Method and device for mining association between pre-sale customer behavior of insurance and post-sale claim result

Abstract

The application relates to the technical field of data mining, and particularly discloses a method and device for mining the correlation between the pre-sale customer behavior and the post-sale claim result of insurance, which comprises the following steps: obtaining a multidimensional data source in an insurance ERP system; constructing a graph network structure containing multiple types of nodes and multiple types of heterogeneous edges based on the multidimensional data source; performing high-order feature learning on the graph network structure based on a multilayer graph attention network to generate a graph attention network model; performing multi-hop causal correlation rule mining based on the graph attention network model to generate a causal correlation prediction result; and performing claim risk prediction based on the causal correlation prediction result to generate a prediction result. The method improves the existing correlation mining method, optimizes the pre-sale and post-sale scenes of insurance, and improves the accuracy of data correlation mining, thereby providing a data basis for further business decision-making of enterprises.

Description

Technical Field

[0001] This invention relates to the field of data mining technology, specifically to a method and apparatus for mining the correlation between pre-sale customer behavior and post-sale claims results in insurance. Background Technology

[0002] Insurance companies' enterprise resource planning systems have accumulated massive amounts of business data resources, covering the entire lifecycle of customers from consultation and insurance application to claims. The effective use of these data assets has important strategic value for improving insurance companies' risk management capabilities, optimizing underwriting and pricing strategies, and developing precise marketing plans. How to extract these deep cross-stage and cross-dimensional relationships from complex business data has become a core technical challenge for the digital operation of the insurance industry.

[0003] Association mining technology based on insurance pre-sales and post-sales data is an important research direction in the current insurance technology field. However, existing technologies have significant shortcomings in data association analysis in specific application scenarios of the insurance industry.

[0004] First, traditional association analysis methods mainly rely on single primary keys such as policy number and customer identifier to associate data. They can only establish first-order associations between customers and policies, and between customers and claims. They cannot capture the deep mapping relationship between customer behavior characteristics and claims results, resulting in a large amount of high-value cross-domain feature information not being effectively utilized.

[0005] Secondly, the insurance business involves complex multi-hop causal paths, such as the entire process from high-frequency online consultations to purchasing insurance through agents, to renewals and additional coverage, and finally to medical insurance claims. Existing technologies lack the ability to model such heterogeneous graph structures composed of multiple types of nodes and edges, and cannot discover complex causal paths that span multiple business stages.

[0006] Furthermore, traditional machine learning methods typically input customer and business features into a classification model by simply concatenating them, neglecting the enhancing effect of structural and neighborhood information between nodes on feature representation. This flattened feature processing approach is difficult to fully express the complex relationships in insurance business, resulting in low accuracy in claims risk prediction.

[0007] Furthermore, most existing prediction models are black-box models, which only output risk scores but cannot provide risk factor rankings or explanations of causal rules. Business personnel find it difficult to understand and apply the model outputs to optimize strategies, resulting in a significant gap between technical solutions and business decisions. Summary of the Invention

[0008] To overcome the aforementioned technical problems in the prior art, embodiments of the present invention provide a method and apparatus for mining the correlation between pre-sales customer behavior and post-sales claims results in insurance. By modifying existing correlation mining methods and optimizing them for pre-sales and post-sales insurance scenarios, the accuracy of data correlation mining is improved, providing data support for further business decisions by enterprises.

[0009] To achieve the above objectives, embodiments of the present invention provide a method for mining the correlation between pre-sales customer behavior and post-sales claims results in insurance. The method includes: acquiring multi-dimensional data sources from an insurance ERP system; constructing a graph network structure containing multiple types of nodes and multiple types of heterogeneous edges based on the multi-dimensional data sources; performing high-order feature learning on the graph network structure based on a multi-layer graph attention network to generate a graph attention network model; performing multi-hop causal association rule mining based on the graph attention network model to generate causal association prediction results; and performing claims risk prediction based on the causal association prediction results to generate prediction results.

[0010] Preferably, the step of constructing a graph network structure containing multiple types of nodes and multiple types of heterogeneous edges based on the multi-dimensional data source includes: obtaining a customer basic information table, a policy information table, a product information table, a channel information table, a service event record table, and a claims record table from the multi-dimensional data source; determining customer basic attribute features and customer node identifiers based on the customer basic information table and the policy information table, and determining customer nodes based on the customer basic attribute features and the customer node identifiers; determining product attribute features and product node identifiers based on the product information table, and determining product nodes based on the product attribute features and the product node identifiers; and determining channels based on the channel information table. The system generates a basic feature and a channel node identifier. Based on these, channel nodes are determined; based on the service event record table, basic event features and service node identifiers are determined, and service event nodes are determined; based on the claims record table, claims fact features and claims record node identifiers are determined, and claims record nodes are generated; corresponding node edges are generated based on the customer node, product node, channel node, service event node, and claims record node; and a graph network structure is constructed based on the node edges and corresponding nodes.

[0011] Preferably, determining the customer's basic attribute characteristics based on the customer basic information table and the policy information table includes: obtaining customer age data, occupation code, residence data, and driving data from the customer basic information table; obtaining historical insurance data, user usage data, and historical claims data from the policy information table; determining customer age characteristics based on the customer age data; determining occupation characteristics based on the occupation code; determining regional characteristics based on the residence data; determining customer behavioral risk characteristics based on the driving data; determining historical insurance frequency characteristics based on the historical insurance data; determining customer activity rating characteristics based on the user usage data; determining risk level label characteristics based on the historical claims data; analyzing the customer basic information table based on a preset customer profile model to generate family annual income level characteristics; and generating customer basic attribute characteristics based on the customer age characteristics, occupation characteristics, regional characteristics, customer behavioral risk characteristics, historical insurance frequency characteristics, customer activity rating characteristics, risk level label characteristics, and family annual income level characteristics.

[0012] Preferably, determining product attribute features based on the product information table includes: extracting product type field, premium amount data, premium period field, historical claims records, product premium data, and target customer data from the product information table; determining product type features based on the product type field, premium range features based on the premium amount data, coverage period features based on the premium period field, historical loss ratio features based on the historical claims records, sales area distribution features based on the product premium data, and target customer profile features based on the target customer data; and generating product attribute features based on the product type features, premium range features, coverage period features, historical loss ratio features, sales area distribution features, and target customer profile features.

[0013] Preferably, determining the basic characteristics of a channel based on the channel information table includes: extracting channel identifier, channel type, historical channel conversion data, channel performance data, historical channel complaint data, and customer evaluation data from the channel information table; generating a channel identifier feature based on the channel identifier, a channel type feature based on the channel type, a conversion rate indicator feature based on the historical channel conversion data, a channel rating feature based on the channel performance data, a historical complaint rate feature based on the historical channel complaint data, and a customer satisfaction index feature based on the customer evaluation data; and generating basic channel characteristics based on the channel identifier feature, the channel type feature, the conversion rate indicator feature, the channel rating feature, the historical complaint rate feature, and the customer satisfaction index feature.

[0014] Preferably, the step of performing high-order feature learning on the graph network structure based on a multi-layer graph attention network to generate a graph attention network model includes: obtaining an initial multi-layer graph attention network; performing initialization processing on the node attributes in the graph network structure to obtain a low-dimensional dense vector; setting multiple parallel attention heads in each layer of the initial multi-layer graph attention network, with each attention head independently calculating the attention coefficients between nodes, and generating attention weights based on the attention coefficients; performing neighbor feature aggregation on the low-dimensional dense vector based on the attention weights to generate an updated node feature vector; setting up multi-layer stacking of the multi-layer graph attention network to obtain a multi-layer stacked structure; and constructing a multi-layer graph attention network model based on the multi-layer stacked structure and the updated node feature vector.

[0015] Preferably, generating the updated node feature vector includes: concatenating the transformed feature vectors of adjacent nodes; inputting the concatenation result into a LeakyReLU activation function for nonlinear transformation to generate a transformation result; normalizing the transformation result to obtain attention weights; performing a weighted summation of the neighbor node features based on the attention weights to obtain an aggregated neighbor feature vector; and inputting the aggregated neighbor feature vector into an activation function for nonlinear transformation to generate the updated node feature vector.

[0016] Preferably, the step of performing multi-hop causal association rule mining based on the graph attention network model to generate causal association prediction results includes: determining constraints; performing a constrained graph traversal operation on the graph attention network model according to the breadth-first search algorithm based on the constraints to obtain a path set; processing the path combinations based on the frequent itemset growth algorithm to obtain the corresponding frequent itemsets; and performing association rule transformation processing on the frequent itemsets to obtain the target association rule.

[0017] Causal association prediction results are generated based on the target association rules.

[0018] Preferably, the process of processing the path combinations based on the frequent itemset growth algorithm to obtain the corresponding frequent itemsets includes: performing format conversion processing on the path set to obtain processed transaction data; traversing the transaction database and counting the support count of each processed transaction data item; constructing an item header table based on the support count; recursively constructing a conditional frequent itemset tree based on the item header table; and filtering the conditional frequent itemset tree based on a preset support threshold to obtain frequent itemsets.

[0019] Preferably, the step of performing association rule transformation processing on the frequent itemsets to obtain target association rules includes: performing a prefix and a suffix partitioning process on each frequent itemset based on trigger influence to obtain a prefix and a suffix; transforming the frequent itemsets into multiple association rules based on the prefix and the suffix; determining the support, confidence, and lift of each association rule; performing a filtering process on each association rule based on the confidence and the lift to obtain filtered rules; and performing a sorting process on the filtered rules to generate target association rules.

[0020] On the other hand, the present invention also provides a device for mining the correlation between pre-sales customer behavior and post-sales claims results in insurance. The device includes: a data acquisition module for acquiring multi-dimensional data sources from an insurance ERP system; a graph construction module for constructing a graph network structure containing multiple types of nodes and heterogeneous edges based on the multi-dimensional data sources; a network training module for performing high-order feature learning on the graph network structure based on a multi-layer graph attention network to generate a graph attention network model; a correlation mining module for performing multi-hop causal association rule mining based on the graph attention network model to generate causal association prediction results; and a risk prediction module for performing claims risk prediction based on the causal association prediction results to generate prediction results.

[0021] On the other hand, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method provided in the embodiments of the present invention.

[0022] The present invention has at least the following technical effects through the technical solution provided by the present invention:

[0023] By constructing a heterogeneous graph structure containing multiple types of nodes and edges, the association analysis dimension is expanded from a single primary key association to multiple feature dimensions, which can fully explore the deep mapping relationship between pre-sales behavior and post-sales results.

[0024] A multi-layer graph attention network is used for high-order feature learning, enabling nodes to perceive multi-level neighbor information, effectively capture multi-hop relationships, and support the discovery of complex causal paths spanning multiple business processes.

[0025] Association rule mining is performed based on the node embedding representation output by graph attention network. Through steps such as constraint graph traversal, frequent pattern mining, and rule generation and filtering, complex causal paths spanning multiple business processes can be discovered. In claims risk prediction, customer nodes are embedded into the fully connected classification layer to achieve node classification tasks. At the same time, the probability of the existence of a link between the customer and the claims record is output through the edge prediction task.

[0026] The graph attention mechanism enables the model to adaptively learn the importance of different neighboring nodes, avoiding the information loss caused by simple feature concatenation in traditional methods. Experiments show that the accuracy of claims risk prediction is significantly improved.

[0027] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0028] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings:

[0029] Figure 1 A flowchart illustrating the specific implementation of the correlation mining method between pre-sales customer behavior and post-sales claims results provided in this embodiment of the invention;

[0030] Figure 2 This is a schematic diagram of the structure of the device for identifying the correlation between pre-sale customer behavior and post-sale claims results in insurance, as provided in an embodiment of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] In this invention, the terms "system" and "network" are used interchangeably. "Multiple" refers to two or more; therefore, in this invention, "multiple" can also be understood as "at least two." "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / ", unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship. Furthermore, it should be understood that in the description of this invention, terms such as "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or order.

[0033] Please see Figure 1 This invention provides a method for mining the correlation between pre-sale customer behavior and post-sale claims results in insurance, the method comprising:

[0034] S10: Obtain multi-dimensional data sources from the insurance ERP system;

[0035] S20, Construct a graph network structure containing multiple types of nodes and multiple types of heterogeneous edges based on the multi-dimensional data source;

[0036] S30, Based on a multi-layer graph attention network, perform high-order feature learning on the graph network structure to generate a graph attention network model;

[0037] S40, perform multi-hop causal association rule mining based on the graph attention network model to generate causal association prediction results;

[0038] S50, perform claims risk prediction based on the causal correlation prediction results, and generate prediction results.

[0039] The specific implementation process of step S10 includes three sub-steps: node definition and creation, node attribute setting, and edge construction and weight calculation.

[0040] In one possible implementation, the first step is to extract raw data from multiple business databases within the insurance ERP system. An insurance ERP system typically comprises five core data sources: a customer management subsystem, a policy management subsystem, a channel management subsystem, a service event management subsystem, and a claims management subsystem. The data structures stored in each subsystem are as follows: The customer management subsystem stores a basic customer information table, including fields for customer identifier, name, ID number, date of birth, gender, occupation code, residential address, and registration time; the policy management subsystem stores a policy information table, including fields for policy identifier, product identifier, customer identifier, channel identifier, premium amount, coverage period, effective date, and termination date; the channel management subsystem stores a channel information table, including fields for channel identifier, channel name, channel type, and parent channel identifier; the service event management subsystem stores a service event record table, including fields for event identifier, customer identifier, product identifier, channel identifier, event type, occurrence time, interaction duration, and event description; and the claims management subsystem stores a claims record table, including fields for claims record identifier, policy identifier, customer identifier, claims type, claims amount, claims result, claims occurrence date, and case closure date.

[0041] After extracting data from the aforementioned data sources, data cleaning and standardization are required. Data cleaning includes: detecting missing values ​​in the customer basic information table; for records with missing age fields, filling them with the average value of customers in the same occupation and age group; marking records with missing occupation fields as "unknown occupation"; detecting outliers in the policy information table; performing data verification and correction on records with premium amounts below 0 or above the upper limit for a specific product type; and detecting outliers in the claims record table; marking records with abnormally high claims amounts as pending review. Data standardization includes: converting time fields in all tables to YYYY-MM-DD HH:MM:SS format; encoding and mapping category fields (such as occupation codes and channel types) to integer category identifiers; and converting amount fields to integer values ​​in cents.

[0042] After data cleaning and standardization are completed, five types of nodes are created based on the business entity type.

[0043] The first type of node is the customer node, which is used to represent insured customers. Customer nodes are created by extracting data from the customer basic information table and the policy information table. Each unique customer identifier corresponds to one customer node. The node identifier of the customer node is set to the customer identifier field in the customer basic information table. The node attributes of the customer node are set as a 64-dimensional feature vector. The first 32 dimensions correspond to the customer basic attribute features, and the last 32 dimensions correspond to the customer behavior features and customer risk features. The specific composition of the customer basic attribute features is as follows: the customer age feature is the difference between the customer's birth date and the current date, normalized to the range of 0 to 1; the occupation feature is a 32-dimensional vector after the occupation code is mapped by the embedding layer; the regional feature is the concatenation of the 16-dimensional vector of the province of residence after the embedding layer mapping and the 8-dimensional vector of the city level after the embedding layer mapping; the historical insurance frequency feature is the normalized value of the number of historical insurance policies of the customer divided by the average number of insurance policies of the customer group; the customer activity score feature is calculated by statistically analyzing the customer's login frequency and function usage frequency in the APP or insurance system within a 90-day statistical period. The login frequency has a weight of 0.6, and the function usage frequency has a weight of 0.4. The calculation results are normalized to the range of 0 to 1. The specific components of customer behavioral risk characteristics are as follows: The average annual vehicle mileage is obtained by parsing the driving area code and mileage estimation code in the customer's vehicle insurance information. Customers with high mileage (annual average greater than 20,000 kilometers) are marked as having high-risk behavioral characteristics. The annual household income level characteristic can be obtained using common customer profiling models (e.g., pre-trained customer profiling models generated based on deep neural network models, such as a three-layer fully connected neural network with 32 input layer neurons (receiving 32 dimensions of features including customer occupation, age, and insurance product type), 64 neurons in the first hidden layer, and ReLU activation function. The second hidden layer neurons... The model has 32 neurons, uses ReLU as the activation function, and has 5 neurons in the output layer, corresponding to 5 bins of annual income level (0-50,000, 50,000-100,000, 100,000-200,000, 200,000-500,000, and over 500,000). The output layer activation function is softmax. The model is trained using the cross-entropy loss function, and the training data is generated by predicting and generating historical customer annual income data. The model input includes features such as customer occupation, age, and type of insurance product, and the output is the binning result of annual income level, which is mapped to a 16-dimensional vector after the embedding layer. The risk level label feature is the risk score calculated based on the customer's historical claims records, which is also mapped to a 16-dimensional vector after the embedding layer.

[0044] The second type of node is the product node, used to represent insurance products. Product nodes are created by extracting data from the product information table, with each unique product identifier corresponding to one product node. The node identifier of a product node is set to the product identifier field in the product information table. The node attributes of a product node are set to a 48-dimensional feature vector, with some segments corresponding to product attribute features and others corresponding to product sales features. The specific components of the product attribute features are as follows: the product type feature is a 16-dimensional vector after embedding the product type field (life insurance / medical insurance / auto insurance / accident insurance); the premium range feature is an 8-dimensional vector after embedding the premium amount range (0 to 3000 yuan / 3000 to 10000 yuan / above 10000 yuan); the coverage period feature is an 8-dimensional vector after embedding the coverage period field (one-year / long-term / lifetime); the historical loss ratio feature is the loss ratio obtained from the statistics of the product's historical claims records, which is normalized and used as a continuous value feature; the sales area distribution feature is the premium proportion vector of the product in each province, which is reduced to 8 dimensions through principal component analysis; the target customer profile feature is generated based on cluster analysis of customer attributes such as age, occupation, and income, with the number of clusters set to 5 group categories, which is then mapped to an 8-dimensional vector through embedding.

[0045] The third type of node is the channel node, used to represent sales channels. Channel nodes are created by extracting data from the channel information table, with each unique channel identifier corresponding to one channel node. The node identifier of a channel node is set to the channel identifier field in the channel information table. The node attributes of a channel node are set to a 40-dimensional feature vector, with some segments corresponding to basic channel characteristics and others corresponding to channel operational characteristics. The basic characteristics of a channel are as follows: the channel identifier is an 8-dimensional vector mapped by the embedding layer; the channel type is a 12-dimensional vector mapped by the embedding layer of the channel type field (online official website / APP / offline counter / agent channel / bank insurance channel); the conversion rate indicator is the ratio of the channel's historical transaction volume to its traffic, normalized and used as a continuous value feature; the channel rating is calculated based on the channel's performance achievement rate and customer renewal rate, with a weight of 0.7 for performance achievement rate and 0.3 for customer renewal rate, and the result is normalized to the range of 0 to 100; the historical complaint rate is obtained by statistically analyzing the ratio of the number of complaint tickets to the total business volume, with a lower complaint rate resulting in a higher score; and the customer satisfaction index is derived from analyzing customer service evaluation data, with a satisfaction score range of 0 to 5 points, normalized to the range of 0 to 1.

[0046] The fourth type of node is the service event node, used to represent the interaction between the customer and the insurance company. Service event nodes are created by extracting data from the service event record table, with each unique event identifier corresponding to one service event node. The node identifier of a service event node is set to the event identifier field in the service event record table. The node attributes of a service event node are set as a 56-dimensional feature vector, with some segments corresponding to basic event features and others corresponding to semantic event features. The specific composition of the basic event features is as follows: the event identifier feature is an 8-dimensional vector after the event identifier has been mapped through an embedding layer; the event type feature is a 16-dimensional vector after the event type field (consultation / insurance intention / product browsing / policy change / renewal reminder) has been mapped through an embedding layer; the occurrence time feature is the concatenation of a 4-dimensional vector after the hourly segment (0 to 6 am / 6 to 12 am / 12 to 18 pm / 18 to 24 pm) and a 2-dimensional vector after the weekday segment (weekday / weekend) has been mapped through an embedding layer; the interaction depth rating feature is a continuous value of the interaction duration normalized to the interval 0 to 1 after logarithmic transformation. The specific components of the event semantic features are as follows: the product type feature associated with the event is obtained by parsing the product identifier field in the service event record, looking up the corresponding product type in the product information table, and mapping it into an 8-dimensional vector through the embedding layer; the channel source feature associated with the event is obtained by parsing the channel identifier field in the service event record, looking up the corresponding channel type in the channel information table, and mapping it into an 8-dimensional vector through the embedding layer; the business value score feature of the event is calculated based on the event type and customer interaction time. The business value weight of the consultation event is 0.2, the business value weight of the insurance intention event is 0.5, the business value weight of the product browsing event is 0.1, the business value weight of the policy change event is 0.3, and the business value weight of the renewal reminder event is 0.4. The formula for calculating the business value score is the event type weight multiplied by the normalized value of the interaction time.

[0047] The fifth type of node is the claims record node, used to represent claims transaction records. Claims record nodes are created by extracting data from the claims record table, with each unique claims record identifier corresponding to one node. The node identifier of a claims record node is set to the claims record identifier field in the claims record table. The node attributes of a claims record node are set as a 52-dimensional feature vector, with some segments corresponding to claims fact features and others corresponding to claims risk features. The specific composition of the claims fact features is as follows: the claims record identifier feature is an 8-dimensional vector after the claims record identifier has been mapped through an embedding layer; the claims type feature is a 12-dimensional vector after the claims type field (critical illness / medical / death / property loss) has been mapped through an embedding layer; the claims amount feature is a continuous value of the claims amount normalized to the range of 0 to 1 after logarithmic transformation; the claims result feature is a 6-dimensional vector after the claims result field (normal / abnormal / rejection) has been mapped through an embedding layer; and the incident time interval feature is the number of days from the incident date to the policy effective date divided by the number of days of the coverage period, normalized to the range of 0 to 1. The specific components of the claims risk characteristics are as follows: the location of the incident feature is determined by parsing the location of the incident field in the claims record to determine the regional category (near home address / near workplace / area outside the place of residence), and mapped into a 6-dimensional vector through the embedding layer; the claims investigation conclusion feature is obtained by parsing the investigation conclusion field in the claims record (normal claims / abnormal pending investigation / suspected insurance fraud), and mapped into a 6-dimensional vector through the embedding layer; the associated policy information feature is obtained by parsing the policy identifier field in the claims record, searching the policy information table to obtain product type and channel type information, and concatenated into an 8-dimensional vector after the embedding layer mapping.

[0048] After creating the five types of nodes, the five types of edges need to be constructed according to the business logic.

[0049] The first type of edge is the customer-product edge, representing a customer's inquiry or insurance purchase behavior. The construction rule for a customer-product edge is as follows: when there are related records in the customer's basic information table and the policy information table, an edge is created between the customer node and the product node. The source node of the edge is the customer node, and the target node is the product node. The weight calculation formula for the edge is: number of interactions multiplied by a time decay coefficient, where the number of interactions is the sum of the number of times the customer inquired about and purchased insurance for the product, and the time decay coefficient is an exponential decay function exp(-λ × number of days), where λ is set to 0.01, and the number of days is the number of days since the interaction occurred. The physical meaning of the time decay coefficient is that the impact of recent interactions on the current risk assessment is weighted higher than that of long-term interactions.

[0050] The second type of edge is the product-channel edge, which represents the sales path of a product. The construction rule for a product-channel edge is as follows: when there is a record in the policy information table that a product is sold through a certain channel, an edge is created between the product node and the channel node. The source node of the edge is the product node, and the target node is the channel node. The weight of the edge is the sales volume through that channel, i.e., the number of policies for that product sold through that channel in the policy information table.

[0051] The third type of edge is the customer-service event edge, representing the customer's business behavior trajectory. The construction rule for customer-service event edges is as follows: when a record of a customer experiencing a certain event exists in the service event record table, an edge is created between the customer node and the service event node. The source node of the edge is the customer node, and the target node is the service event node. The weight of the edge represents the importance level of the event, with the following specific values: consultation event weight is 1, insurance intention event weight is 2, policy change event weight is 3, and renewal reminder event weight is 4.

[0052] The fourth type of edge is the service event-claims record edge, representing a potential causal relationship between an action and its outcome. The construction rules for service event-claims record edges are as follows: edges are created based on the temporal relationships and business logic rules in the service event record table and the claims record table. Specifically, the edge creation conditions are: the service event occurred earlier than the claim record's incident time, and the time interval is within a preset range. The edge weight is calculated based on the temporal interval and business logic: when the service event and the claim record are associated with the same policy, and the service event occurred within 180 days before the claim incident, the edge weight is set to 0.8; when the service event occurred between 180 and 365 days before the claim incident, the edge weight is set to 0.5; and when the service event occurred more than 365 days before the claim incident, the edge weight is set to 0.3.

[0053] The fifth type of edge is the customer-claims record edge, representing a customer's claim fact. The construction rule for a customer-claims record edge is as follows: when a customer's claim record exists in the claims record table, an edge is created between the customer node and the claim record node. The source node of the edge is the customer node, and the target node is the claim record node. The weight of the edge is the ratio of the claim amount to the historical average claim amount, that is, the claim amount divided by the average claim amount of all claim records in the claims record table.

[0054] Through the above process of constructing nodes and edges, a heterogeneous graph structure is created. This heterogeneous graph structure includes five types of nodes: customer nodes, product nodes, channel nodes, service event nodes, and claims record nodes, as well as five types of edges: customer-product edges, product-channel edges, customer-service event edges, service event-claims record edges, and customer-claims record edges, realizing a multi-level association representation between pre-sales behavior and post-sales results.

[0055] In this embodiment of the invention, a graph network structure is constructed based on multi-source data to realize multi-level correlation between pre-sales behavior and post-sales results, providing a data foundation for subsequent in-depth data mining and comprehensive analysis.

[0056] After completing the construction of the graph network structure, the graph attention network model is further trained. The specific implementation process includes five sub-steps: node feature initialization, implementation of multi-head attention mechanism, neighbor feature aggregation, multi-layer stacking structure setting, and training target definition.

[0057] In one possible implementation, during the node feature initialization sub-step, the attributes of each type of node need to be mapped to a low-dimensional dense vector through an embedding layer, serving as the initial feature representation of the node. The embedding layer is implemented using a learnable lookup table structure, with the size of the embedding table determined based on the number of node types and the feature dimension. The size of the customer node embedding table is 64 times the total number of customer nodes, and it is initialized using uniformly distributed random initialization, with an initialization range of -0.1 to 0.1. The size of the product node embedding table is 48 times the total number of product nodes, and its initialization method is the same as for customer nodes. The size of the channel node embedding table is 40 times the total number of channel nodes, and its initialization method is the same. The size of the service event node embedding table is 56 times the total number of service event nodes, and its initialization method is the same. The size of the claims record node embedding table is 52 times the total number of claims record nodes, and its initialization method is the same.

[0058] During embedding queries, the corresponding low-dimensional dense vector is obtained by searching the embedding table based on the original attribute features of the node. For categorical attributes (such as occupation code, product type, and channel type), they are first converted into integer category identifiers through encoding mapping, and then the category identifiers are used as indexes to search for the corresponding vectors in the embedding table. For continuous attributes (such as age, premium amount, and claim amount), they need to be normalized and then concatenated with the embedding vector as continuous value features. For textual attributes (such as event descriptions), textual feature vectors need to be extracted using a pre-trained natural language processing model and then concatenated with the embedding vector. The resulting initial feature vectors have the following dimensions: 64 dimensions for customer nodes, 48 ​​dimensions for product nodes, 40 dimensions for channel nodes, 56 dimensions for service event nodes, and 52 dimensions for claim record nodes.

[0059] In the multi-head attention mechanism implementation sub-step, K parallel attention heads need to be set up in each layer of the graph attention network, and each attention head independently calculates the attention coefficients between nodes. The calculation process of the attention coefficients includes the following steps: First, for any pair of adjacent nodes i and j in the graph, the input feature vector of the attention mechanism is calculated. The current feature vector of node i is then... With the current feature vector of node j The weight matrix of the attention head respectively Multiplying them yields the transformed eigenvectors. and Then, the two transformed feature vectors are concatenated to obtain a concatenated vector. Next, the concatenated vector is input into the LeakyReLU activation function for nonlinear transformation to obtain intermediate variables. Finally, softmax normalization is performed on the intermediate variables to obtain the attention weights of node i to node j. The mathematical expression for this calculation process is:

[0060]

[0061] in, This represents the attention weight of node i to node j calculated by the k-th attention head. Let be the weight matrix of the k-th attention head. , Let be the current feature vectors of nodes i and j. LeakyReLU is a linear rectified activation function with leakage. The slope of the LeakyReLU activation function in the negative interval is set to 0.2. Weight matrix. The dimensions are determined based on the input feature dimensions and the output feature dimensions, the first layer The dimension is the input dimension multiplied by the output dimension divided by the number of attention heads.

[0062] In the neighbor feature aggregation sub-step, it is necessary to aggregate the features of neighbor nodes based on attention weights and update the representation of the current node. The feature aggregation process includes the following steps: First, for node i, obtain the set N(i) of all its neighbor nodes j. Then, based on the attention weights calculated in the previous step... The feature vectors of each neighbor node j are weighted and summed to obtain the aggregated neighbor feature representation. Finally, the aggregated neighbor feature vectors are input into an activation function for nonlinear transformation to obtain the updated node feature vectors. The mathematical expression for this aggregation process is:

[0063]

[0064] in, This represents the updated feature vector of node i. `k` is the activation function. In a multi-head attention mechanism, each attention head independently executes the aggregation process described above, obtaining K aggregation result vectors. These K result vectors are then concatenated or averaged to obtain the final node update representation. When using the concatenation strategy, the output dimension is the single-head output dimension multiplied by K; when using the averaging strategy, the output dimension is the same as the single-head output dimension.

[0065] In the multi-layer stacked structure setting sub-step, a multi-layer graph attention network (GAT) structure is used to achieve high-order feature learning. By stacking L layers of GAT, nodes can perceive information from L-order neighbors and capture multi-hop relationships. This invention uses a three-layer graph attention network structure with the following specific parameter settings: the input node feature dimension of the first GAT layer is 64, the number of attention heads K is set to 4, and the output feature dimension after feature transformation is 128; the input feature dimension of the second GAT layer is 128, the number of attention heads K is set to 4, and the output feature dimension after feature transformation is 256; the input feature dimension of the third GAT layer is 256, the number of attention heads K is set to 4, and the output feature dimension after feature transformation is 512. The activation function of each GAT layer is the exponential linear unit (ELU) function, and the parameter α of the ELU activation function is set to 1.0. Weight matrix. The dimension settings match the input and output dimensions of each layer. First layer The dimension is 64×32 (input 64 dimensions, output 128 dimensions divided by 4 heads), second layer The dimensions are 128×64, the third layer The dimensions are 256×128. Layer normalization operations are added between layers to stabilize the training process.

[0066] In the training objective definition sub-step, the training objective includes two parts: node classification loss and edge prediction loss. The node classification loss is used to learn the risk level representation of customers, and the loss function adopted is the cross-entropy loss function. The calculation process of the cross-entropy loss is as follows: First, the final embedded representation of the customer node is input into a fully connected classification layer. The fully connected classification layer contains two hidden layers. The first hidden layer has 256 neurons with a modified linear unit function (MRU) activation function, and the second hidden layer has 128 neurons with a MRU activation function. The output layer contains 3 neurons, corresponding to the three risk levels: high risk, medium risk, and low risk. Then, the output of the output layer is softmax normalized to obtain the probability distribution of the customer belonging to each risk level. Finally, the cross-entropy loss between the probability distribution and the true label is calculated.

[0067] The edge prediction loss is used to learn the strength of associations between nodes, and the loss function is the binary cross-entropy loss. The implementation process of the edge prediction task is as follows: First, for each pair of customer nodes and claims record nodes, the probability of the existence of an associated edge is calculated based on the node embedding representation. The calculation method is as follows: the customer node embedding and the claims record node embedding are concatenated and then input into the edge predictor. The edge predictor contains two fully connected layers. The first layer has a dimension of 1024 and uses a modified linear unit function as the activation function. The second layer has a dimension of 1 and uses a sigmoid function as the activation function. The output is a probability value between 0 and 1. Then, for samples where the edge actually exists in the graph, the label is set to 1; for samples where the edge does not exist in the graph, it is generated through negative sampling and the label is set to 0. Finally, the binary cross-entropy loss between the predicted probability and the true label is calculated.

[0068] The total loss is defined as the weighted sum of the main loss and the auxiliary loss, calculated as follows: the total loss equals the product of the node classification loss, the edge prediction loss, and the weight coefficients, where the weight coefficients are set to 0.5. During training, an adaptive moment estimation optimizer is used, with a learning rate of 0.001, a first-order momentum estimation decay coefficient of 0.9, a second-order momentum estimation decay coefficient of 0.999, and an epsilon parameter of 1e-8. The training run consists of 200 epochs, with model performance evaluated on the validation set every 10 epochs. The model with the lowest validation set loss is saved as the final model.

[0069] In this embodiment of the invention, by employing a multi-layer graph attention network to perform high-order feature learning on heterogeneous graphs, and by using a multi-head attention mechanism and neighbor feature aggregation to achieve adaptive updating of node features, the dynamic response of the entire model to real-time insurance data can be effectively improved, thereby enhancing the accuracy and real-time performance of data mining.

[0070] After completing the construction of the graph attention network model, multi-hop causal association rule mining is further carried out. The specific implementation process includes four sub-stages: constraint graph traversal, frequent pattern mining, rule generation and filtering, and business interpretation generation.

[0071] In one possible implementation, in the constrained graph traversal sub-step, it is necessary to start from the target node, set the maximum number of hops, and use a breadth-first search algorithm to extract the set of paths from the customer node to the claims record node. The constraints include the following three aspects: First, the starting node of the traversal is the high-risk claims customer node, that is, the customer nodes with a risk score higher than 75 points in the model prediction results of step 2 constitute the seed set; Second, the maximum number of hops of the traversal is set to 4, that is, starting from the starting customer node, a maximum of 4 edges are traversed to reach the claims record node; Third, only specific types of edges are allowed during the traversal process, in the following order: customer-service event edge, service event-product edge (product associated through the service event node), product-channel edge, channel-service event edge (service event associated through the channel node), and service event-claims record edge. The complete valid path is customer → service event → product → channel → claims record, or a combination of customer → service event → product → claims record.

[0072] The specific execution process of the breadth-first search algorithm is as follows: Initialize a first-in, first-out (FIFO) queue, and add all customer nodes from the seed set as starting search nodes to the queue. Each queue element contains the current node, the path to the current node, and the set of visited nodes. The algorithm iteratively executes the following steps: Take an element from the head of the queue. If the current node is a claims record node and the path length is greater than or equal to 2 hops, add the current path to the path set; otherwise, obtain all neighboring nodes of the current node. For each neighboring node, if it has not been visited by the current path, create a new queue element. The current node of the new element is that neighboring node, the path of the new element is the current path concatenated with that neighboring node, the set of visited nodes of the new element is added to that neighboring node, and the new element is added to the tail of the queue. The loop terminates when the queue is empty or the path set reaches a preset upper limit of 1 million.

[0073] During the path extraction process, the type, identifier, and attribute information of each intermediate node are recorded. The data structure of each path record includes the following fields: path number, starting customer identifier, ending claim record identifier, path length, node type sequence of each node in the path, node identifier sequence of each node in the path, and edge weight sequence of each edge in the path.

[0074] In the frequent pattern mining sub-stage, the frequent itemset growth algorithm is used to process the path set, setting a minimum support threshold, and outputting path patterns that meet the threshold as frequent itemsets. The implementation process of the frequent itemset growth algorithm includes the following steps:

[0075] The first step is to convert the path set into a transactional database format. Each path is converted into a transaction, and the items in the transaction correspond to the node types or node attribute values ​​in the path. Item types include two categories: node type items and node attribute items. The value range for node type items is {customer node, service event node, product node, channel node, claims record node}. The value range for node attribute items is determined based on the node type: Customer node attribute items include high consultation frequency (consultations greater than or equal to 3 times), low consultation frequency (consultations less than 3 times), high activity (activity score greater than or equal to 0.7), low activity (activity score less than 0.7), high risk level, medium risk level, and low risk level; Product node attribute items include auto insurance products (product type: vehicle damage). Insurance products include: compulsory traffic accident liability insurance, commercial third-party liability insurance, theft insurance, life insurance, medical insurance, accident insurance, high-premium products (premium greater than 10,000 yuan), and low-premium products (premium less than or equal to 10,000 yuan); channel node attributes include online channels (channel type is official website / APP), offline channels (channel type is counter / agent / bancassurance), telemarketing channels, online sales channels, and car dealer channels; claims record node attributes include high claims amount (claim amount greater than 10,000 yuan), low claims amount (claim amount less than or equal to 10,000 yuan), normal claims, abnormal claims, and rejected claims.

[0076] The second step is to traverse the transaction database and count the support of each item. The support of an item is the number of transactions containing that item divided by the total number of transactions. In this embodiment of the invention, the minimum support threshold is set to 0.05, that is, items with a support greater than or equal to 0.05 are frequent 1-itemsets.

[0077] The third step is to construct an item header table based on frequent itemsets, with the item header table sorted in descending order of support. For each frequent itemset, a conditional schema base is created, which is the set of prefix paths preceding the frequent itemset in all transactions containing that frequent itemset.

[0078] The fourth step is to recursively construct a conditional frequent itemset tree for each frequent itemset and extract frequent itemsets from it. The recursion terminates when the conditional schema base is empty or there is only one transaction in the conditional schema base. The format of the frequent itemset output is the itemset and its support.

[0079] The fifth step is to filter path patterns from the frequent itemset that meet the minimum support threshold. The filtering criteria are: the frequent itemset contains at least one customer node attribute item and at least one claims record node attribute item, and the items in the path pattern can form a complete path from the customer node to the claims record node.

[0080] Examples of output results from frequent pattern mining are as follows: Frequent itemset 1 {high consultation frequency, vehicle damage insurance products, car dealer channels}, support 0.12; Frequent itemset 2 {high insurance intention, commercial third-party liability insurance products, online sales channels}, support 0.08; Frequent itemset 3 {high claim amount, abnormal claims, renewal and additional insurance}, support 0.06.

[0081] In this embodiment of the invention, frequent path patterns need to be converted into association rules, the confidence and lift of each rule need to be calculated, and low-confidence rules need to be filtered out. The process of generating association rules includes the following steps:

[0082] The first step is to divide each frequent itemset into antecedent and consequent. The antecedent is a condition that may lead to the occurrence of the consequent, and the consequent is the outcome to be predicted. The division method is as follows: the last item in the frequent itemset is the consequent, and any non-empty subset of the remaining items is the antecedent. For example, the frequent itemset {high consultation frequency, vehicle damage insurance products, car dealer channels, high claim amount} can generate the following association rules: Rule 1 Antecedent {high consultation frequency, vehicle damage insurance products, car dealer channels} Consequence {high claim amount}; Rule 2 Antecedent {high consultation frequency, vehicle damage insurance products} Consequence {high claim amount, car dealer channels}.

[0083] The second step is to calculate the support, confidence, and lift of each rule. Support is the probability that all items in a frequent itemset occur simultaneously; that is, the support of the frequent itemset. Confidence is the probability that a subsequent item occurs given that the preceding item has occurred. The formula is: the number of transactions where both the preceding and subsequent items occur simultaneously divided by the number of transactions where the preceding item occurs. Lift is the confidence divided by the support of the subsequent item occurring alone. A lift greater than 1 indicates a positive correlation between the preceding and subsequent items.

[0084] The third step is to set confidence and lift thresholds for rule filtering. The confidence threshold is set to 0.6, and the lift threshold is set to 1.5. The filtering criteria are: rules with a confidence level greater than or equal to 0.6 and a lift greater than or equal to 1.5 are retained; all other rules are filtered out.

[0085] The fourth step is to sort the retained association rules in descending order of lift and output a rule list. The rule list includes the following fields: rule number, preceding item description, following item description, support, confidence, and lift.

[0086] Examples of the output results of rule generation and filtering are as follows: Rule 1 is described as "the customer has consulted more than or equal to 3 times and has purchased vehicle damage insurance through a car dealer channel", and the following is described as "the probability of claims is greater than 0.7", with a support of 0.12, a confidence of 0.75, and an lift of 2.3; Rule 2 is described as "the customer has purchased insurance through online sales channels", and the following is described as "high coverage increases the risk of claims", with a support of 0.08, a confidence of 0.68, and an lift of 1.9; Rule 3 is described as "the customer has compared products more than or equal to 2 times", and the following is described as "the probability of renewal or additional coverage is high", with a support of 0.06, a confidence of 0.82, and an lift of 3.1.

[0087] In the business interpretation generation sub-step, the association rules need to be converted into a business-understandable descriptive format, including the rule's antecedent description, consequent description, support value, confidence value, and lift value. Simultaneously, the weighted ranking of risk factors involved in each node of the rule should be output. The specific implementation process of business interpretation generation is as follows:

[0088] First, for each retained association rule, it is converted into a natural language description format. The conversion rule is as follows: connect the conditions in the preceding description according to the logical relationship of "AND", and convert the following description into a prediction result description. For example, the business interpretation of rule 1 is: "When a customer consults 3 or more times before purchasing insurance, selects vehicle damage insurance, and completes the insurance through a car dealer channel, the probability of this customer having a high-value claim (claim amount exceeding 10,000 yuan) in the future is 75%. This rule appears in 12% of all high-risk customers, and its prediction improvement is 2.3, indicating that this combination of conditions can improve the accuracy of claim risk prediction to 2.3 times the baseline risk."

[0089] Secondly, the risk factor weights for each node in the calculation rules are determined. The weight calculation method is as follows: for each condition in the preceding clause of the rule, calculate the decrease in the rule's confidence level after removing that condition. The larger the decrease, the greater the risk contribution of that condition. The risk contribution values ​​of each condition are then normalized to obtain the risk factor weights for each node. For example, the risk factor weights for each node in Rule 1 are: customer consultation frequency weight 0.4 (the largest decrease in confidence level after removal), vehicle damage insurance product weight 0.3, car dealer channel weight 0.2, and other implicit factors weight 0.1.

[0090] Finally, all related rules, their business interpretations, and risk factor weights are integrated into a business report document. The business report document consists of two parts: a rule summary table and detailed explanations. The rule summary table lists the key metrics for all retained rules, including rule number, antecedent, consequent, support, confidence, and lift. The detailed explanations provide a natural language description of each rule and list the risk factor weights for that rule in order of priority.

[0091] In this embodiment of the invention, by mining multi-hop causal association rules, it is possible to effectively mine the changing trends and correlation characteristics of various data from massive amounts of data, thereby providing better support for the company's subsequent business decisions and providing data support for the accurate output of subsequent marketing strategies.

[0092] After generating the causal correlation prediction results, the claims risk prediction is performed and the corresponding prediction results are generated. In this embodiment of the invention, the specific implementation process of claims risk prediction includes four sub-stages: node classification task, edge prediction task, risk warning and business application.

[0093] In the node classification task sub-stage, customer nodes need to be embedded into the fully connected classification layer to output the customer's risk level. The fully connected classification layer's network structure consists of multiple layers, specifically: the input layer has 512 neurons, matching the final embedding dimension of the customer nodes; the first hidden layer has 256 neurons, using the modified linear unit function (MRU) as the activation function; the second hidden layer has 128 neurons, also using the MRU as the activation function; and the output layer contains 3 neurons, corresponding to high-risk, medium-risk, and low-risk levels, using the softmax function as the activation function.

[0094] The specific execution process of the node classification task is as follows: First, extract the final embedding representations of all customer nodes from the GAT model trained in step 2. The embedding representations have a dimension of 512. Then, input the embedding representation of each customer node into a fully connected classification layer. After a nonlinear transformation through two hidden layers, intermediate representations of 256 and 128 dimensions are obtained. Finally, input the 128-dimensional intermediate representation into the output layer. After softmax normalization, the predicted probability distribution of the customer belonging to three risk levels (high risk, medium risk, and low risk) is obtained. The level with the highest probability value is taken as the predicted risk level of the customer.

[0095] Simultaneously, the predicted probability values ​​are mapped to preset risk scoring thresholds to generate a customer's risk score. The risk score is calculated as follows: high risk level corresponds to a score range of 75 to 100 points, with the score equal to the high-risk predicted probability multiplied by 25 plus 75; medium risk level corresponds to a score range of 60 to 75 points, with the score equal to the medium-risk predicted probability multiplied by 15 plus 60; low risk level corresponds to a score range of 0 to 60 points, with the score equal to the low-risk predicted probability multiplied by 60. Finally, a risk score is output for each customer, ranging from 0 to 100 points.

[0096] In the edge prediction task sub-stage, it is necessary to predict the link between the customer and the claims record to output the probability of the customer making a future claim. The edge predictor adopts a multi-layer edge predictor structure, specifically configured as follows: the input layer has 1024 neurons, which is composed of customer node embeddings (512-dimensional) and claims record node embeddings (512-dimensional); the hidden layer has a dimension of 512, and the activation function is the modified linear unit function; the output layer has a dimension of 1, and the activation function is the sigmoid function, outputting a probability value between 0 and 1.

[0097] The specific execution process of the edge prediction task is as follows: First, extract the final embedding representations of all customer nodes and claim record nodes from the GAT model trained in step 2. Then, for each customer node, calculate the predicted probability that there is an associated edge between it and all claim record nodes. The calculation method is as follows: concatenate the customer node embedding and the claim record node embedding to obtain a 1024-dimensional concatenated vector; input the concatenated vector into the edge predictor, and after nonlinear transformation of the hidden layer, obtain a 512-dimensional intermediate representation; input the intermediate representation into the output layer, and after activation by a sigmoid function, obtain the probability value of the existence of a link between the customer and the claim record. Finally, for each customer, aggregate the predicted probabilities of its links with all claim record nodes to generate a comprehensive probability value of the customer's future claim. The aggregation method is as follows: calculate the weighted average of all link predicted probabilities, with the weights being the historical weights of the corresponding edges.

[0098] The output of the side forecasting task is a predicted claim probability value for each customer, ranging from 0 to 1. The relationship between the predicted claim probability value and the risk score is as follows: the higher the predicted claim probability value, the higher the predicted risk level for the corresponding customer.

[0099] In the risk warning sub-stage, a list of high-risk customers needs to be generated based on the prediction results and pushed to the business system for underwriting intervention or claims monitoring. The specific implementation process of risk warning is as follows:

[0100] First, risk levels are determined based on risk scores. The risk level classification rules are as follows: high-risk customers score 75 points or higher, medium-risk customers score between 60 and 75 points, and low-risk customers score below 60 points. According to these rules, high-risk customers account for approximately 5% of all customers, medium-risk customers account for approximately 15%, and low-risk customers account for approximately 80%.

[0101] Next, a high-risk customer list is generated. The data structure of the high-risk customer list includes the following fields: customer identifier, risk score, predicted claim probability, and key risk factors and their weights. The customer identifier is the customer identifier field in the customer basic information table. The risk score is the score output by the node classification task. The predicted claim probability is the probability value output by the edge prediction task. The key risk factors and their weights are obtained from the output of the business interpretation generation sub-step in step 3. Based on the association rules associated with this customer, the risk factors and their weights involved in the rules are extracted.

[0102] The data format for the high-risk customer list is as follows: Customer ID 10086, risk score 89, predicted claim probability 0.82, main risk factors and their weights are high consultation frequency (weight 0.4), vehicle damage insurance products (weight 0.3), first-year insurance for a new car (weight 0.2), and young driving experience (weight 0.1); Customer ID 20001, risk score 82, predicted claim probability 0.75, main risk factors and their weights are high coverage of commercial third-party liability insurance (weight 0.5), telemarketing channels (weight 0.3), and older vehicles (weight 0.2).

[0103] Finally, the list of high-risk customers is pushed to the underwriting system and claims monitoring system via an API. The API call is a synchronous RESTful API call, the request address is the risk warning data receiving interface provided by the business system, the request method is POST, and the request body is the high-risk customer list data in JSON format, including fields such as customer identifier list, customer score list, and main risk factor list.

[0104] In the business application sub-stage, it is necessary to formulate and implement underwriting intervention strategies and targeted marketing recommendation strategies based on the high-risk customer list. The underwriting intervention strategy includes the following: For high-risk customers, increase the deductible or premium coefficient during the underwriting process. The specific increase is determined based on the risk score. For customers with a risk score between 75 and 85, the deductible is increased by 10% to 20%, and the premium coefficient is increased by 0.1 to 0.2; for customers with a risk score between 85 and 100, the deductible is increased by 20% to 30%, and the premium coefficient is increased by 0.2 to 0.3. For high-risk customers, restrict the purchase of high-risk supplementary insurance, including deductible waiver insurance, passenger liability insurance, and theft insurance. For high-risk customers, designate key claims investigations, and conduct focused reviews of their claims applications during the claims process, increasing the frequency and depth of on-site investigations.

[0105] The targeted marketing recommendation strategy includes the following: For customers with high inquiry frequency but no insurance coverage, a safety service package is attached when recommending vehicle damage insurance products. This package includes vehicle safety inspections, chauffeur services, and emergency roadside assistance, raising customers' risk awareness. For customers who frequently compare products, detailed product comparison reports are sent, including information such as the coverage scope, premium prices, and historical claims ratios of each product, helping customers make more informed insurance decisions and improving conversion rates.

[0106] In this embodiment of the invention, by predicting claims risks based on causal correlation prediction results, the quantitative assessment and early warning of claims risks are effectively realized, which helps enterprises avoid abnormal claims behavior, prevents business losses, and improves business efficiency.

[0107] Please see Figure 2 Based on the same inventive concept, embodiments of the present invention provide a device for mining the correlation between pre-sales customer behavior and post-sales claims results in insurance. The device includes: a data acquisition module for acquiring multi-dimensional data sources from an insurance ERP system; a graph construction module for constructing a graph network structure containing multiple types of nodes and heterogeneous edges based on the multi-dimensional data sources; a network training module for performing high-order feature learning on the graph network structure based on a multi-layer graph attention network to generate a graph attention network model; a correlation mining module for performing multi-hop causal association rule mining based on the graph attention network model to generate causal association prediction results; and a risk prediction module for performing claims risk prediction based on the causal association prediction results to generate prediction results.

[0108] Furthermore, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the methods described in the embodiments of the present invention.

[0109] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above embodiments. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention.

[0110] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not describe the various possible combinations separately.

[0111] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0112] Furthermore, various different implementations of the present invention can be combined arbitrarily, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed in the present invention.

Claims

1. A method for mining the correlation between pre-sale customer behavior and post-sale claims results in insurance, characterized in that, The method includes: Obtain multi-dimensional data sources from the insurance ERP system; Based on the aforementioned multi-dimensional data source, a graph network structure containing multiple types of nodes and multiple types of heterogeneous edges is constructed; A graph attention network model is generated by performing high-order feature learning on the graph network structure based on a multi-layer graph attention network. Based on the graph attention network model, multi-hop causal association rule mining is performed to generate causal association prediction results; Based on the causal correlation prediction results, a claims risk prediction is performed, and a prediction result is generated.

2. The method according to claim 1, characterized in that, The construction of a graph network structure containing multiple types of nodes and multiple types of heterogeneous edges based on the multi-dimensional data source includes: The customer basic information table, policy information table, product information table, channel information table, service event record table, and claims record table are obtained from the multi-dimensional data sources. Based on the customer basic information table and the policy information table, customer basic attribute characteristics and customer node identifiers are determined, and customer nodes are determined based on the customer basic attribute characteristics and customer node identifiers; Based on the product information table, product attribute features and product node identifiers are determined, and product nodes are determined based on the product attribute features and product node identifiers. Based on the channel information table, determine the basic characteristics of the channel and the channel node identifier, and determine the channel node based on the basic characteristics of the channel and the channel node identifier; Based on the service event record table, determine the basic characteristics of the event and the service node identifier, and determine the service event node based on the basic characteristics of the event and the service node identifier; Based on the claims record table, the characteristics of the claims facts and the identifiers of the claims record nodes are determined, and claims record nodes are generated based on the characteristics of the claims facts and the identifiers of the claims record nodes. Based on the customer node, the product node, the channel node, the service event node, and the claims record node, corresponding node edges are generated; A graph network structure is constructed based on the node edges and their corresponding nodes.

3. The method according to claim 2, characterized in that, The process of determining product attribute characteristics based on the product information table includes: Extract the product type field, premium amount data, premium term field, historical claims records, product premium data, and target customer data from the product information table; Based on the product type field, product type characteristics are determined; based on the premium amount data, premium range characteristics are determined; based on the premium term field, coverage period characteristics are determined; based on the historical claims records, historical loss ratio characteristics are determined; based on the product premium data, sales area distribution characteristics are determined; and based on the target customer data, target customer profile characteristics are determined. Product attribute features are generated based on the product type features, premium range features, coverage period features, historical loss ratio features, sales area distribution features, and target customer profile features.

4. The method according to claim 2, characterized in that, The determination of basic channel characteristics based on the channel information table includes: Extract the channel identifier, channel type, historical channel transaction conversion data, channel performance data, historical channel complaint data, and customer evaluation data from the channel information table. Based on the channel identifier, a channel identifier feature is generated; based on the channel type, a channel type feature is generated; based on the channel's historical transaction conversion data, a conversion rate indicator feature is generated; based on the channel performance data, a channel level rating feature is generated; based on the channel's historical complaint data, a historical complaint rate feature is generated; and based on customer evaluation data, a customer satisfaction index feature is generated. Basic channel characteristics are generated based on the channel identification characteristics, channel type characteristics, conversion rate indicator characteristics, channel level rating characteristics, historical complaint rate characteristics, and customer satisfaction index characteristics.

5. The method according to claim 1, characterized in that, The step of performing high-order feature learning on the graph network structure based on a multi-layer graph attention network to generate a graph attention network model includes: Obtain the initial multi-layer graph attention network; Initialize the node attributes in the graph network structure to obtain a low-dimensional dense vector; In the initial multi-layer graph attention network, multiple parallel attention heads are set in each layer of the graph attention network. Each attention head independently calculates the attention coefficients between nodes and generates attention weights based on the attention coefficients. Based on the attention weights, a neighbor feature aggregation operation is performed on the low-dimensional dense vector to generate an updated node feature vector; The multi-layer graph attention network is stacked in multiple layers to obtain a multi-layer stacked structure; A multi-layer graph attention network model is constructed based on the multi-layer stacked structure and the updated node feature vectors.

6. The method according to claim 5, characterized in that, The generation of the updated node feature vector includes: The transformed feature vectors of adjacent nodes are concatenated, and the concatenation result is input into the LeakyReLU activation function for nonlinear transformation to generate the transformation result; The attention weights are obtained by normalizing the transformation results. Based on the attention weights, the neighbor node features are weighted and summed to obtain the aggregated neighbor feature vector; The aggregated neighbor feature vectors are input into an activation function for nonlinear transformation to generate updated node feature vectors.

7. The method according to claim 1, characterized in that, The step of performing multi-hop causal association rule mining based on the graph attention network model to generate causal association prediction results includes: Define the constraints; Based on the constraints, a constrained graph traversal operation is performed on the graph attention network model according to the breadth-first search algorithm to obtain a path set; The path combination is processed using a frequent itemset growth algorithm to obtain the corresponding frequent itemsets. Perform association rule transformation processing on the frequent itemsets to obtain the target association rule; Causal association prediction results are generated based on the target association rules.

8. The method according to claim 7, characterized in that, The process of combining the paths based on the frequent itemset growth algorithm to obtain the corresponding frequent itemsets includes: Perform format conversion processing on the path set to obtain processed transaction data; Traverse the transaction database and count the support score of each processed transaction data. Construct an item header table based on the support count; Based on the item header table, a conditional frequent itemset tree is recursively constructed. The frequent itemset tree is filtered based on a preset support threshold to obtain frequent itemsets.

9. The method according to claim 8, characterized in that, The step of performing association rule transformation on the frequent itemsets to obtain the target association rule includes: Based on the triggering effect, perform front and back item partitioning for each frequent itemset to obtain the front and back items; Based on the preceding and following terms, the frequent itemset is transformed into multiple association rules; Determine the support, confidence, and lift of each association rule; Each association rule is filtered based on the confidence level and the lift level to obtain the filtered rules. The filtered rules are sorted to generate target association rules.

10. A device for identifying the correlation between pre-sale customer behavior and post-sale claims results in insurance, characterized in that, The device includes: The data acquisition module is used to acquire multi-dimensional data sources from the insurance ERP system; The graph construction module is used to construct a graph network structure containing multiple types of nodes and multiple types of heterogeneous edges based on the multi-dimensional data source. The network training module is used to perform high-order feature learning on the graph network structure based on a multi-layer graph attention network to generate a graph attention network model. The association mining module is used to perform multi-hop causal association rule mining based on the graph attention network model and generate causal association prediction results. The risk prediction module is used to perform claims risk prediction based on the causal correlation prediction results and generate prediction results.

Images