Personalized insurance product customization methods and risk pricing methods based on customer profiles

Abstract

This application belongs to the field of intelligent pricing technology for insurance products, involving personalized insurance product customization methods and risk pricing methods based on customer profiles. By collecting basic customer information, social network data, insurance history data, and occupational information, a rule-based recommendation engine recommends relevant insurance products based on the customer's basic information and historical data. Graph neural networks and social network data are introduced to further filter insurance products, assigning a level to each product. Corresponding customer data is selected based on the characteristics of different insurance products, and risk assessment models are trained and applied separately. The price weighting factors of the products are adjusted based on the assessment results. Finally, combining the benchmark price corresponding to each product level and the adjusted price weighting factors, the final personalized product price is output, ensuring that each customer receives a tailor-made insurance product at a reasonable price. This makes product recommendations more accurate and pricing more flexible.

Description

Technical Field

[0001] This application belongs to the field of intelligent pricing technology for insurance products, and more specifically, it relates to personalized insurance product customization methods and risk pricing methods based on customer profiles. Background Technology

[0002] With the development of technology, especially the application of big data and artificial intelligence, the traditional insurance industry is undergoing a profound transformation. Traditional insurance products typically employ a fixed, standardized pricing model, meaning customers can only choose from existing products on the market, rather than customizing them according to their specific needs or characteristics. This "one-size-fits-all" approach not only limits the diversity of customer choices but also makes it difficult for insurance companies to differentiate their products in a highly competitive market.

[0003] In the traditional process of insurance product customization and pricing, product design and pricing often rely on experience and market research. Insurance products are typically designed according to product categories (such as life insurance, car insurance, health insurance, etc.), and each product has a fixed scope of coverage and coverage amount. Pricing is set based on the standard characteristics of these products. Many insurance companies use simple rule engines to recommend products based on basic customer information, but due to the limitations of the rules, the recommended products often differ significantly from the customer's actual needs and are difficult to meet personalized requirements. Summary of the Invention

[0004] This invention provides a method for customizing personalized insurance products and a risk pricing method based on customer profiles, aiming to solve the problem that current pricing differs greatly from customers' actual needs and is difficult to meet personalized requirements.

[0005] Personalized insurance product customization methods and customer profile-based risk pricing methods include the following steps:

[0006] Step 1: Collect basic customer information, social network data, insurance history data, and occupational information, and preprocess the collected data;

[0007] Step 2: Use a rule-based recommendation engine to recommend relevant insurance products based on preprocessed basic customer information and insurance history data. Then, filter insurance products based on the recommended insurance products, graph neural network and preprocessed social network data, and determine the level of each insurance product.

[0008] Step 3: Select corresponding customer characteristics to segment the data according to different insurance products, train and apply the corresponding risk assessment models for different insurance products, and adjust the price weight factor of each insurance product based on the results of the risk assessment models; the customer characteristics are preprocessed customer basic information, insurance history data and occupation information;

[0009] Step 4: Output the final product price based on the benchmark price corresponding to each insurance product level and the price weighting factor of each insurance product.

[0010] This invention collects and preprocesses basic customer information, social network data, insurance history data, and occupational information to lay a data foundation for subsequent product customization and pricing. A rule-based recommendation engine recommends relevant insurance products based on customer information and historical data. Graph Neural Networks (GNNs) and social network data are used to further filter insurance products and assign a level to each product, ensuring that the recommended products better match customer needs. By selecting corresponding customer data based on the characteristics of different insurance products, risk assessment models are trained and applied to accurately assess customer risk. The price weighting factors of the products are adjusted based on the assessment results, achieving personalized price adjustments. Finally, by combining the base price corresponding to each product level and the adjusted price weighting factors, the final personalized product price is output, ensuring that each customer receives a tailor-made insurance product at a reasonable price. By comprehensively utilizing multi-dimensional data sources and combining graph neural networks and deep learning technologies, product recommendations are more accurate and pricing is more flexible, fundamentally solving the problem that traditional insurance pricing cannot accurately reflect customer risk characteristics and promoting the personalized development of the insurance industry.

[0011] Preferably, step 2 includes the following steps:

[0012] Preliminary matching of insurance products: Based on the pre-processed basic customer information, a rule-based recommendation engine is used to initially screen insurance products that match the customer's needs;

[0013] Extracting product recommendation levels: Based on the constructed GNN model and customer social network data, insurance products are filtered out to obtain a list of filtered insurance products;

[0014] Insurance product rating determination: The appropriate insurance product rating is selected based on the customer's pre-processed insurance history.

[0015] Preferably, the specific steps for screening insurance products are as follows:

[0016] Graph-structured representation of social network data:

[0017] Social network data is transformed into a graph structure, which includes nodes, edges, and edge weights. Each customer represents a node in the graph, and the characteristics of the node include the customer's basic information and insurance history data. Each edge represents the relationship between two nodes in the social network. The edge weight is used to define the strength of the relationship between two nodes.

[0018] Graph neural networks perform message passing and node updates:

[0019] Each node i receives a message from its neighbor node j. The message contains the feature vector and edge weights of node j.

[0020] m ij =w ij ·h j ;

[0021] Where: m ij This represents the message received by node i from node j; w ij h represents the weight of the edge. j The feature vector of node j represents the basic information and insurance history data of customer j.

[0022] Each node aggregates all messages from its neighboring nodes to obtain the aggregated information for node i:

[0023]

[0024] In the formula: h represents the set of neighboring nodes of node i; i ′ This represents the aggregation information of node i;

[0025] The features of each node are updated based on the received aggregated message:

[0026]

[0027] In the formula: This represents the feature vector of node i at iteration t+1; W represents the feature vector of node i at iteration t; σ represents the Sigmoid activation function; (t) Let represent the learnable weight matrix for round t; Indicates feature concatenation operation;

[0028] Iteration and Termination:

[0029] In the graph structure described above, a node includes two layers of neighbors. The graph neural network performs two rounds of iteration. In the first round of iteration, message passing and node updates are performed on the first layer of neighbors. In the second round of iteration, message passing and node updates are performed on the second layer of neighbors. Finally, the feature representation of the node is output.

[0030] Product interest rating calculation:

[0031] The feature representation of the output node after the last iteration Includes customer social network information and personal characteristics; calculates an interest score for each customer i and each insurance product p:

[0032]

[0033] In the formula h represents the feature vector of customer i after the last iteration; p Let represent the feature vector of insurance product p; ‖·‖ represents the L2 norm of the vector;

[0034] A list of recommended insurance products is generated based on interest scores and set thresholds.

[0035] Preferably, the steps for determining the weight of the edge are as follows:

[0036] Social relationship similarity calculation:

[0037]

[0038] In the formula: This represents the social feature vector of customer i; Let represent the social feature vector of customer j; Let L2 norm represent the social feature vector of customer i; Let L2 norm represent the social feature vector of customer j; Indicates the similarity of social relationships;

[0039] Behavioral similarity calculation:

[0040]

[0041] In the formula: b i b represents the behavioral feature vector of customer i; j Represents the behavioral feature vector of customer j; ||b i || represents the L2 norm of the behavioral feature vector of customer i; || b j || represents the L2 norm of the behavioral feature vector of customer j;

[0042] Weight calculation of the combined edges:

[0043]

[0044] In the formula: α and β represent weight coefficients; after calculating the weights of the composite edge, the weights are normalized.

[0045] Preferably, the specific steps for determining the insurance product level are as follows:

[0046] Iterate through the customer's historical insurance records to find the level of each insurance product and count the frequency of each product level; select the level with the highest frequency as the insurance product level to be determined this time.

[0047] Preferably, the risk assessment model includes an input layer, a feature selection and attention mechanism layer, a task feature extraction layer, and a task-specific layer;

[0048] The input layer is used to receive preprocessed basic customer information, insurance history data, and occupational information.

[0049] The number of feature selection and attention mechanism layers is adapted to the number of insurance products. Each feature selection and attention mechanism layer includes feature selection and self-attention mechanism. Feature selection selects features related to the corresponding insurance product according to the needs of the insurance product. The self-attention mechanism assigns a weight to each feature to obtain a weighted feature.

[0050] The number of task feature extraction layers is adapted to the number of insurance products. Different fully connected layers are used in the task feature extraction layers for different insurance products to mine weighted features and obtain the deep features of the corresponding insurance products.

[0051] The number of task-specific layers is adapted to the number of insurance products. Each task-specific layer makes predictions based on the deep features output by the task feature extraction layer, corresponding to the price weight factor of the insurance product.

[0052] Preferably, the specific steps for feature selection are as follows:

[0053] Correlation quantification: Calculate f for each feature i Correlation with target variable y:

[0054] RelevanceScore(f i ) = corr(f i ,y);

[0055] In the formula: corr(f i ,y) represents the Peel correlation coefficient between the feature and the target variable;

[0056] Feature selection combined with sparse constraints:

[0057]

[0058] In the formula: y i x represents the i-th observation of the target variable; i λ represents the corresponding feature vector; w represents the feature weight vector; λ represents the regularization parameter. Represents the eigenvector x i The dot product of the feature weight vector w represents the predicted value of the target variable; w j Representation and feature f jThe associated weights; μ represents the hyperparameter of the task relevance score; m represents the number of features, i.e., the feature dimension of each sample in the dataset; n represents the number of samples, i.e., the total number of samples in the dataset.

[0059] By combining feature selection with sparse constraints, the weights |w are obtained. j Features with |>0 will have weights |w j Features with a value greater than 0 are used as the corresponding insurance product features and fed into the attention mechanism for weighting.

[0060] Preferably, the formula for calculating the final product price is as follows:

[0061] P final (p)=P base (p)·(1+W p );

[0062] In the formula: P base (p) represents the benchmark price of the insurance product; W p P represents the price weighting factor; final (p) represents the final product price.

[0063] The beneficial effects of this invention include:

[0064] This invention collects and preprocesses basic customer information, social network data, insurance history data, and occupational information to lay a data foundation for subsequent product customization and pricing. Using a rule-based recommendation engine, relevant insurance products are recommended based on customer information and historical data. Graph Neural Networks (GNNs) and social network data are further used to filter insurance products, assigning a level to each product to ensure the recommendations better match customer needs. By selecting corresponding customer data based on the characteristics of different insurance products, risk assessment models are trained and applied to accurately assess customer risk. The price weighting factors of the products are adjusted based on the assessment results, achieving personalized price adjustments. Finally, combining the base price corresponding to each product level and the adjusted price weighting factors, the final personalized product price is output, ensuring that each customer receives a tailor-made insurance product at a reasonable price. By comprehensively utilizing multi-dimensional data sources and combining graph neural networks and deep learning technologies, product recommendations are more accurate and pricing is more flexible, fundamentally solving the problem that traditional insurance pricing cannot accurately reflect customer risk characteristics and promoting the personalized development of the insurance industry. Attached Figure Description

[0065] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0066] Figure 1 This is an overall step diagram provided for an embodiment of the present invention.

[0067] Figure 2 This is a schematic diagram illustrating the graph-structured thinking of social network data provided in an embodiment of the present invention. Detailed Implementation

[0068] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0069] See Figure 1 As shown, the personalized insurance product customization method and the risk pricing method based on customer profile include the following steps:

[0070] Step 1: Collect basic customer information, social network data, insurance history data, and occupational information, and preprocess the collected data;

[0071] Step 2: Use a rule-based recommendation engine to recommend relevant insurance products based on preprocessed basic customer information and insurance history data. Then, filter insurance products based on the recommended insurance products, graph neural network and preprocessed social network data, and determine the level of each insurance product.

[0072] As one possible implementation of this embodiment, step 2 includes the following steps:

[0073] Preliminary matching of insurance products: Based on the pre-processed basic customer information, a rule-based recommendation engine is used to initially screen insurance products that match the customer's needs;

[0074] An example rule is as follows:

[0075] Rule 1: Customers under the age of 18 are excluded from all adult insurance products.

[0076] Rule 2: If the customer is over 60 years old, all health insurance products (such as health insurance, critical illness insurance, etc.) targeting young people are excluded.

[0077] Rule 3: Depending on the product requirements, some insurance products may only be suitable for people in specific age groups. For example, some life insurance products are only for people aged 30-50, so they will be excluded if the customer's age does not fall within this range.

[0078] Rule 4: Customers with poor health conditions (such as chronic diseases like heart disease or diabetes) are excluded from high-risk health insurance products.

[0079] Rule 5: If the customer has already purchased "Life Insurance A" product, products related to "Life Insurance A" will not be recommended.

[0080] Rule 6: If a customer has already purchased a health insurance product, other products of the same insurance category (such as health insurance) are excluded.

[0081] The above rules are merely examples, and the rules are set based on actual application scenarios. The setting of the above rules is not intended to limit the present invention.

[0082] In this embodiment, a rule-based recommendation engine is used to initially screen insurance products that meet the customer's needs, and then further screening is performed based on the GNN model. The reason for this is that the rule engine can handle hard rules, while the graph application network can handle complex data-based personalized needs. The combination of the two can ensure that the recommendation system not only meets business rules, but also achieves personalization.

[0083] For example, there are product design limitations. Different insurance products have different underwriting conditions and restrictions (such as age and health requirements). These conditions are difficult to fully learn from data because they are part of the business rules. Therefore, we can use a rule engine to identify insurance products that customers do not meet the underwriting requirements. Furthermore, we can use a rule engine to eliminate duplicate insurance product items. For example, if a user has already purchased a certain insurance product, the rule engine can ensure that other recommended products do not overlap with their existing products. Therefore, our approach of combining a rule engine and a GNN model can provide higher accuracy, interpretability, and flexibility in various situations.

[0084] Extracting product recommendation levels: Based on the constructed GNN model and customer social network data, insurance products are filtered out to obtain a list of filtered insurance products;

[0085] The specific steps for selecting insurance products are as follows:

[0086] Graph-structured representation of social network data:

[0087] Social network data is transformed into a graph structure, which includes nodes, edges, and edge weights. Each customer represents a node in the graph, and the characteristics of the node include the customer's basic information and insurance history data. Each edge represents the relationship between two nodes in the social network. The edge weight is used to define the strength of the relationship between two nodes.

[0088] Graph neural networks perform message passing and node updates:

[0089] Each node i receives a message from its neighbor node j. The message contains the feature vector and edge weights of node j.

[0090] m ij =w ij ·h j ;

[0091] Where: m ij This represents the message received by node i from node j; w ij h represents the weight of the edge. j The feature vector of node j represents the basic information and insurance history data of customer j.

[0092] Each node aggregates all messages from its neighboring nodes to obtain the aggregated information for node i:

[0093]

[0094] In the formula: h represents the set of neighboring nodes of node i; i ′ This represents the aggregation information of node i;

[0095] The features of each node are updated based on the received aggregated message:

[0096]

[0097] In the formula: This represents the feature vector of node i at iteration t+1; W represents the feature vector of node i at iteration t; σ represents the Sigmoid activation function; (t) Let represent the learnable weight matrix for round t; Indicates feature concatenation operation;

[0098] Iteration and Termination:

[0099] In the graph structure described above, a node includes two layers of neighbors. The graph neural network performs two rounds of iteration. In the first round of iteration, message passing and node updates are performed on the first layer of neighbors. In the second round of iteration, message passing and node updates are performed on the second layer of neighbors. Finally, the feature representation of the node is output.

[0100] Product interest rating calculation:

[0101] The feature representation of the output node after the last iteration Includes customer social network information and personal characteristics; calculates an interest score for each customer i and each insurance product p:

[0102]

[0103] In the formula: h represents the feature vector of customer i after the last iteration; p Let represent the feature vector of insurance product p; ‖·‖ represents the L2 norm of the vector;

[0104] A list of recommended insurance products is generated based on interest scores and set thresholds.

[0105] The loss function of the GNN model is as follows:

[0106] L BPR =-∑ i ∑ p+∈Pos(i) ∑ p-∈Neg(i) logσ(S ip + -S ip- );

[0107] In the formula: This indicates that customer i is interested in insurance product p. + Interest rating (products selected by customers); This indicates that customer i is interested in insurance product p. - Interest rating (products not selected by the customer); Pos(i) represents the set of products already selected by customer i; Neg(i) represents the set of products not selected by customer i; σ represents the sigmoid activation function;

[0108] The loss is calculated based on the loss function described above, and backpropagation is performed using the gradient descent algorithm to update the learnable parameters of the model.

[0109] As one possible implementation of this embodiment, the steps for determining the weight of the edge are as follows:

[0110] Social relationship similarity calculation:

[0111]

[0112] In the formula: Represents the social feature vector of customer i; Let represent the social feature vector of customer j; Let L2 norm represent the social feature vector of customer i; Let L2 norm represent the social feature vector of customer j; This indicates the similarity of social relationships; the social characteristics may include customer interaction frequency, mutual friends, etc.

[0113] Exemplary social feature vectors:

[0114] Where 3 represents the number of mutual friends; 10 represents the interaction frequency; the interaction frequency is the frequency of interaction between customers on the social platform, which is represented by analyzing the number of messages, comments and other behaviors between customers, and calculating the number of interactions between customer i and customer j within a certain period of time, such as the number of messages, comments, etc.

[0115] Behavioral similarity calculation:

[0116]

[0117] In the formula: b i b represents the behavioral feature vector of customer i; j Represents the behavioral feature vector of customer j; ||b i || represents the L2 norm of the behavioral feature vector of customer i; || b j ‖ represents the L2 norm of the behavioral feature vector of customer j; the behavioral features may include the customer's purchase history, click behavior, search history, etc.

[0118] Exemplary behavioral feature vectors include:

[0119] Product categories purchased: Records the different categories of products a customer has purchased. For example, if a customer has purchased a certain type of insurance (such as health insurance, car insurance, etc.), this can be represented as part of the feature vector by the category numbers or vectors.

[0120] Purchase frequency: How often a customer purchases a particular type of product can be considered a characteristic. For example, the number of times a customer buys insurance each month, or the number of times they purchase a specific type of insurance product.

[0121] Purchase amount: The total amount spent by a customer within a specific period of time. It can be represented as a continuous feature.

[0122] Click count: Records the number of times a customer clicks on a certain type of product. Click count can be used to indicate a customer's interest in that category.

[0123] Click Time: Records the time distribution of customer clicks on a particular product category. If a customer frequently clicks on a certain product category within a specific time period, it indicates a high level of interest in that product category.

[0124] Click-through rate (CTR): The probability that a customer clicks on a product after browsing it; it can be used to measure the strength of a customer's interest in a particular category.

[0125] Specific examples of behavioral feature vectors are as follows:

[0126] b i =[1,5,100,10,3]; where 1 represents purchasing health insurance once, with a purchase frequency of 5, a consumption amount of 100, a click history of 10, and a search history of 3;

[0127] Weight calculation of the combined edges:

[0128]

[0129] In the formula: α and β represent weight coefficients; after calculating the weights of the composite edge, the weights are normalized.

[0130] In this embodiment, a social feature vector is constructed using information such as customer behavior, interaction frequency, and mutual friends in their social networks to reflect their social network relationships. Secondly, a behavioral feature vector is constructed using behavioral data such as purchase history, click behavior, and search records to reflect customer consumption preferences and interests. Based on this, a weighted sum is performed to obtain the edge weights, which can more accurately represent the social relationships and behavioral similarities between customers, thus making the social recommendation system more personalized, accurate, and influential. In social network-based recommendation systems, social relationships often have a significant impact on customer decisions; for example, customers may be influenced by friends or people in their social circles. Therefore, weighted social network edges can better reflect this social influence; the closer the relationship between a customer and the members of their social circle, the greater the impact of their purchasing behavior and consumption preferences on the customer; weighted edges can effectively enhance this influence, thereby improving the effectiveness of recommendations; and the present invention can handle the sparsity problem through the above technical solution, such as many customers not having direct social relationships or behavioral overlaps. By weighting based on information such as customer interaction frequency and mutual friends, this sparsity problem can be alleviated. By inferring potential interests and behavioral preferences between customers through indirect connections in social networks, the model's learning and generalization capabilities for sparse data are enhanced.

[0131] Insurance product rating determination: The appropriate insurance product rating is selected based on the customer's pre-processed insurance history.

[0132] In this embodiment, the initial feature vector of each node along the way contains the customer's individual information and historical insurance data, for example:

[0133] Basic customer information: such as age, gender, income, occupation, medical history, etc.;

[0134] Customer's insurance history data: such as the insurance products purchased in the past, the types of insurance, the frequency of purchase, etc.;

[0135] This intrinsic information is the feature vector at the initial moment of the node, representing the characteristics of the customer itself, which helps the model to understand the customer's preferences even without any social network information.

[0136] Then, through message passing and node updates in the graph neural network, each node can not only understand its own information, but also adjust and update its feature representation through the information of its neighboring nodes. This information fusion method enables the GNN model to capture the comprehensive influence of customer and personal characteristics on social network information, thereby providing more accurate predictions for insurance product recommendations.

[0137] See Figure 2 As shown, it's important to note that in this embodiment, we employ two iterations, meaning we have two layers of neighbors: the first neighbor nodes (e.g., friends and family) of the central node (customer) and the second neighbor nodes (e.g., friends and family) of the first neighbor nodes. The main reason for using two iterations is to simulate the influence of direct and indirect relationships in real social networks. For example, if friends and family recommend an insurance product, the customer may be influenced by these individuals; this layer's influence is direct and significant. Customer i obtains information through friends of friend j (i.e., the second neighbor nodes). While this influence may be relatively weak, it still has some effect, especially in certain situations where the customer's decision-making may be limited by the entire social circle (including distant relationships). Through this second iteration, node i can combine the information from the second neighbor nodes (i.e., the information obtained by node i in the first iteration) with the information from the second neighbor nodes. The message passing and aggregation between node i and its first neighbor node (and the message passing and aggregation between node i and its second neighbor node in the second iteration) further enriches the feature representation and captures the indirect influence in the social network. If more iteration layers are added, although the model can capture social networks at greater distances, this influence tends to become weaker and weaker, and the possibility of this happening in real social relationships is almost zero. Therefore, it is not considered in this embodiment. Thus, adding more layers of neighbors may not bring substantial information gain, especially when the factors influencing customer decisions gradually weaken, further increasing the layers of the social network may become meaningless. The two-layer neighbor system in this embodiment can capture both the direct influence (first neighbor) and the indirect influence (second neighbor) in the social network without making the model too complex.

[0138] It should also be noted that in this embodiment, the execution order of the rule engine and GNN model is not mandatory. They can be executed in parallel or sequentially. The execution order can be planned according to the actual application scenario constraints, such as CPU performance requirements. After deployment, if the CPU performance is sufficient, they can be executed in parallel; if the CPU performance is insufficient, they can be executed sequentially.

[0139] The following are exemplary technical solutions for the parallel execution:

[0140] Simultaneously, insurance products are selected using a rule engine and a GNN model. For example, the rule engine outputs critical illness insurance and accident insurance; the GNN model outputs critical illness insurance and medical insurance. The intersection of these two output insurance products is found, and since only critical illness insurance is present, the final insurance product obtained is critical illness insurance. The reason for finding the intersection here is that the medical insurance output by the GNN might be an insurance product filtered by the rule engine, and it may not conform to the relevant rules of the rule engine. Secondly, accident insurance is filtered out because we considered the customer's own situation and the influence of social networks through the GNN model, and since the output is critical illness insurance and medical insurance, accident insurance is filtered out.

[0141] The following are exemplary technical solutions for sequential execution:

[0142] We can first execute the GNN model to output the corresponding insurance products, and then use a rule engine to filter out the insurance products output by the GNN model to obtain the final insurance products.

[0143] Secondly, if the rule engine is executed first and then the GNN model is executed, the rule engine outputs the corresponding insurance products through the GNN model, and then the intersection is calculated.

[0144] Insurance products and insurance product levels appear multiple times in this invention. The following is a detailed description of insurance products and insurance product levels.

[0145] Currently, insurance products are diversified, and each insurance product involves multiple product levels, with different base insurance premiums and coverage amounts for each product level.

[0146] For example, insurance products include: critical illness insurance, whole life insurance, medical insurance, critical illness insurance, accidental death insurance, etc.

[0147] Different insurance products correspond to different tiers, and each tier corresponds to different premiums and coverage amounts, for example:

[0148] Critical Illness Insurance:

[0149] Basic version: Covers 10 common diseases, with a coverage amount of 500,000 yuan and an annual premium of 1,000 yuan.

[0150] Standard version: Covers 25 diseases, with a sum insured of 1 million yuan and an annual premium of 2,500 yuan.

[0151] Premium version: Covers 50 diseases, with a coverage amount of 2 million yuan and an annual premium of 5,000 yuan.

[0152] Medical insurance:

[0153] Basic version: Hospitalization reimbursement, coverage amount of 300,000 yuan, annual premium of 800 yuan.

[0154] Standard version: Inpatient reimbursement + outpatient reimbursement, coverage amount of 600,000 yuan, annual premium of 2,000 yuan.

[0155] Premium version: Inpatient reimbursement + outpatient reimbursement + high-end medical services, coverage of 1 million, annual premium of 4,000 yuan.

[0156] The specific steps for determining the insurance product level are as follows:

[0157] Iterate through the customer's historical insurance records to find the level of each insurance product and count the frequency of each product level; select the level with the highest frequency as the insurance product level to be determined this time.

[0158] As mentioned earlier, different insurance products correspond to different levels, so we need to select the level that matches the customer's actual situation. We select the highest-frequency level by statistically analyzing the frequency of product levels in the customer's historical insurance records.

[0159] Example: Seven types of insurance were purchased, including the standard version 4 times, the basic version 3 times, and the advanced version 0 times. It can be seen that the standard version appeared the most frequently, so the standard version is used as the level of this insurance product.

[0160] Step 3: Select corresponding customer characteristics to segment the data according to different insurance products, train and apply the corresponding risk assessment models for different insurance products, and adjust the price weight factor of each insurance product based on the results of the risk assessment models; the customer characteristics are preprocessed customer basic information, insurance history data and occupation information;

[0161] The risk assessment model includes an input layer, a feature selection and attention mechanism layer, a task feature extraction layer, and a task-specific layer.

[0162] The input layer is used to receive preprocessed basic customer information, insurance history data, and occupational information.

[0163] The number of feature selection and attention mechanism layers is adapted to the number of insurance products. Each feature selection and attention mechanism layer includes feature selection and self-attention mechanism. Feature selection selects features related to the corresponding insurance product according to the needs of the insurance product. The self-attention mechanism assigns a weight to each feature to obtain a weighted feature.

[0164] The specific steps for feature selection are as follows:

[0165] Correlation quantification: Calculate f for each feature i Correlation with target variable y:

[0166] RelevanceScore(f i ) = corr(f i ,y);

[0167] In the formula: corr(f i ,y) represents the Peel correlation coefficient between the feature and the target variable;

[0168] Feature selection combined with sparse constraints:

[0169]

[0170] In the formula: y i x represents the i-th observation of the target variable; i λ represents the corresponding feature vector; w represents the feature weight vector; λ represents the regularization parameter. Represents the eigenvector x i The dot product of the feature weight vector w represents the predicted value of the target variable; w j Representation and feature f j The associated weights; μ represents the hyperparameter of the task relevance score; m represents the number of features, i.e., the feature dimension of each sample in the dataset; n represents the number of samples, i.e., the total number of samples in the dataset.

[0171] By combining feature selection with sparse constraints, the weights |w are obtained. j Features with |>0 will have weights |w j Features with a value greater than 0 are used as the corresponding insurance product features and fed into the attention mechanism for weighting.

[0172] The number of task feature extraction layers is adapted to the number of insurance products. Different fully connected layers are used in the task feature extraction layers for different insurance products to mine weighted features and obtain the deep features of the corresponding insurance products.

[0173] The number of task-specific layers is adapted to the number of insurance products. Each task-specific layer makes predictions based on the deep features output by the task feature extraction layer, corresponding to the price weight factor of the insurance product.

[0174] Example technical solution for critical illness insurance risk assessment model:

[0175] Input layer: Customer basic information: age, gender, health status (such as whether there is a family history of hereditary diseases), smoking and drinking habits, weight, height, etc.

[0176] Insurance history data: Health insurance policies purchased by the customer, claims records (such as whether there have been critical illness claims), and whether there is a history of renewal of critical illness insurance, etc.

[0177] Occupational information: Whether the client is engaged in a high-risk occupation (such as miner, construction worker, etc.).

[0178] For critical illness insurance, the relevance of the following features will be considered when selecting features:

[0179] Health status (directly related to the probability of developing serious illnesses);

[0180] Age (the older you are, the higher the probability of developing a serious illness);

[0181] Smoking and drinking habits (smoking and drinking are closely related to many serious diseases such as cancer and cardiovascular disease);

[0182] Weight (being overweight may increase the risk of diseases such as diabetes and heart disease);

[0183] Feature selection and attention mechanism layer: After feature selection (based on the feature selection formula described above), each feature is assigned a weight, and the weighted feature is calculated through a self-attention mechanism; it should be noted that using a self-attention mechanism to calculate the weighted feature is a conventional technique in this field, so it will not be described in detail in this embodiment.

[0184] Task Feature Extraction Layer: Uses a fully connected layer (FC Layer) to extract weighted features and mine deep features related to critical illness insurance;

[0185] For example, the specific structure of the task feature extraction layer is as follows:

[0186] The first fully connected layer takes a weighted feature vector as input and outputs a 256-neuron dimension to capture the non-linear relationships between features; the activation function uses ReLU (Rectified Linear Unit) to increase non-linearity.

[0187] h1 = ReLU(W1x + b1);

[0188] In the formula: x represents the weighted feature vector (i.e., the feature vector after processing by the attention mechanism layer); W1 and b1 represent the weights and biases of the first fully connected layer, respectively; h1 represents the output of the first fully connected layer.

[0189] The second fully connected layer further passes the output of the first layer to the second layer, with an output dimension of 128 neurons. It uses the ReLU activation function to further extract deeper features.

[0190] h2 = ReLU(W2h1 + b2);

[0191] In the formula: h2 represents the output of the second fully connected layer; W2 and b2 represent the weights and biases of the second fully connected layer, respectively;

[0192] The third fully connected layer further reduces the feature dimension to 64 neurons; the activation function is the sigmoid function, which outputs a value between 0 and 1 to adapt to the final weighting factor.

[0193] h3 = σ(W3h2 + b3);

[0194] In the formula: h3 represents the output of the third fully connected layer; W3 and b3 represent the weights and biases of the third fully connected layer, respectively; σ represents the sigmoid activation function;

[0195] Task-specific layer and price weighting factor adjustment: Risk assessment of critical illness insurance is performed based on the deep features output from the task feature extraction layer; the price weighting factor for critical illness insurance is calculated through the task-specific layer.

[0196] For example:

[0197] The task-specific layer outputs a single numerical value representing the customer's risk of illness through a regression model. The risk value is assumed to range from 0 to 1, where 0 represents the lowest risk and 1 represents the highest risk. An example structure for the task-specific layer is as follows:

[0198] First fully connected layer: Input is h3, output dimension is 32, activation function is ReLU:

[0199] h4 = ReLU(W4h3 + b4);

[0200] In the formula: h4 represents the output of the first fully connected layer in the task-specific layer; W4 and b4 represent the weights and biases of the first fully connected layer in the task-specific layer;

[0201] The second fully connected layer has an output of h4 and an output dimension of 16; the activation function is ReLU.

[0202] h5 = ReLU(W5h4 + b5);

[0203] In the formula: h5 represents the output of the second fully connected layer in the task-specific layer; W5 and b5 represent the weights and biases of the second fully connected layer in the task-specific layer;

[0204] Output layer: Finally, a single-neuron output layer is used for risk prediction. The output value r represents the customer's risk of illness (it should be noted that this risk of illness in other models may represent the risk of accidental death, that is, r is the same as the corresponding risk prediction type. Therefore, it should be noted that the same model structure can be used for prediction in this invention, only the input data and the risk type described in the result are different). The sigmoid activation function is used to compress the output value to the [0,1] interval.

[0205] r = σ(W6h5 + b6);

[0206] In the formula: r represents the predicted risk value; σ represents the sigmoid activation function; W6 and b6 represent the weights and biases of the output layer, respectively;

[0207] Output price weighting factors:

[0208] W p =α·r;

[0209] In the formula: α represents the degree to which risk affects prices.

[0210] The price weight of insurance products is calculated by using the risk value output from the task-specific layer. Therefore, according to the model design, the higher the risk, the higher the price weight factor will be.

[0211] Example technical solution for accidental death insurance risk assessment model:

[0212] Input layer: Customer basic information: age, gender, past medical history (such as whether there are diseases that may affect the occurrence of accidents, such as epilepsy, osteoporosis, etc.).

[0213] Insurance history data: whether the customer has a history of accident insurance claims, whether they have other accident insurance policies, etc.

[0214] Occupational information: Whether the client works in a high-risk occupation (such as driver, construction worker, mountain guide, etc.).

[0215] For accidental death insurance, the relevance of the following features will be considered when selecting features:

[0216] Age (the older the person, the higher the risk of accidental death, especially for the elderly);

[0217] Occupations (clients in high-risk occupations (such as drivers, miners, etc.) face a higher risk of accidents);

[0218] A history of pre-existing medical conditions (such as osteoporosis) may increase the probability of an accident.

[0219] Feature selection and attention mechanism layer: After feature selection (based on the feature selection formula described above), each feature is weighted by a self-attention mechanism; for accidental death insurance, occupation and age may receive higher weights.

[0220] Task Feature Extraction Layer: Deep features are extracted from the weighted features through different fully connected layers to obtain deep features related to accidental death insurance; the specific structure of the task feature extraction layer is the same as the specific structure of the critical illness insurance risk assessment model.

[0221] Task-specific layer and price weight factor adjustment: Based on the deep features output by the task feature extraction layer, the risk assessment of accidental death insurance is carried out, and the adjusted price weight factor is finally obtained.

[0222] The specific methods for adjusting the task feature extraction layer and the task-specific layer along with the price weight factor can refer to the exemplary technical solution of the critical illness insurance risk assessment model; that is, we can use the same model structure, only the input parameters and the output risk categories are different.

[0223] Step 4: Output the final product price based on the benchmark price corresponding to each insurance product level and the price weighting factor of each insurance product.

[0224] The formula for calculating the final product price is as follows:

[0225] P final (p)=P base (p)·(1+W p );

[0226] In the formula: P base (p) represents the benchmark price of the insurance product; W p P represents the price weighting factor; final (p) represents the final product price.

[0227] This invention collects and preprocesses basic customer information, social network data, insurance history data, and occupational information to lay a data foundation for subsequent product customization and pricing. A rule-based recommendation engine recommends relevant insurance products based on customer information and historical data. Graph Neural Networks (GNNs) and social network data are used to further filter insurance products and assign a level to each product, ensuring that the recommended products better match customer needs. By selecting corresponding customer data based on the characteristics of different insurance products, risk assessment models are trained and applied to accurately assess customer risk. The price weighting factors of the products are adjusted based on the assessment results, achieving personalized price adjustments. Finally, by combining the base price corresponding to each product level and the adjusted price weighting factors, the final personalized product price is output, ensuring that each customer receives a tailor-made insurance product at a reasonable price. By comprehensively utilizing multi-dimensional data sources and combining graph neural networks and deep learning technologies, product recommendations are more accurate and pricing is more flexible, fundamentally solving the problem that traditional insurance pricing cannot accurately reflect customer risk characteristics and promoting the personalized development of the insurance industry.

[0228] The above are merely preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for customizing personalized insurance products and a risk pricing method based on customer profiles, characterized in that, Includes the following steps: Step 1: Collect basic customer information, social network data, insurance history data, and occupational information, and preprocess the collected data; Step 2: Using a rule-based recommendation engine, recommend relevant preliminary insurance products based on preprocessed customer basic information and insurance history data; then construct a graph structure based on preprocessed social network data, and iterate the graph structure through a graph neural network to filter out insurance products that meet customer needs based on the preliminary insurance products, resulting in an insurance product list; finally, select the appropriate insurance product level based on the customer's preprocessed insurance history. The graph structure of the graph neural network includes two layers of neighbors. The graph neural network performs two rounds of iteration. In the first round of iteration, message passing and node updates are performed on the first layer of neighbors. In the second round of iteration, message passing and node updates are performed on the second layer of neighbors. Finally, the feature representation of the node is output, which includes the customer's social network information and personal characteristics. Then, interest scores are calculated based on the feature representation of the nodes, and the list of insurance products is obtained based on the interest scores and the set thresholds. The weights between two adjacent neighbor nodes in the graph structure are calculated based on a weighted sum of social relationship similarity and behavioral similarity; Step 3: Select corresponding customer characteristics to segment the data according to different insurance products, train and apply the corresponding risk assessment models for different insurance products, and adjust the price weight factor of each insurance product based on the results of the risk assessment models; the customer characteristics are preprocessed customer basic information, insurance history data and occupation information; Step 4: Output the final product price based on the benchmark price corresponding to each insurance product level and the price weighting factor of each insurance product.

2. The personalized insurance product customization method and the risk pricing method based on customer profiles according to claim 1, characterized in that, The specific steps to obtain the insurance product list are as follows: Graph-structured representation of social network data: Social network data is transformed into a graph structure, which includes nodes, edges, and edge weights. Each customer represents a node in the graph, and the characteristics of the node include the customer's basic information and insurance history data. Each edge represents the relationship between two nodes in the social network. The edge weight is used to define the strength of the relationship between two nodes. Graph neural networks perform message passing and node updates: Each node i receives a message from its neighbor node j. The message contains the feature vector and edge weights of node j. ； In the formula: This represents the message received by node i from node j; Indicates the weight of the edge; The feature vector representing node j contains the basic information and insurance history data of customer j. Each node aggregates all messages from its neighboring nodes to obtain the aggregated information for node i: ； In the formula: Represents the set of neighboring nodes of node i; This represents the aggregation information of node i; The features of each node are updated based on the received aggregated message: ； In the formula: This represents the feature vector of node i at iteration t+1; This represents the feature vector of node i at iteration t; This represents the Sigmoid activation function; Let represent the learnable weight matrix for round t; Indicates feature concatenation operation; Iteration and Termination: In the graph structure described above, a node includes two layers of neighbors. The graph neural network performs two rounds of iteration. In the first round of iteration, message passing and node updates are performed on the first layer of neighbors. In the second round of iteration, message passing and node updates are performed on the second layer of neighbors. Finally, the feature representation of the node is output. Product interest rating calculation: The feature representation of the output node after the last iteration It includes the customer's social network information and personal characteristics; for each customer i and each insurance product p, an interest score is calculated: ； In the formula: This represents the feature vector of customer i after the last iteration; This represents the feature vector of insurance product p; The L2 norm of a vector; A list of recommended insurance products is generated based on interest scores and set thresholds.

3. The personalized insurance product customization method and the risk pricing method based on customer profiles according to claim 2, characterized in that, The steps for determining the weight of the edge are as follows: Social relationship similarity calculation: ； In the formula: Represents the social feature vector of customer i; Let represent the social feature vector of customer j; Let L2 norm represent the social feature vector of customer i; Let L2 norm represent the social feature vector of customer j; Indicates the similarity of social relationships; Behavioral similarity calculation: ； In the formula: This represents the behavioral feature vector of customer i; This represents the behavioral feature vector of customer j; The L2 norm of the behavioral feature vector of customer i; Let L2 norm represent the behavioral feature vector of customer j; Weight calculation of the combined edges: ； In the formula: and This represents the weight coefficient; after calculating the weight of the composite edge, the weight is normalized.

4. The personalized insurance product customization method and the risk pricing method based on customer profile as described in claim 1, characterized in that, The steps for determining the insurance product level are as follows: Iterate through the customer's historical insurance records to find the level of each insurance product and count the frequency of each product level; select the level with the highest frequency as the insurance product level to be determined this time.

5. The personalized insurance product customization method and the risk pricing method based on customer profile as described in claim 1, characterized in that, The risk assessment model includes an input layer, a feature selection and attention mechanism layer, a task feature extraction layer, and a task-specific layer. The input layer is used to receive preprocessed basic customer information, insurance history data, and occupational information. The number of feature selection and attention mechanism layers is adapted to the number of insurance products. Each feature selection and attention mechanism layer includes feature selection and self-attention mechanism. Feature selection selects features related to the corresponding insurance product according to the needs of the insurance product. The self-attention mechanism assigns a weight to each feature to obtain a weighted feature. The number of task feature extraction layers is adapted to the number of insurance products. Different fully connected layers are used in the task feature extraction layers for different insurance products to mine weighted features and obtain the deep features of the corresponding insurance products. The number of task-specific layers is adapted to the number of insurance products. Each task-specific layer makes predictions based on the deep features output by the task feature extraction layer, corresponding to the price weight factor of the insurance product.

6. The personalized insurance product customization method and the risk pricing method based on customer profile as described in claim 5, characterized in that, The specific steps for feature selection are as follows: Correlation quantification: Calculate each feature Correlation with target variable y: ； In the formula: The Peel correlation coefficient represents the relationship between the feature and the target variable; Feature selection combined with sparse constraints: ； In the formula: This represents the i-th observation of the target variable; represents the corresponding feature vector; w represents the feature weight vector; Represents the regularization parameter; Representing the eigenvector Weight vector of features The inner product represents the predicted value of the target variable; Representation and Features Associated weights; The hyperparameter representing the impact of task relevance score; m represents the number of features, i.e., the feature dimension of each sample in the dataset; n represents the number of samples, i.e., the total number of samples in the dataset. By combining feature selection with sparse constraints, weights are obtained. The characteristics will weight The features are used as the corresponding features of the insurance products and are then weighted by the attention mechanism.

7. The personalized insurance product customization method and the risk pricing method based on customer profile as described in claim 1, characterized in that, The formula for calculating the final product price is as follows: ； In the formula: This indicates the benchmark price of the insurance product; Indicates the price weighting factor; This indicates the final product price.

Images