Behavior prediction method and device, computer equipment and storage medium

Abstract

The invention provides a behavior prediction method and device, computer equipment and a storage medium, and the method comprises the steps: obtaining the user information of a target user and a constructed target network diagram, the target network diagram comprises a plurality of nodes and connection edges among the nodes, each node corresponds to a node feature vector, and the connection edges of the nodes correspond to the node feature vectors; the types of the nodes comprise historical user nodes, insurance agent nodes and geographic nodes, node feature vectors of the historical user nodes are determined based on user information of historical users, and node feature vectors of the insurance agent nodes and the geographic nodes are determined through embedded layer learning when the target neural network is obtained through training; based on the user information of the target user, updating the target network diagram to generate an updated target network diagram; and utilizing the target neural network to perform feature extraction on the updated target network diagram, and determining the insurance buying behavior probability of the target user. According to the invention, accurate identification of the insured customer can be improved, and the prediction accuracy of the insured behavior probability is improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and more specifically, to a behavior prediction method, apparatus, computer device, and storage medium. Background Technology

[0002] In the insurance industry, predicting a user's insurance consumption behavior—that is, determining the probability of a user purchasing insurance—can facilitate insurance promotion. Generally, traditional econometrics and machine learning models can be used to determine the probability of a user's insurance purchase behavior. Traditional models, such as Logistic Regression (LR) and ensemble models (such as XGBoost), typically rely on carefully designed feature engineering. Researchers need to manually extract individual attributes (such as age) and network features (such as node PageRank values ​​or degree centrality) as input variables to the model, attempting to establish a statistical correlation between these features and insurance purchase decisions.

[0003] However, traditional models rely heavily on manual feature engineering and cannot model complex network relationships end-to-end. Furthermore, these models require analysts or engineers to predefine and extract network features, a process that is not only time-consuming and labor-intensive but also highly dependent on personal experience and domain knowledge. More importantly, this method cannot capture the nonlinear, dynamic, and deep interaction patterns inherent in the complex networks behind insurance consumer behavior, resulting in low accuracy in predicting behavioral probabilities. Summary of the Invention

[0004] In view of this, this application provides a behavior prediction method, apparatus, computer equipment, and storage medium, which significantly improves the accurate identification of insured customers and increases the accuracy of the probability of insurance behavior.

[0005] Specifically, this application is implemented through the following technical solution: In a first aspect, embodiments of this application provide a behavior prediction method, including: The system acquires user information of target users and constructs a target network graph, wherein the target network graph includes multiple nodes and connecting edges between multiple nodes, each node has a corresponding node feature vector, the types of nodes include historical user nodes, insurance agent nodes and geographical nodes, and the node feature vector of the historical user node is determined based on the user information of the historical user, and the node feature vectors of the insurance agent node and the geographical node are determined by learning through the embedding layer when training the target neural network used for insurance probability prediction; Based on the user information of the target user, the target network graph is updated to generate an updated target network graph; Using the target neural network, feature extraction is performed on the updated target network graph to determine the probability of the target user's insurance purchase behavior.

[0006] In one optional implementation, the user information includes a user identifier, the identifier of the insurance agent matched with the user, age information, the number of visits between the user and the insurance agent, the activity level of the application, and the geolocation code of the user.

[0007] In one optional implementation, the target neural network includes multiple graph convolutional layers, a first graph convolutional layer for processing the positive social edges corresponding to the target user node, a second graph convolutional layer for processing the negative social edges corresponding to the target user node, a third graph convolutional layer for processing the positive geographic edges corresponding to the target user node, and a fourth graph convolutional layer for processing the negative geographic graph edges corresponding to the target user node. The positive social edge refers to an edge from a user node to another user node and / or an edge from a user node to an insurance agent node; the negative social edge refers to an edge from another user node to a user node and / or an edge from an insurance agent node to a user node; the positive geographic edge refers to an edge from a user node to a geographic node, and the negative geographic edge refers to an edge from a geographic node to a user node.

[0008] In one optional implementation, the step of using the target neural network to extract features from the updated target network graph and determine the probability of the target user's insurance purchase behavior includes: The multiple graph convolutional layers in the target neural network are used to extract features from the multiple edges corresponding to the target user nodes in the updated target network graph, thereby generating multiple intermediate feature data corresponding to the target user. The multiple intermediate feature data corresponding to the target user are aggregated to generate target feature data corresponding to the target user. The target feature data is processed using a classifier in the target neural network to determine the probability of the target user's insurance purchase behavior.

[0009] In one optional implementation, the aggregation operation includes one of the following: taking the maximum value, taking the average value, summing, and weighted summing using an attention mechanism.

[0010] In one optional implementation, the target neural network is trained according to the following steps: Obtain a sample dataset, wherein the sample dataset includes historical user information, insurance agent information, and geographic region information; Based on the sample dataset, an initial network graph is constructed, wherein the initial network graph includes: historical user nodes, insurance agent nodes, and geographical nodes, as well as bidirectional edges between different nodes; The initial feature vectors of the historical user nodes in the initial network graph are determined based on historical user information; Using the first embedding layer of the neural network to be trained, the learnable initial feature vector of the insurance agent node in the initial network graph is determined, and using the second embedding layer, the learnable initial feature vector of the geographic node is determined. By utilizing multiple graph convolutional layers of the neural network to be trained, features are extracted from the initial network graph, which includes the initial feature vectors of each node, to generate target feature data corresponding to multiple historical users. The target feature data corresponding to the multiple historical users are processed using the classifier of the neural network to be trained, and the predicted insurance probability corresponding to the historical users in the sample dataset is generated. Based on the predicted insurance probability and insurance status information corresponding to the multiple historical users, the parameters of the neural network to be trained are adjusted to generate the target neural network.

[0011] In one optional implementation, the initial feature vector of the historical user node is determined as the node feature vector of the historical user node; the target neural network includes a first embedding matrix of a first embedding layer and a second embedding matrix of a second embedding layer obtained through training, wherein the first embedding matrix is ​​used to determine the node feature vector corresponding to the insurance agent node based on the node identifier of the insurance agent node, and the second embedding matrix is ​​used to determine the node feature vector corresponding to the geographical node based on the node identifier of the geographical node.

[0012] Secondly, embodiments of this application also provide a behavior prediction device, comprising: The acquisition module is used to acquire user information of the target user and a constructed target network graph, wherein the target network graph includes multiple nodes and connecting edges between the nodes, each node has a corresponding node feature vector, the types of the nodes include historical user nodes, insurance agent nodes and geographical nodes, and the node feature vector of the historical user node is determined based on the user information of the historical user, and the node feature vectors of the insurance agent node and the geographical node are determined by learning through the embedding layer when training the target neural network for predicting the probability of insurance. The update module is used to update the target network graph based on the user information of the target user, and generate an updated target network graph. The determination module is used to extract features from the updated target network graph using the target neural network to determine the probability of the target user's insurance purchase behavior.

[0013] This application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described behavior prediction method.

[0014] This application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described behavior prediction method.

[0015] The method provided in this application obtains user information of the target user and constructs a target network graph. The target network graph includes multiple nodes and connecting edges between the nodes. Each node corresponds to a node feature vector. The types of nodes include historical user nodes, insurance agent nodes, and geographical nodes. The node feature vector of historical user nodes is determined based on the user information of historical users. The node feature vectors of insurance agent nodes and geographical nodes are determined through embedding layer learning when training the target neural network used for insurance probability prediction. It can be seen that different nodes in the target network graph of this application have different node feature vectors, explicitly distinguishing node and edge information of different semantic types. In order to update the target network graph using the user information of the target user and generate the updated target network graph, the target neural network can be used to extract features from the updated target network graph to more accurately determine the probability of the target user's insurance behavior. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of this specification and form part of this specification, illustrate exemplary embodiments and are used to explain this specification, but do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating a behavior prediction method in an exemplary embodiment of this application; Figure 2 This is a schematic diagram of the structure of a target neural network shown in an exemplary embodiment of this application; Figure 3 This is a schematic diagram of a behavior prediction device shown in an exemplary embodiment of this application; Figure 4 This is a schematic diagram of the structure of a computer device provided in this application. Detailed Implementation

[0017] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0018] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0019] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0020] One approach using related technologies involves determining the probability of a user's insurance purchase behavior using traditional econometrics, machine learning models, and other methods. Traditional models, such as Logistic Regression (LR) and ensemble models (like XGBoost), typically rely on carefully designed feature engineering. Researchers need to manually extract individual attributes (such as age) and network features (such as node PageRank values ​​or degree centrality) as input variables to the model, attempting to establish a statistical correlation between these features and insurance purchase decisions.

[0021] However, traditional models heavily rely on manual feature engineering, making it impossible to model complex network relationships end-to-end. These models require analysts or engineers to predefine and extract network features, a process that is not only time-consuming and labor-intensive but also highly dependent on personal experience and domain knowledge. More importantly, this approach fails to capture the nonlinear, dynamic, and deep interaction patterns inherent in the complex networks behind insurance consumption behavior, resulting in low accuracy in predicting behavioral probabilities. For example, the models cannot automatically learn the potential synergistic or competitive effects between geographical proximity (such as living in the same town) and online social interactions (such as app communication). Traditional models simplify the network into flat feature vectors, making it difficult to effectively distinguish and integrate the two fundamentally different influence paths and their interaction mechanisms: geographical modalities (reflecting regional economic levels and offline trust building) and social modalities (reflecting information diffusion and online influence). This is precisely the core of understanding the "peer effect" in insurance consumption. Experimental results show that even with carefully selected network features, the best traditional model (such as XGBoost) only achieves a macro-F1 score of 53%, far lower than the innovative solutions to be introduced later.

[0022] Another approach involves using homogeneous graph neural network models to predict the probability of user insurance purchase behavior. However, homogeneous graph neural networks obfuscate heterogeneous information, limiting performance and making it difficult to identify critical minority classes. Specifically, models like GraphSAGE and GAT treat all nodes and edges as homogeneous, leading to the indiscriminate mixing of node and edge information of different semantic types during message transmission. For example, messages from neighbors in the same town (potentially reflecting regional commonalities) and messages from neighbors with active online interactions (potentially reflecting social influence) are processed and aggregated in the same way. This information obfuscation severely blurs the essential characteristics of geographical proximity and social interaction—two differentiated influence paths—making it impossible for the model to design specific and optimized information propagation and aggregation mechanisms for different modalities (such as spatial embedding needs at the geographical level and the directionality of social relationships). Crucially, in the scenario of insurance consumption prediction, where there is a severe class imbalance (i.e., insured customers are a minority), this deficiency leads to serious consequences. Experimental data shows that while homogeneous GNN models (GAT and GraphSAGE) achieved an accuracy of approximately 90%, their macro-F1 score, a key metric for identifying insured customers (a minority class), was as low as 47%. This indicates that in pursuit of overall accuracy, the models heavily favored predicting all samples as the majority "uninsured" category, failing to effectively utilize network structure information to solve core business problems. The fundamental reason is that homogeneous models cannot effectively extract discriminative information with strong differentiation, especially sensitivity to geographical modalities, from heterogeneous network data.

[0023] Based on this, this application proposes a behavior prediction method. This method acquires user information of the target user and constructs a target network graph. The target network graph includes multiple nodes and connecting edges between them. Each node has a corresponding node feature vector. Node types include historical user nodes, insurance agent nodes, and geographical nodes. The node feature vector of historical user nodes is determined based on the historical user's user information. The node feature vectors of insurance agent nodes and geographical nodes are determined through embedding layer learning during the training of the target neural network used for insurance probability prediction. It is evident that different nodes in the target network graph of this application have different node feature vectors, explicitly distinguishing node and edge information of different semantic types. This allows for updating the target network graph using the target user's user information. After generating the updated target network graph, the target neural network can be used to extract features from the updated target network graph, accurately determining the insurance probability of the target user.

[0024] Furthermore, due to the display distinction of different nodes and edges in the target network graph, when using the target neural network to extract features from the updated target network graph, it is possible to explicitly distinguish and efficiently integrate heterogeneous information from both geographic (such as the interaction between the target user and geographic nodes) and social (such as the interaction between the target user and the insurance agent) modalities. In particular, it makes full use of the strong distinguishing ability of the geographic modality for low-frequency categories, significantly improving the accurate identification of insured customers and increasing the accuracy of the probability of insurance behavior.

[0025] To facilitate understanding of this embodiment, a behavior prediction method disclosed in this disclosure will first be described in detail. The execution entity of the behavior prediction method provided in this disclosure is generally a computer device with a certain computing capability. This computer device may include, for example, a terminal device, a server, or other processing devices. The terminal device may be a user equipment (UE), a mobile device, a user terminal, a personal digital assistant (PDA), a handheld device, a computing device, an embedded device, etc. In some possible implementations, this behavior prediction method can be implemented by a processor calling computer-readable instructions stored in memory.

[0026] See Figure 1 The diagram shows a flowchart of a behavior prediction method provided in an embodiment of this disclosure. The method includes steps S101 to S103, wherein: S101. Obtain user information of the target user and construct a target network graph, wherein the target network graph includes multiple nodes and connecting edges between multiple nodes, each node corresponds to a node feature vector, the types of the nodes include historical user nodes, insurance agent nodes and geographical nodes, and the node feature vector of the historical user node is determined based on the user information of the historical user, and the node feature vectors of the insurance agent node and the geographical node are determined by learning through the embedding layer when training the target neural network used for insurance probability prediction.

[0027] The target user's information includes, but is not limited to, user identifier, the identifier of the insurance agent matched with the user, age information, the number of visits between the user and the insurance agent, the activity level of the application, and the geolocation code of the user.

[0028] Each insurance agent has a unique identifier, and these identifiers differ between agents. The insurance agent matched to a user can be the one who will handle business with the target user. The number of visits between a user and an insurance agent can be measured by the number of times they meet in person.

[0029] Application activity can be measured by factors such as the frequency of user logins / use of the application. This application may include one or more applications related to insurance business, where application activity reflects the intensity of user online interaction.

[0030] The user's geolocation code can be an identifier corresponding to the user's geographical location. This geographical location can be detailed down to the specific county or town. For example, the geographical location can include at least three levels: city, county / district, and town / street.

[0031] During implementation, a target network graph can be obtained, which includes multiple nodes and the edges connecting them. This target network graph is constructed based on an acquired sample dataset, which includes user information for multiple historical users (i.e., historical user information), information for multiple insurance agents (i.e., insurance agent information), and geographical region information. The types of nodes include historical user nodes, insurance agent nodes, and geographical nodes. Specifically, nodes corresponding to historical users are historical user nodes, nodes corresponding to insurance agents are insurance agent nodes, and one or more geographical nodes are constructed for each geographical location. For example, if the geographical location is: xx city xx county xx town, then three geographical nodes can be constructed: one city-level geographical node for xx city, one county-level geographical node for xx county, and one town-level geographical node for xx town.

[0032] In a network of nodes, if any two nodes have an interactive relationship, then there is a connecting edge between them. For example, if historical user 1 and historical user 2 have an online communication relationship, then there is a connecting edge between the node corresponding to historical user 1 (i.e., the historical user node) and the node corresponding to historical user 2. Similarly, if historical user 1 is maintained by insurance agent 1, then there is a connecting edge between the node corresponding to historical user 1 and the node corresponding to insurance agent 1. Furthermore, if historical user 1 belongs to xx county in xx city, then there is a connecting edge between the node corresponding to historical user 1 and the geographical node corresponding to xx county.

[0033] In this application, to more accurately extract node features, bidirectional connecting edges are set between different nodes. During implementation, the source and target nodes of a forward edge can be interchanged to generate a corresponding directional edge, ensuring the symmetry of the connecting edges in the target network graph. This means ensuring that each forward edge has a corresponding reverse edge, forming a complete bidirectional graph structure. For example, when determining the social relationship between historical user 1 and insurance agent 1, there are two connecting edges between the nodes of historical user 1 and insurance agent 1: one connecting edge from the node corresponding to historical user 1 to the node corresponding to insurance agent 1, and another connecting edge from the node corresponding to insurance agent 1 to the node corresponding to historical user 1.

[0034] Furthermore, each connection edge in this application corresponds to a connection edge type, so that different convolutional modules can be used to extract features from different types of connection edges. Connection edge types include social edges and geographical edges. Social edges include connection edges between user nodes and insurance agent nodes, and connection edges between different user nodes; geographical edges refer to connection edges between user nodes and geographical nodes, connection edges between insurance agents and geographical nodes, etc.

[0035] For example, in the target network graph, geographical nodes at different levels can be connected by connecting edges, such as bidirectional connecting edges between town nodes and county nodes, bidirectional connecting edges between county nodes and city nodes, etc. That is, a town node can be connected to its county node, a county node can be connected to its city node, a city node is connected to its included county nodes, a county node is connected to its included town nodes, etc., forming a tree-like spatial structure, so that the node feature vector of each node can be determined more accurately in the future, and spatial feature aggregation can be achieved.

[0036] Each node in the target network graph corresponds to a node feature vector. The node feature vector of historical user nodes is determined based on historical user information, such as age, number of visits, and application activity level. The node feature vectors of insurance agent nodes and geographic nodes are determined through embedding layers during the training of the target neural network. The target neural network is used for predicting insurance application probabilities.

[0037] S102. Based on the user information of the target user, update the target network graph to generate an updated target network graph.

[0038] A new user node corresponding to the target user is added to the target network graph. The node feature vector of the user node is determined based on the user information, as well as the connection edge information between the user node and the node in the target network graph, thus generating an updated target network graph.

[0039] S103. Using the target neural network, feature extraction is performed on the updated target network graph to determine the probability of the target user's insurance purchase behavior.

[0040] During implementation, the updated target neural network is input into the target neural network for feature extraction to determine the target feature data corresponding to the target user, and the probability of the target user's insurance behavior is determined based on the target feature data corresponding to the target user.

[0041] Optionally, the target neural network includes multiple graph convolutional layers, a first graph convolutional layer for processing the positive social edges corresponding to the target user node, a second graph convolutional layer for processing the negative social edges corresponding to the target user node, a third graph convolutional layer for processing the positive geographic edges corresponding to the target user node, and a fourth graph convolutional layer for processing the negative geographic graph edges corresponding to the target user node.

[0042] The positive social edge refers to an edge from a user node to another user node and / or an edge from a user node to an insurance agent node; the negative social edge refers to an edge from another user node to a user node and / or an edge from an insurance agent node to a user node; the positive geographic edge refers to an edge from a user node to a geographic node, and the negative geographic edge refers to an edge from a geographic node to a user node.

[0043] The target neural network comprises multiple graph convolutional layers, such as SAGEConv. Each graph convolutional layer processes a type of connection edge. Specifically, connection edges include social edges and geographic edges. Social edges include forward social edges and reverse social edges, while geographic edges include forward geographic edges and reverse geographic edges. Social edges can represent service or interaction relationships between users and agents, and social relationships between users; geographic edges can represent membership relationships between users and geographic nodes, as well as membership and inclusion relationships between geographic levels.

[0044] Positive social edges can be used to represent relationships between users and agents or other users, physically reflecting a user actively seeking services or contacting other users. A positive social edge can be an edge pointing from a user node to another user node, or an edge pointing from a user node to an insurance agent node. When a positive social edge points from a user node to another user node, it represents a social relationship between the user and other users, reflecting the user actively contacting other users. When a positive social edge points from a user node to an insurance agent node, it represents that the user is served or contacted by the insurance agent, reflecting the user actively contacting the insurance agent. This connection edge can be used to determine the spread of the agent's influence to the user.

[0045] Reverse social edges refer to edges pointing from other user nodes to a user node, or from insurance agent nodes to user nodes. They represent the relationship where the agent is contacted by a user; physically, this means the agent passively receives contact requests. Reverse social edges are the reverse of forward social edges and are primarily used to model potential social diffusion paths between different users through a common insurance agent, i.e., the indirect influence that may exist between different users of the same insurance agent. Furthermore, reverse social edges can also reflect the potential influence of other users on other users.

[0046] A forward geographic edge is an edge pointing from a user node to a geographic node, representing the geographic region to which the user belongs, i.e., the user's geographic affiliation. A reverse geographic edge is an edge pointing from a geographic node to a user node, i.e., the opposite edge of a forward geographic edge, representing that the region contains users, physically reflecting the user composition of that geographic region (i.e., the users included in the geographic region).

[0047] The user nodes include target user nodes and historical user nodes. Specifically, the positive social edge corresponding to the target user node is the connection edge from the target user node to the insurance agent node, and the positive social edge corresponding to the historical user node is the connection edge from the historical user node to the insurance agent node.

[0048] The target network graph in this application contains three types of heterogeneous nodes: users, agents, and geographic nodes, as well as two types of bidirectional connection edges: social modality edges (such as forward social edge contact and reverse social edge inverse_contact) and geographic modality edges (such as forward geographic edge located_in, reverse geographic edge inverse_location, and bidirectional hierarchical edges). It explicitly distinguishes the two influence paths with different physical meanings: geographic proximity and social interaction, thus providing support for the subsequent processing of the target neural network.

[0049] This application sets up bidirectional edges to describe the same event (i.e., the corresponding positive and negative social edges are generated based on the same contact record), so as to realize the transmission of information in the target network graph. That is, information can be transmitted through two connecting edges in different directions, so that the target neural network can learn richer features based on the target network graph, thereby improving the expressive power of the target neural network model and enabling more accurate prediction of the probability of users' insurance behavior.

[0050] For example, a positive social edge enables information to flow from user nodes to agent nodes. In the message passing of the target neural network, the agent node can aggregate information from its users (for example, through this connection edge, agent A can learn that multiple users have been contacted). A negative social edge enables information to flow from agent nodes to user nodes. The user node can aggregate information from the agents it has contacted (for example, through this connection edge, user B can learn about the conversion rate of the agents it has contacted).

[0051] In this application, the key to setting bidirectional edges in heterogeneous graph neural networks is that the direction of the connecting edges can determine the direction of information flow and the content of aggregation, rather than describing different events. Therefore, the setting of bidirectional edges provides strong support for the subsequent processing of the target neural network.

[0052] Specifically, the purpose of setting bidirectional edges is to ensure that information can flow bidirectionally between different node types in a heterogeneous graph, playing a crucial role in the learning of neural networks. Specifically, it ensures the integrity of information propagation. On the one hand, it avoids information silos; that is, setting only unidirectional edges would cause information to flow only from the source node to the target node, unable to propagate back, forming a dead end in information propagation. On the other hand, it supports bidirectional message passing. Subsequently, in the message passing framework of the target neural network, each node needs to aggregate information from neighbors in all directions. Bidirectional edges ensure that user nodes can aggregate information from agents and geographical regions, agent nodes can aggregate information from users, and geographical nodes can aggregate information from users.

[0053] By setting up bidirectional edges, the capabilities of the target neural network model are enhanced. On the one hand, the feature representation is enriched, and bidirectional edges enable each node to obtain more comprehensive contextual information. For example, user features not only include their own attributes, but also the features of their connected agents and the statistical features of their geographical region. For example, agent features can include the group features of the users they serve. On the other hand, it supports complex propagation patterns, allowing information to spread between multi-hop neighbors, such as indirect connections between users, agents, and other users.

[0054] By setting up multiple graph convolutional layers and utilizing them to extract features from different types of connection edges, a bimodal heterogeneous message passing mechanism is implemented. Specifically, for the social modality, a bidirectional graph convolutional path is designed (customer -> agent aggregates agent influence + agent -> customer captures potential social diffusion); for the geographical modality, a hierarchical spatial embedding propagation path is designed (multi-scale spatial feature aggregation is achieved through customer -> town and town -> county -> city and backpropagation). This modality-specific message passing design effectively avoids the information confusion problem in homogeneous graph models and can effectively extract the features of connection edges in the target network graph, so as to accurately determine the probability of user insurance behavior in the future.

[0055] In practice, multiple graph convolutional layers in the target neural network can be used to extract features from various connection edges corresponding to the target user, so as to determine the target feature data corresponding to the target user based on the feature data extracted by multiple graph convolutional layers.

[0056] This application sets up independent convolutional layers, with each edge in each direction using an independent graph convolutional layer such as the SAGEConv layer, to learn different information transmission patterns in different directions; and sets up multiple graph convolutional layers to achieve parameter separation, that is, the convolutional layer parameters of the graph convolutional layers corresponding to the forward connection edge and the reverse connection edge are not shared, so as to achieve feature extraction in different semantics. For example, the forward connection edge reflects "the customer contacts the salesperson" and the reverse connection edge reflects "the salesperson is contacted by the customer", which are semantically different in the semantic feature extraction of the target neural network.

[0057] The bidirectional edge configuration in this application enables an information propagation mechanism. On the one hand, during forward propagation, each node receives neighbor information through different types of edges. For example, a user node can receive agent-related information through a reverse social edge and send information to a geographic node through a forward geographic edge. Similarly, an agent node can receive user information through a forward social edge, and a geographic node can receive user information through a reverse geographic edge. On the other hand, after multiple iterations, the information can be fully propagated in the target network graph. For example, information about directly connected neighbors can be obtained through one-hop propagation, while information about indirectly connected neighbors can be obtained through multi-hop propagation (e.g., similar users of a user can be obtained through the path "user → agent → user").

[0058] Optionally, the step of using the target neural network to extract features from the updated target network graph and determine the probability of the target user's insurance purchase behavior includes: Step a1: Use the multiple graph convolutional layers in the target neural network to extract features from the multiple edges corresponding to the target user node in the updated target network graph, and generate multiple intermediate feature data corresponding to the target user.

[0059] Step a2: Aggregate the multiple intermediate feature data corresponding to the target user to generate target feature data corresponding to the target user.

[0060] Step a3: Use the classifier in the target neural network to process the target feature data and determine the probability of the target user's insurance purchase behavior.

[0061] In implementation, multiple graph convolutional layers in the target neural network can extract features from multiple edges corresponding to the target user node in the updated target network graph. Specifically, the first graph convolutional layer extracts features from the positive social edges corresponding to the target user node, the second graph convolutional layer extracts features from the negative social edges corresponding to the target user node, the third graph convolutional layer extracts features from the positive geographic edges corresponding to the target user node, and the fourth graph convolutional layer extracts features from the negative geographic edges corresponding to the target user node, thus obtaining four intermediate feature data corresponding to the target user.

[0062] Then, the multiple intermediate feature data corresponding to the target user are aggregated, such as by taking the maximum value, to generate the target feature data corresponding to the target user. The aggregation operation includes one of the following: taking the maximum value, taking the average value, summing, or using an attention mechanism for weighted summation.

[0063] The maximum value operation determines the maximum value of multiple element values ​​in the same feature dimension from multiple intermediate feature data, and generates the target feature data based on the maximum values ​​corresponding to each feature dimension. For example, a maximum value function can be used to determine the maximum value of element values ​​in the same feature dimension from multiple intermediate feature data.

[0064] The averaging operation determines the average value of multiple element values ​​of the same feature dimension in multiple intermediate feature data, and generates target feature data based on the average value corresponding to each feature dimension.

[0065] The summation operation generates target feature data by determining the sum of multiple element values ​​of the same feature dimension in multiple intermediate feature data.

[0066] Weighted summation using attention mechanisms involves summing multiple intermediate feature data using attention weights to obtain the target feature data.

[0067] After generating the target feature data corresponding to the target user, the target feature data is processed by the classifier in the target neural network. For example, the classifier may include a fully connected layer (Linear Layer) and a Sigmoid activation function. The fully connected layer maps the high-dimensional embedding to a scalar value, and the Sigmoid function compresses this scalar value to the interval [0, 1] to determine the probability of the target user's insurance behavior.

[0068] In this application, multiple graph convolutional layers in the target neural network are used to extract features from multiple edges corresponding to the target user nodes in the updated target network graph, generating multiple intermediate feature data corresponding to the target user. These intermediate feature data are then aggregated to fuse multiple feature representations into a unified representation, enabling collaborative reasoning of geographic and social information. This allows the target neural network to automatically learn the comprehensive impact of the interaction between the two modalities on insurance decisions, and subsequently, the aggregated target feature data can be used to more accurately determine the probability of the target user's insurance behavior.

[0069] This application's target network graph based on heterogeneous bidirectional edges, utilizing a target neural network to predict the probability of a user's insurance purchase behavior, has the following advantages: First, optimized gradient propagation. Specifically, the bidirectional connection edge configuration ensures that gradients can propagate in any direction, avoiding gradient vanishing or exploding; and it supports deeper network structures, capturing more complex graph patterns. Second, enhanced representation learning. Specifically, node embeddings (i.e., node feature data) include contextual information from all relevant nodes, and the representations learned by the network can simultaneously encode the node's own features and its structural role in the network. Third, improved predictive ability. For the probability of a user's insurance purchase behavior, the network model can consider the user's direct features (i.e., their own attributes), the user's social environment (such as the characteristics and success rate of contacting agents), and the user's geographical environment (the economic level of their region). It is evident that this design of a bidirectional heterogeneous graph neural network physically corresponds to a real bidirectional relationship, achieving comprehensive information flow and effective aggregation at the neural network level, thus improving the network model's representational ability and predictive performance.

[0070] The process of training the target neural network is explained below.

[0071] Optionally, the target neural network is trained according to the following steps: Step b1: Obtain a sample dataset, wherein the sample dataset includes historical user information, insurance agent information, and geographic area information.

[0072] Step b2: Based on the sample dataset, construct an initial network graph, wherein the initial network graph includes: historical user nodes, insurance agent nodes, and geographical nodes, as well as bidirectional edges between different nodes.

[0073] Step b3: Determine the initial feature vector of the historical user node in the initial network graph based on historical user information.

[0074] Step b4: Using the first embedding layer of the neural network to be trained, determine the learnable initial feature vector of the insurance agent node in the initial network graph, and using the second embedding layer, determine the learnable initial feature vector of the geographic node.

[0075] Step b5: Using multiple graph convolutional layers of the neural network to be trained, perform feature extraction on the initial network graph, which includes the initial feature vectors of each node, to generate target feature data corresponding to multiple historical users.

[0076] Step b6: Use the classifier of the neural network to be trained to process the target feature data corresponding to the multiple historical users, and generate the predicted insurance probability corresponding to the historical users in the sample dataset.

[0077] Step b7: Based on the predicted insurance probability and insurance status information corresponding to the multiple historical users, adjust the parameters of the neural network to be trained to generate the target neural network.

[0078] The sample dataset includes historical user information, insurance agent information, and geographic region information. Historical user information includes, but is not limited to: historical user identifiers, identifiers of insurance agents matched with the user, age information, number of visits between the user and the insurance agent, application activity level, geographic location code of the user, and user insurance status. Insurance agent information may include insurance agent identifiers. Geographic region information includes, but is not limited to: the geographic region served by the insurance agent, and the geographic location code of the historical user (at least including three levels: city, county / district, and town / street).

[0079] Based on the sample dataset, multiple nodes are identified, with each historical user and each insurance agent in the dataset treated as a historical user node and an insurance agent node, respectively. Geographic information is then converted into hierarchical geographic nodes, including city, county / district, and town / street nodes. Furthermore, bidirectional connections between these nodes are determined using the sample dataset. These bidirectional connections include forward social edges, reverse social edges, forward geographic edges, and reverse geographic edges. The multiple nodes and their connections constitute the initial network graph.

[0080] Based on historical user information, initial feature vectors for historical user nodes in the initial network graph are determined. These initial feature vectors could include age, number of visits, and application activity. Considering that agent attributes may be missing or inconsistent, the first embedding layer of the neural network to be trained can be used to determine learnable initial feature vectors for insurance agent nodes, enabling the neural network to learn node feature vectors for insurance agent nodes during training. Similarly, the second embedding layer of the neural network to be trained is used to determine learnable initial feature vectors for geographic nodes, enabling the neural network to learn node feature vectors for geographic nodes during training.

[0081] After determining the initial feature vectors of each node, multiple graph convolutional layers of the neural network to be trained are used to extract features from various connection edges corresponding to historical users in the initial network graph, generating target feature data for historical users. In practice, the multiple graph convolutional layers can extract features from the forward social edges, reverse social edges, forward geographic edges, and reverse geographic edges corresponding to historical users, respectively. For example, they can extract features from the node information connected by the connection edges, generating multiple intermediate feature data corresponding to historical users. These multiple intermediate feature data are then aggregated to obtain the target feature data for historical users.

[0082] Specifically, for positive social edges, i.e. the edge type (contact) from a user to an insurance agent, a graph convolution operator based on neighbor sampling (i.e., the first graph convolution layer, such as SAGEConv) is used to aggregate the information of the agent's neighbor nodes, update the influence features of the user node on the insurance agents it has contacted, and obtain the intermediate feature data related to the positive social edge of the user.

[0083] For reverse social edges, i.e., the type of edge from the insurance agent to the user (inverse_contact), a graph convolution operator (i.e., a second graph convolutional layer, such as SAGEConv) is also used to aggregate information from the user's neighbor nodes and update the features of the agent node. This step aims to capture the potential social diffusion effect between different users connected through the agent, and can determine the intermediate feature data related to the reverse social edge corresponding to the user.

[0084] For positive geographic edges, i.e., edges pointing from users to geographic nodes (such as town nodes) (located_in): a graph convolution operator (i.e., a third graph convolution layer, such as SAGEConv) is used to aggregate geographic location features and update the representation of the user node. This step can capture the feature information of the user's geographic location, and thus determine the intermediate feature data related to the user's corresponding positive geographic edge.

[0085] For inverse geographic edges, i.e., edges pointing from geographic nodes (such as town nodes) to users (inverse_location): A graph convolution operator (i.e., the fourth graph convolution layer, such as SAGEConv) is used to aggregate information from the customer's neighboring nodes (belonging to the same town) and update the representation of the geographic node (town). This step allows town-level nodes to aggregate the features of customers within their jurisdiction, thus determining the intermediate feature data related to the user's corresponding forward geographic edge.

[0086] In implementation, this application enables hierarchical spatial embedding propagation between geographic nodes, meaning the target neural network can pass messages through geographic hierarchical edges (town to county, county to city). For example, using graph convolution operators, information from subordinate town nodes is aggregated to update the representation of county nodes; then, information from subordinate county nodes is aggregated to update the representation of city nodes. Information can then flow in reverse (or via skip connections); for example, the updated city / county node information can propagate back to town nodes, and even ultimately back to user nodes. This process allows user nodes to obtain multi-scale (town, county, city) spatial context information, improving the feature richness of user nodes and enabling more accurate prediction of users' insurance purchase behavior probabilities.

[0087] After bimodal independent graph convolution computation, each historical user node can obtain updated representations of two modalities: a social modality representation and a geographic modality representation. Furthermore, the social modality representation and the geographic modality representation (i.e., multiple intermediate feature data) can be aggregated to generate the final embedded representation of the historical user node, which integrates geographic and social information (i.e., target feature data).

[0088] Then, the target feature data of multiple historical users are input into the classifier of the neural network to be trained, and the predicted insurance probability corresponding to the historical users in the sample dataset is output.

[0089] Optionally, to avoid overfitting of the neural network, after obtaining the target feature data of historical users, a dropout process can be performed on the target feature data of multiple historical users to obtain processed target feature data for multiple historical users. The dropout rate p can be set as needed, for example, p can be set to 0.5. Then, the processed target feature data of multiple historical users is input into the classifier of the neural network to be trained, and the output is the predicted insurance probability for the historical users in the sample dataset.

[0090] After obtaining the predicted insurance purchase probability of historical users, the loss function value can be determined based on the predicted insurance purchase probability and insurance status information of the historical users. For example, the binary cross-entropy loss function can be used as the loss function for the neural network to be trained. Based on the loss function value, the gradient is calculated using the backpropagation algorithm, and the network parameters of the neural network to be trained (including the weights of the graph convolutional layers, the embedding matrix, the weights of the fully connected layers, etc.) are updated to obtain the adjusted neural network. Then, the process returns to step b4 to iteratively adjust the adjusted neural network again until the training cutoff condition is met. The training cutoff condition can be set according to the actual situation. For example, it can be determined that the training cutoff condition is met when the number of training iterations exceeds a set threshold, or it can be determined that the training cutoff condition is met when the adjusted neural network converges. Finally, the adjusted neural network obtained from the last training iteration can be determined as the target neural network.

[0091] Optionally, the initial feature vector of the historical user node is determined as the node feature vector of the historical user node. After training the target neural network, the target neural network includes a first embedding matrix of the first embedding layer and a second embedding matrix of the second embedding layer, wherein the first embedding matrix is ​​used to determine the node feature vector corresponding to the insurance agent node based on the node identifier of the insurance agent node, and the second embedding matrix is ​​used to determine the node feature vector corresponding to the geographical node based on the node identifier of the geographical node.

[0092] For example, the first embedding matrix can include the node feature vector corresponding to the insurance agent node. The node feature vector of the insurance agent node can be determined from the first embedding matrix using its node identifier. Similarly, the node feature vector of a geographic node can be determined from the second embedding matrix using its node identifier.

[0093] After obtaining the target neural network, this application can also analyze the target neural network to determine the influence strength of geographical areas and the centrality of agent nodes in social networks. This can then guide agents to deploy more effectively in key areas or customer communities, thereby improving the success rate of insurance business promotion.

[0094] See Figure 2 As shown, combined with Figure 2The training process of the neural network to be trained is illustrated exemplarily. During training, an initial network graph can be constructed based on the sample dataset. Initial feature vectors (i.e., Customer Features) of historical user nodes in the initial network graph are determined based on historical user information; these can be 3-dimensional feature vectors. Furthermore, the first embedding layer of the neural network to be trained (e.g., Agent Embedding) is used to determine the learnable initial feature vectors of insurance agent nodes through an embedding lookup method. Similarly, the second embedding layer (e.g., Location Embedding) is used to determine the learnable initial feature vectors of geographic nodes through an embedding lookup method. The dimensions of the initial feature vectors for insurance agent nodes and geographic nodes can be flexibly set, for example, to 256 dimensions.

[0095] For example, according to Figure 2 As can be seen from the embedding search process, the Agent Embedding layer can store the feature vectors corresponding to each insurance agent identifier, so that the node feature vector corresponding to the insurance agent can be determined by inputting the insurance agent identifier ID.

[0096] After obtaining the initial feature vectors of the nodes, the initial network graph, including the initial feature vectors, can be input into multiple graph convolutional layers to extract features from the positive social edges (i.e., user-to-agent), negative social edges (i.e., agent-to-user), positive geographic edges (i.e., user-to-geography), and negative geographic edges (i.e., geography-to-user) of historical users. This yields the social and geographic features of historical users. These features are then aggregated to obtain the target feature data for historical users. A regularization layer is used to dropout the target feature data from multiple historical users, and the processed target feature data is then input into a fully connected layer for further processing. An activation function is used to normalize the feature information output from the fully connected layer, resulting in the predicted insurance probability for multiple historical users.

[0097] This application constructs and applies a specially designed Geosocial HeteroGNN model (i.e., the target neural network) to accurately predict individual users' insurance consumption (insurance purchase) decision-making behavior in an end-to-end manner. This method fully leverages the networked relationships of two key modalities in insurance consumption scenarios: geographical proximity (e.g., offline agent visits, regional clustering) and social interaction (e.g., online application communication). Through heterogeneous graph modeling, a dual-modal collaborative message passing mechanism, and representation learning, it significantly improves the ability to identify insured customers (usually a minority group) and enhances the accuracy of predicting the probability of user insurance purchase behavior.

[0098] The method proposed in this application has the following technical effects: First, this application can improve the identification ability of key minority groups (such as insured customers) (macro-F1).

[0099] In related technologies, especially homogeneous graph neural network models, in scenarios with severe class imbalance (where non-insured customers constitute the vast majority), the pursuit of overall accuracy often sacrifices the ability to identify the key business objective, namely "insured customers" (a minority class), resulting in extremely low macro-F1 scores (experimental results show: GAT / GraphSAGE ~47%). This application significantly improves the accuracy of identifying insured customers by explicitly distinguishing and efficiently fusing heterogeneous information from both geographical and social modalities, particularly by fully utilizing the strong discriminative power of the geographical modality for low-frequency categories. Experimental data demonstrates that the proposed solution achieves a macro-F1 score of 84% on the same dataset, a 37 percentage point improvement compared to homogeneous graph neural networks and a 31 percentage point improvement compared to XGBoost (macro-F1 53%) which incorporates network features. Therefore, this application can more accurately locate potential policyholders, significantly improve the success rate of insurance promotion, and effectively solve business pain points that existing technologies cannot address.

[0100] Second, this application realizes the collaborative modeling of dual-modal information of geographical proximity and social interaction, and captures complex influence mechanisms.

[0101] In related technologies, such as traditional econometric models and homogeneous graph neural networks (GNNs), it is impossible to effectively distinguish and model the two fundamentally different yet synergistic influence paths: the geographical proximity modality (offline agent visits, peer effect in the same region, which reflects regional economy, culture, and offline trust establishment) and the social interaction modality (online application communication, indirect connections generated by co-agents, which reflects information diffusion and online influence). The heterogeneous message passing mechanism (a two-layer spatial embedding propagation path for the geographical modality and a two-way social information diffusion path for the social modality) and feature aggregation strategy (such as maximum value aggregation) of this application enable the model to automatically learn the characteristics of each modality end-to-end (e.g., the geographical modality focuses on hierarchical regional commonalities, while the social modality focuses on the direction of influence propagation) and their interaction effects (e.g., how online activity and regional economic level jointly drive insurance decisions). This provides a powerful technical tool for a deeper understanding of the "peer effect" in insurance consumption behavior, which is not available in related technologies.

[0102] Third, this application avoids tedious and inefficient manual feature engineering and achieves end-to-end learning.

[0103] Traditional econometric models in related technologies (such as LR and XGBoost) heavily rely on analysts pre-defining, calculating, and inputting complex network features (such as PageRank and degree centrality). This process is time-consuming, labor-intensive, and highly dependent on expert experience, making it difficult to capture deep-seated nonlinear dynamic relationships within the network. The heterogeneous graph neural network model (i.e., the target neural network) constructed in this application can automatically learn low-dimensional, high-discriminative embedding representations of nodes through graph convolutional layers, starting from the original graph structure and node attributes. The entire process is completely end-to-end, requiring no manual design or feature extraction steps. This application not only significantly reduces the technical implementation threshold and cost but also avoids information loss or bias caused by improper feature engineering, improving the model's versatility and robustness.

[0104] Fourth, this application achieves high-performance and highly stable prediction results (Accuracy & F1).

[0105] In addition to a significant improvement in the key metric macro-F1, this application also demonstrates superior performance in overall accuracy and weighted F1 score. Experimental data show that the accuracy reaches 95% and the weighted F1 score reaches 95%, both significantly outperforming the benchmark model. Furthermore, the target neural network in this application exhibits good convergence and generalization capabilities, avoiding overfitting and ensuring stable performance across different datasets.

[0106] Corresponding to the embodiments of the aforementioned behavior prediction method, this application also provides embodiments of a behavior prediction device. Figure 3 A schematic diagram of the behavior prediction device provided in this application specifically includes: The acquisition module 301 is used to acquire user information of the target user and a constructed target network graph, wherein the target network graph includes multiple nodes and connection edges between the multiple nodes, each node corresponds to a node feature vector, the types of the nodes include historical user nodes, insurance agent nodes and geographical nodes, and the node feature vector of the historical user node is determined based on the user information of the historical user, and the node feature vectors of the insurance agent node and the geographical node are determined by learning through the embedding layer when training the target neural network used for insurance probability prediction; The update module 302 is used to update the target network graph based on the user information of the target user, and generate an updated target network graph. The determination module 303 is used to extract features from the updated target network graph using the target neural network to determine the probability of the target user's insurance behavior.

[0107] In one optional implementation, the user information includes a user identifier, the identifier of the insurance agent matched with the user, age information, the number of visits between the user and the insurance agent, the activity level of the application, and the geolocation code of the user.

[0108] In one optional implementation, the target neural network includes multiple graph convolutional layers, a first graph convolutional layer for processing the positive social edges corresponding to the target user node, a second graph convolutional layer for processing the negative social edges corresponding to the target user node, a third graph convolutional layer for processing the positive geographic edges corresponding to the target user node, and a fourth graph convolutional layer for processing the negative geographic graph edges corresponding to the target user node. The positive social edge refers to an edge from a user node to another user node and / or an edge from a user node to an insurance agent node; the negative social edge refers to an edge from another user node to a user node and / or an edge from an insurance agent node to a user node; the positive geographic edge refers to an edge from a user node to a geographic node, and the negative geographic edge refers to an edge from a geographic node to a user node.

[0109] In an optional implementation, the determining module 303, when using the target neural network to extract features from the updated target network graph and determine the probability of the target user's insurance purchase behavior, is used to: The multiple graph convolutional layers in the target neural network are used to extract features from the multiple edges corresponding to the target user nodes in the updated target network graph, thereby generating multiple intermediate feature data corresponding to the target user. The multiple intermediate feature data corresponding to the target user are aggregated to generate target feature data corresponding to the target user. The target feature data is processed using a classifier in the target neural network to determine the probability of the target user's insurance purchase behavior.

[0110] In one optional implementation, the aggregation operation includes one of the following: taking the maximum value, taking the average value, summing, and weighted summing using an attention mechanism.

[0111] In an optional embodiment, the apparatus further includes a training module 304 for training the target neural network according to the following steps: Obtain a sample dataset, wherein the sample dataset includes historical user information, insurance agent information, and geographic region information; Based on the sample dataset, an initial network graph is constructed, wherein the initial network graph includes: historical user nodes, insurance agent nodes, and geographical nodes, as well as bidirectional edges between different nodes; The initial feature vectors of the historical user nodes in the initial network graph are determined based on historical user information; Using the first embedding layer of the neural network to be trained, the learnable initial feature vector of the insurance agent node in the initial network graph is determined, and using the second embedding layer, the learnable initial feature vector of the geographic node is determined. By utilizing multiple graph convolutional layers of the neural network to be trained, features are extracted from the initial network graph, which includes the initial feature vectors of each node, to generate target feature data corresponding to multiple historical users. The target feature data corresponding to the multiple historical users are processed using the classifier of the neural network to be trained, and the predicted insurance probability corresponding to the historical users in the sample dataset is generated. Based on the predicted insurance probability and insurance status information corresponding to the multiple historical users, the parameters of the neural network to be trained are adjusted to generate the target neural network.

[0112] In one optional implementation, the initial feature vector of the historical user node is determined as the node feature vector of the historical user node; the target neural network includes a first embedding matrix of a first embedding layer and a second embedding matrix of a second embedding layer obtained through training, wherein the first embedding matrix is ​​used to determine the node feature vector corresponding to the insurance agent node based on the node identifier of the insurance agent node, and the second embedding matrix is ​​used to determine the node feature vector corresponding to the geographical node based on the node identifier of the geographical node.

[0113] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0114] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0115] This application also provides a computer-readable storage medium storing a computer program that can be used to perform the behavior prediction method described in the above embodiments.

[0116] This application also provides a computer device, see [link to relevant documentation] Figure 4 The diagram shown illustrates the structure of the computer device provided in this application. At the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for various business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to implement the behavior prediction method described in the above embodiments. Of course, besides software implementation, this specification does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution entity of the following processing flow is not limited to individual logic units, but can also be hardware or logic devices.

[0117] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of one or more computer-usable storage media (including, but not limited to, disk storage, etc.) containing computer-usable program code. It takes the form of a computer program product implemented on (such as optical memory, etc.).

[0118] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0119] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0120] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the functions specified in one or more boxes. In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0121] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0122] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, and optical disc read-only memory (ODROM). Computer-readable media may be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transient media such as modulated data signals and carrier waves.

[0123] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0124] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0125] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0126] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A behavior prediction method, characterized in that, The method includes: The system acquires user information of target users and constructs a target network graph, wherein the target network graph includes multiple nodes and connecting edges between multiple nodes, each node has a corresponding node feature vector, the types of nodes include historical user nodes, insurance agent nodes and geographical nodes, and the node feature vector of the historical user node is determined based on the user information of the historical user, and the node feature vectors of the insurance agent node and the geographical node are determined by learning through the embedding layer when training the target neural network used for insurance probability prediction; Based on the user information of the target user, the target network graph is updated to generate an updated target network graph; Using the target neural network, feature extraction is performed on the updated target network graph to determine the probability of the target user's insurance purchase behavior.

2. The method according to claim 1, characterized in that, The user information includes the user identifier, the identifier of the insurance agent matched with the user, age information, the number of visits between the user and the insurance agent, the activity level of the application, and the geolocation code of the user.

3. The method according to claim 1, characterized in that, The target neural network includes multiple graph convolutional layers. The first graph convolutional layer is used to process the positive social edges corresponding to the target user node, the second graph convolutional layer is used to process the negative social edges corresponding to the target user node, the third graph convolutional layer is used to process the positive geographic edges corresponding to the target user node, and the fourth graph convolutional layer is used to process the negative geographic graph edges corresponding to the target user node. The positive social edge refers to an edge from a user node to another user node and / or an edge from a user node to an insurance agent node; the negative social edge refers to an edge from another user node to a user node and / or an edge from an insurance agent node to a user node; the positive geographic edge refers to an edge from a user node to a geographic node, and the negative geographic edge refers to an edge from a geographic node to a user node.

4. The method according to claim 3, characterized in that, The step of using the target neural network to extract features from the updated target network graph and determining the probability of the target user's insurance purchase behavior includes: The multiple graph convolutional layers in the target neural network are used to extract features from the multiple edges corresponding to the target user nodes in the updated target network graph, thereby generating multiple intermediate feature data corresponding to the target user. The multiple intermediate feature data corresponding to the target user are aggregated to generate target feature data corresponding to the target user. The target feature data is processed using a classifier in the target neural network to determine the probability of the target user's insurance purchase behavior.

5. The method according to claim 4, characterized in that, The aggregation operation includes one of the following: maximum value operation, average value operation, summation operation, and weighted summation operation using an attention mechanism.

6. The method according to any one of claims 1-5, characterized in that, The target neural network is trained according to the following steps: Obtain a sample dataset, wherein the sample dataset includes historical user information, insurance agent information, and geographic region information; Based on the sample dataset, an initial network graph is constructed, wherein the initial network graph includes: historical user nodes, insurance agent nodes, and geographical nodes, as well as bidirectional edges between different nodes; The initial feature vectors of the historical user nodes in the initial network graph are determined based on historical user information; Using the first embedding layer of the neural network to be trained, the learnable initial feature vector of the insurance agent node in the initial network graph is determined, and using the second embedding layer, the learnable initial feature vector of the geographic node is determined. By utilizing multiple graph convolutional layers of the neural network to be trained, features are extracted from the initial network graph, which includes the initial feature vectors of each node, to generate target feature data corresponding to multiple historical users. The target feature data corresponding to the multiple historical users are processed using the classifier of the neural network to be trained, and the predicted insurance probability corresponding to the historical users in the sample dataset is generated. Based on the predicted insurance probability and insurance status information corresponding to the multiple historical users, the parameters of the neural network to be trained are adjusted to generate the target neural network.

7. The method according to claim 6, characterized in that, The initial feature vector of the historical user node is determined as the node feature vector of the historical user node; the target neural network includes a first embedding matrix of the first embedding layer and a second embedding matrix of the second embedding layer obtained by training, wherein the first embedding matrix is ​​used to determine the node feature vector corresponding to the insurance agent node according to the node identifier of the insurance agent node, and the second embedding matrix is ​​used to determine the node feature vector corresponding to the geographical node according to the node identifier of the geographical node.

8. A behavior prediction device, characterized in that, The device includes: The acquisition module is used to acquire user information of the target user and a constructed target network graph, wherein the target network graph includes multiple nodes and connecting edges between the nodes, each node has a corresponding node feature vector, the types of the nodes include historical user nodes, insurance agent nodes and geographical nodes, and the node feature vector of the historical user node is determined based on the user information of the historical user, and the node feature vectors of the insurance agent node and the geographical node are determined by learning through the embedding layer when training the target neural network for predicting the probability of insurance. The update module is used to update the target network graph based on the user information of the target user, and generate an updated target network graph. The determination module is used to extract features from the updated target network graph using the target neural network to determine the probability of the target user's insurance purchase behavior.

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

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor performs the steps of the method according to any one of claims 1-7.

Images