Client guarantee demand analysis and insurance product recommendation method based on data elements

Abstract

The embodiment of the invention discloses customer guarantee demand analysis and insurance product recommendation based on data elements. A specific embodiment of the method comprises the steps of obtaining a user information set, an insurance policy information set, a user comment information set and an insurance product information set; performing associated user identification processing on the user information set to obtain an associated user information set; generating a user portrait label; performing knowledge graph construction on the insurance policy information set and the insurance product information set to obtain an insurance policy knowledge graph and an insurance product knowledge graph; determining user value information; performing user emotion recognition on the user comment information set to obtain user demand information; determining a user guarantee demand information set; performing insurance product recommendation on the target user to generate insurance product recommendation information; and compressing and pushing the insurance product recommendation information to a target display end for display. According to the embodiment, accurate recognition of target user preferences is improved, recommendation accuracy is improved, waste of transmission resources is reduced, and user experience is improved.

Description

Technical Field

[0001] The embodiments disclosed herein relate to the field of computer technology, and more specifically to a method for analyzing customer protection needs and recommending insurance products based on data elements. Background Technology

[0002] Currently, with the development of information technology, people's demand for insurance policies is increasing. How to accurately select the desired insurance policy from a wide variety of options has become a growing concern. The typical approach to policy recommendation is to use a collaborative filtering algorithm based on user-policy interaction to recommend insurance products to the target user, resulting in a recommended insurance product set. This recommended insurance product set is then pushed to the user's device and displayed.

[0003] However, in practice, it has been found that when recommending insurance products using the above method, the following technical problems often occur: First, because the number of policies that users have interacted with accounts for only a very small portion of the total number of policies and insurance products, and there is information on insurance products that new users have not interacted with, there is very little information that can be used as a reference. The insurance product recommendation information contains a large amount of erroneous and redundant information, resulting in a low accuracy rate of insurance product recommendations. Second, the quality of the transmission network is not taken into account, resulting in transmission delays and data packet loss, wasting a lot of transmission resources and prolonging the transmission time of recommendation information.

[0004] The information disclosed in this background section is only intended to enhance the understanding of the background of the present disclosure concept, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The summary portion of this disclosure is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description portion. This summary portion is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0006] Some embodiments of this disclosure propose a method for analyzing customer protection needs and recommending insurance products based on data elements, in order to solve one or more of the technical problems mentioned in the background section above.

[0007] In a first aspect, some embodiments of this disclosure provide a method for analyzing customer protection needs and recommending insurance products based on data elements, including: obtaining a user information set, a policy information set, and a user review information set of a target user from a user information storage database, and obtaining an insurance product information set, wherein the aforementioned user information storage database is obtained by processing and storing multi-source heterogeneous user information; performing associated user identification processing on the aforementioned user information set to obtain an associated user information set; generating user profile tags for the aforementioned target user based on the associated user information set and the aforementioned user information set; constructing knowledge graphs for the aforementioned policy information set and the aforementioned insurance product information set respectively to obtain a policy knowledge graph and an insurance product knowledge graph; determining user value information based on the aforementioned user information set and the aforementioned policy information set; and for the aforementioned user... User sentiment recognition is performed on the comment information set to obtain user demand information; the aforementioned associated user information set, user profile tags, policy knowledge graph, insurance product knowledge graph, user value information, and user demand information are determined as the user protection demand information set; based on the aforementioned user protection demand information set, insurance products are recommended to the aforementioned target users to generate insurance product recommendation information; in response to determining that the data volume of the aforementioned insurance product recommendation information is greater than or equal to a preset transmission data volume threshold, network performance testing is performed on the server sending the aforementioned insurance product recommendation information to obtain network performance testing information; and in response to determining that the transmission delay duration represented by the aforementioned network performance testing information is greater than or equal to a preset delay duration threshold, the aforementioned insurance product recommendation information is compressed and pushed to the target display terminal to display the insurance product push information in a personalized manner.

[0008] Secondly, some embodiments of this disclosure provide a data-based customer protection needs analysis and insurance product recommendation apparatus, comprising: an acquisition unit configured to acquire a user information set, a policy information set, and a user review information set of a target user from a user information storage database, and to acquire an insurance product information set, wherein the aforementioned user information storage database is obtained by processing and storing multi-source heterogeneous user information; an associated user identification unit configured to perform associated user identification processing on the aforementioned user information set to obtain an associated user information set; a generation unit configured to generate user profile tags for the target user based on the aforementioned associated user information set and the aforementioned user information set; a knowledge graph construction unit configured to construct knowledge graphs for the aforementioned policy information set and the aforementioned insurance product information set respectively to obtain a policy knowledge graph and an insurance product knowledge graph; and a first determination unit configured to determine user value information based on the aforementioned user information set and the aforementioned policy information set; The first unit is configured to perform user sentiment recognition on the aforementioned user comment information set to obtain user demand information; the second unit is configured to determine the aforementioned associated user information set, the aforementioned user profile tags, the aforementioned policy knowledge graph, the aforementioned insurance product knowledge graph, user value information, and the aforementioned user demand information as a user protection demand information set; the second unit is configured to recommend insurance products to the aforementioned target user based on the aforementioned user protection demand information set to generate insurance product recommendation information; and the third unit is configured to, in response to determining that the data volume of the aforementioned insurance product recommendation information is greater than or equal to a preset transmission data volume threshold, perform network performance testing on the server sending the aforementioned insurance product recommendation information to obtain network performance testing information, and in response to determining that the transmission delay duration represented by the aforementioned network performance testing information is greater than or equal to a preset delay duration threshold, compress and push the aforementioned insurance product recommendation information to the target display terminal to display the insurance product push information in a personalized manner.

[0009] Thirdly, some embodiments of this disclosure provide an electronic device, including: one or more processors; and a storage device having one or more programs stored thereon, such that when the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any implementation of the first aspect.

[0010] Fourthly, some embodiments of this disclosure provide a computer-readable medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method as described in any implementation of the first aspect.

[0011] The above embodiments of this disclosure have the following beneficial effects: The customer protection needs analysis and insurance product recommendation method based on data elements in some embodiments of this disclosure improves the data quality and accuracy of user information sets and the identification of associated users by processing multi-source heterogeneous user information sets, thereby improving the accurate identification of target user preferences, increasing the accuracy of policy recommendations, reducing the waste of transmission resources, and improving user experience. Specifically, the reasons for the low accuracy of related insurance product recommendations, the waste of a large amount of transmission resources, and the extended transmission time of recommendation information are as follows: Since the number of policies that users have interacted with accounts for only a very small portion of the total number of policies and insurance products, and there is insurance product information that new users have not interacted with, there is limited information that can be used as a reference. The insurance product recommendation information contains a large amount of erroneous and redundant information, resulting in a low accuracy of insurance product recommendations. Furthermore, the quality of the transmission network is not considered, leading to transmission delays and data packet loss, wasting a large amount of transmission resources and extending the transmission time of recommendation information. Based on this, some embodiments of the policy recommendation method disclosed herein can first obtain the target user's user information set, policy information set, and user review information set, as well as the insurance product information set, from a user information storage database. The aforementioned user information storage database is obtained by processing and storing multi-source heterogeneous user information. Processing and storing diverse heterogeneous data across multiple nodes reduces storage resource waste and facilitates subsequent processing of the user information set, policy information set, and user review information set, as well as product insurance recommendations. Secondly, the aforementioned user information set undergoes associated user identification processing to obtain an associated user information set. This associated user identification identifies users closely related to the target user, enabling precise control over the target user's preferences and avoiding cold start and data sparsity issues in user recommendations. Thirdly, based on the associated user information set and the aforementioned user information set, user profile tags for the target user are generated. These user profile tags accurately identify user behavior and preferences, improving the accuracy of subsequent policy recommendations. Next, knowledge graphs are constructed for the aforementioned policy information set and insurance product information set, resulting in policy knowledge graphs and insurance product knowledge graphs, respectively. This enhances the target user's understanding of policy information and insurance products, and facilitates subsequent policy recommendations based on the user's historical insurance product selections, thereby improving the accuracy of insurance product recommendations and enhancing user experience. Subsequently, user value information is determined based on the aforementioned user information set and policy information set. Determining user value information helps identify the user's importance, allowing for more targeted resource allocation and minimizing resource waste. Finally, user sentiment analysis is performed on the aforementioned user comment information set to obtain user demand information. This allows for a precise understanding of the user's needs regarding policy information, further improving the accuracy of policy recommendations.Then, the aforementioned set of associated user information, user profile tags, policy knowledge graph, insurance product knowledge graph, user value information, and user demand information are defined as the user protection demand information set. Here, analyzing user protection needs from multiple dimensions based on data elements improves the accuracy and comprehensiveness of the user protection demand information set. Next, based on the aforementioned user protection demand information set, insurance products are recommended to the target users to generate insurance product recommendation information. Here, policy recommendations using multi-dimensional data improve the comprehensiveness and accuracy of the recommendations. Finally, in response to the determination that the data volume of the aforementioned insurance product recommendation information is greater than or equal to a preset transmission data volume threshold, network performance testing is performed on the server sending the insurance product recommendation information to obtain network performance test information. Furthermore, in response to the determination that the transmission delay duration represented by the aforementioned network performance test information is greater than or equal to a preset delay duration threshold, the aforementioned insurance product recommendation information is compressed and pushed to the target display terminal for personalized display of the insurance product push information. Here, the data volume of transmitted insurance product recommendation information can be reduced, transmission efficiency improved, transmission resource waste and data transmission time reduced, thereby improving the user experience. Therefore, this policy push method improves the data quality and accuracy of user information sets by processing multi-source heterogeneous user information sets, as well as the identification of related users, thereby improving the accurate identification of target user preferences, improving the accuracy of policy recommendations, reducing the waste of transmission resources, and improving the user experience. Attached Figure Description

[0012] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0013] Figure 1 This is a flowchart of some embodiments of the data element-based customer protection needs analysis and insurance product recommendation method according to this disclosure;

[0014] Figure 2 This is a schematic diagram of a policy knowledge graph displayed in the data element-based customer protection needs analysis and insurance product recommendation method disclosed herein;

[0015] Figure 3 This is a schematic diagram showing user value information in the data element-based customer protection needs analysis and insurance product recommendation method disclosed herein;

[0016] Figure 4 This is a schematic diagram of the structure of some embodiments of the data element-based customer protection needs analysis and insurance product recommendation device according to the present disclosure;

[0017] Figure 5 This is a schematic diagram of the structure of an electronic device suitable for implementing some embodiments of the present disclosure. Detailed Implementation

[0018] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0019] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0020] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0021] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0022] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0023] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0024] Figure 1 A flow 100 of some embodiments of the policy push method according to this disclosure is shown. The policy push method includes the following steps:

[0025] Step 101: Obtain the target user's user information set, policy information set, and user review information set, as well as the insurance product information set, from the user information storage database.

[0026] In some embodiments, the executing entity (e.g., an electronic device) of the above-mentioned data element-based customer protection needs analysis and insurance product recommendation method can obtain the target user's user information set, policy information set, and user review information set, as well as the insurance product information set, from a user information storage database via wired or wireless connection. The user information storage database is obtained by processing and storing multi-source heterogeneous user information. The user information in the user information set can be information related to the target user. For example, the user information set may include, but is not limited to, at least one of the following: basic user information, user policy purchase behavior information, health status information, and consumption level information. The policy information in the policy information set can be information on the contract certificate formed between the user and a third-party company after the user purchases an insurance product. For example, the policy information set may include, but is not limited to, at least one of the following: policy type, policy premium information, policy number, and policy terms information. The user review information in the user review information set can characterize the target user's emotional inclination towards the policies they have interacted with. The user information storage database can be a database that transmits user information through multiple data interfaces, processes and privatizes the transmitted user information, and then stores it. The insurance product information in the aforementioned insurance product information collection can be contractual products that provide intangible services to protect users in dealing with risks.

[0027] In some optional implementations of certain embodiments, the aforementioned user information storage database may be obtained through the following steps:

[0028] The first step is to acquire a multi-source heterogeneous user information set from different sensors. These different sensors can be storage devices from different sources that store data related to users and policies.

[0029] The second step involves preprocessing each piece of multi-source heterogeneous user information in the aforementioned multi-source heterogeneous user information set according to a preset rule engine, resulting in a preprocessed multi-source heterogeneous user information set. The preset rule engine can be a pre-defined engine that executes data preprocessing rules. The rules included in the preset rule engine may include, but are not limited to, at least one of the following: converting ID card numbers to 18 digits, converting full-width characters to half-width characters, converting letters to uppercase, removing leading and trailing spaces; for name data, removing leading and trailing spaces for English names, removing leading and trailing spaces for Chinese names, removing punctuation marks, converting full-width characters to half-width characters, ensuring non-empty strings, and ensuring the length is greater than one Chinese character.

[0030] The third step involves generating global identifier information for each preprocessed heterogeneous user information in the aforementioned preprocessed multi-source user information set, resulting in a global identifier information set. This global identifier information can uniquely represent the data. In practice, the executing entity can first, in response to the ID card type and ID number conforming to the 15- or 18-digit ID card verification rule, convert the ID number into an 18-digit ID card number and extract the user's gender and date of birth from the 18-digit ID card number. Then, through a preset merging rule, the user information set is integrated to obtain the integrated user data. This preset merging rule can be: matching fields based on user name, ID card type, ID card number, gender, and date of birth, or CIF (Customer Information File), name, ID card number, gender, and date of birth. If all fields match or all fields except the ID card number match, the user is identified as the same user, and multiple data entries corresponding to that user are integrated. Finally, a 30-character string obtained by concatenating the current timestamp and sequence number is used as the global identifier information set. The aforementioned serial number can be determined using a database sequence (PostgreSQL SEQUENCE), reset daily at midnight, or continuously incremented using the Snowflake algorithm.

[0031] The fourth step involves performing semantic disambiguation on the preprocessed multi-source heterogeneous user information set to obtain a semantically disambiguated multi-source heterogeneous user information set. This semantic disambiguation can be performed using a dictionary-based method, such as the Lesk algorithm.

[0032] Fifth, based on the aforementioned preset rule engine, determine the user data quality of each piece of preprocessed multi-source heterogeneous user information in the preprocessed multi-source heterogeneous user information set, thus obtaining a user data quality set. The user data quality characterizes the accuracy, completeness, consistency, and reliability of the preprocessed multi-source heterogeneous user information. The aforementioned preset rule engine can determine user data quality through a weighted sum of the credibility of the data source and the data acquisition time. The credibility of the data source can be determined by the highest quality data obtained from legitimate websites, followed by data obtained from other channels. The rule for determining user quality based on acquisition time can be that the closer the acquisition time is to the current time, the higher the data quality.

[0033] Step 6: Based on the aforementioned global identifier information set and user data quality set, perform data integration processing on the semantically disambiguated multi-source heterogeneous user information set to obtain an integrated user information set. In practice, the executing entity can first integrate the data of each semantically disambiguated multi-source heterogeneous user information in the aforementioned semantically disambiguated multi-source heterogeneous user information set using the global identifier information set to obtain an initial integrated user information set. Then, using the user data quality set, perform data overlay on the initial integrated user information set to obtain the integrated user information set. This data overlay can involve high-quality user data overlaying low-quality user data, or non-empty data overlaying empty data.

[0034] The seventh step is to perform multi-node balanced storage on the integrated user information set to obtain a user information storage database. In practice, the executing entity can first determine the business scenario for each new piece of information about the integrated user in the integrated user information set, thus obtaining a business scenario set. Then, using the business scenario set, the integrated user information set is stored in the corresponding nodes to obtain the user information storage database.

[0035] In addressing the first technical problem mentioned above, a second technical problem often arises: how to achieve balanced load distribution of storage resources across multiple data nodes when storing multi-source heterogeneous user information sets on multiple nodes, thereby reducing storage resource waste and improving data node stability. A conventional solution for this second technical problem is to use the bat optimization algorithm to perform balanced storage across multiple nodes on the integrated user information set, resulting in a user policy storage database. However, this conventional solution still suffers from the following issues: the bat optimization algorithm uses random initialization to obtain the initial population, leading to significant randomness and potentially causing the algorithm's optimization efficiency to decrease with increasing execution counts, resulting in local optima; unreasonable resource allocation among multiple data storage nodes, leading to significant storage resource waste; uneven load distribution among multiple data storage nodes; unstable cluster performance; increased server node damage; and reduced user experience. Considering the shortcomings of conventional solutions and leveraging the advantages and current state of multi-node data storage technology within our company, we have decided to adopt the following solution:

[0036] Optionally, the above-mentioned multi-node balanced storage of the integrated user information set to obtain the user policy storage database may include the following steps:

[0037] The first step is to obtain the node resource usage status information of each data storage node in the multiple data storage nodes included in the aforementioned user policy storage database, thus obtaining a node resource usage status information set. This node usage status information characterizes the resource usage of the data storage nodes, specifically the usage of resources such as memory and CPU.

[0038] The second step involves performing chaotic initialization processing on the multiple data storage nodes based on the aforementioned node resource usage status information set, resulting in an initial node population. Each initial node in this population comprises: an initial node position vector, an initial node velocity vector, an initial node acoustic frequency vector, an initial node impulse loudness vector, and an initial node impulse frequency vector. These initial nodes can represent a multi-node balanced storage scheme involving multiple data storage nodes. The initial node position vector represents the search position of each storage node in the multi-node balanced storage scheme. The initial node velocity vector represents the movement speed of the initial node in the search space. The initial node acoustic frequency vector represents the frequency of the acoustic waves emitted by the initial node, used to control the search range and accuracy. The initial node impulse loudness vector represents the intensity of the acoustic waves emitted by the initial node, used to measure the satisfaction level of the node's initial position or the motivation to explore new positions. The initial node impulse frequency vector represents the frequency with which the node emits acoustic waves. As an example, the aforementioned execution entity uses a chaotic algorithm to perform chaotic initialization processing on the aforementioned multiple data storage nodes based on the aforementioned node resource usage status information set, thereby obtaining an initial node population.

[0039] The third step involves generating a node fitness function for the initial node population. This function includes a target fitness function and a set of constraint functions. The target fitness function can be a function that minimizes the weighted sum of load balancing, resource utilization, data response time, and system energy consumption across multiple data storage nodes. The set of constraint functions can include: data response time cannot exceed a preset maximum tolerance period; the proportion of channel bandwidth resources cannot exceed 1; the proportion of allocated CPU, memory, and disk resources cannot exceed 1; and the space required for the integrated user information set cannot exceed the total space corresponding to the user policy storage database.

[0040] Fourth, based on the initial node population, perform the following node population update steps:

[0041] Sub-step 1: Input the initial position set of nodes included in the initial node population into the above node fitness function to obtain the initial node fitness value set.

[0042] Sub-step 2: Select the initial node fitness value with the smallest value from the initial node fitness value set, and use it as the initial target node fitness value.

[0043] Sub-step 3: For each initial node individual included in the initial node population, perform the following node individual update steps:

[0044] The first sub-step involves updating the initial node individuals based on the initial target node fitness value and the initial node fitness value set, resulting in updated node individuals. These updated node individuals can include those obtained by updating the velocity and position of the initial nodes. For example, the execution entity can first determine the initial position of the optimal node individual corresponding to the initial target node fitness value and the position difference vector between the initial positions of the nodes corresponding to the initial node fitness value, as the node individual position difference vector. Secondly, it determines the sum of the difference between the maximum and minimum value vectors of the node's initial acoustic frequency, multiplied by a random acoustic frequency vector, and summed with the minimum value vector, as the updated node acoustic frequency. This random acoustic frequency vector can be a uniformly distributed random vector between 0 and 1. Thirdly, it determines the sum of the node individual position difference vector, the updated node acoustic frequency vector, and the node's initial velocity vector, as the updated node individual velocity vector. Finally, it determines the sum of the chaotic control parameters and the node individual search range length, multiplied by the updated node individual velocity vector and the node's initial position vector, as the target node individual vector. The aforementioned chaos control parameters can be used to control the properties of chaotic noise added during the global search process. Finally, the target node vector is input into the chaotic mapping function to obtain the updated node.

[0045] The second sub-step involves generating a first random value and a second random value. These first and second random values ​​can be random values ​​generated using a randomization algorithm.

[0046] The third sub-step, in response to determining that the first random value is greater than the initial pulse frequency of the node corresponding to the initial node, performs a global and local combined random walk search on the updated node's position vector, obtaining a locally updated node position vector. This locally updated node position vector can be achieved by performing a local auxiliary update after a global update of the position vectors in the updated node, avoiding the possibility of getting trapped in local optima due to insufficient random perturbation when relying solely on global position updates. The local search, by introducing an additional perturbation mechanism, can generate new solutions near the local optimum, increasing the probability of escaping the local optimum. In practice, the aforementioned execution entity can determine the product of the third and fourth random numbers and add it to the updated node's position vector to obtain the locally updated node position vector. The third random number can be a random number within the range [0, 1]. The fourth random number can be a random number within the range [0, initial pulse loudness].

[0047] The fourth sub-step, in response to determining that the second random value is less than the size of the initial impulse loudness vector corresponding to the initial node, and the updated fitness value is less than the initial target node fitness value, updates the initial impulse loudness vector and initial impulse frequency vector corresponding to the initial node individual, obtaining the updated node impulse loudness vector and updated node impulse frequency vector. The updated node impulse loudness vector can be the product of a preset impulse constant and the initial node impulse loudness vector. The preset impulse constant can be a pre-defined constant less than 1. The updated node impulse frequency vector can be obtained through the following steps: First, determine an exponential function with base e and the negative of the product of the impulse frequency enhancement coefficient and the number of executions as the exponent, as the impulse frequency exponential function. The impulse frequency enhancement coefficient can be a constant ranging from [0, 1], used to control the increase of the impulse emission rate over time. A smaller value indicates a larger search range, and the initial node tends to perform a local search around the optimal solution; a larger value indicates a smaller search range, and the initial node tends to perform a global search. Then, the difference between 1 and the aforementioned pulse frequency exponential function is determined as the pulse frequency difference. Finally, the product of the node's initial pulse frequency vector and the aforementioned pulse frequency difference is determined as the updated node pulse frequency vector.

[0048] The fifth sub-step involves inputting the updated node velocity vector, locally updated node position vector, updated node impulse loudness vector, and updated node impulse frequency vector—all included in the updated node individual—into the node fitness function to obtain the updated node fitness value set.

[0049] The sixth sub-step involves selecting the smallest value from the updated set of node fitness values ​​and the initial set of node fitness values, and using this value as the selected node fitness value.

[0050] The seventh sub-step is to determine the number of times the above node population update steps have been executed.

[0051] Sub-step 4, in response to determining that the number of executions is greater than or equal to a preset execution count threshold, performs multi-node balanced storage on the integrated user information set according to the weight values ​​corresponding to the fitness values ​​of the filtered nodes, resulting in a user policy storage database. The preset execution count threshold can be a pre-set maximum number of executions. For example, the preset execution count threshold could be 100 times.

[0052] Fifth, in response to the determination that the number of executions is less than a preset execution threshold, the updated node fitness value set is determined as the initial node fitness value set, and the sum of the number of executions and the preset value is determined as the number of executions, so as to execute the above node population update step again. The preset value can be a pre-defined value. For example, the preset value can be 1.

[0053] The above-described technical solution and its related content, as an inventive point of this disclosure, solve the second technical problem mentioned in the background: "Because the bat optimization algorithm uses random initialization to obtain the initial population, it has a high degree of randomness, which easily leads to lower optimization efficiency as the number of executions increases, causing it to get trapped in local optima. Furthermore, unreasonable resource allocation among multiple data storage nodes results in a large waste of storage resources, uneven load on multiple data storage nodes, unstable cluster performance, increased damage to server nodes, and a reduced user experience." The factors leading to a large waste of storage resources, uneven load on multiple data storage nodes, unstable cluster performance, increased damage to server nodes, and a reduced user experience are often as follows: Because the bat optimization algorithm uses random initialization to obtain the initial population, it has a high degree of randomness, which easily leads to lower optimization efficiency as the number of executions increases, causing it to get trapped in local optima. Furthermore, unreasonable resource allocation among multiple data storage nodes results in a large waste of storage resources, uneven load on multiple data storage nodes, unstable cluster performance, increased damage to server nodes, and a reduced user experience. Solving the above factors can reduce storage resource waste, improve load balancing across multiple data storage nodes, enhance cluster performance stability, reduce server node failures, and improve user experience. To achieve this, this disclosure proposes an energy-minimizing optimization problem by jointly considering migration decisions, caching decisions, channel bandwidth allocation, and the distribution of computing resources among fog nodes, ensuring load balancing across multiple data storage nodes. Based on the Bat Algorithm, chaotic sequences and random mutation strategies are incorporated to enhance the algorithm's scalability and flexibility, improve its global convergence and local search capabilities, reduce storage resource waste, improve load balancing across multiple data storage nodes, enhance cluster performance stability, reduce server node failures, and improve user experience.

[0054] Step 102: Perform associated user identification processing on the user information set to obtain the associated user information set.

[0055] In some embodiments, the executing entity may perform associated user identification processing on the aforementioned user information set to obtain an associated user information set. The associated user information in the associated user information set may be users who are associated with the aforementioned target user.

[0056] In addressing the first technical problem mentioned above, a third technical problem often arises: how to determine the social network graph of the target user to more accurately identify their preferences and improve policy recommendation accuracy. A conventional solution to this third technical problem is to use graph contrastive learning to embed each node in the target user's social network with equal edge weights, resulting in a set of node embedding vectors. Then, a clustering algorithm is used to cluster these vectors to obtain associated user information. However, this conventional solution still suffers from the following problems: because the social network graph contains a large number of nodes and edges, embedding with equal edge weights through graph contrastive learning can lead to inconsistent influences between users with minimal and significant interactions with the target user, resulting in significant noise and an inability to accurately extract crucial semantic information. Furthermore, the clustering algorithm only considers node influence, leading to redundant errors, consuming substantial computational resources, reducing computational efficiency, and lowering the accuracy of user association identification. Considering the shortcomings of conventional solutions and leveraging the advantages and current state of multi-node data storage technology within our company, we have decided to adopt the following solution:

[0057] In some optional implementations of certain embodiments, performing associated user identification processing on the above-mentioned user information set to obtain an associated user information set may include the following steps:

[0058] The first step is to generate a social network relationship graph for the target user. In this graph, nodes represent users associated with the target user, and edges represent the relationships between users and the target user. Each node includes a set of node attribute information. This set of attribute information may include, but is not limited to, at least one of the following: user name, gender, address, and behavioral preferences. The social network relationship graph can be a graph representing other users who are associated with the target user.

[0059] The second step is to determine the set of node degrees and node traversal weights for each social node in the aforementioned social network graph. The node degree in the degree set can be the number of edges connecting the social node. The node traversal weights in the degree centrality set can be node weights determined using the degree centrality formula. These node traversal weights characterize the importance of the social nodes.

[0060] The third step involves generating an edge sampling probability set for each of the aforementioned social edges, based on the node degree set and the node traversal weight set. The edge sampling probability in this set represents the probability that a social edge will be sampled.

[0061] As an example, the aforementioned execution entity can perform the following determination steps for each social edge in each social edge: First, determine the average of the node degree in the node degree set and the node walk weight value in the node walk weight value set, thus defining the social node value set of the social node set. Then, determine the average of the two social node values ​​of the two social nodes associated with the social edge, as the initial social edge value. Finally, normalize the initial social edge value to obtain a normalized value, which is then determined as the edge sampling probability.

[0062] The fourth step is to determine the data type set of the attribute information of each node in the aforementioned social network relationship graph. The data types in this set can include: sparse one-hot types and dense floating-point types.

[0063] Fifth, based on the aforementioned data type set, generate an attribute sampling probability set for each node's attribute information. The attribute sampling probability in this set can be the probability value of whether the attribute information included in the social node has been truncated.

[0064] As an example, the aforementioned execution entity can perform the following processing steps for each social node: First, in response to determining that the data type is a sparse one-hot type, the importance of the node attribute information of the aforementioned social node in the aforementioned social network relationship graph is determined using the TF-IDF algorithm, as the attribute sampling probability. Then, in response to determining that the data type is a dense floating-point type, the L2 norm of the node attribute information of the aforementioned social node is determined, as the attribute sampling probability.

[0065] Step 6: Based on the aforementioned edge sampling probability set and attribute sampling probability set, the aforementioned social network graph is augmented to obtain an edge-enhanced social network graph and an attribute-enhanced social network graph. The edge-enhanced social network graph can be a social graph obtained by pruning the edges of the social network graph using the edge sampling probability set. The attribute-enhanced social network graph can be a social graph obtained by pruning the attributes of the social network graph using the attribute sampling probability set.

[0066] As an example, the aforementioned execution entity can first create an initial node mask matrix for the aforementioned social network graph. This initial node mask matrix can be a matrix where all values ​​are 1. Next, for each social edge included in the aforementioned social network graph, a random number with a value between [0, 1] is generated to obtain the social edge random number. In response to determining that the social edge random number is greater than the corresponding edge sampling probability, the social edge is deleted, thus obtaining an edge-enhanced social network graph. Then, for each social node included in the aforementioned social network graph, a random number with a value between [0, 1] is generated to obtain the node random number. In response to determining that the node random number is greater than the corresponding attribute sampling probability, the random part of the row vector of the social node in the aforementioned initial node mask matrix is ​​set to 0, obtaining an attribute mask vector matrix. The attribute information corresponding to the attribute mask values ​​with values ​​of 0 in the attribute mask vector matrix is ​​deleted from the aforementioned social network graph, resulting in an attribute-enhanced social network graph.

[0067] Step 7: Perform graph embedding processing on the edge-enhanced social network graph and the attribute-enhanced social network graph respectively to obtain the edge-enhanced graph embedding vector set and the attribute-enhanced graph embedding vector set. The graph embedding processing can be performed using GraphSAGE (Graph Sample and AggreGatE).

[0068] Step 8: Input the aforementioned edge-enhanced graph embedding vector set and attribute-enhanced graph embedding vector set into the contrastive learning model to obtain a user social network cluster set. The user social network clusters in this set can be clusters formed by users with close relationships. The contrastive learning model can be a model that performs node-level and community-level comparisons on the input edge-enhanced graph embedding vector set and attribute-enhanced graph embedding vector set to perform community detection processing on the social network relationship graph and output user clusters. The node-level comparison can be a comparison of the node-level feature vector set obtained by inputting the aforementioned edge-enhanced graph embedding vector set and attribute-enhanced graph embedding vector set into the node-level comparison model. The node-level comparison model can be a multilayer perceptron. The multilayer perceptron can be a network model including two linear layers and a ReLU activation function. The community-level comparison can be a comparison learning model corresponding to the attribute-enhanced graph embedding vector set, where the input node-level feature vector set is input into the community-level comparison model to output user clusters. The community-level comparison model can also be a model that first passes through a multilayer perceptron for projection and then inputs it into a Softmax layer for probability distribution calculation and prediction.

[0069] The ninth step is to filter out the user social network clusters that include the target users from the above user social network clusters, obtain the target user social network clusters, and determine the information of each user included in the target user social network clusters as the associated user information set.

[0070] The above-described technical solution and its related content, as an inventive point of this disclosure, solve the third technical problem mentioned in the background: "Because social network relationship graphs contain a large number of nodes and edges, embedding learning through graph comparison learning with edge weights in the same way will result in users with very little interaction with the target user and users with a lot of interaction having the same influence on the target user, resulting in a large amount of noise, failing to accurately extract important semantic information, and clustering algorithms only considering the influence of nodes, leading to a large amount of redundant error information, consuming a large amount of computing resources, reducing computing efficiency, and resulting in low accuracy in identifying user relationships." This leads to a large amount of redundant error information, consumes a large amount of computing resources, reduces computing efficiency, and results in low accuracy in identifying user relationships. If the above factors are solved, it is possible to reduce a large amount of redundant error information, reduce the waste of computing resources, and improve computing efficiency and accuracy in identifying user relationships. To achieve this effect, this disclosure first uses generated edge sampling probability sets and attribute sampling probability sets to enhance the social network relationship graph, which can reduce a large amount of redundant error data in the social network relationship graph. Then, through a contrastive learning model, node-level and social-level comparative learning is performed on the edge-enhanced graph embedding vector set and the attribute-enhanced graph embedding vector set. They jointly learn node-level and community-level information, dynamically adjusting the weights of the edge-enhanced and attribute-enhanced social network graphs. This enables accurate user community discovery, improves the identification of user relationships, reduces computational resource waste, and enhances the efficiency and accuracy of user relationship identification. Finally, by associating user information sets with target users for insurance product recommendations and terminal push notifications, transmission resource waste can be reduced and user experience improved.

[0071] Step 103: Generate user profile tags for the target user based on the associated user information set and the user information set.

[0072] In some embodiments, the aforementioned executing entity can generate user profile tags for the target user based on the aforementioned associated user information set and the aforementioned user information set. These user profile tags can represent the user's multi-dimensional characteristics, behaviors, and preferences. The user profile tags can include, but are not limited to, at least one of the following: basic user information, user purchasing behavior, source of the user's policy acquisition, and user risk information. The aforementioned basic user information can include, but is not limited to, at least one of the following: user name, identification number, whether the user is an insured, and whether the user has been a bancassurance customer within the past year. The aforementioned user purchasing behavior can include the type of bancassurance policy purchased by the user, the source of the policy, cross-purchasing of different types of policies, and the number of policies purchased. The aforementioned user risk information can represent the target user's risk level. The aforementioned user risk information can be information obtained by inputting the aforementioned user information set into a scoring card model.

[0073] As an example, the aforementioned execution entity can first perform clustering processing on the aforementioned associated user information set and the aforementioned user information set to obtain a user cluster set. Then, it can filter out the cluster to which the target user belongs from the aforementioned user clusters to obtain the target user cluster. Finally, it can use an autoencoder to tag the aforementioned target user cluster to obtain the user profile tags of the target user.

[0074] Step 104: Construct knowledge graphs for the policy information set and the insurance product information set respectively to obtain the policy knowledge graph and the insurance product knowledge graph.

[0075] In some embodiments, the executing entity may construct knowledge graphs for the policy information set and the insurance product information set respectively, resulting in a policy knowledge graph and an insurance product knowledge graph. The policy knowledge graph may be a graph representing the relationships between policies and the relationships between policy attribute information. Figure 2 As shown, Figure 2 It illustrates the relationships between different insurance policies and users, as well as between users themselves. The aforementioned product knowledge graph can be presented in graphical form as a diagram showing the relationships between insurance products and the relationships between their attribute information.

[0076] In some optional implementations of certain embodiments, the above-mentioned construction of knowledge graphs for the policy information set and the insurance product information set to obtain the policy knowledge graph and the insurance product knowledge graph may include the following steps:

[0077] The first step is to construct a policy ontology model for the aforementioned policy information set. This policy ontology model can represent concepts, attributes, and the logical relationships between concepts. Specifically, the policy ontology model can be a pre-constructed relationship graph related to policy information. The aforementioned concepts can be abstract representations of domain knowledge. For example, the aforementioned concepts can include, but are not limited to, at least one of the following: insurance product, policyholder, policy company, and claims terms. The aforementioned attributes can be specific descriptive information about the concepts. For example, the aforementioned attributes can include, but are not limited to, at least one of the following: policy type, policy term, policy number, policyholder name, and policyholder contact information.

[0078] The second step involves performing the following knowledge fusion steps for each policy information in the aforementioned policy information set:

[0079] Sub-step 1 involves performing a format conversion process on the aforementioned policy information to obtain policy text information. This format conversion process can involve converting all types of policy information into text format.

[0080] Sub-step 2 involves inputting the aforementioned policy text information into the feature representation layer of the policy knowledge extraction model to obtain a set of semantic feature vectors for the policy text. This policy knowledge extraction model further includes a context feature representation layer, a semantic temporal extraction layer, and a label prediction layer. The policy knowledge extraction model can be a model that extracts entity and attribute words from the input policy text information. The semantic feature vectors in the aforementioned policy text semantic feature vector set can represent the semantic information of the policy. The feature representation layer can be a BERT model. This feature representation layer can fully learn the meaning of each word segment and can also solve the problem of the model's inability to learn semantics well due to the small training set size during feature extraction in the context feature representation layer and the semantic temporal extraction layer. The context feature representation layer can be a network layer using an IDCNN model to learn the global semantic information of the policy information through dilated convolution. The semantic temporal extraction layer can be a network layer using a recurrent neural network model with long short-term memory to learn the global semantic information of the policy information. The aforementioned label prediction layer can be a network layer that uses a CRF (Conditional Random Field) model to predict the feature vectors output by the semantic temporal extraction layer.

[0081] Sub-step 3 involves inputting the aforementioned policy text semantic feature vector set into the aforementioned context feature representation layer to obtain the policy semantic context feature vector set.

[0082] Sub-step 4: Input the above policy semantic context feature vector set into the above semantic temporal extraction layer to obtain the semantic temporal feature vector set.

[0083] Sub-step 5 involves inputting the aforementioned semantic temporal feature vector set into the label prediction layer to obtain the entity segmentation label sequence. It should be noted that the label prediction layer considers the adjacency relationships between labels, which can address the constraint information ignored by the context feature representation layer and the semantic temporal extraction layer, thereby improving the accuracy of entity segmentation label sequence prediction.

[0084] Sub-step 6 involves determining the entity similarity value between each entity segment in the aforementioned entity segmentation tag sequence and each policy entity in the preset policy knowledge graph's policy entity set, thus obtaining an entity similarity value set. The entity similarity values ​​in this set can be obtained by weighted summation of the Jaccard coefficient, the Levenstein distance, and the feature similarity coefficients obtained based on the FASPell model.

[0085] Sub-step 7: Based on the above entity similarity value set, perform entity alignment on the above entity segmentation tag sequence to obtain the aligned entity segmentation set.

[0086] As an example, the aforementioned execution entity can filter out the entity similarity value with the highest value from the aforementioned entity similarity value set as the target entity similarity value. Then, the entity segment corresponding to the target entity similarity value is identified as the policy entity, and entity alignment is performed on the aforementioned entity segmentation tag sequence to obtain the aligned entity segmentation set.

[0087] Sub-step 8 involves performing attribute imputation prediction processing on the aforementioned policy text information to obtain the entity attribute set for the aligned entity segmentation set. In practice, the execution entity can first use the Bootstrapping algorithm to extract candidate attribute tags from the policy text information. Then, based on the candidate attribute tags, the attribute paragraph corresponding to each attribute tag is determined. Finally, the BERT model is used to classify the candidate attribute paragraphs to obtain the entity attribute set.

[0088] Sub-step 9 involves extracting entity relationships from the aforementioned policy text information to obtain an entity relationship set. The entity relationships in this set represent the associations between entities.

[0089] Sub-step 10: Input the above aligned entity word segmentation set, the above entity attribute set, and the above entity relationship set into the above policy ontology model to obtain the policy ontology knowledge graph.

[0090] The third step is to perform knowledge graph fusion processing on the multiple policy ontology knowledge graphs obtained to obtain the policy knowledge graph.

[0091] The fourth step is to construct a knowledge graph from the aforementioned insurance product information set, resulting in an insurance product knowledge graph. The steps for constructing this insurance product knowledge graph can be referenced from the process of constructing a policy knowledge graph.

[0092] Step 105: Determine user value information based on the user information set and the policy information set.

[0093] In some embodiments, the aforementioned implementing entity may determine user value information based on the aforementioned user information set and the aforementioned policy information set. This user value information may represent the importance of the user. For example... Figure 3 As shown, Figure 3 This displays user value information, which can be composed of the user's current direct value, user loyalty, and potential direct value. The user's current direct value represents the value derived from policies already purchased by the user. User loyalty indicates whether the user will continue to purchase policies in the future. The user's potential direct value represents the value generated by policies the user may purchase in the future.

[0094] As an example, the aforementioned implementing entity can use the RFM (User Relationship Management Model) model to determine user value information based on the aforementioned user information set and the aforementioned policy information set.

[0095] In some optional implementations of certain embodiments, determining user value information based on the aforementioned user information set and the aforementioned policy information set may include the following steps:

[0096] The first step is to perform multi-table joins on the aforementioned user information set and policy information set to obtain a user value association data table. This user value association data table can be obtained by integrating the aforementioned user information set and policy information set into a single data table.

[0097] The second step involves performing field statistical processing on the aforementioned user value association data table to obtain the following data: cumulative paid policy attribute value, pending paid policy attribute value, policy target attribute value, policy payment date, number of associated persons, cumulative years, policy type richness, first-type policy classification information, and second-type policy classification information. Specifically, the cumulative paid policy attribute value can be the accumulated premiums of at least one policy purchased by the target user from the start of policy purchase to the current moment. The pending paid policy attribute value can be the premiums of policies awaiting payment. The policy target attribute value can be the policy's profit (NBP profit). The policy payment date can be the date the target user purchased the policy. The number of associated persons can be the number of family members of the target user. The cumulative years can be the years from the start of policy purchase to the current moment. The policy type richness represents the ratio of the number of policy types corresponding to at least one policy purchased by the target user to the total number of policy types. The total number of policy types can be the number of policy types in the policy knowledge graph corresponding to the policy set. The first type of policy classification information can be the repurchase rate and estimated profit of protection-type policies. The second type of policy classification information can be the repurchase rate and estimated profit of planning-type policies.

[0098] The third step involves generating target user value information based on the aforementioned cumulative payment policy attribute values, pending payment policy attribute values, and policy target attribute values. This target user value information represents the profit value reflected in the policies already purchased by the target user.

[0099] As an example, the aforementioned executing entity can determine the first, second, and third scores of the cumulative payment policy attribute value, the pending payment policy attribute value, and the policy target attribute value using preset interval division rules. These preset interval division rules can be pre-defined rules that determine the score for each of the cumulative payment policy attribute value, pending payment policy attribute value, and policy target attribute value falling within a preset interval. Then, the first, second, and third scores are weighted and summed to obtain the target user value value. Finally, the score identifier of the interval containing the target user value value is determined as the target user value information.

[0100] The fourth step involves generating user loyalty value information based on the aforementioned policy payment dates, the number of associated individuals, the cumulative years, and the richness of policy types. This user loyalty value information characterizes the likelihood that the target user will continue to purchase policies. The implementation method is illustrated in step three and will not be repeated here.

[0101] The fifth step involves generating user potential value information based on the first and second types of policy classification information mentioned above. This user potential value information represents the probability that the target user will continue to purchase policies in the future. The implementation of this fifth step can be found in the example of step three, and will not be repeated here.

[0102] The sixth step is to combine the target user value information, user loyalty value information, and user potential value information to generate the user value information.

[0103] Step 106: Perform user sentiment recognition on the user comment information set to obtain user demand information.

[0104] In some embodiments, the aforementioned executing entity may perform user sentiment recognition on the aforementioned user comment information set to obtain user demand information. This user demand information may be information representing the user's needs and interests regarding the insurance policy.

[0105] In some optional implementations of certain embodiments, the above-mentioned user sentiment recognition of the user comment information set to obtain user demand information may include the following steps:

[0106] The first step is to perform text preprocessing on the aforementioned user comment information set to obtain a preprocessed comment information set. The user comment information in this set can be information about users' emotional inclinations towards the insurance policy. Text preprocessing may include text standardization and noise removal.

[0107] The second step involves performing word segmentation and part-of-speech tagging on the preprocessed comment information set to obtain a user comment word segmentation set and a corresponding part-of-speech tagging set. The part-of-speech tags in the part-of-speech tagging set can be information about the grammatical classification attributes of the user comment word segments.

[0108] The third step involves selectively filtering the user comment word segments based on the aforementioned part-of-speech tag set to obtain the target comment word segment set. This target comment word segment set can be the set of words after removing meaningless stop words from the user comment word segments. For example, the executing entity can remove meaningless stop words from the user comment word segments and retain at least one user comment word segment with the part-of-speech tag of negation, degree adverb, noun, or adjective as the target comment word segment set.

[0109] The fourth step is to determine the word frequency of each target comment word in the aforementioned target comment word segmentation set, thus obtaining a word frequency set. The word frequency in this set represents the ratio of the number of times the target comment word appears to the total number of words, indicating the importance of the word segment.

[0110] The fifth step involves performing joint theme sentiment identification on the target comment word segmentation set based on the aforementioned part-of-speech tag set, word frequency set, and user comment information set, to obtain an initial sentiment tendency information set. This initial sentiment tendency information set can characterize the initially identified user sentiment towards the policy, such as positive or negative. As an example, the executing entity can utilize an LDA (Latent Dirichlet Allocation) topic model to perform joint theme sentiment identification on the target comment word segmentation set based on the part-of-speech tag set, word frequency set, and user comment information set, thereby obtaining the initial sentiment tendency information set.

[0111] The sixth step involves performing hierarchical clustering on the target comment word segmentation set to obtain a multi-level sentiment information set. This multi-level sentiment information set can be obtained by clustering based on the similarity of the target comment word segmentation sets and then stratifying them according to a pre-defined sentiment hierarchy. This pre-defined sentiment hierarchy can be based on the hierarchy of different functional modules of the policy, or it can be based on the hierarchy of different issues related to the policy.

[0112] The seventh step involves quantifying each level of sentiment tendency information in the aforementioned multi-level sentiment tendency information set to obtain a quantified multi-level sentiment tendency information set. In practice, the implementing entity can utilize the matter-element model in extension theory to quantify each level of sentiment tendency information in the aforementioned multi-level sentiment tendency information set to obtain a quantified multi-level sentiment tendency information set.

[0113] Step 8: Determine the emotional intensity and frequency of each multi-level emotional tendency information in the aforementioned multi-level emotional tendency information set, resulting in an emotional intensity information set and an emotional frequency set. The emotional intensity information in the emotional intensity information set guarantees the intensity of the emotion. The emotional frequency in the emotional frequency set can be the ratio of the total number of occurrences of different target comment segments representing the same emotion to the total number of occurrences of those segments.

[0114] Step nine involves performing context-based fine-grained sentiment recognition on the aforementioned multi-level sentiment information set to obtain the target multi-level sentiment information set. The multi-level sentiment information in the target multi-level sentiment information set can be obtained by fine-tuning the sentiment information based on context. This context-based fine-grained sentiment recognition can be performed using the BERT model.

[0115] Step 10: Based on the aforementioned quantified multi-level sentiment tendency information set, the aforementioned sentiment intensity information set, and the aforementioned sentiment frequency information set, optimize the aforementioned target multi-level sentiment tendency information set to obtain user demand information. As an example, the executing entity can first sum and average the aforementioned quantified multi-level sentiment tendency information set, the aforementioned sentiment intensity information set, and the aforementioned sentiment frequency information set to obtain a sentiment weight coefficient set. Then, using the sentiment weight coefficient set, perform a weighted summation on the aforementioned target multi-level sentiment tendency information set to obtain user demand information.

[0116] Step 107: The associated user information set, user profile tags, policy knowledge graph, insurance product knowledge graph, user value information, and user demand information are identified as the user protection demand information set.

[0117] In some embodiments, the aforementioned executing entity may determine the aforementioned associated user information set, the aforementioned user profile tags, the aforementioned policy knowledge graph, the aforementioned insurance product knowledge graph, the aforementioned user value information, and the aforementioned user demand information as a user protection demand information set.

[0118] Step 108: Based on the user's protection needs information set, recommend insurance products to the target user to generate insurance product recommendation information.

[0119] In some embodiments, the aforementioned implementing entity may recommend insurance products to the aforementioned target users based on the aforementioned user protection needs information set, thereby generating insurance product recommendation information. This insurance product recommendation information may be a recommendation based on the preferences and behaviors of the target users, comprising the names of at least one insurance product that meets the user needs of the target users.

[0120] In some optional implementations of certain embodiments, the process of recommending insurance products to the target user based on the aforementioned user protection needs information set to generate insurance product recommendation information may include the following steps:

[0121] The first step, in response to determining that the user value information included in the aforementioned user protection demand information set is the first user value information, involves fusing the associated user social network graph corresponding to the associated user information set included in the aforementioned user protection demand information set with the aforementioned policy knowledge graph to obtain a user-policy fusion graph. This user-policy fusion graph can be a heterogeneous graph containing user nodes and policy nodes. The aforementioned first user value information can be the highest-level user value information.

[0122] The second step involves determining the policy weight set for each policy-related edge and the user weight set for each user-related edge in the user-policy fusion graph, based on the associated user information set and the policy knowledge graph included in the aforementioned user protection needs information set. Here, the policy-related edge can be an edge associated with a policy node. The user-related edge can be an edge whose two nodes are both user nodes. The policy weight values ​​in the policy weight set represent the similarity between policies when both nodes associated with a policy-related edge are policies, and the frequency of user interaction with policies when the nodes associated with a policy-related edge are both policies and users. The user weight values ​​in the user weight set represent the influence of adjacent users on the target user. As an example, the executing entity can determine the user preference information set corresponding to the associated user set included in the aforementioned associated user information set. Next, each user preference information in the aforementioned user preference information set and the user profile label of the target user are vectorized to obtain an associated preference feature vector set and a target preference feature vector. Next, the cosine distance between each associated preference feature vector and the target preference feature vector in the aforementioned associated preference feature vector set is determined as the preference similarity value, resulting in a preference similarity value set. Then, the relationship closeness value set between the target user and the aforementioned associated user set is determined, where the relationship closeness value can be the ratio of the intersection and union of the associated users of the associated user and the target user. Subsequently, the sum of each preference similarity value in the preference similarity value set and the corresponding relationship closeness value in the aforementioned relationship closeness value set is determined as the user weight value set. Then, in response to determining that a policy-related edge is an edge where all associated nodes are policies, the similarity value of the associated policies is determined as the policy weight value. Finally, in response to determining that a policy-related edge is an edge connecting a user and a policy, the ratio of the number of interactions between the user and the policy to the target number of interactions is determined as the policy weight value. Here, the target number of interactions can be the number of interactions between the target user and the policy set included in the policy knowledge graph.

[0123] The third step involves performing weighted pruning on the aforementioned policy weight set and user weight set to obtain a weighted user-policy graph. This weighted user-policy graph can be obtained by removing at least one policy-related edge and one user-related edge from the aforementioned user-policy fusion graph that are less than a preset weight threshold. The preset weight can be a pre-defined threshold value for determining whether to retain a weight. For example, the executing entity can determine at least one policy weight value and at least one user weight value from the aforementioned policy weight set and user weight set that are less than or equal to the preset weight threshold. Then, the policy-related edges and user-related edges corresponding to at least one policy weight value and at least one user weight value are removed from the user-policy fusion graph to obtain the weighted user-policy graph.

[0124] The fourth step involves embedding node types into the aforementioned user-policy weighted graph to obtain a set of user node feature vectors and a set of policy node feature vectors. The user node feature vectors in the user node feature vector set represent the preference features of a user node influenced by its neighboring nodes. The policy node feature vectors in the policy node feature vector set represent the policy features of a policy node influenced by its neighboring nodes. In practice, the executing entity can input the aforementioned user-policy weighted graph into a HetGNN (Heterogeneous Graph Neural Network) to obtain the user node feature vector set and the policy node feature vector set.

[0125] The fifth step involves inputting the aforementioned user node feature vector set and policy node feature vector set into the spatial channel attention mechanism layer to obtain the user node weight feature vector set and the policy node weight feature vector set. The user node weight feature vector in the user node weight feature vector set represents the weight of the user node. The policy node weight feature vector in the policy node weight feature vector set represents the weight of the policy node. The spatial channel attention mechanism layer can be a network layer that determines attention weights from both channel and spatial perspectives. This layer can include a channel attention module and a spatial attention module. The channel attention module can process the initial input feature vector through global max pooling and global average pooling to obtain two 1*1*C feature vectors. These two 1*1*C feature vectors are then input into a multilayer perceptron (MLP) with shared weights. The module sums the two 1*1*C feature vectors element-wise to output channel weight coefficients. Finally, the channel weight coefficients are multiplied by the initial feature vectors to obtain the output feature vector. Here, C can be a predetermined threshold for the number of channels in the output feature vector. The aforementioned spatial attention module can be a convolutional network that performs channel-based max pooling and average pooling on the output feature vector of the channel attention module to obtain a pooled feature vector. Then, the pooled feature vector is compressed to 1 by passing it through a convolutional layer with a 7*7 kernel. The spatial weight coefficients are obtained by passing it through a Sigmoid activation function layer. Finally, the spatial weight coefficients and the output feature vector are multiplied together to output a convolutional feature vector.

[0126] The sixth step involves fusing the aforementioned user node weight feature vector set and the aforementioned policy node weight feature vector set to obtain the user's long-term preference feature vector. This long-term preference feature vector can represent the user's long-term preferences.

[0127] Step 7: Perform session modeling on the user profile tags included in the aforementioned user assurance demand information set to obtain a short-term user preference feature vector. This short-term preference feature vector represents the user's preference information after the most recent change in preference. In practice, the executing entity can first convert the user behavior information in the aforementioned user profile tags into a session graph, obtaining a user behavior session graph. Secondly, input the user behavior session graph into the LESSR model to obtain the short-term user preference feature vector. The LESSR model can be a session-based recommendation algorithm.

[0128] Step 8: Based on the insurance product knowledge graph included in the aforementioned user protection needs information set, a gating mechanism is applied to the aforementioned long-term user preference feature vector and the aforementioned short-term user preference feature vector to obtain a predicted insurance product information set. The predicted insurance product information in this set can be policies obtained through dynamic and static predictions of users' long-term and short-term preferences. In practice, the executing entity can input the aforementioned long-term user preference feature vector, the insurance feature vector set corresponding to the aforementioned insurance product knowledge graph, and the aforementioned short-term user preference feature vector into the forgetting gate and cell gating mechanism included in the Long Short-Term Memory neural network model to obtain the predicted policy information set.

[0129] Step nine involves filtering the predicted insurance product information set based on the user demand information included in the aforementioned user protection demand information set, obtaining filtered policy information, which serves as insurance product recommendation information. As an example, the executing entity can first match the predicted insurance product information set with the aforementioned user demand information to obtain user demand information for each predicted insurance product, which serves as the target demand information set. Then, it can filter at least one predicted insurance product from the predicted insurance product information set whose corresponding target demand information is positive, using this as insurance product recommendation information.

[0130] Step 109: In response to determining that the data volume of the insurance product recommendation information is greater than or equal to a preset data volume threshold, network performance testing is performed on the server sending the insurance product recommendation information to obtain network performance testing information. In response to determining that the transmission delay duration represented by the network performance testing information is greater than or equal to a preset delay duration threshold, the insurance product recommendation information is compressed and pushed to the target display terminal to display the insurance product push information in a personalized manner.

[0131] In some embodiments, the executing entity may, in response to determining that the data volume of the insurance product recommendation information is greater than or equal to a preset transmission data volume threshold, perform network performance testing on the server sending the insurance product recommendation information to obtain network performance testing information; and in response to determining that the transmission delay duration represented by the network performance testing information is greater than or equal to a preset delay duration threshold, compress the insurance product recommendation information and push it to the target display terminal to display the insurance product push information in a personalized manner. The target display terminal may be the terminal corresponding to the target user. The personalized display may be a display of the recommended policy information set according to the user profile tags to match the target user's preferences. The preset transmission data volume threshold may be a pre-set threshold value for the transmission volume of data transmitted in batches. The preset delay duration threshold may be the maximum value of the time required for the insurance product recommendation information to be sent from the server to the target client.

[0132] Further reference Figure 4 As an implementation of the methods shown in the above figures, this disclosure provides some embodiments of a device for analyzing customer protection needs and recommending insurance products based on data elements. These device embodiments are similar to... Figure 1 Corresponding to the method embodiments shown, this data element-based customer protection needs analysis and insurance product recommendation device can be specifically applied to various electronic devices.

[0133] like Figure 4 As shown, a customer protection needs analysis and insurance product recommendation device 400 based on data elements includes: an acquisition unit 401, an associated user identification unit 402, a generation unit 403, a knowledge graph construction unit 404, a first determination unit 405, an emotion recognition unit 406, a second determination unit 407, an insurance product recommendation unit 408, and a compression and push unit 409.

[0134] It is understandable that the various units and references recorded in the data element-based customer protection needs analysis and insurance product recommendation device 400 are... Figure 1 The steps described in the method correspond to each other. Therefore, the operations, features, and beneficial effects described above for the method also apply to the data element-based customer protection needs analysis and insurance product recommendation device 400 and the units contained therein, and will not be repeated here.

[0135] The following is for reference. Figure 5 It shows a schematic diagram of the structure of an electronic device (e.g., an electronic device) 500 suitable for implementing some embodiments of the present disclosure. Figure 5 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this disclosure.

[0136] like Figure 5As shown, the electronic device 500 may include a processing unit (e.g., a central processing unit, a graphics processor, etc.) 501, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 502 or a program loaded from a storage device 508 into a random access memory (RAM) 503. The RAM 503 also stores various programs and data required for the operation of the electronic device 500. The processing unit 501, ROM 502, and RAM 503 are interconnected via a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0137] Typically, the following devices can be connected to I / O interface 505: input devices 506 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 507 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 508 including, for example, magnetic tapes, hard disks, etc.; and communication devices 509. Communication device 509 allows electronic device 500 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 5 An electronic device 500 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively. Figure 5 Each box shown can represent a device or multiple devices as needed.

[0138] In particular, according to some embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, some embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication device 509, or installed from storage device 508, or installed from ROM 502. When the computer program is executed by processing device 501, it performs the functions defined in the methods of some embodiments of this disclosure.

[0139] It should be noted that, in some embodiments of this disclosure, the computer-readable medium described above may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In some embodiments of this disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In some embodiments of this disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0140] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and end-to-end networks (e.g., ad hoc end-to-end networks), as well as any currently known or future-developed networks.

[0141] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the content corresponding to steps 101 to 109.

[0142] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0143] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0144] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A method for analyzing customer protection needs and recommending insurance products based on data elements, comprising: The system retrieves the target user's user information set, policy information set, and user review information set from the user information storage database, as well as the insurance product information set. The user information storage database is obtained by processing and storing multi-source heterogeneous user information. Perform associated user identification processing on the user information set to obtain an associated user information set; Based on the associated user information set and the user information set, generate user profile tags for the target user; Knowledge graphs are constructed on the policy information set and the insurance product information set respectively to obtain the policy knowledge graph and the insurance product knowledge graph. Based on the user information set and the policy information set, determine the user value information; User sentiment recognition is performed on the user comment information set to obtain user demand information; The associated user information set, the user profile tags, the policy knowledge graph, the insurance product knowledge graph, the user value information, and the user demand information are determined as the user protection demand information set; Based on the user protection needs information set, insurance products are recommended to the target user to generate insurance product recommendation information; In response to determining that the data volume of the insurance product recommendation information is greater than or equal to a preset data volume threshold, network performance testing is performed on the server sending the insurance product recommendation information to obtain network performance testing information. In response to determining that the transmission delay duration represented by the network performance testing information is greater than or equal to a preset delay duration threshold, the insurance product recommendation information is compressed and pushed to the target display terminal to display the insurance product push information in a personalized manner.

2. The method according to claim 1, wherein, The user information storage database is obtained through the following steps: Acquire multi-source heterogeneous user information sets from different sensors; According to the preset rule engine, each multi-source heterogeneous user information in the multi-source heterogeneous user information set is preprocessed to obtain the preprocessed multi-source heterogeneous user information set. Generate global identifier information for each preprocessed multi-source heterogeneous user information in the preprocessed multi-source heterogeneous user information set to obtain a global identifier information set; The preprocessed multi-source heterogeneous user information set is subjected to semantic disambiguation processing to obtain a semantically disambiguated multi-source heterogeneous user information set. Based on the preset rule engine, the user data quality of each preprocessed multi-source heterogeneous user information in the preprocessed multi-source heterogeneous user information set is determined to obtain the user data quality set. Based on the global identifier information set and the user data quality set, the semantically disambiguated multi-source heterogeneous user information set is subjected to data integration processing to obtain the integrated user information set. The integrated user information set is then stored in a multi-node balanced manner to obtain a user information storage database.

3. The method according to claim 1, wherein, The step of determining user value information based on the user information set and the policy information set includes: By performing multi-table association on the user information set and the policy information set, a user value association data table is obtained; The user value association data table is subjected to field statistical processing to obtain the cumulative paid policy attribute value, pending payment policy attribute value, policy target attribute value, policy payment date, number of users associated with the user, cumulative years, policy type richness, first type policy classification information and second type policy classification information; Based on the cumulative paid policy attribute values, the pending paid policy attribute values, and the policy target attribute values, target user value information is generated. User loyalty value information is generated based on the policy payment date, the number of users associated with the user, the cumulative years, and the richness of policy types. Based on the first type of policy classification information and the second type of policy classification information, generate user potential value information; The target user value information, the user loyalty value information, and the user potential value information are concatenated to generate the user value information.

4. The method according to claim 1, wherein, The step of recommending insurance products to the target user based on the user's protection needs information set, to generate insurance product recommendation information, includes: In response to determining that the user value information included in the user protection demand information set is the first user value information, the associated user social network graph corresponding to the associated user information set included in the user protection demand information set and the policy knowledge graph are fused to obtain a user policy fusion graph. Based on the associated user information set and policy knowledge graph included in the user protection needs information set, determine the policy weight value set of the policy association edge and the user weight value set of the user association edge in the user policy fusion graph; Based on the policy weight value set and the user weight value set, the user policy fusion graph is subjected to weighted pruning to obtain the user policy weighted graph. The user policy weighted graph is represented by node type embedding to obtain the user node feature vector set and the policy node feature vector set; The user node feature vector set and the policy node feature vector set are input into the spatial channel attention mechanism layer to obtain the user node weight feature vector set and the policy node weight feature vector set. The user node weight feature vector set and the policy node weight feature vector set are fused to obtain the user's long-term preference feature vector. The user profile tags included in the user protection needs information set are subjected to session modeling processing to obtain the user short-term preference feature vector; Based on the insurance product knowledge graph included in the user protection needs information set, a gating mechanism is applied to the user's long-term preference feature vector and the user's short-term preference feature vector to obtain a predicted insurance product information set. Based on the user demand information included in the user protection demand information set, the predicted insurance product information set is filtered to obtain a filtered insurance product information set, which serves as insurance product recommendation information.

5. The method according to claim 1, wherein, The step of constructing knowledge graphs for the policy information set and the insurance product information set respectively, to obtain a policy knowledge graph and an insurance product knowledge graph, includes: Construct a policy ontology model for the policy information set, wherein the policy ontology model represents a model of concepts, attributes, and logical relationships between concepts; For each policy information in the policy information set, perform the following knowledge fusion steps: The policy information is converted to obtain policy text information; The policy text information is input into the feature representation layer of the policy knowledge extraction model to obtain a set of semantic feature vectors of the policy text. The policy knowledge extraction model also includes: a context feature representation layer, a semantic temporal extraction layer and a label prediction layer. The policy text semantic feature vector set is input into the context feature representation layer to obtain the policy semantic context feature vector set; The policy semantic context feature vector set is input into the semantic temporal extraction layer to obtain the semantic temporal feature vector set; The semantic temporal feature vector set is input into the label prediction layer to obtain the entity segmentation label sequence; Determine the entity similarity value between each entity segment in the entity segmentation tag sequence and each policy entity in the policy entity set included in the preset policy knowledge graph, and obtain the entity similarity value set; Based on the entity similarity value set, the entity segmentation tag sequence is aligned to obtain the aligned entity segmentation set; The policy text information is subjected to attribute filling prediction processing to obtain an entity attribute set for the aligned entity word segmentation set; Entity relations are extracted from the policy text information to obtain an entity relation set; The aligned entity word segmentation set, the entity attribute set, and the entity relationship set are input into the policy ontology model to obtain the policy ontology knowledge graph; The multiple policy ontology knowledge graphs obtained are subjected to knowledge graph fusion processing to obtain a policy knowledge graph; A knowledge graph is constructed from the insurance product information set to obtain an insurance product knowledge graph.

6. The method according to claim 1, wherein, The step of performing user sentiment recognition on the user comment information set to obtain user demand information includes: The user comment information set is preprocessed to obtain a preprocessed comment information set; The preprocessed comment information set is segmented and labeled with parts of speech to obtain a user comment segmentation set and a corresponding part-of-speech tagging set; Based on the part-of-speech tag set, the user comment word segmentation set is selectively filtered to obtain the target comment word segmentation set; Determine the word frequency of each target comment word in the target comment word segmentation set to obtain the word frequency set; Based on the word segmentation part-of-speech set, the word segmentation frequency set, and the user comment information set, the target comment word segmentation set is subjected to joint theme sentiment recognition to obtain an initial sentiment tendency information set. Hierarchical clustering is performed on the initial sentiment tendency information set to obtain a multi-level sentiment tendency information set; Each multi-level sentiment tendency information in the multi-level sentiment tendency information set is quantified to obtain the quantified multi-level sentiment tendency information set. Determine the emotional intensity information and emotional frequency of each multi-level emotional tendency information in the multi-level emotional tendency information set to obtain the emotional intensity information set and the emotional frequency set; Context-based fine-grained sentiment recognition is performed on the multi-level sentiment tendency information set to obtain the target multi-level sentiment tendency information set; Based on the quantified multi-level sentiment tendency information set, the sentiment intensity information set, and the sentiment frequency information set, the target multi-level sentiment tendency information set is optimized to obtain user demand information.

7. A device for analyzing customer protection needs and recommending insurance products based on data elements, comprising: The acquisition unit is configured to acquire a set of user information, a set of policy information, and a set of user review information of the target user from a user information storage database, as well as an insurance product information set. The user information storage database is obtained by processing and storing multi-source heterogeneous user information. The associated user identification unit is configured to perform associated user identification processing on the user information set to obtain the associated user information set. The generation unit is configured to generate user profile tags for the target user based on the associated user information set and the user information set; The knowledge graph construction unit is configured to construct knowledge graphs for the policy information set and the insurance product information set respectively, to obtain a policy knowledge graph and an insurance product knowledge graph. The first determining unit is configured to determine user value information based on the user information set and the policy information set; The emotion recognition unit is configured to perform user emotion recognition on the user comment information set to obtain user demand information; The second determining unit is configured to determine the associated user information set, the user profile tags, the policy knowledge graph, the insurance product knowledge graph, the user value information, and the user demand information as a user protection demand information set; The insurance product recommendation unit is configured to recommend insurance products to the target user based on the user's protection needs information set, so as to generate insurance product recommendation information; The compression push unit is configured to, in response to determining that the data volume of the insurance product recommendation information is greater than or equal to a preset transmission data volume threshold, perform network performance testing on the server sending the insurance product recommendation information to obtain network performance testing information, and in response to determining that the transmission delay duration represented by the network performance testing information is greater than or equal to a preset delay duration threshold, compress and push the insurance product recommendation information to the target display terminal to display the insurance product push information in a personalized manner.

8. An electronic device, comprising: One or more processors; Storage device, on which one or more programs are stored, When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-6.

9. A computer-readable medium having a computer program stored thereon, wherein, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-6.

Images