature are determined. The values ​​of each continuous feature are mapped to their corresponding discrete bin indices. At the start of each iteration, a new decision tree is initialized as the current tree, and the maximum number of leaf nodes and / or maximum depth parameters of the current tree are set. During the growth of the current tree, the following leaf node splitting steps are repeated until the current tree reaches the maximum number of leaf nodes or the maximum depth: all leaf nodes in the current tree are obtained; for each leaf node, a histogram is constructed based on the bin indices of the feature values ​​of the training samples falling into that leaf node; the histogram is traversed to search for candidate split points that maximize the splitting gain, and the candidate split points and their corresponding splitting gains are recorded; after traversal, the target leaf node with the maximum splitting gain is selected from all leaf nodes, and the splitting operation is performed according to the candidate split points corresponding to the target leaf node to generate two new child leaf nodes. After each iteration, the current model is used to predict the validation sample set and calculate the preset evaluation index value. When the improvement of the current evaluation index value relative to the historical best evaluation index value is less than the predetermined minimum improvement threshold, the counter for consecutive unimproved rounds is incremented by 1. When the value of the counter reaches the predetermined maximum tolerable rounds, the iteration stops and the model parameters are rolled back to the iteration round corresponding to the historical best evaluation index value. The model parameters corresponding to the iteration round with the best performance on the validation sample set are used as the trained gradient boosting decision tree model.

[0070] In an optional embodiment, the prediction module is further configured to determine the satisfaction label according to the following manner: The first satisfaction score is determined based on the recommendation ratings given by customers after historical service interactions, wherein the recommendation rating is a quantitative rating given by customers within a predetermined rating range; A second satisfaction score is determined based on the historical complaint records and service handling records of historical customers, and the number of complaints and / or complaint resolution rate are mapped to the satisfaction score. The third satisfaction score is determined based on the customer's historical renewal behavior data. The third satisfaction score is the first value when the historical customer completes the renewal within the predetermined time after the policy expires, and the second value otherwise. The fourth satisfaction rating is determined based on the number of times insurance was added and the premium growth rate in the historical consumption behavior data of customers. The satisfaction label is determined based on the first satisfaction score, the second satisfaction score, the third satisfaction score, and the fourth satisfaction score.

[0071] In an optional embodiment, the partitioning module is specifically used for: Obtain the percentile ranking of each customer's predicted satisfaction value among all customers, and obtain the current life cycle stage label of each customer, which includes the new customer stage, growth stage, maturity stage, decline stage, and churn stage. The hierarchical division will be executed in the following priority order: Those whose percentile ranking falls within the top 100th percentile are classified into the first tier; Customers whose percentile ranking falls between the first percentile and the second percentile, or whose lifecycle stage tag belongs to the first pre-booking stage set, are classified into the second tier. The first pre-booking stage set includes the growth stage and the new customer stage. Customers whose percentile ranking falls between the second percentile and the third percentile, or whose lifecycle stage label belongs to the second pre-sales stage set, are classified into the third tier. The second pre-sales stage set includes the new customer period and the decline period. The remaining customers who do not meet any of the above classification criteria will be classified into the fourth tier.

[0072] In an optional embodiment, the partitioning module is further configured to: When the level is not a preset reference level: The group feature vector of the level is compared with the group feature vector of the preset reference level using a feature-by-feature difference calculation to obtain a difference feature sequence, wherein the group feature vector is composed of the average contribution value of each customer feature type within the level; based on the difference feature sequence, customer feature types whose absolute difference value exceeds a predetermined threshold are marked as significant difference feature types of the level; customer feature types among the significant difference feature types whose average contribution direction is negative are identified as key bottleneck feature types of the customer group at the level. When the level is a preset reference level, the customer feature types whose average contribution direction is positive and whose average contribution value ranks in the top K positions in the group feature vector of the level are obtained as positive advantage feature types.

[0073] In an optional embodiment, the filtering module is specifically used for: For each level that is not a preset reference level, based on the key constraint bottleneck feature type of the level, the insurance product tag corresponding to the key constraint bottleneck feature type is matched from the feature-policy mapping rule base; For the preset reference level, based on the positive advantage feature type, the insurance product label corresponding to the positive advantage feature type is matched from the feature-policy mapping rule base.

[0074] In an optional embodiment, the recommendation module is specifically used for: Obtain the first insurance category and first coverage amount of the customer's currently valid insurance policy, and obtain the second insurance category and second coverage amount of each candidate insurance product in the candidate insurance product set corresponding to the customer's level; Each candidate insurance product's insurance category is compared with the insurance category of the valid policies. If no valid policy has the same insurance category as the candidate product, the candidate product is marked with the highest recommended score. If a valid policy has the same insurance category as the candidate product, the existing coverage amount of that policy is obtained. The difference between the candidate product's coverage amount and the existing coverage amount is calculated. If the difference is greater than zero, the recommended score for the candidate product is determined based on the magnitude of the difference. If the difference is less than or equal to zero, the candidate product is eliminated. Based on the recommendation scores of each candidate insurance product, the candidate insurance products are ranked, and at least one candidate insurance product with the highest ranking is selected as the insurance product recommendation result for the customer.

[0075] It should be noted that other corresponding descriptions of the functional units involved in the insurance product recommendation device provided in this application embodiment can be found by referring to... Figures 1 to 2 The corresponding descriptions in the method will not be repeated here.

[0076] This application also provides a computer device, which may specifically be a personal computer, a server, a network device, etc. Figure 4 As shown, the computer device includes a bus, a processor, memory, and a communication interface, and may also include an input / output interface and a display device. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores location information. The network interface allows communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the various method embodiments.

[0077] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0078] In one embodiment, a computer-readable storage medium is provided, which may be non-volatile or volatile, having stored thereon a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0079] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0080] It should be noted that the user personal information involved in the embodiments of this application is all authorized (with the knowledge and consent) by the relevant parties or fully authorized by all parties, and the executing entity can obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with the relevant laws and regulations of the relevant countries and regions, and do not violate public order and good morals. It should be noted that if any software tools or components other than those of this company appear in the embodiments of this application, they are merely illustrative examples and do not represent actual use.

[0081] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, graphics processors, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0082] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0083] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for recommending insurance products, characterized in that, The method includes: The system acquires customer features from multiple customers, inputs each customer's features into the trained gradient boosting decision tree model, and infers and outputs the predicted satisfaction value of each customer within a predetermined time period. Attribution analysis is then used to calculate the contribution value and direction of each customer feature type to the predicted satisfaction value of each customer, generating a feature contribution vector for each customer. Based on the predicted satisfaction values ​​of each customer, each customer is divided into multiple levels. For each level, the feature contribution vectors of customers belonging to that level are aggregated. The average contribution value and average contribution direction of each customer feature type within that level are calculated. Based on the average contribution value with a negative average contribution direction, the key bottleneck feature type of the customer group in that level is determined. Based on the key constraint bottleneck feature types corresponding to each level, the corresponding insurance product tags are matched from the feature-policy mapping rule base, and a set of candidate insurance products matching the insurance product tags is selected from the insurance product database; Based on each customer's historical insurance records and the set of candidate insurance products corresponding to each customer's level, the recommended insurance products for each customer are determined.

2. The method according to claim 1, characterized in that, The training methods for the gradient boosting decision tree model include: Acquire historical customer sample data, and divide the historical customer sample data into a training sample set and a validation sample set. The historical customer sample data includes historical customer characteristics and corresponding satisfaction tags. A gradient boosting decision tree model is trained using a gradient boosting decision algorithm. Before training begins, each continuous feature in the historical customer features is binned, and multiple bin boundaries for each continuous feature are determined. The values ​​of each continuous feature are mapped to their corresponding discrete bin indices. At the start of each iteration, a new decision tree is initialized as the current tree, and the maximum number of leaf nodes and / or maximum depth parameters of the current tree are set. During the growth of the current tree, the following leaf node splitting steps are repeated until the current tree reaches the maximum number of leaf nodes or the maximum depth: all leaf nodes in the current tree are obtained; for each leaf node, a histogram is constructed based on the bin indices of the feature values ​​of the training samples falling into that leaf node; the histogram is traversed to search for candidate split points that maximize the splitting gain, and the candidate split points and their corresponding splitting gains are recorded; after traversal, the target leaf node with the maximum splitting gain is selected from all leaf nodes, and the splitting operation is performed according to the candidate split points corresponding to the target leaf node to generate two new child leaf nodes. After each iteration, the current model is used to predict the validation sample set and calculate the preset evaluation index value. When the improvement of the current evaluation index value relative to the historical best evaluation index value is less than the predetermined minimum improvement threshold, the counter for consecutive unimproved rounds is incremented by 1. When the value of the counter reaches the predetermined maximum tolerable rounds, the iteration stops and the model parameters are rolled back to the iteration round corresponding to the historical best evaluation index value. The model parameters corresponding to the iteration round with the best performance on the validation sample set are used as the trained gradient boosting decision tree model.

3. The method according to claim 2, characterized in that, The satisfaction rating is determined according to the following method: The first satisfaction score is determined based on the recommendation ratings given by customers after historical service interactions, wherein the recommendation rating is a quantitative rating given by customers within a predetermined rating range; A second satisfaction score is determined based on the historical complaint records and service handling records of historical customers, and the number of complaints and / or complaint resolution rate are mapped to the satisfaction score. The third satisfaction score is determined based on the customer's historical renewal behavior data. The third satisfaction score is the first value when the historical customer completes the renewal within the predetermined time after the policy expires, and the second value otherwise. The fourth satisfaction rating is determined based on the number of times insurance was added and the premium growth rate in the historical consumption behavior data of customers. The satisfaction label is determined based on the first satisfaction score, the second satisfaction score, the third satisfaction score, and the fourth satisfaction score.

4. The method according to claim 1, characterized in that, Based on the predicted satisfaction values ​​of each customer, customers are divided into multiple levels, including: Obtain the percentile ranking of each customer's predicted satisfaction value among all customers, and obtain the current life cycle stage label of each customer, which includes the new customer stage, growth stage, maturity stage, decline stage, and churn stage. The hierarchical division will be executed in the following priority order: Those whose percentile ranking falls within the top 100th percentile are classified into the first tier; Customers whose percentile ranking falls between the first percentile and the second percentile, or whose lifecycle stage tag belongs to the first pre-booking stage set, are classified into the second tier. The first pre-booking stage set includes the growth stage and the new customer stage. Customers whose percentile ranking falls between the second percentile and the third percentile, or whose lifecycle stage label belongs to the second pre-sales stage set, are classified into the third tier. The second pre-sales stage set includes the new customer period and the decline period. The remaining customers who do not meet any of the above classification criteria will be classified into the fourth tier.

5. The method according to claim 4, characterized in that, The step of determining the key bottleneck characteristic types of the customer group at each level based on the average contribution value with a negative average contribution direction in each customer characteristic type includes: When the level is not a preset reference level: The group feature vector of the level is compared with the group feature vector of the preset reference level using a feature-by-feature difference calculation to obtain a difference feature sequence, wherein the group feature vector is composed of the average contribution value of each customer feature type within the level; based on the difference feature sequence, customer feature types whose absolute difference value exceeds a predetermined threshold are marked as significant difference feature types of the level; customer feature types among the significant difference feature types whose average contribution direction is negative are identified as key bottleneck feature types of the customer group at the level. When the level is a preset reference level, the customer feature types whose average contribution direction is positive and whose average contribution value ranks in the top K positions in the group feature vector of the level are obtained as positive advantage feature types.

6. The method according to claim 5, characterized in that, The step of matching corresponding insurance product tags from the feature-policy mapping rule base based on the key constraint bottleneck feature types corresponding to each level includes: For each level that is not a preset reference level, based on the key constraint bottleneck feature type of the level, the insurance product tag corresponding to the key constraint bottleneck feature type is matched from the feature-policy mapping rule base; For the preset reference level, based on the positive advantage feature type, the insurance product label corresponding to the positive advantage feature type is matched from the feature-policy mapping rule base.

7. The method according to claim 1, characterized in that, The process of determining the insurance product recommendation result for any given customer includes: Obtain the first insurance category and first coverage amount of the customer's currently valid insurance policy, and obtain the second insurance category and second coverage amount of each candidate insurance product in the candidate insurance product set corresponding to the customer's level; Each candidate insurance product's insurance category is compared with the insurance category of the valid policies. If no valid policy has the same insurance category as the candidate product, the candidate product is marked with the highest recommended score. If a valid policy has the same insurance category as the candidate product, the existing coverage amount of that policy is obtained. The difference between the candidate product's coverage amount and the existing coverage amount is calculated. If the difference is greater than zero, the recommended score for the candidate product is determined based on the magnitude of the difference. If the difference is less than or equal to zero, the candidate product is eliminated. Based on the recommendation scores of each candidate insurance product, the candidate insurance products are ranked, and at least one candidate insurance product with the highest ranking is selected as the insurance product recommendation result for the customer.

8. An insurance product recommendation device, characterized in that, The device includes: The prediction module is used to acquire customer features of multiple customers, input each customer's customer features into the trained gradient boosting decision tree model, infer and output the predicted satisfaction value of each customer in the future within a predetermined time period, and use the attribution analysis method to calculate the contribution value and direction of each customer feature type to the predicted satisfaction value of each customer, generating the feature contribution vector of each customer. The segmentation module is used to divide each customer into multiple levels based on the predicted satisfaction value of each customer, and for each level, aggregate the feature contribution vectors of customers belonging to the level, calculate the average contribution value and average contribution direction of each customer feature type within the level, and determine the key constraint bottleneck feature type of the customer group in the level based on the average contribution value with a negative average contribution direction. The filtering module is used to match the corresponding insurance product tags from the feature-policy mapping rule base according to the key constraint bottleneck feature types corresponding to each level, and to filter the set of candidate insurance products that match the insurance product tags from the insurance product database. The recommendation module is used to determine the recommended insurance products for each customer based on their historical insurance records and the set of candidate insurance products corresponding to their respective insurance tier.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

10. A computer device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

Images