Insurance product push method and system based on user interest modeling

Abstract

The present application provides a method and system for pushing insurance products using user interest modeling. By constructing target interactive behavior graph structure data, the present invention can accurately capture the correlation and characteristics between user behaviors, and use logistic regression networks and random forest tree networks to model user interests respectively, and fuse the two modeling data. The present invention can more comprehensively reflect the user's interest preferences and potential needs. In addition, by using different sample product interactive behavior data sequences for network parameter learning, the present invention further improves the accuracy and robustness of user interest modeling. Finally, based on the target user interest modeling data for pushing insurance products, the present invention can significantly improve the accuracy of push and user satisfaction, and provide data support for personalized and intelligent recommendations of insurance products.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and more specifically, to a method and system for pushing insurance products using user interest modeling. Background Art

[0002] In the insurance industry, with the rapid development of the Internet and the increasing demand for personalized services from users, how to accurately and efficiently push insurance products has become an urgent problem to be solved. Traditional insurance product push methods are often based on the analysis of users' basic information or simple behavioral data, which makes it difficult to deeply explore users' real interests and potential needs, resulting in unsatisfactory push results.

[0003] In recent years, with the continuous advancement of big data and artificial intelligence technology, user interest modeling has gradually become a research hotspot. By building a user interest model, we can more accurately grasp the user's preferences and needs, thereby improving the pertinence and effectiveness of product push. However, existing user interest modeling methods often have some problems, such as a single data processing method and weak model generalization ability, which makes it difficult to meet the actual needs of the insurance product push field. Summary of the invention

[0004] In view of this, the purpose of this application is to provide an insurance product push method and system using user interest modeling.

[0005] According to a first aspect of the present application, a method for pushing insurance products using user interest modeling is provided, the method comprising:

[0006] Acquire product interaction behavior data of target users in insurance product service applications, and construct a graph structure for the product interaction behavior data according to a target graph neural network to generate target interaction behavior graph structure data of the product interaction behavior data;

[0007] Performing user interest modeling on the product interaction behavior data according to the logistic regression network and the target interaction behavior graph structure data to generate first user interest modeling data;

[0008] Performing user interest modeling on the product interaction behavior data according to the random forest tree network and the target interaction behavior graph structure data to generate second user interest modeling data, wherein the logistic regression network is generated by learning network parameters according to the first sample product interaction behavior data sequence, the random forest tree network is generated by learning network parameters according to the second sample product interaction behavior data sequence, the second sample product interaction behavior data sequence is a branch sequence of the first sample product interaction behavior data sequence, and the first sample product interaction behavior data sequence and the second sample product interaction behavior data sequence are different sequences;

[0009] Determining target user interest modeling data for the product interaction behavior data based on the first user interest modeling data and the second user interest modeling data;

[0010] Insurance products are pushed to the target user based on the target user interest modeling data.

[0011] According to the second aspect of the present application, an insurance service system is provided, which includes a machine-readable storage medium and a processor, wherein the machine-readable storage medium stores machine-executable instructions, and when the processor executes the machine-executable instructions, the insurance service system implements the aforementioned insurance product push method using user interest modeling.

[0012] According to a third aspect of the present application, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed, the aforementioned insurance product push method using user interest modeling is implemented.

[0013] According to any one of the above aspects, the technical effect of the present application is:

[0014] The embodiment of the present application realizes the deep mining and efficient utilization of the product interaction behavior data of the target users in the insurance product service application by combining the target graph neural network, the logistic regression network and the random forest tree network. By constructing the target interaction behavior graph structure data, the present invention can accurately capture the correlation and characteristics between user behaviors, and provide a solid foundation for user interest modeling. By using the logistic regression network and the random forest tree network to model user interests respectively, and fusing the two modeling data, the present invention can more comprehensively reflect the user's interest preferences and potential needs. In addition, by using different sample product interaction behavior data sequences for network parameter learning, the present invention further improves the accuracy and robustness of user interest modeling. Finally, based on the target user interest modeling data for insurance product push, the present invention can significantly improve the accuracy of push and user satisfaction, and provide data support for the personalized and intelligent recommendation of insurance products. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0016] Figure 1A flow chart of a method for pushing insurance products using user interest modeling provided in an embodiment of the present application is shown;

[0017] Figure 2 A schematic diagram of the component structure of an insurance service system provided in an embodiment of the present application for implementing the above-mentioned insurance product push method using user interest modeling is shown. DETAILED DESCRIPTION

[0018] The embodiments of the present application are described below in conjunction with the drawings in the present application. It should be understood that the implementation methods described below in conjunction with the drawings are exemplary descriptions for explaining the technical solutions of the embodiments of the present application and do not constitute a limitation on the technical solutions of the embodiments of the present application.

[0019] It will be understood by those skilled in the art that, unless specifically stated, the singular forms "one", "said", and "the" used herein may also include plural forms. It should be further understood that the terms "including" and "comprising" used in the embodiments of the present application refer to that the corresponding features can be implemented as the presented features, information, data, steps, operations, elements and / or components, but do not exclude the implementation as other features, information, data, steps, operations, elements, components and / or combinations thereof supported by the technical field. It should be understood that when an element is "connected" or "coupled" to another element, the element may be directly connected or coupled to the other element, or may refer to the element and the other element establishing a connection relationship through an intermediate element. In addition, the "connection" or "coupling" used herein may include wireless connection or wireless coupling, and the term "and / or" used herein indicates at least one of the items defined by the term, such as "A and / or B" may be implemented as "A", or as "B", or as "A and B".

[0020] In order to make the purpose, technical solution and advantages of the present application clearer, the implementation mode of the present application will be further described in detail with reference to the accompanying drawings. The technical solution of the embodiment of the present application and the technical effect produced by the technical solution of the present application are explained below by describing several exemplary implementation modes. It should be pointed out that the following implementation modes can refer to, draw lessons from or combine with each other, and the same terms, similar features and similar implementation steps in different implementation modes are not described repeatedly.

[0021] Figure 1 The flowchart of the insurance product push method and system using user interest modeling provided by the embodiment of the present application is shown. It should be understood that in other embodiments, the order of some steps in the insurance product push method using user interest modeling of this embodiment can be shared with each other according to actual needs, or some steps can be omitted or maintained. The detailed steps of the insurance product push method using user interest modeling include:

[0022] Step S110, obtaining product interaction behavior data of a target user in an insurance product service application, and constructing a graph structure for the product interaction behavior data based on a target graph neural network to generate target interaction behavior graph structure data of the product interaction behavior data.

[0023] In this embodiment, many users interact with the insurance company through the insurance product service application. Assume that the target user is a customer named Li.

[0024] The server obtains Li's product interaction behavior data from the database of the insurance product service application. These product interaction behavior data contain Li's various operation records in the application. For example, Li queried the detailed terms of a certain critical illness insurance product on the application. This query operation includes the timestamp of the operation (10:00 on May 10, 2023), the critical illness insurance product identifier queried (product number 12345), the operation type is a query operation, and the originator is Li's user account (account number LXS123). In addition, Li also applied for insurance for a certain accident insurance product. The application time was 15:00 on June 15, 2023. The accident insurance product identifier was 67890, the operation type was an insurance application, and the originator was also his account LXS123.

[0025] Then, the server constructs the graph structure of these product interaction behavior data based on the target graph neural network. First, different interaction behavior record units are identified from these product interaction behavior data. For example, the two behaviors of querying critical illness insurance and insuring accident insurance are different record units. For the record unit of querying critical illness insurance, a decomposition operation is performed to separate the origin of the interaction behavior (Li's account LXS123), the identifier of the insurance product targeted by the interaction behavior (critical illness insurance product number 12345), the specific type of interaction behavior (query operation) and the timestamp information of the interaction behavior (May 10, 2023 10:00), and these elements are combined into an interaction behavior element set. Similarly, the record unit of insuring accident insurance is also decomposed to obtain another interaction behavior element set.

[0026] Then, according to the pre-set classification rules, classification operations are performed on these interactive behavior element sets. For user identification elements, Li's account LXS123 is defined as an independent node, which represents Li's subject identity in the entire interactive behavior graph; for product information elements, a node is defined for the critical illness insurance product number 12345 and the accident insurance product number 67890 respectively, which represent the insurance products in the graph structure; for behavior type elements, the query operation and the insurance application operation are defined as a node respectively, indicating the existence of different types of interactive behaviors in the graph structure.

[0027] After that, the relationship between each node is mined based on the interactive behavior element set of predefined node information. For example, there is a query relationship between Li's account number LXS123 and the critical illness insurance product number 12345. The frequency of this relationship is determined by the number of times Li has queried the product. Assuming that he has queried 3 times, the initial quantization value can be set to 3. Then, this initial quantization value is further adjusted and weighted according to the pre-set quantization rules. For example, considering that the query operation has a lower weight than the insurance operation in reflecting the degree of interest, the weight of the query relationship is set to 0.3, and the adjusted quantization value is 0.9.

[0028] The predefined nodes are connected according to the node relationships with quantitative relationships and weights to generate a preliminary graph structure. In this preliminary graph structure, there is a connection edge between Li's account node and the critical illness insurance product node, and the edge weight is 0.9, which represents the quantitative value of the query relationship.

[0029] Finally, the preliminary graph structure is checked for integrity and optimized. Check whether each node has a relationship connection with at least one other node. It is found that the critical illness insurance product number 12345 node has no other relationship connection except the query relationship connection with Li's account, but this is reasonable because it is a relationship based on actual data and there is no isolated node problem. At the same time, check whether there are unreasonable relationship connections. After inspection, no connections between nodes that do not conform to business logic or edges with business errors in weight settings are found. Then, according to the type of node, the nature of the node relationship, and the business logic, the optimized graph structure is hierarchically divided. For example, Li's account node is placed in the user layer, the insurance product node is placed in the product layer, and the operation type node is placed in the behavior layer. The interactive behavior graph structure after hierarchical division is feature enhanced, such as adding node attribute information, etc., and finally the target interactive behavior graph structure data of Li's product interactive behavior data is generated.

[0030] Step S120, performing user interest modeling on the product interaction behavior data according to the logistic regression network and the target interaction behavior graph structure data to generate first user interest modeling data.

[0031] In this embodiment, the logistic regression network is generated by learning network parameters based on the first sample product interaction behavior data sequence. Assume that the first sample product interaction behavior data sequence contains the interaction behavior data of many other users in the insurance product service application and their corresponding training supervision data. For example, sample user A inquired about a life insurance product and finally insured it. The corresponding training supervision data is active interest point data (indicating that he has a high interest in the life insurance product and has taken insurance actions); sample user B inquired about a medical insurance product but did not perform any further operations. The corresponding training supervision data is passive interest point data (indicating that he has a certain interest but has not gone deep).

[0032] The server loads Li's target interaction behavior graph structure data into the logistic regression network. The logistic regression network analyzes and models Li's interaction behavior data based on the parameters learned previously. For example, within the network, Li's query operation on critical illness insurance, the number of queries, the query time interval, and other factors are combined with the weight parameters stored in the network for calculation. If the number of queries is large and the time interval is short, the network may determine that Li has a high potential interest in critical illness insurance. At the same time, for Li's operation of insuring accident insurance, the network will also conduct a comprehensive analysis based on the completeness of his operation (such as whether the complete insurance information is filled in), the insured amount, and other factors. After a complex calculation and model reasoning process, the first user interest modeling data is finally generated. This first user interest modeling data may contain Li's interest level scores for different types of insurance products, such as a score of 0.6 for critical illness insurance and a score of 0.8 for accident insurance.

[0033] Step S130, performing user interest modeling on the product interaction behavior data based on the random forest tree network and the target interaction behavior graph structure data to generate second user interest modeling data, wherein the logistic regression network is generated by learning network parameters based on the first sample product interaction behavior data sequence, and the random forest tree network is generated by learning network parameters based on the second sample product interaction behavior data sequence, and the second sample product interaction behavior data sequence is a branch sequence of the first sample product interaction behavior data sequence, and the first sample product interaction behavior data sequence and the second sample product interaction behavior data sequence are different sequences.

[0034] In this embodiment, the random forest tree network is generated by learning network parameters based on the second sample product interaction behavior data sequence, and the second sample product interaction behavior data sequence is a branch sequence of the first sample product interaction behavior data sequence. For example, the first sample product interaction behavior data sequence contains interaction behavior data of 1,000 sample users, and the second sample product interaction behavior data sequence may be interaction behavior data of 500 sample users under specific conditions selected from these 1,000 samples.

[0035] Assume that the screening rule of the second sample product interaction behavior data sequence is to select the interaction behavior data of sample users who have queried insurance products between 3 and 5 times. Among these data, sample user C has queried a pension insurance product 3 times, and the corresponding training supervision data is passive tendency interest point data; sample user D has queried a health insurance product 4 times, and the corresponding training supervision data is active tendency interest point data, etc.

[0036] The server inputs Li's target interaction behavior graph structure data into the random forest tree network. The random forest tree network consists of multiple decision trees, each of which makes an independent judgment on Li's interaction behavior data. For example, a decision tree may first determine whether Li has queried critical illness insurance more than three times. If so, it will enter the next node to determine whether Li has queried the product again within a short period of time. Another decision tree may start with Li's behavior of purchasing accident insurance to determine whether the insured amount is within a specific range. After comprehensive judgments from multiple decision trees, the random forest tree network finally generates the second user interest modeling data. This second user interest modeling data may differ from the first user interest modeling data in some aspects, such as a score of 0.5 for the interest level in critical illness insurance and a score of 0.7 for the interest level in accident insurance.

[0037] Step S140, determining target user interest modeling data of the product interaction behavior data based on the first user interest modeling data and the second user interest modeling data.

[0038] In this embodiment, in the first user interest modeling data, Li's interest in critical illness insurance is scored as 0.6, and his interest in accident insurance is scored as 0.8; in the second user interest modeling data, his interest in critical illness insurance is scored as 0.5, and his interest in accident insurance is scored as 0.7. The server can use a weighted average method to determine the target user interest modeling data. Assume that the weight of the first user interest modeling data generated by the logistic regression network is 0.6, and the weight of the second user interest modeling data generated by the random forest tree network is 0.4.

[0039] For critical illness insurance, the interest level score in the target user interest modeling data is calculated as: (0.6×0.6 + 0.5×0.4) = 0.56; for accident insurance, it is calculated as: (0.8×0.6+ 0.7×0.4) = 0.76. In addition to this weighted average method, other fusion methods can also be used according to specific business needs and data characteristics, such as taking the maximum value, minimum value or more complex fusion algorithms, and finally obtaining Li's target user interest modeling data. This target user interest modeling data accurately reflects Li's comprehensive interest level in different insurance products.

[0040] Step S150: Push insurance products to the target user based on the target user interest modeling data.

[0041] In this embodiment, first, the server obtains information about candidate insurance products from the product database in the insurance product service application. For example, the candidate insurance products include a newly launched cancer prevention insurance, an upgraded medical insurance, and a long-term savings insurance. For each insurance product, the server obtains its product data, such as the terms and conditions of the cancer prevention insurance, including the types of cancer covered, the specific conditions for claims, and the payment method of annual payment and installment for 10 years; the coverage data of the medical insurance covers the reimbursement ratio of treatment expenses for common diseases, and the claims conditions data includes the grade requirements of the hospital for medical treatment; the income model data of the long-term savings insurance shows the expected annualized rate of return, etc.

[0042] Then, the product data of these insurance products are classified according to multiple pre-set classification dimensions to generate the classified insurance product feature set corresponding to each insurance product. For cancer insurance, after classification according to the dimensions of coverage, claim conditions, and payment methods, its feature set may include features such as coverage for specific types of cancer, meeting specific claim conditions (such as compensation after diagnosis), and annual payment for 10 years.

[0043] Next, the user interest points in Li's target user interest modeling data are matched one by one with the corresponding classified insurance product feature set of each insurance product. Assuming that Li's target user interest modeling data shows that he has a high interest in disease protection products, then cancer insurance and medical insurance match his interest points in this regard, and a preliminary matching reference insurance product list is generated, which includes cancer insurance and medical insurance.

[0044] Afterwards, the range of Li's spending power was mined based on his user portrait data. Li's historical consumption records in the insurance product service application showed that he had purchased a medium-value accident insurance with a premium of about 2,000 yuan per year; his income level estimation data showed that his annual income was between 100,000 and 150,000 yuan; his asset status data showed that he had a certain amount of savings and stable real estate. Combining this information, the server determined that Li's spending power was at a medium level.

[0045] Finally, according to Li's spending power and the list of reference insurance products that were initially matched, insurance products were pushed to Li. Since Li's spending power was at a medium level, when the server pushed cancer insurance to him, it emphasized the cost-effectiveness of the cancer insurance, such as the relatively low premium for a higher coverage; for medical insurance, it emphasized that it could provide more comprehensive disease protection under medium spending power, in order to meet Li's interest in disease protection products and his spending power.

[0046] Based on the above steps, the embodiment of the present application realizes the deep mining and efficient utilization of the product interaction behavior data of the target users in the insurance product service application by combining the target graph neural network, the logistic regression network and the random forest tree network. By constructing the target interaction behavior graph structure data, the present invention can accurately capture the correlation and characteristics between user behaviors, and provide a solid foundation for user interest modeling. By using the logistic regression network and the random forest tree network to model user interests respectively, and fusing the two modeling data, the present invention can more comprehensively reflect the user's interest preferences and potential needs. In addition, by using different sample product interaction behavior data sequences for network parameter learning, the present invention further improves the accuracy and robustness of user interest modeling. Finally, based on the target user interest modeling data for insurance product push, the present invention can significantly improve the accuracy of push and user satisfaction, and provide data support for the personalized and intelligent recommendation of insurance products.

[0047] In a possible implementation, the method further includes:

[0048] Step A110, obtaining the first sample product interaction behavior data sequence, wherein the first sample product interaction behavior data sequence includes multiple first sample product interaction behavior data and corresponding first training supervision data, and the first training supervision data is active tendency interest point data or passive tendency interest point data.

[0049] Step A120, optimizing the basic graph neural network and the first basic interest modeling network according to the first sample product interaction behavior data sequence, and generating the target graph neural network and the logistic regression network.

[0050] Step A130: extracting the second sample product interaction behavior data sequence from the first sample product interaction behavior data sequence.

[0051] Step A140, locking the target graph neural network, and optimizing the second basic interest modeling network according to the second sample product interaction behavior data sequence to generate the random forest tree network.

[0052] In a possible implementation, step A120 includes:

[0053] Step A121: load multiple first sample product interaction behavior data in the first sample product interaction behavior data sequence into the basic graph neural network to generate sample interaction behavior graph structure data corresponding to each of the first sample product interaction behavior data.

[0054] Step A122: Load the sample interaction behavior graph structure data corresponding to each of the first sample product interaction behavior data into the first basic interest modeling network to generate interest modeling data.

[0055] Step A123, calculating the interest modeling error based on the first training supervision data of each of the first sample product interaction behavior data and the interest modeling data.

[0056] Step A124, calculating the interest modeling comparison error based on the sample interaction behavior graph structure data corresponding to each of the first sample product interaction behavior data and the first sample product interaction behavior data sequence.

[0057] Step A125, determining a global error parameter according to the interest modeling error and the interest modeling comparison error.

[0058] Step A126, cyclically optimize the basic graph neural network and the first basic interest modeling network according to the global error parameter to generate the target graph neural network and logistic regression network.

[0059] In a possible implementation, step A130 includes:

[0060] Step A131: extracting a passive tendency interest behavior data sequence from the first sample product interaction behavior data that includes passive tendency interest point data in the first sample product interaction behavior data sequence.

[0061] Step A132: extracting an active tendency interest behavior data sequence from the second sample product interaction behavior data that includes active tendency interest point data in the first sample product interaction behavior data sequence.

[0062] Step A133, summarizing the active tendency interest behavior data sequence and the passive tendency interest behavior data sequence to generate the second sample product interaction behavior data sequence.

[0063] In this embodiment, during the training process, the server collects a large amount of first-instance product interaction behavior data from the historical database of the insurance product service application. These first-instance product interaction behavior data are records of the interaction between many users and insurance products. For example, user A inquired about a long-term critical illness insurance product at 10:00 on March 1, 2023, and the query duration was 5 minutes. Then, on March 5, 2023, he inquired about the claim cases related to the product again. This series of operations constitutes a complete first-instance product interaction behavior data. At the same time, each first-instance product interaction behavior data has corresponding first training supervision data. For example, user A's query behavior for long-term critical illness insurance products, after subsequent analysis, it is found that user A has inquired many times but has not insured, then the corresponding first training supervision data is passive tendency interest point data, indicating that the user has a certain interest but has not reached the level of active action; while for user B, he inquired about a medical insurance product and soon carried out the insurance operation, and the corresponding first training supervision data is active tendency interest point data, indicating that the user has a strong interest in the product and takes action.

[0064] Next, the basic graph neural network and the first basic interest modeling network are optimized according to the first sample product interaction behavior data sequence to generate the target graph neural network and the logistic regression network. The server loads multiple first sample product interaction behavior data in the first sample product interaction behavior data sequence into the basic graph neural network. For example, the query interaction behavior data of user A for long-term critical illness insurance is loaded, and the basic graph neural network constructs a graph structure according to each element in the interaction behavior, such as user identification, product identification, interaction type, timestamp, etc., to generate the sample interaction behavior graph structure data corresponding to the first sample product interaction behavior data of user A. This graph structure data accurately reflects the interactive relationship network between user A and the long-term critical illness insurance product. Then, the server loads the sample interaction behavior graph structure data corresponding to each first sample product interaction behavior data into the first basic interest modeling network to generate interest modeling data. The first basic interest modeling network calculates the node relationship, weight and other information in the graph structure data to obtain the preliminary modeling result about the user's interest, that is, the interest modeling data. Afterwards, the interest modeling error is calculated based on the first training supervision data and interest modeling data of each first sample product interaction behavior data. For example, for user A's passive tendency interest point data, it is compared with the interest modeling data generated by the first basic interest modeling network. If the interest modeling data shows that user A has a high interest in long-term critical illness insurance, and the actual passive tendency interest point data shows a low interest, then a certain error will be generated. At the same time, the interest modeling comparison error is calculated based on the sample interaction behavior graph structure data corresponding to each first sample product interaction behavior data and the first sample product interaction behavior data sequence. This step is to compare at the overall structural level, considering whether the difference between the graph structure data and the original data sequence will affect the interest modeling. For example, if some node relationships are incorrectly constructed when constructing the sample interaction behavior graph structure data, it may cause a discrepancy with the actual situation reflected by the original data sequence, resulting in an interest modeling comparison error. The server determines the global error parameter based on the interest modeling error and the interest modeling comparison error. This global error parameter comprehensively considers the errors of the model in data fitting and structural rationality. Finally, the basic graph neural network and the first basic interest modeling network are cyclically optimized according to the global error parameters. During the multiple cycles of optimization, the node relationship construction method and weight distribution in the basic graph neural network and the calculation parameters in the first basic interest modeling network are continuously adjusted until the predetermined optimization goal is reached, thereby generating the target graph neural network and logistic regression network.

[0065] Then, the second sample product interaction behavior data sequence is extracted from the first sample product interaction behavior data sequence. The server first extracts the passive tendency interest behavior data sequence from the first sample product interaction behavior data that covers the passive tendency interest point data in the first sample product interaction behavior data sequence. For example, among the numerous first sample product interaction behavior data, those interaction behavior data with passive tendency interest point data are screened out, such as user C has inquired about a pension insurance product many times but has not insured it, user D has viewed the terms of an accident insurance but has not taken further action, etc., and these data constitute the passive tendency interest behavior data sequence. Next, the active tendency interest behavior data sequence is extracted from the second sample product interaction behavior data that covers the active tendency interest point data in the first sample product interaction behavior data sequence. For example, user E immediately insured a life insurance product after inquiring about it, and user F quickly completed the insurance process after viewing the details of a health insurance product, etc., and these data constitute the active tendency interest behavior data sequence. Finally, the active tendency interest behavior data sequence and the passive tendency interest behavior data sequence are summarized to generate the second sample product interaction behavior data sequence. This second sample product interaction behavior data sequence contains user interaction behavior data with different interest tendencies, providing a data basis for the subsequent optimization of the second basic interest modeling network.

[0066] Finally, the target graph neural network is locked, and the second basic interest modeling network is optimized according to the second sample product interaction behavior data sequence to generate a random forest tree network. The server first locks the optimized target graph neural network, which has been optimized in the previous step and can accurately construct the graph structure of the user's interaction behavior with the insurance product. Then, the second sample product interaction behavior data sequence is loaded into the second basic interest modeling network. For example, the interaction behavior data sequence containing user C, user D, user E, user F, etc. is loaded. The second basic interest modeling network is learned and optimized according to the graph structure constructed by the target graph neural network and the specific interaction behavior information in the data sequence. In this process, the decision rules, parameters, etc. within the network are continuously adjusted to better model user interests. For example, for user C's pension insurance query uninsured behavior, the network learns the user interest characteristics under this behavior mode, and for user E's life insurance query after insurance behavior, the network also learns the interest factors behind this positive behavior. After learning and optimizing a large amount of data, a random forest tree network is finally generated. This random forest tree network can accurately model user interests based on user interaction behaviors, providing an important basis for insurance product recommendations.

[0067] In a possible implementation, step A131 includes:

[0068] Step A1311: load the first sample product interaction behavior data in the first sample product interaction behavior data sequence into the target graph neural network to generate passive tendency interest behavior graph structure data of the first sample product interaction behavior data.

[0069] Step A1312: outputting the fusion graph structure data of the passive tendency interest behavior graph structure data of all the first sample product interaction behavior data in the first sample product interaction behavior data sequence as passive tendency graph structure data.

[0070] Step A1313, respectively calculating the characteristic distance between each of the passive tendency interest behavior graph structure data and the passive tendency graph structure data.

[0071] Step A1314: output the first X first sample product interaction behavior data in ascending order of feature distance with the passive tendency graph structure data as a first extracted sample data sequence, wherein X is smaller than the sample size of the first sample product interaction behavior data in the first sample product interaction behavior data sequence.

[0072] Step A1315: randomly extract Y of the first sample product interaction behavior data and the corresponding first training supervision data from the first extracted sample data sequence as the passive tendency interest behavior data sequence, where Y is less than X.

[0073] In this embodiment, taking the scenario of user interaction with insurance products mentioned above as an example, the first example product interaction behavior data contains interaction records between many users and different insurance products. For example, user A has queried a medical insurance product twice. The query time, query duration, and specific content of each query are all included in this interaction behavior data. After loading these data into the target graph neural network, the target graph neural network constructs a graph structure for these interaction behavior data based on its own construction rules and algorithms. For user A's medical insurance query behavior, a specific graph structure will be constructed based on elements such as user identification, product identification, query operation type, and query timestamp. This graph structure is the passive tendency interest behavior graph structure data of the first example product interaction behavior data. This graph structure accurately reflects the behavior pattern of user A in the process of querying medical insurance and its relationship with other related elements.

[0074] Next, the server fuses the passive tendency interest behavior graph structure data of all the first example product interaction behavior data in the first example product interaction behavior data sequence, generates fused graph structure data, and outputs it as passive tendency graph structure data. For example, in addition to the graph structure corresponding to user A's medical insurance query behavior, there are also graph structure data corresponding to interaction behaviors such as user B viewing the details of a certain accident insurance product and user C querying a certain pension insurance product but not taking further action. These graph structure data are fused, and this fusion process involves the integration and processing of information such as nodes, node relationships, and weights in each graph structure. The fused passive tendency graph structure data can reflect the common characteristics of user interaction behaviors with passive tendency interest point data as a whole.

[0075] Then, the server calculates the characteristic distance between each passive tendency interest behavior graph structure data and the passive tendency graph structure data respectively. This calculation process is based on a specific distance calculation algorithm, such as the Euclidean distance algorithm or other distance calculation methods suitable for graph structure features. Taking the passive tendency interest behavior graph structure data corresponding to the medical insurance query behavior of user A as an example, calculate the characteristic distance between it and the passive tendency graph structure data. This distance reflects the degree of difference between the interactive behavior graph structure of user A and the overall passive tendency graph structure data. Such calculations are performed for each passive tendency interest behavior graph structure data.

[0076] Afterwards, the server outputs the first X first-example product interaction behavior data in ascending order of feature distance with the passive tendency graph structure data as the first extracted sample data sequence. Assume that there are 100 first-example product interaction behavior data containing passive tendency interest point data in the first-example product interaction behavior data sequence, and set X to 30. Then the server selects the first 30 first-example product interaction behavior data with the smallest distance based on the calculated feature distance. These data are most similar to the overall passive tendency graph structure data to some extent, and they constitute the first extracted sample data sequence. The data in this sequence has a high similarity and can represent the typical characteristics of user interaction behavior with passive tendency interest point data.

[0077] Finally, the first sample product interaction behavior data and the corresponding first training supervision data of Y are randomly extracted from the first extracted sample data sequence as a passive tendency interest behavior data sequence. Assuming that Y is 15, in the first extracted sample data sequence consisting of 30 data, the server randomly extracts 15 first sample product interaction behavior data and their corresponding first training supervision data. For example, the data related to user A's query on medical insurance, the data of user D's query on a specific insurance product, etc., and their corresponding first training supervision data representing passive tendency interest points are randomly extracted. These randomly extracted data and their corresponding supervision data constitute a passive tendency interest behavior data sequence, which will be used for subsequent operations, such as optimizing the second basic interest modeling network.

[0078] In a possible implementation, the first sample product interaction behavior data sequence includes a plurality of first sample product interaction behavior data, a plurality of first training supervision data corresponding to the first sample product interaction behavior data, and second training supervision data of the first sample product interaction behavior data covering passive tendency interest point data, the second training supervision data being one of a plurality of passive interest tendency patterns, and step A140 includes:

[0079] Step A141, obtaining a third sample product interaction behavior data sequence, wherein the third sample product interaction behavior data sequence includes multiple second sample product interaction behavior data, multiple first training supervision data corresponding to the second sample product interaction behavior data, and second training supervision data of the second sample product interaction behavior data covering passive tendency interest point data, and the second training supervision data in the third sample product interaction behavior data sequence is different from the second training supervision data of the first sample product interaction behavior data sequence.

[0080] In this embodiment, this process needs to filter out data that meets specific conditions from the database of the insurance product service application. Taking the scenario of user interaction with insurance products mentioned above as an example, the first sample product interaction behavior data sequence contains the interaction behavior data of many users and the corresponding supervision data. The third sample product interaction behavior data sequence also contains multiple second sample product interaction behavior data, the first training supervision data corresponding to these data, and the second training supervision data of the second sample product interaction behavior data covering the passive tendency point of interest data. For example, in the first sample product interaction behavior data sequence, for some users who did not take further action after querying the insurance product (passive tendency point of interest data), their second training supervision data may be expressed as a specific passive interest tendency pattern, such as because they were hesitant about the product price and did not take action. The difference between the second training supervision data in the third sample product interaction behavior data sequence and the first sample product interaction behavior data sequence may indicate that no action was taken because the product's coverage was not satisfactory. The acquisition of these data requires analysis and screening of a large number of user interaction records to ensure the accuracy and representativeness of the third sample product interaction behavior data sequence.

[0081] Step A142, summarizing the second sample product interaction behavior data sequence and the third sample product interaction behavior data sequence to generate a fourth sample product interaction behavior data sequence, wherein the sample size of the sample product interaction behavior data in the fourth sample product interaction behavior data sequence is smaller than that in the first sample product interaction behavior data sequence.

[0082] In this embodiment, this aggregation operation is not just a simple data merger, but also requires data integration and preprocessing. For example, the second sample product interaction behavior data sequence may contain user A who did not purchase insurance after querying a medical insurance product. The corresponding first training supervision data is passive tendency interest point data, and the second training supervision data indicates that the claim terms are unclear and no action was taken; the third sample product interaction behavior data sequence may contain user B who did not take further action after querying a pension insurance product. The corresponding supervision data is also passive tendency interest point data, but the second training supervision data indicates that the payment method is not satisfactory and no action was taken. When aggregating these two sequences, it is necessary to ensure that the data format is unified and the relevant information corresponds accurately. The sample size of the sample product interaction behavior data in the generated fourth sample product interaction behavior data sequence is smaller than that of the first sample product interaction behavior data sequence. This is because it is aggregated from two specific subsequences, aiming to focus on a specific type of user interaction behavior data, and provide more targeted data for the subsequent optimization of the second basic interest modeling network.

[0083] Step A143, lock the target graph neural network, and optimize the second basic interest modeling network according to the fourth sample product interaction behavior data sequence to generate the random forest tree network.

[0084] In a possible implementation, step A143 includes:

[0085] Step A1431, split the fourth sample product interaction behavior data sequence into Z branch sequences on average.

[0086] Step A1432: output the Z branch sequences as test data sequences respectively, and output the branch sequences of the Z branch sequences except the test data sequences as target learning data sequences.

[0087] Step A1433, obtaining multiple sets of reference network weight parameters of the second basic interest modeling network.

[0088] Step A1434, lock the target graph neural network, use the reference network weight parameters as the stable weight parameters of the second basic interest modeling network, and optimize the second basic interest modeling network according to the first sample product interaction behavior data sequence to generate a third basic interest modeling network corresponding to the target learning data sequence.

[0089] Step A1435, testing the third basic interest modeling network corresponding to the target learning data sequence according to the test data sequence corresponding to the target learning data sequence, and generating a network test result of the third basic interest modeling network corresponding to the target learning data sequence.

[0090] Step A1436, determining the global test results of the reference network weight parameters based on the network test results of the third basic interest modeling network corresponding to the multiple target learning data sequences.

[0091] Step A1437, outputting the reference network weight parameters whose global test results meet the target requirements as target network weight parameters.

[0092] Step A1438, lock the target graph neural network, use the target network weight parameters as the stable weight parameters of the second basic interest modeling network, and optimize the second basic interest modeling network according to the fourth sample product interaction behavior data sequence to generate the random forest tree network.

[0093] In this embodiment, the fourth sample product interaction behavior data sequence is evenly split into Z branch sequences. Assuming Z is 5, the server evenly divides the fourth sample product interaction behavior data sequence into 5 parts according to a certain rule, and each part becomes a branch sequence. This splitting is for subsequent data cross-validation and model optimization.

[0094] Next, the server outputs the Z branch sequences as test data sequences, and outputs the branch sequences other than the test data sequences as target learning data sequences. For example, if the first branch sequence is used as the test data sequence, then the remaining four branch sequences are used as target learning data sequences. This is a common model training and testing strategy, which can better evaluate the performance of the model by using a portion of the data for testing and the rest of the data for learning.

[0095] Afterwards, the server obtains multiple sets of reference network weight parameters of the second basic interest modeling network. These weight parameters may be some pre-set initial values ​​or a set of candidate parameters obtained by other means. For example, these weight parameters may be based on some empirical values ​​or the results obtained by preliminary analysis of similar data before.

[0096] Then, the server locks the target graph neural network, uses the reference network weight parameters as the stable weight parameters of the second basic interest modeling network, and optimizes the second basic interest modeling network based on the first sample product interaction behavior data sequence to generate a third basic interest modeling network corresponding to the target learning data sequence. In this process, since the target graph neural network has been locked, it provides a fixed graph structure foundation for the second basic interest modeling network. Taking the previous user interaction data as an example, the data such as user query insurance products and insurance operations in the first sample product interaction behavior data sequence are loaded into this second basic interest modeling network with stable weight parameters. According to the algorithm and structure inside the network, the data is processed and learned, and other parameters in the network are adjusted to generate the third basic interest modeling network corresponding to the target learning data sequence. This network can perform preliminary modeling of user interests based on the input user interaction behavior data, under the structural foundation and stable weight parameters of the target graph neural network.

[0097] Next, the server tests the third basic interest modeling network corresponding to the target learning data sequence according to the test data sequence corresponding to the target learning data sequence, and generates a network test result of the third basic interest modeling network corresponding to the target learning data sequence. For example, the branch sequence previously used as the test data sequence is loaded into the generated third basic interest modeling network, and the output result of the network is compared with the actual user interest (measured by the first training supervision data and the second training supervision data), and relevant evaluation indicators such as accuracy and recall are calculated, which constitute the network test result.

[0098] Afterwards, the server determines the global test results of the reference network weight parameters based on the network test results of the third basic interest modeling network corresponding to the multiple target learning data sequences. Since the fourth sample product interaction behavior data sequence was previously split into Z branch sequences, the above-mentioned test operation was performed on each branch sequence to obtain multiple network test results. By comprehensively analyzing these results, such as calculating the average of all accuracies, the weighted average of recall rates, etc., the global test results of the reference network weight parameters are obtained. This global test result can reflect the overall performance of the reference network weight parameters under different combinations of target learning data sequences and test data sequences.

[0099] Finally, the server outputs the reference network weight parameters whose global test results meet the target requirements as the target network weight parameters. For example, if the set target requirements are an accuracy rate of more than 80% and a recall rate of more than 70%, then the parameters that meet this condition are selected from all the reference network weight parameters as the target network weight parameters. Then, the target graph neural network is locked again, and the target network weight parameters are used as the stable weight parameters of the second basic interest modeling network, and the second basic interest modeling network is optimized according to the fourth sample product interaction behavior data sequence to generate a random forest tree network. In this final optimization process, all data in the fourth sample product interaction behavior data sequence are loaded into the second basic interest modeling network using the determined target network weight parameters, and comprehensive learning and adjustment are performed according to the internal algorithm and structure of the network to finally generate a random forest tree network. This random forest tree network can accurately model user interests based on the user's product interaction behavior data.

[0100] In a possible implementation, step S110 includes:

[0101] Step S111, identifying different interactive behavior record units from the product interactive behavior data, performing a decomposition operation on each interactive behavior record unit to separate the individual interactive behavior elements in each interactive behavior record unit, and generating an interactive behavior element set containing all interactive behavior elements, wherein each interactive behavior element includes an initiator of the interactive behavior, an identifier of the insurance product targeted by the interactive behavior, a specific type of the interactive behavior, and timestamp information of the occurrence of the interactive behavior.

[0102] In this embodiment, different interactive behavior record units are first identified from the product interactive behavior data. In the scenario of insurance product service application, the product interactive behavior data contains the interaction records of many users and insurance products. For example, a user queries the detailed terms of a critical illness insurance product, another user performs an insurance operation on an accident insurance product, and a user modifies the payment method of a life insurance product, etc. These different operations constitute different interactive behavior record units. For each interactive behavior record unit, the server performs a decomposition operation, separates the various interactive behavior elements therein, and then generates an interactive behavior element set containing all interactive behavior elements. Taking the interactive behavior record unit in which a user queries the detailed terms of a critical illness insurance product as an example, the initiator of the interactive behavior is the user's account identifier (such as USER123), the identifier of the insurance product targeted by the interactive behavior is the number of the critical illness insurance product (such as PROD1001), the specific type of the interactive behavior is a query operation, and the timestamp information of the interactive behavior is 10:00 on May 10, 2023. In this way, each interactive behavior element in the interactive behavior record unit is completely decomposed, and the elements obtained after decomposing all interactive behavior record units are combined to form an interactive behavior element set.

[0103] Step S112: performing a classification operation on the interactive behavior element set according to a preset classification rule to mark the category to which each interactive behavior element belongs, and generating classification marking data of the interactive behavior element set.

[0104] For example, for user identification elements (such as USER123), they are marked as categories representing user subject identities; for product information elements (such as PROD1001), they are marked as insurance product categories; for specific types of interactive behaviors such as query operations, insurance operations, and payment method modification, they are marked as behavior type categories. Through this classification operation, each interactive behavior element has a clear category affiliation, providing a basis for the nodes in the subsequent predefined graph structure.

[0105] Step S113, generating predefined node information according to the nodes in the predefined graph structure of the classification tag data of the interactive behavior element set, wherein, for user identification type elements, each user identification is defined as an independent node, and the node is used to represent the subject identity of the user in the entire interactive behavior graph; for product information type elements, a node is defined for each different insurance product, and the node serves as a representative of the insurance product in the graph structure; for behavior type type elements, each interactive behavior type is defined as a node, which is used to represent the existence of different types of interactive behaviors in the graph structure.

[0106] For user identification elements, each user identification is defined as an independent node, which represents the user's subject identity in the entire interactive behavior graph. For example, the user identification USER123 becomes a node, which represents the user subject USER123 in the process of interacting with the insurance product. For product information elements, a node is defined for each different insurance product, and this node serves as a representative of the insurance product in the graph structure. For example, the critical illness insurance product PROD1001 has a corresponding node, which reflects the existence of the product in the entire interactive behavior graph. For behavior type elements, each interactive behavior type is defined as a node to represent the existence of different types of interactive behaviors in the graph structure. Operations such as query operations and insurance operations each become a node, indicating that there are such different types of interactive behaviors in the graph structure.

[0107] Step S114, based on the interactive behavior element set of predefined node information, the relationship between each node is mined, and the node relationship is quantified according to the frequency of occurrence of the mined node relationship to assign different initial quantization values, and the initial quantization value is further adjusted and weighted according to a pre-set quantization rule to generate a node relationship carrying a quantized relationship and a weight, wherein the weight reflects the relative importance of the node relationship in the graph structure.

[0108] For example, there is a query relationship between the user node USER123 and the critical illness insurance product node PROD1001. Assuming that USER123 queries the product PROD1001 three times within a period of time, the initial quantization value can be set to 3 based on the frequency of this query relationship. Then, according to the pre-set quantization rules, if the query operation has a lower weight than the insurance operation in reflecting the user's interest in the product, for example, the weight of the query operation is set to 0.3, then the adjusted quantization value is 0.9, thus generating a node relationship with a quantization relationship (query relationship) and a weight (0.3), and this weight reflects the relative importance of the node relationship in the graph structure.

[0109] Step S115, connecting the predefined nodes according to the node relationship carrying the quantized relationship and the weight, to generate a preliminary graph structure, in which the connecting edges between the nodes represent the node relationship between the nodes, and the weight of the edge reflects the quantized value of the node relationship.

[0110] For example, there is a connection edge between the USER123 node and the PROD1001 node, and the edge weight is 0.9, which indicates the query relationship of USER123 to PROD1001 and its quantified importance.

[0111] Step S116, performing integrity check and optimization on the preliminary graph structure to generate an optimized graph structure, wherein it is specifically checked whether each node has a relationship connection with at least one other node; if isolated nodes are detected, the reasons for the isolation of the isolated nodes are analyzed; if the isolation reason is caused by missing data or data errors, repair is performed according to the corresponding interactive behavior logic and data features to supplement the node relationship information; if the isolation reason is caused by setting business logic, the isolation reason is marked; and, it is checked whether there are unreasonable relationship connections in the preliminary graph structure; if there are connections between nodes that do not conform to the business logic or edges with business errors in weight settings, the unreasonable relationship connections are adjusted and corrected.

[0112] For example, it is found that a newly defined insurance product node (assuming it is PROD2001) has no connection relationship with any user node or behavior type node. If the isolation is caused by missing data or data errors, such as the partial loss of the interactive behavior data of the product in the data record, resulting in no relationship with other nodes, then repair it according to the corresponding interactive behavior logic and data characteristics to supplement the node relationship information. If the isolation is caused by setting business logic, for example, the product is a special high-end customized insurance product that can only be accessed by a specific user group, and it is reasonable that there is no interactive behavior record for the time being, then mark this isolation reason. At the same time, the server checks whether there are unreasonable relationship connections in the preliminary graph structure. If there are connections between nodes that do not conform to the business logic or edges with business errors in the weight setting, these unreasonable relationship connections are adjusted and corrected. For example, it is found that the connection weight between a user node and a behavior type node is set too high, which does not conform to the degree of influence of the behavior type on the relationship between users and products in actual business, then the weight is adjusted.

[0113] Step S117, hierarchically dividing the optimized graph structure according to the node type, the nature of the node relationship and the business logic, and performing feature enhancement on the hierarchically divided interactive behavior graph structure to generate target interactive behavior graph structure data of the product interactive behavior data.

[0114] For example, user nodes are divided into one layer according to the node type, representing the user subject; insurance product nodes are divided into one layer, representing the insurance product object; and behavior type nodes are divided into one layer, representing the interactive behavior type. Then, the hierarchy is further subdivided or the hierarchical relationship is adjusted according to the nature of the node relationship and the business logic. When performing feature enhancement, more attribute information can be added to the node, such as adding basic information such as the user's age and gender to the user node (if this information is important in the interaction of insurance products), and adding detailed information such as the product's risk level and guarantee period to the insurance product node, thereby generating richer and more accurate target interactive behavior graph structure data, providing a more effective data foundation for subsequent operations such as user interest modeling.

[0115] In a possible implementation, step S150 includes:

[0116] Step S151, obtain candidate insurance product information from the product database in the insurance product service application, and obtain product data of each insurance product in the candidate insurance product information, wherein the product data includes one or more combinations of clause content data, coverage data, claim condition data, payment method data, and revenue model data.

[0117] For example, for a critical illness insurance product, its terms and conditions data specifies in detail the scope of insurance liability, such as the types of covered diseases (such as common major diseases such as cancer and cardiovascular disease), the insurance period (such as the coverage period is lifetime or a certain number of years), and exemption clauses (such as certain congenital diseases are not covered). The coverage data may specify the compensation ratio of the disease (such as 100% of the basic insurance amount will be paid if a certain critical illness is diagnosed), whether compensation for minor illnesses is included, and the compensation ratio for minor illnesses. The claim conditions data includes the supporting materials required for the claim (such as the hospital's diagnosis certificate, pathology report, etc.), the claim application process (such as submitting a claim application within a certain period of time after the diagnosis), and the claim review standards. The payment method data may include a one-time payment, annual payment, or monthly payment, and stipulates the payment period (such as a 20-year payment period). The income model data (if it is a critical illness insurance with financial management characteristics) may include the growth method of the cash value, the calculation method of dividends (if there are dividends), etc. For an accident insurance product, its terms and conditions data will explain the definition of accident (such as accidental death, accidental disability), the types of accidents covered (such as traffic accidents, general accidents, etc.); the coverage data includes the compensation amount for accidental death and disability, the reimbursement ratio of accidental medical expenses, etc.; the claim conditions data covers the reporting time requirements after the accident occurs, the accident proof materials required for claims, etc.; the payment method data can be a short-term one-time payment or an annual payment, etc.; the income model data (if any) includes the use of funds after the accidental disability is paid, etc. Similarly, for other types of insurance products, such as medical insurance, life insurance, etc., the server will also obtain their corresponding product data, which includes one or more combinations of terms and conditions data, coverage data, claim conditions data, payment method data, and income model data.

[0118] Step S152, classifying the product data of the insurance product according to a plurality of pre-set different classification dimensions, and generating a classified insurance product feature set corresponding to the insurance product.

[0119] Taking critical illness insurance products as an example, according to the classification dimension of coverage, it can be divided into subcategories such as disease type coverage (such as specific disease coverage, multiple major disease coverage), compensation ratio range (such as minor disease compensation ratio, serious disease compensation ratio), etc.; according to the classification dimension of claim conditions, it can be divided into subcategories such as claim material requirements (such as the type of medical certificate documents required), claim review speed (such as fast claims, regular claims), etc.; according to the classification dimension of payment method, it can be divided into subcategories such as short-term payment (such as one-time payment, 3-5 years payment), long-term payment (such as 10-20 years payment), etc.; according to the classification dimension of income model (if applicable), it can be divided into subcategories such as fixed income (such as fixed insurance amount growth) and dividend income (such as dividends based on the company's operating conditions). Combining the subcategories under these different classification dimensions, a classified insurance product feature set corresponding to critical illness insurance products is formed. For other insurance products such as accident insurance, medical insurance, and life insurance, the server also classifies them according to similar classification dimensions to obtain their corresponding classified insurance product feature sets.

[0120] Step S153, matching the user interest points in the target user interest modeling data with the classified insurance product feature set corresponding to each insurance product one by one, and generating a preliminary matching reference insurance product list based on the matching results with each insurance product.

[0121] Assume that the target user interest modeling data shows that the user has a high interest in disease protection products, especially products that include multiple major disease protection, have a high major disease compensation ratio, a fast claim review speed, and a long-term payment method (such as a 20-year payment period). When matching, the server checks whether the classified insurance product feature set of the previously obtained major disease insurance products meets these user interest points. If a certain major disease insurance product covers multiple major diseases, has a high major disease compensation ratio, a relatively fast claim review process, and a 20-year payment period, then this major disease insurance product matches the user's interest points. For accident insurance products, if the user interest modeling data shows that the user is interested in products with high accidental death and disability compensation amounts, high accidental medical expense reimbursement ratios, and a short-term one-time payment method, the server will search for products that meet these conditions in the classified insurance product feature set of accident insurance products. By matching all candidate insurance products one by one, the insurance products that meet the user's interest points are screened out, and a preliminary matching reference insurance product list is generated.

[0122] Step S154, mining the consumption capacity range of the target user based on the user portrait data of the target user, the user portrait data including the historical consumption records of the target user in the insurance product service application, the estimated income level data of the target user, and the asset status data of the target user.

[0123] For example, the target user's historical consumption records in the insurance product service application show that he had previously purchased a critical illness insurance with a coverage of 500,000 yuan, with an annual premium of 5,000 yuan, and also purchased an accident insurance with an annual premium of 500 yuan. The target user's income level estimation data shows that his annual income is approximately between 200,000 and 300,000 yuan, and the asset status data shows that he has a certain amount of savings (such as a bank deposit of 500,000 yuan), real estate (such as a real estate worth 2 million yuan), and a small amount of investment (such as stock investment worth 100,000 yuan). Combining this information, the server can determine the target user's consumption capacity range. If we look at the historical consumption records, he can accept an insurance consumption of about 5,000 to 5,500 yuan per year; from the perspective of income level and asset status, he has a certain financial strength to bear the cost of relatively high coverage and more complex insurance products, but considering his other living expenses and financial planning, his consumption capacity also has an upper limit. By comprehensively analyzing these factors, the server accurately mines the target user's consumption capacity range.

[0124] Step S155: Push insurance products to the target user based on the target user's spending capacity and the preliminarily matched reference insurance product list.

[0125] For example, in the list of reference insurance products that are initially matched, there is a high-end critical illness insurance product with a very comprehensive coverage and a high insured amount, but the annual premium is also relatively high, reaching 15,000 yuan; there is also a mid-range critical illness insurance product with a wide coverage, moderate insured amount, and an annual premium of 8,000 yuan; and a basic critical illness insurance product that covers basic major diseases, with a relatively low insured amount and an annual premium of 3,000 yuan. Combined with the target user's consumption capacity range, if the target user's consumption capacity is more suitable for products with an annual premium of 5,000-8,000 yuan, then the server will focus on pushing mid-range critical illness insurance products to users. For this product, the server can introduce in detail its coverage (such as which major diseases are included, the protection of minor diseases, etc.), claim conditions (such as the convenience of claim settlement, the required materials, etc.) and payment methods that match the user's consumption capacity (such as how to pay the annual premium of 8,000 yuan in installments, etc.), so as to improve the user's acceptance of this insurance product. For other types of insurance products, such as accident insurance, medical insurance, etc., the server also pushes targeted products in the same way according to the user's consumption capacity and the results of preliminary matching, ensuring that the pushed insurance products both meet the user's interests and are within the user's consumption capacity.

[0126] Figure 2 An insurance service system 100 provided in an embodiment of the present application is shown, including a processor 1001 and a memory 1003 and program code stored in the memory 1003, and the processor 1001 executes the above program code to implement the steps of the insurance product push method using user interest modeling.

[0127] Figure 2 The insurance service system 100 shown includes: a processor 1001 and a memory 1003. The processor 1001 and the memory 1003 are connected, such as through a bus 1002. Optionally, the insurance service system 100 may also include a transceiver 1004, which can be used for data interaction between the insurance service system and other insurance service systems, such as data transmission and / or data reception. It should be noted that in actual scheduling, the transceiver 1004 is not limited to one, and the structure of the insurance service system 100 does not constitute a limitation on the embodiments of the present application.

[0128] Processor 1001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. It may implement or execute various exemplary logic blocks, modules and circuits described in conjunction with the disclosure of this application. Processor 1001 may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0129] The bus 1002 may include a path to transmit information between the above components. The bus 1002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus. The bus 1002 may be divided into an address bus, a data bus, a control bus, and the like.

[0130] The memory 1003 may be a ROM (Read Only Memory) or other types of static storage devices that can store static information and instructions, a RAM (Random Access Memory) or other types of dynamic storage devices that can store information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory) or other optical disk storage, optical disk storage (including compressed optical disk, laser disk, optical disk, digital versatile disk, Blu-ray disk, etc.), magnetic disk storage medium, other magnetic storage devices, or any other medium that can be used to have or store program code and can be read by a computer, without limitation herein.

[0131] The memory 1003 is used to store program codes for executing the embodiments of the present application, and the execution is controlled by the processor 1001. The processor 1001 is used to execute the program codes stored in the memory 1003 to implement the steps shown in the above method embodiments.

[0132] An embodiment of the present application provides a computer-readable storage medium having program code stored thereon. When the program code is executed by a processor, the steps and corresponding contents of the aforementioned method embodiment can be implemented.

[0133] It should be understood that, although each operation step is indicated by arrows in the flowchart of the embodiment of the present application, the implementation order of these steps is not limited to the order indicated by the arrows. Unless there is clear explanation in this article, in some implementation scenarios of the embodiment of the present application, the implementation steps in each flowchart can be performed in other orders based on demand. In addition, some or all of the steps in each flowchart may include multiple sub-steps or multiple stages according to the actual implementation scenario, and some or all of these sub-steps or stages may be executed at the same time, and each sub-step or stage in these sub-steps or stages may also be executed at different times respectively. In different scenarios at the execution time, the execution order of these sub-steps or stages can be flexibly configured based on demand, and the embodiment of the present application does not limit this.

[0134] The above is only an optional implementation method for some implementation scenarios of the present application. It should be pointed out that for ordinary technicians in this technical field, without departing from the technical concept of the scheme of the present application, other similar implementation methods based on the technical ideas of the present application are also within the protection scope of the embodiments of the present application.

Claims

1. A method for pushing insurance products using user interest modeling, characterized in that: The method comprises: Acquire product interaction behavior data of target users in insurance product service applications, and construct a graph structure for the product interaction behavior data according to a target graph neural network to generate target interaction behavior graph structure data of the product interaction behavior data; Performing user interest modeling on the product interaction behavior data according to the logistic regression network and the target interaction behavior graph structure data to generate first user interest modeling data; Performing user interest modeling on the product interaction behavior data according to the random forest tree network and the target interaction behavior graph structure data to generate second user interest modeling data, wherein the logistic regression network is generated by learning network parameters according to the first sample product interaction behavior data sequence, the random forest tree network is generated by learning network parameters according to the second sample product interaction behavior data sequence, the second sample product interaction behavior data sequence is a branch sequence of the first sample product interaction behavior data sequence, and the first sample product interaction behavior data sequence and the second sample product interaction behavior data sequence are different sequences; Determining target user interest modeling data for the product interaction behavior data based on the first user interest modeling data and the second user interest modeling data; Insurance products are pushed to the target user based on the target user interest modeling data.

2. The insurance product push method using user interest modeling according to claim 1, characterized in that: The method further comprises: Acquire the first sample product interaction behavior data sequence, wherein the first sample product interaction behavior data sequence includes a plurality of first sample product interaction behavior data and corresponding first training supervision data, and the first training supervision data is active tendency interest point data or passive tendency interest point data; Optimizing the basic graph neural network and the first basic interest modeling network according to the first sample product interaction behavior data sequence to generate the target graph neural network and the logistic regression network; Extracting the second sample product interaction behavior data sequence from the first sample product interaction behavior data sequence; The target graph neural network is locked, and the second basic interest modeling network is optimized according to the second sample product interaction behavior data sequence to generate the random forest tree network.

3. The insurance product push method using user interest modeling according to claim 2, characterized in that: The step of optimizing the basic graph neural network and the first basic interest modeling network according to the first sample product interaction behavior data sequence to generate the target graph neural network and the logistic regression network includes: Loading a plurality of first sample product interaction behavior data in the first sample product interaction behavior data sequence into the basic graph neural network, and generating sample interaction behavior graph structure data corresponding to each of the first sample product interaction behavior data; Loading the sample interaction behavior graph structure data corresponding to each of the first sample product interaction behavior data into the first basic interest modeling network to generate interest modeling data; Calculating an interest modeling error based on the first training supervision data of each of the first sample product interaction behavior data and the interest modeling data; Calculating an interest modeling comparison error based on the sample interaction behavior graph structure data corresponding to each of the first sample product interaction behavior data and the first sample product interaction behavior data sequence; Determining a global error parameter according to the interest modeling error and the interest modeling comparison error; The basic graph neural network and the first basic interest modeling network are cyclically optimized according to the global error parameter to generate the target graph neural network and the logistic regression network.

4. The insurance product push method using user interest modeling according to claim 2, characterized in that: The step of extracting the second sample product interaction behavior data sequence from the first sample product interaction behavior data sequence comprises: Extracting a passive tendency interest behavior data sequence from the first sample product interaction behavior data covering the passive tendency interest point data in the first sample product interaction behavior data sequence; Extracting an active tendency interest behavior data sequence from the second sample product interaction behavior data covering active tendency interest point data in the first sample product interaction behavior data sequence; The active tendency interest behavior data sequence and the passive tendency interest behavior data sequence are aggregated to generate the second sample product interaction behavior data sequence.

5. The insurance product push method using user interest modeling according to claim 4 is characterized in that: The step of extracting the passive tendency interest behavior data sequence from the first sample product interaction behavior data covering the passive tendency interest point data in the first sample product interaction behavior data sequence comprises: Loading the first sample product interaction behavior data in the first sample product interaction behavior data sequence into the target graph neural network to generate passive tendency interest behavior graph structure data of the first sample product interaction behavior data; Outputting the fusion graph structure data of the passive tendency interest behavior graph structure data of all the first sample product interaction behavior data in the first sample product interaction behavior data sequence as passive tendency graph structure data; Respectively calculating the characteristic distance between each of the passive tendency interest behavior graph structure data and the passive tendency graph structure data; Outputting the first X pieces of the first sample product interaction behavior data in ascending order of characteristic distances from the passive tendency graph structure data as a first extracted sample data sequence, wherein X is less than the sample size of the first sample product interaction behavior data in the first sample product interaction behavior data sequence; Randomly extract Y of the first sample product interaction behavior data and the corresponding first training supervision data from the first extracted sample data sequence as the passive tendency interest behavior data sequence, where Y is less than X.

6. The insurance product push method using user interest modeling according to claim 2, characterized in that: The first sample product interaction behavior data sequence includes a plurality of first sample product interaction behavior data, a plurality of first training supervision data corresponding to the first sample product interaction behavior data, and a second training supervision data of the first sample product interaction behavior data covering passive tendency interest point data, wherein the second training supervision data is one of a plurality of passive interest tendency patterns, and the target graph neural network is locked, and the second basic interest modeling network is optimized according to the second sample product interaction behavior data sequence to generate the random forest tree network, including: Acquire a third sample product interaction behavior data sequence, wherein the third sample product interaction behavior data sequence includes a plurality of second sample product interaction behavior data, a plurality of first training supervision data corresponding to the second sample product interaction behavior data, and second training supervision data of the second sample product interaction behavior data covering passive tendency point of interest data, and the second training supervision data in the third sample product interaction behavior data sequence is different from the second training supervision data of the first sample product interaction behavior data sequence; Aggregating the second sample product interaction behavior data sequence and the third sample product interaction behavior data sequence to generate a fourth sample product interaction behavior data sequence, wherein the sample size of the sample product interaction behavior data in the fourth sample product interaction behavior data sequence is smaller than that in the first sample product interaction behavior data sequence; The target graph neural network is locked, and the second basic interest modeling network is optimized according to the fourth sample product interaction behavior data sequence to generate the random forest tree network.

7. The insurance product push method using user interest modeling according to claim 6, characterized in that: The step of locking the target graph neural network and optimizing the second basic interest modeling network according to the fourth sample product interaction behavior data sequence to generate the random forest tree network includes: Splitting the fourth sample product interaction behavior data sequence into Z branch sequences on average; Outputting the Z branch sequences as test data sequences respectively, and outputting the branch sequences of the Z branch sequences other than the test data sequences as target learning data sequences; Obtaining multiple groups of reference network weight parameters of the second basic interest modeling network; Locking the target graph neural network, using the reference network weight parameters as stable weight parameters of the second basic interest modeling network, and optimizing the second basic interest modeling network according to the first sample product interaction behavior data sequence to generate a third basic interest modeling network corresponding to the target learning data sequence; Testing a third basic interest modeling network corresponding to the target learning data sequence according to the test data sequence corresponding to the target learning data sequence, and generating a network test result of the third basic interest modeling network corresponding to the target learning data sequence; Determine a global test result of the reference network weight parameter according to the network test results of the third basic interest modeling network corresponding to the plurality of target learning data sequences; Outputting the reference network weight parameters whose global test results meet the target requirements as target network weight parameters; The target graph neural network is locked, the target network weight parameters are used as stable weight parameters of the second basic interest modeling network, and the second basic interest modeling network is optimized according to the fourth sample product interaction behavior data sequence to generate the random forest tree network.

8. The insurance product push method using user interest modeling according to claim 1, characterized in that: The step of constructing a graph structure for the product interaction behavior data according to a target graph neural network to generate target interaction behavior graph structure data for the product interaction behavior data includes: Identify different interactive behavior record units from the product interactive behavior data, perform a decomposition operation on each interactive behavior record unit to separate each interactive behavior element in each interactive behavior record unit, and generate an interactive behavior element set containing all interactive behavior elements, wherein each interactive behavior element includes an initiator of the interactive behavior, an identifier of the insurance product targeted by the interactive behavior, a specific type of the interactive behavior, and timestamp information of the occurrence of the interactive behavior; According to a preset classification rule, a classification operation is performed on the interactive behavior element set to mark the category to which each interactive behavior element belongs, and classification marking data of the interactive behavior element set is generated; Predefine the nodes in the graph structure according to the classification tag data of the interactive behavior element set, and generate predefined node information, wherein, for user identification elements, each user identification is defined as an independent node, and the node is used to represent the subject identity of the user in the entire interactive behavior graph; for product information elements, a node is defined for each different insurance product, and the node serves as a representative of the insurance product in the graph structure; for behavior type elements, each interactive behavior type is defined as a node, which is used to represent the existence of different types of interactive behaviors in the graph structure; Based on the interactive behavior element set of predefined node information, the relationship between each node is mined, and the node relationship is quantified according to the frequency of occurrence of the mined node relationship to assign different initial quantization values, and the initial quantization value is further adjusted and weighted according to a pre-set quantization rule to generate a node relationship carrying a quantization relationship and a weight, wherein the weight reflects the relative importance of the node relationship in the graph structure; Connecting predefined nodes according to the node relationships carrying the quantized relationships and weights to generate a preliminary graph structure, in which the connecting edges between the nodes represent the node relationships between the nodes, and the edge weights reflect the quantized values ​​of the node relationships; Performing integrity check and optimization on the preliminary graph structure to generate an optimized graph structure, wherein specifically checking whether each node has a relational connection with at least one other node; if an isolated node is detected, analyzing the reason for the isolation of the isolated node; if the isolation reason is caused by data missing or data error, repairing it according to the corresponding interactive behavior logic and data features to supplement the node relationship information; if the isolation reason is caused by setting business logic, marking the isolation reason; and checking whether there are unreasonable relational connections in the preliminary graph structure; if there are connections between nodes that do not conform to business logic or edges with business errors in weight settings, adjusting and correcting the unreasonable relational connections; The optimized graph structure is hierarchically divided according to the node type, the nature of the node relationship and the business logic, and the interactive behavior graph structure after the hierarchical division is feature enhanced to generate target interactive behavior graph structure data of the product interactive behavior data.

9. The insurance product push method using user interest modeling according to claim 1, characterized in that: The step of pushing insurance products to the target user based on the target user interest modeling data includes: Obtaining candidate insurance product information from a product database in the insurance product service application, and obtaining product data of each insurance product in the candidate insurance product information, wherein the product data includes one or more combinations of clause content data, coverage data, claim condition data, payment method data, and revenue model data; Classifying the product data of the insurance product according to a plurality of pre-set different classification dimensions, and generating a classified insurance product feature set corresponding to the insurance product; Matching the user interest points in the target user interest modeling data with the classified insurance product feature set corresponding to each insurance product one by one, and generating a preliminary matching reference insurance product list based on the matching results with each insurance product; Mining the target user's spending capacity range based on the target user's user portrait data, the user portrait data including the target user's historical spending records in the insurance product service application, the target user's income level estimation data, and the target user's asset status data; Insurance products are pushed to the target user based on the target user's spending capacity and the preliminarily matched reference insurance product list.

10. An insurance service system, characterized in that: It includes a processor and a computer-readable storage medium, wherein the computer-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are executed by the processor, the insurance product push method using user interest modeling as described in any one of claims 1 to 9 is implemented.

Images