Big model-based information pushing strategy generation method and device, storage medium and computer device

Abstract

The application discloses a large model-based information pushing strategy generation method and device, a storage medium and computer equipment. The method comprises the following steps: first, obtaining customer original behavior data and dividing time and space dimensions, and constructing a time-space fusion similarity matrix according to the time and space dimensions; second, constructing a graph structure with the customer as a node and extracting customer fusion features; third, searching and matching knowledge information in a preset knowledge base; and finally, constructing prompt words based on the knowledge information and the customer fusion features, and generating an information pushing strategy by a strategy generation large model. The method can be applied to the information pushing strategy generation scene of financial technology and medical health, can help to deeply mine the potential law of customer behavior, accurately grasp the needs of customers in different scenes, improve the accuracy and individuality of the information pushing strategy, improve the efficiency and effect of information pushing, and further improve the satisfaction of customers. The method can be applied to the financial technology and medical health scene.

Description

Technical Field

[0001] This application relates to the field of data analysis technology, and in particular to a method, apparatus, storage medium and computer equipment for generating information push strategies based on large models. Background Technology

[0002] In today's digital age, push notifications have become an important means for businesses to communicate with customers, improve service quality, and promote business growth. Accurate and personalized push notifications are crucial for meeting customer needs, enhancing customer loyalty, and increasing business conversion rates.

[0003] However, existing methods for generating information push strategies have many shortcomings. In the fintech field, traditional methods often formulate push strategies based solely on basic customer information or simple transaction data, ignoring the complexity and dynamism of customer behavior across time and space. For example, some banks, when recommending wealth management products, only consider the customer's account balance and past investment records, failing to take into account changes in the customer's investment preferences over different time periods and differences in consumption behavior across different geographical locations. This results in recommended products not matching the customer's actual needs, leading to poor push effectiveness. In the healthcare field, existing technologies typically push information based on patients' medical records. However, medical records are often static and cannot comprehensively reflect changes in a patient's health status and needs across different times and places. For instance, when hospitals push health management advice to patients, they rely solely on a single medical visit record, without considering the patient's daily activity patterns and lifestyle habits, making the recommendations lack specificity and practicality.

[0004] Current methods for generating information push strategies struggle to deeply uncover potential patterns in customer behavior, resulting in low accuracy and personalization of these strategies, and making it difficult to guarantee the efficiency and effectiveness of information push. Summary of the Invention

[0005] In view of this, embodiments of this application provide a method, apparatus, storage medium, and computer device for generating information push strategies based on a large model, which helps to deeply explore the potential patterns of customer behavior, accurately grasp customer needs in different scenarios, improve the accuracy and personalization of information push strategies, improve the efficiency and effectiveness of information push, and thus improve customer satisfaction.

[0006] According to one aspect of this application, a method for generating an information push strategy based on a large model is provided, the method comprising: Acquire raw behavioral data from multiple customers and divide the raw behavioral data into time-dimensional data and spatial-dimensional data; A time-dimensional behavior similarity matrix is ​​constructed based on the time-dimensional data of different customers, and a spatial-dimensional behavior similarity matrix is ​​constructed based on the spatial-dimensional data of different customers. The time-dimensional behavior similarity matrix and the spatial-dimensional behavior similarity matrix are then fused to obtain a spatiotemporal fusion similarity matrix. Using each customer as a node, the edge weights between nodes are determined based on the spatiotemporal fusion similarity between customers represented by the spatiotemporal fusion similarity matrix. A graph structure containing nodes and edges between nodes is constructed, and high-order features are extracted from each node based on the graph structure to obtain the customer fusion features corresponding to each customer. Retrieve knowledge information that matches the customer fusion characteristics of each customer from a preset knowledge base, wherein the knowledge information includes service introduction information and service success script information; Based on the knowledge information and the customer fusion characteristics, prompt words are constructed, and a large-scale model is generated to produce an information push strategy based on the prompt words.

[0007] According to another aspect of this application, an apparatus for generating an information push strategy based on a large model is provided, the apparatus comprising: The data acquisition module is used to acquire raw behavioral data from multiple customers and divide the raw behavioral data into time-dimensional data and spatial-dimensional data. The matrix construction module is used to construct a time-dimensional behavior similarity matrix based on the time-dimensional data of different customers, construct a spatial-dimensional behavior similarity matrix based on the spatial-dimensional data of different customers, and fuse the time-dimensional behavior similarity matrix and the spatial-dimensional behavior similarity matrix to obtain a spatiotemporal fusion similarity matrix. The feature fusion module is used to determine the edge weights between nodes based on the spatiotemporal fusion similarity between customers represented by the spatiotemporal fusion similarity matrix, with each customer as a node, construct a graph structure containing nodes and edges between nodes, and extract high-order features from each node based on the graph structure to obtain the customer fusion features corresponding to each customer. The knowledge retrieval module is used to retrieve knowledge information that matches the customer fusion characteristics of each customer from a preset knowledge base, wherein the knowledge information includes service introduction information and service success script information; The strategy generation module is used to construct prompt words based on the knowledge information and the customer fusion features, and generate information push strategies based on the prompt words through the strategy generation big model.

[0008] According to another aspect of this application, a storage medium is provided that stores a computer program thereon, which, when executed by a processor, implements the above-described method for generating information push strategies based on large models.

[0009] According to another aspect of this application, a computer device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the program to implement the above-described method for generating an information push strategy based on a large model.

[0010] By employing the above technical solutions, this application provides a method, apparatus, storage medium, and computer device for generating information push strategies based on a large model. First, it acquires original customer behavior data and divides it into time and spatial dimensions, thereby constructing a spatiotemporal fusion similarity matrix. Then, it constructs a graph structure with customers as nodes and extracts customer fusion features. Next, it retrieves matching knowledge information from a preset knowledge base. Finally, it constructs prompt words based on the knowledge information and customer fusion features, and generates an information push strategy from a large model. This application helps to deeply explore the potential patterns of customer behavior, accurately grasp customer needs in different scenarios, improve the accuracy and personalization of information push strategies, enhance the efficiency and effectiveness of information push, and ultimately improve customer satisfaction.

[0011] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0012] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This illustration shows an application environment diagram of an information push strategy generation method based on a large model provided in an embodiment of this application; Figure 2 The illustration shows a flowchart of an information push strategy generation method based on a large model provided in an embodiment of this application; Figure 3 This paper illustrates a flowchart of another method for generating information push strategies based on a large model, as provided in an embodiment of this application. Figure 4 This illustration shows a structural schematic diagram of an information push strategy generation device based on a large model, according to an embodiment of this application. Figure 5 A schematic diagram of the device structure of a computer device provided in an embodiment of this application is shown. Detailed Implementation

[0013] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0014] The information push strategy generation method based on a large model provided in this application can be applied to, for example... Figure 1 In this application environment, the client communicates with the server via a network. The server can obtain the client's raw behavioral data from the client and divide the raw behavioral data into time-dimensional data and spatial-dimensional data. A time-dimensional behavioral similarity matrix is ​​constructed based on the time-dimensional data of different clients, and a spatial-dimensional behavioral similarity matrix is ​​constructed based on the spatial-dimensional data of different clients. The time-dimensional behavioral similarity matrix and the spatial-dimensional behavioral similarity matrix are then fused to obtain a spatiotemporal fusion similarity matrix. Using each client as a node, the edge weights between nodes are determined based on the spatiotemporal fusion similarity between clients represented by the spatiotemporal fusion similarity matrix. A graph structure containing nodes and edges between nodes is constructed, and high-order feature extraction is performed on each node based on the graph structure to obtain the client fusion features corresponding to each client. This is done within a preset knowledge base. The system retrieves knowledge information matching the customer fusion features of each customer, including service introduction information and service success scripts. Based on the knowledge information and customer fusion features, prompt words are constructed. A strategy generation model generates an information push strategy based on the prompt words, and information is fed back to the client based on the information push strategy. In this application, the original customer behavior data is first acquired and divided into time and space dimensions to construct a spatiotemporal fusion similarity matrix. Then, a graph structure is constructed with customers as nodes, and customer fusion features are extracted. Next, matching knowledge information is retrieved from a preset knowledge base. Finally, prompt words are constructed based on the knowledge information and customer fusion features, and an information push strategy is generated by a strategy generation model, which helps improve the accuracy and personalization of the information push strategy. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a separate server or a server cluster composed of multiple servers. The following detailed description of specific embodiments further illustrates this application.

[0015] This embodiment provides a method for generating information push strategies based on large models, such as... Figure 2 As shown, the method includes: Step 101: Obtain raw behavioral data from multiple customers and divide the raw behavioral data into time-dimensional data and spatial-dimensional data.

[0016] In this embodiment, raw behavioral data generated by multiple customers in various scenarios is first collected. This data can originate from multiple online and offline channels. Then, this data is categorized according to two dimensions: time and space. The raw behavioral data specifically includes: Behavioral data: including social media interactions (likes, comments, shares), transaction records (time, amount, category), page browsing paths, etc. Spatiotemporal data: including location information such as GPS coordinates, dwell time, and access frequency, and constructing behavioral trajectories chronologically. Interaction data: including unstructured text such as customer service call records (transcribed via ASR), online chat logs, and product reviews. For example, in the fintech field, banks collect customers' transaction records over a period of time, including transaction time and location (such as ATM location, POS machine transaction location, etc.). The time dimension data can be the transaction frequency at different times of the day, while the spatial dimension data is the transaction distribution in different geographical locations. In the healthcare field, hospitals collect patients' medical records. The time dimension data can be the time interval between each patient's visit and the visit situation in different seasons, while the spatial dimension data can be the patient's visit path to different departments and the distribution of visits in different medical institutions (such as primary hospitals and tertiary hospitals).

[0017] Step 102: Construct a time-dimensional behavior similarity matrix based on the time-dimensional data of different customers, construct a spatial-dimensional behavior similarity matrix based on the spatial-dimensional data of different customers, and fuse the time-dimensional behavior similarity matrix and the spatial-dimensional behavior similarity matrix to obtain a spatiotemporal fusion similarity matrix.

[0018] In this embodiment, for time-dimensional data, the behavioral similarity between different customers in the time dimension can be calculated using methods such as the cosine similarity algorithm, constructing a time-dimensional behavioral similarity matrix. For example, in fintech, this compares the similarity of transaction frequencies between two customers in different time periods. For spatial-dimensional data, a suitable algorithm is also used to calculate the behavioral similarity between different customers in the spatial dimension, constructing a spatial-dimensional behavioral similarity matrix. For example, in the healthcare field, this compares the similarity of the distribution of departments visited and the geographical locations of visits by two patients. These two matrices are then fused, comprehensively considering both time and spatial factors, to obtain a spatiotemporal fusion similarity matrix, which more comprehensively reflects the behavioral similarity between customers.

[0019] It should be noted that in step 102, the temporal and spatial behavioral similarity matrices can be fused directly according to empirically set weights, or the weights can be determined first and then the matrices can be fused. For example, in step 102, fusing the temporal and spatial behavioral similarity matrices to obtain a spatiotemporal fusion similarity matrix includes: constructing an optimization objective function based on maximizing mutual information and minimizing information redundancy; optimizing the weights of temporal and spatial features based on the optimization objective function to obtain temporal feature weights and spatial feature weights; and fusing the temporal and spatial behavioral similarity matrices according to the temporal and spatial feature weights to obtain the spatiotemporal fusion similarity matrix.

[0020] Mutual information measures the degree of interdependence between two variables; that is, the reduction in uncertainty about one variable after knowing information about the other. In customer behavior analysis, time-dimensional and spatial-dimensional features may be inherently related. For example, in fintech, customers may be more inclined to make purchases in specific areas (such as shopping malls) during specific time periods (e.g., weekends); in healthcare, patients may visit certain departments (e.g., respiratory departments) more frequently during specific seasons (e.g., winter). By maximizing mutual information, the potential connections between time and spatial features can be uncovered, allowing the fused matrix to retain more valuable information and avoiding information loss. Minimizing information redundancy: Information redundancy refers to duplicate information contained in two variables. If there is a large amount of redundant information in the time and spatial dimensions, direct fusion may lead to repeated data calculations, reducing the fusion effect. For example, in fintech, there may be some overlap between a customer's weekday transaction records and transaction records near their workplace; in healthcare, there may be overlap between a patient's routine check-ups at fixed times each day and their visits to specific departments. Minimizing information redundancy can remove this duplicate information, making the fused matrix more concise and efficient. In summary, the constructed optimization objective function aims to find a balance point that fully utilizes the complementary information of temporal and spatial features while avoiding interference from redundant information, thereby improving the quality of the fusion matrix. Based on the constructed optimization objective function, specific methods can be used to optimize the weights of temporal and spatial features. Taking gradient descent as an example, it continuously adjusts the weight values, searching in the opposite direction of the gradient of the objective function, gradually reducing the value of the objective function until the optimal weight combination that minimizes the objective function is found. For example, in a fintech scenario, after multiple iterations, the weight of the temporal feature might be 0.6, and the weight of the spatial feature might be 0.4; in a healthcare scenario, the weight of the temporal feature might be 0.4, and the weight of the spatial feature might be 0.6. These weights reflect the relative importance of temporal and spatial features in the fusion process. Furthermore, based on the temporal and spatial feature weights obtained through weight optimization, the temporal behavioral similarity matrix and the spatial behavioral similarity matrix are weighted and fused. The specific formula is: Spatiotemporal fusion similarity matrix = Time dimension feature weight × Time dimension behavior similarity matrix + Spatial dimension feature weight × Spatial dimension behavior similarity matrix. In this way, information from both time and space dimensions is combined to obtain a spatiotemporal fusion similarity matrix that more comprehensively reflects the similarity of customer behavior.

[0021] Step 103: Using each customer as a node, determine the edge weights between nodes based on the spatiotemporal fusion similarity between customers represented by the spatiotemporal fusion similarity matrix, construct a graph structure containing nodes and edges between nodes, and extract high-order features from each node based on the graph structure to obtain the customer fusion features corresponding to each customer.

[0022] In this embodiment, each customer is treated as a node in a graph structure. The weights of edges between nodes are determined based on a spatiotemporal fusion similarity matrix; the higher the similarity, the greater the edge weight. For example, in fintech, the more similar the transaction behaviors of two customers in time and space, the greater the edge weight between their corresponding nodes. After constructing the graph structure, high-order features are extracted from each node using graph neural networks and other algorithms to uncover deeper relationships and features between customers, thereby obtaining customer fusion features for each customer. In the healthcare field, graph structures can be used to uncover potential connections between patients, such as shared departments or similar treatment paths, thereby extracting more comprehensive patient fusion features.

[0023] In some examples, step 103 includes: Step 1031: Taking each customer as a node, based on a preset spatiotemporal fusion similarity threshold, identify target elements in the spatiotemporal fusion similarity matrix that are greater than the preset spatiotemporal fusion similarity threshold, construct edges between nodes based on the target elements, and determine the edge weights between nodes based on the values ​​of the target elements, thereby constructing a graph structure containing nodes and edges between nodes. Step 1032: Based on the original behavioral data of each customer, construct the original features corresponding to each customer; Step 1033: For each node in the graph structure, calculate the betweenness centrality of the node as the structural importance of the node. Concatenate the structural importance with the original features of the node to obtain the node features corresponding to the node. Calculate the topological distance between the node and other nodes using a network topological distance algorithm. Determine the feature fusion node corresponding to the node based on the topological distance. Then, extract higher-order features from the node features of the node based on the node features and edge weights corresponding to the feature fusion node to obtain the customer fusion features corresponding to each customer.

[0024] First, based on a spatiotemporal fusion similarity matrix, customer pairs (target elements) with high similarity are selected by setting a preset spatiotemporal fusion similarity threshold, constructing a graph structure with customers as nodes and similarity as edge weights. Next, the original features of each customer are extracted from the original behavioral data, and the betweenness centrality of each node is calculated as its structural importance in the graph, concatenating it with the original features to form node features. Finally, network topological distance (such as shortest path) is used to determine the feature fusion nodes for each node (e.g., nodes with a distance less than the threshold), and by aggregating the features and edge weights of these fusion nodes, higher-order node feature extraction is achieved, resulting in customer fusion features that integrate graph structure information and node importance. The beneficial effects of this scheme are that it effectively captures the spatiotemporal behavioral relationships between customers through the graph structure, strengthens the influence of key customers through betweenness centrality, and enables the mining of implicit information from multi-hop neighbors through higher-order feature aggregation, improving the expressive power and robustness of customer features, and providing a more accurate data foundation for subsequent knowledge retrieval and strategy generation. In practice, the Dijkstra algorithm can be used to calculate the shortest path between nodes as the topological distance, and a distance threshold can be set or the K nearest neighbor nodes can be selected as the fusion nodes. Then, the features of the fusion nodes can be weighted and summed with the edge weights as coefficients, or the graph attention mechanism can be used to dynamically allocate weights to complete the extraction of higher-order features.

[0025] In some examples, step 1033, which involves extracting higher-order features from the node features of the fusion node based on the node features and edge weights corresponding to the feature fusion node, to obtain the customer fusion features corresponding to each customer, includes: Step 10331: Determine the low-order fusion feature corresponding to the node based on the product of the edge weight between the node and its neighboring nodes and the node feature corresponding to the neighboring node; determine the low-order feature corresponding to the node based on the low-order fusion feature and the node feature corresponding to the node. Step 10332: Determine the high-order fusion feature corresponding to the node based on the product of the edge weight between the node and the neighboring node and the low-order feature corresponding to the neighboring node. Determine the high-order feature corresponding to the node based on the low-order feature and the high-order fusion feature. Use the high-order feature as the customer fusion feature of the customer corresponding to the node.

[0026] High-order customer features can be extracted through a two-stage feature fusion process. First, a node aggregates the node features of its neighbors (neighbor features multiplied by corresponding edge weights) to form a low-order fusion feature. This low-order feature is then fused with the node's own features. For example, in a financial scenario, a customer node integrates the transaction pattern features of its similar trading partners (directly connected neighbor nodes) with its own transaction features. The node further performs a second weighted aggregation based on the low-order features of its neighbors (edge ​​weights × neighbor low-order features) to form a high-order fusion feature. Finally, the low-order and high-order fusion features are fused together to form the high-order customer fusion feature. For example, in a medical scenario, a patient node integrates the health behavior features (low-order) of patients with similar treatment paths and the deep-seated treatment patterns (high-order) of these patients. The benefits are: through two-stage neighborhood information propagation, indirect connections and group behavior patterns between customers can be effectively mined, ensuring that feature representation not only includes individual behavior but also incorporates collaborative information from multi-hop neighbors, thereby enhancing the robustness and discriminative power of the features; simultaneously, edge weights, as a similarity metric, ensure the rationality of information aggregation, providing a more comprehensive and accurate customer representation for subsequent knowledge retrieval and strategy generation. In the fintech field, by capturing multi-layered behavioral correlations among customers (such as the direct similarity and indirect influence of transaction habits), more accurate financial recommendation strategies can be generated. For example, identifying customers who are influenced by their own investment preferences and driven by the behavior of similar customer groups can improve marketing effectiveness. In the healthcare field, by integrating patients' direct health characteristics (such as visit frequency) and indirect group characteristics (such as the disease development trends of similar patient groups), more forward-looking health management plans can be developed. For example, for patients who are simultaneously influenced by their own chronic disease characteristics and the recovery patterns of patients with the same disease, rehabilitation guidance strategies can be optimized. In specific implementation, it is necessary to first determine the neighbor set of a node (based on a spatiotemporal similarity threshold), then aggregate the neighbor features layer by layer through a weighted summation method, and finally fuse the multi-layer aggregation results with the node's own features to form a multi-dimensional customer fusion feature that includes direct features, features influenced by nearby neighbors, and features transmitted from distant neighbors.

[0027] Step 104: Retrieve knowledge information that matches the customer fusion characteristics of each customer from the preset knowledge base, wherein the knowledge information includes service introduction information and service success script information.

[0028] In this embodiment, the pre-stored knowledge base can contain a large amount of knowledge information related to business services, including detailed introductions of various services and successful service examples. Based on each customer's customer fusion characteristics, a search is performed in the knowledge base to find matching knowledge information. In the fintech field, if customer fusion characteristics indicate a potential demand for high-yield wealth management products, detailed introductions of relevant high-yield wealth management products and successful recommendation scripts for such products to similar customers can be retrieved. In the healthcare field, if patient fusion characteristics indicate a chronic disease and interest in rehabilitation treatment, service introductions for rehabilitation treatment of that chronic disease and successful guidance scripts for patients undergoing rehabilitation treatment can be retrieved.

[0029] Step 105: Construct prompt words based on the knowledge information and the customer fusion features, and generate an information push strategy based on the prompt words through a strategy generation model.

[0030] In this embodiment, retrieved knowledge information and customer fusion features are integrated to construct prompt words suitable for a large-scale strategy generation model. These prompt words provide contextual information for the large-scale model, enabling it to generate information push strategies that better meet customer needs. For example, in fintech, prompt words can include customer trading habits, risk preferences, demand for wealth management products, and introductions and successful sales pitches for related wealth management products. Based on these prompt words, the large-scale model generates personalized wealth management product recommendation strategies, improving the accuracy and effectiveness of information push, and helping to enhance customer investment satisfaction and the business conversion rate of financial institutions. In the healthcare field, prompt words can cover patients' conditions, treatment needs, rehabilitation intentions, and introductions and successful sales pitches for related medical services. Based on this, the large-scale model generates targeted health management information push strategies.

[0031] It should be noted that the training method of the strategy generation large model in this application embodiment may specifically include: First, constructing a supervised fine-tuning dataset, where each sample contains customer fusion features, relevant knowledge information retrieved from the knowledge base, and manually annotated optimal information push strategy text. These data are concatenated into prompt words according to a fixed template (e.g., "Based on the following customer features and knowledge information, a push strategy is generated: features: [feature description], knowledge: [knowledge information]"), and the strategy text is used as the target output. Then, a general large language model (such as GPT, LLaMA, etc.) is used as the base, and supervised fine-tuning is performed on the constructed dataset. The model parameters are optimized by minimizing the cross-entropy loss between the generated text and the annotated text. During fine-tuning, instruction fine-tuning methods can be introduced, multiple prompt templates can be used to enhance generalization ability, and LoRA and other parameter efficient fine-tuning techniques can be used to reduce training costs. Through supervised learning from historical success cases, the model can effectively learn the complex mapping relationship between customer characteristics, knowledge information, and effective strategies, thereby generating personalized, accurate, and business-logical push strategies in new scenarios. At the same time, the large model's generation capability ensures the naturalness and diversity of strategy text, which helps to improve user acceptance and conversion rates.

[0032] In some examples, after extracting high-order features from each node based on the graph structure to obtain customer fusion features corresponding to each customer, the method further includes: performing customer group clustering based on the customer fusion features corresponding to each customer, and determining the customer group features of each customer group based on the obtained customer fusion features corresponding to each customer in each customer group. The step of retrieving knowledge information that matches the customer's fusion characteristics from the preset knowledge base includes: retrieving knowledge information that matches the customer group characteristics from the preset knowledge base; The step of constructing prompt words based on the knowledge information and the customer fusion features, and generating an information push strategy based on the prompt words through a strategy generation model, includes: for each customer in any customer group, constructing prompt words based on the knowledge information and the customer group features corresponding to the customer group, and generating an information push strategy corresponding to each customer in the customer group through a strategy generation model.

[0033] Among these methods, customer group clustering and hierarchical knowledge retrieval can improve the accuracy and efficiency of information push strategies. Specifically, after extracting customer fusion features, customers are first clustered based on feature similarity, and the group characteristics of each customer group are calculated (such as the average risk tolerance of the customer group and the recovery period of common diseases). In the knowledge retrieval stage, the applicable information of the group in the knowledge base is matched with the customer group characteristics as the core (such as retrieving wealth management white papers for "high-net-worth wealth management customers" and blood glucose management guidelines for "diabetic chronic disease customers"). Finally, when generating strategies, prompt words are constructed by combining customer group characteristics and individual customer characteristics (such as "This customer belongs to a highly active investment customer group, the customer group characteristics show a preference for technology stocks, and the customer has a recent need to increase their holdings"), so that the large model generates strategies that meet both the common needs of the group and adapt to individual differences. For example, unsupervised clustering algorithms (such as K-means) can be used to segment customers, customer characteristics can be determined by feature aggregation (such as mean and centroid), and knowledge graphs or semantic matching techniques can be used to achieve the association retrieval of customer characteristics and knowledge information. Finally, customer and individual information can be integrated into a large model through prompt word templates.

[0034] In some examples, the method further includes: acquiring original behavior update data for each customer based on a first time period; constructing original update features for each customer based on the original behavior update data; updating the node features of each node in the graph structure based on the original update features to update the customer fusion features and customer group features corresponding to each node; and / or, re-clustering the customer group based on the customer fusion features of each node in the graph structure based on a second time period to determine the updated customer group features.

[0035] Furthermore, a dual-time-cycle dynamic update mechanism can be used to continuously optimize customer characteristics and customer groups. Specifically, in the first time cycle (e.g., daily), updated customer behavior data (e.g., new transaction records for financial customers, latest medical information for medical patients) is acquired, original updated features (e.g., changes in transaction frequency, fluctuations in symptom indicators) are constructed, and graph node features are updated based on these features. Then, customer fusion features and customer group features (e.g., average spending power of customer groups, typical disease development stages) are recalculated through higher-order feature extraction. In the second time cycle (e.g., monthly), customer groups are re-clustered based on the latest distribution of customer fusion features, and customer group feature descriptions are updated. In specific application scenarios, updated data can be acquired by setting up a timed data collection module, and new and old features can be integrated using feature fusion algorithms (e.g., weighted average, feature concatenation). Dynamic clustering algorithms (e.g., adaptive K-means) can be used to re-divide customer groups, and finally, personalized push strategies adapted to the latest customer status are dynamically generated through knowledge base association retrieval and prompt word templates.

[0036] By applying the technical solution of this embodiment, the original customer behavior data is first acquired and divided into time and spatial dimensions. Based on this, a spatiotemporal fusion similarity matrix is ​​constructed. Then, a graph structure is constructed with customers as nodes, and customer fusion features are extracted. Next, matching knowledge information is retrieved from a preset knowledge base. Finally, prompt words are constructed based on the knowledge information and customer fusion features, and an information push strategy is generated by a strategy generation model. This embodiment helps to deeply explore the potential patterns of customer behavior, accurately grasp customer needs in different scenarios, improve the accuracy and personalization of information push strategies, improve the efficiency and effectiveness of information push, and thus improve customer satisfaction.

[0037] Furthermore, as a refinement and extension of the specific implementation of the above embodiments, to fully illustrate the specific implementation process of this embodiment, another information push strategy generation method based on a large model is provided. The prompt words include structured strategy guidance information, which guides the strategy generation large model to generate communication key point strategies, recommendation service and recommendation reason strategies, recommendation channel strategies, and recommendation timing strategies for the customer based on the knowledge information and the customer fusion characteristics; such as... Figure 3 As shown, after generating an information push strategy based on the prompt words using a large model through strategy generation, the method includes: Step 201: Generate content recommendation information through a content generation agent based on the communication key point strategy, the recommendation service, and the recommendation reason strategy.

[0038] Step 202: Deliver the intelligent agent through the channel, and deliver the content recommendation information to the customer according to the recommendation channel strategy and the recommendation timing strategy.

[0039] Step 203: When the customer interacts based on the received content recommendation information, the dialogue agent generates an interactive dialogue according to the communication key point strategy and the customer's interaction information, and updates the original features of the customer in the graph structure according to the interactive dialogue with the customer.

[0040] Step 204: Obtain the reach effect information of the content recommendation information, score the content recommendation information based on the reach effect information, and if the score is higher than a preset score threshold, perform at least one of the following: extract new service success dialogue information based on the content recommendation information, and add the customer fusion features corresponding to the customer and the new service success dialogue information to the preset knowledge base; optimize and train the strategy generation model based on the prompt words and the content recommendation information.

[0041] In the above embodiments, precise iteration of push strategies is achieved through structured strategy guidance and closed-loop collaboration among multiple agents. The multi-dimensional strategy guidance information embedded in the prompts, including key communication points, recommended services and reasons, channels, and timing, provides a clear framework for the large model to generate strategies. For example, in a financial scenario, "the key communication points should include a risk-reward ratio explanation, the recommended service should be a customized index fund, the reason should be based on the customer's recent high-frequency trading behavior in technology stocks, the channel should be mobile banking push, and the timing should be set on days with unusual activity in the technology sector." The content generation agent generates specific copy based on this, the channel delivery agent selects the optimal reach path based on the customer profile (e.g., automatically selecting based on the customer's historical channel preferences and active periods), so as to reach users through channels such as SMS and App push at the specified time. The dialogue agent captures real-time feedback during customer interactions (e.g., replying with keywords, clicking links) and dynamically updates the customer's original characteristics. The execution effect of each strategy is comprehensively tracked, including business metrics such as exposure, click-through rate, conversion rate, and purchase amount, as well as real-time customer feedback (e.g., satisfaction rating, dialogue sentiment). The outreach effectiveness scoring mechanism evaluates the effectiveness of the strategy through the aforementioned feedback indicators. When the score meets the target, the scripts and customer characteristics of successful cases are back-injected into the knowledge base. Successful strategy cases are automatically accumulated back into the enterprise knowledge base and strategy base, forming a self-reinforcing closed loop. This allows the system to accumulate experience over time, becoming more intelligent with use. Furthermore, it can fine-tune the strategy generation model through reinforcement learning algorithms, forming a closed loop of "strategy generation - effect feedback - model optimization".

[0042] Furthermore, as Figure 2 In terms of specific implementation, this application provides an information push strategy generation device based on a large model, such as... Figure 4 As shown, the device includes: The data acquisition module is used to acquire raw behavioral data from multiple customers and divide the raw behavioral data into time-dimensional data and spatial-dimensional data. The matrix construction module is used to construct a time-dimensional behavior similarity matrix based on the time-dimensional data of different customers, construct a spatial-dimensional behavior similarity matrix based on the spatial-dimensional data of different customers, and fuse the time-dimensional behavior similarity matrix and the spatial-dimensional behavior similarity matrix to obtain a spatiotemporal fusion similarity matrix. The feature fusion module is used to determine the edge weights between nodes based on the spatiotemporal fusion similarity between customers represented by the spatiotemporal fusion similarity matrix, with each customer as a node, construct a graph structure containing nodes and edges between nodes, and extract high-order features from each node based on the graph structure to obtain the customer fusion features corresponding to each customer. The knowledge retrieval module is used to retrieve knowledge information that matches the customer fusion characteristics of each customer from a preset knowledge base, wherein the knowledge information includes service introduction information and service success script information; The strategy generation module is used to construct prompt words based on the knowledge information and the customer fusion features, and generate information push strategies based on the prompt words through the strategy generation big model.

[0043] Optionally, in this embodiment of the application, the feature fusion module is specifically used for: Using each customer as a node, target elements with a value greater than the preset spatiotemporal fusion similarity threshold are identified in the spatiotemporal fusion similarity matrix according to the preset spatiotemporal fusion similarity threshold. Edges between nodes are constructed based on the target elements, and the edge weights between nodes are determined based on the values ​​of the target elements, thus constructing a graph structure containing nodes and edges between nodes. Based on the original behavioral data of each customer, construct the original characteristics corresponding to each customer; For each node in the graph structure, the betweenness centrality of the node is calculated as the structural importance of the node. The structural importance is concatenated with the original features of the node to obtain the node features. The topological distance between the node and other nodes is calculated using a network topological distance algorithm. Based on the topological distance, the feature fusion node corresponding to the node is determined. Based on the node features and edge weights corresponding to the feature fusion node, higher-order feature extraction is performed on the node features to obtain the customer fusion features corresponding to each customer.

[0044] Optionally, in this embodiment of the application, the feature fusion module is further used for: The low-order fusion feature corresponding to the node is determined by the product of the edge weight between the node and its neighboring nodes and the node feature corresponding to the neighboring node. The low-order feature corresponding to the node is determined by the low-order fusion feature and the node feature corresponding to the node. The higher-order fusion feature corresponding to the node is determined by multiplying the edge weight between the node and its neighboring nodes with the lower-order feature corresponding to the neighboring node. The higher-order feature corresponding to the node is determined by combining the lower-order feature and the higher-order fusion feature. The higher-order feature is then used as the customer fusion feature of the customer corresponding to the node.

[0045] Optionally, in this embodiment, the device further includes: a customer group clustering module, used for: After extracting high-order features from each node based on the graph structure to obtain the customer fusion features corresponding to each customer, customer group clustering is performed based on the customer fusion features corresponding to each customer, and the customer group features of each customer group are determined based on the customer fusion features corresponding to each customer in each customer group. The knowledge retrieval module is specifically used for: Retrieve knowledge information that matches the characteristics of the customer group from a preset knowledge base; The strategy generation module is specifically used for: For each customer in any customer group, prompt words are constructed based on the knowledge information and the customer group characteristics corresponding to the customer group. A strategy generation model is then used to generate an information push strategy corresponding to each customer in the customer group based on the prompt words.

[0046] Optionally, in this embodiment of the application, the customer group clustering module is further used for: Based on the first time period, acquire the original behavior update data of each customer, and construct the original update features corresponding to each customer based on the original behavior update data; update the node features of each node in the graph structure based on the original update features, so as to update the customer fusion features and customer group features corresponding to each node; and / or, Based on the second time period, customer groups are re-clustered according to the customer fusion characteristics of each node in the graph structure to determine the updated customer group characteristics.

[0047] Optionally, in this embodiment, the prompt words include structured strategy guidance information, which guides the strategy generation model to generate communication point strategies, recommendation service and recommendation reason strategies, recommendation channel strategies, and recommendation timing strategies for the customer based on the knowledge information and the customer fusion characteristics; the device further includes: an interaction module, used for: After constructing prompt words based on the knowledge information and the customer fusion features, and generating an information push strategy based on the prompt words through a strategy generation big model, a content generation intelligent agent generates content recommendation information according to the communication key point strategy, the recommendation service, and the recommendation reason strategy. The intelligent agent delivers content recommendation information to the customer based on the recommended channel strategy and the recommended timing strategy. When the customer interacts based on the received content recommendation information, the dialogue agent generates an interactive dialogue according to the communication key point strategy and the customer's interaction information, and updates the customer's original features in the graph structure based on the interactive dialogue with the customer.

[0048] Optionally, in this embodiment, the apparatus further includes: an update module, configured to: After delivering the content recommendation information to the customer, the delivery effect information of the content recommendation information is obtained. The content recommendation information is scored based on the delivery effect information, and if the score is higher than a preset score threshold, at least one of the following is performed: New service success script information is extracted based on the content recommendation information, and the customer integration characteristics corresponding to the customer and the new service success script information are added to the preset knowledge base; Based on the prompt words and the content recommendation information, the strategy generation model is optimized and trained.

[0049] It should be noted that other corresponding descriptions of the functional units involved in the information push strategy generation device based on a large model provided in this application embodiment can be found in the following references. Figures 2 to 3 The corresponding descriptions in the method will not be repeated here.

[0050] This application also provides a computer device, which may specifically be a personal computer, a server, a network device, etc. Figure 5 As shown, the computer device includes a bus, a processor, memory, and a communication interface, and may also include an input / output interface and a display device. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores location information. The network interface allows communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the various method embodiments.

[0051] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0052] In one embodiment, a computer-readable storage medium is provided, which may be non-volatile or volatile, having stored thereon a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0053] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0054] It should be noted that the user personal information involved in the embodiments of this application is all authorized (with the knowledge and consent) by the relevant parties or fully authorized by all parties, and the executing entity can obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with the relevant laws and regulations of the relevant countries and regions, and do not violate public order and good morals. It should be noted that if any software tools or components other than those of this company appear in the embodiments of this application, they are merely illustrative examples and do not represent actual use.

[0055] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, graphics processors, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0056] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0057] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for generating information push strategies based on large models, characterized in that, The method includes: Acquire raw behavioral data from multiple customers and divide the raw behavioral data into time-dimensional data and spatial-dimensional data; A time-dimensional behavior similarity matrix is ​​constructed based on the time-dimensional data of different customers, and a spatial-dimensional behavior similarity matrix is ​​constructed based on the spatial-dimensional data of different customers. The time-dimensional behavior similarity matrix and the spatial-dimensional behavior similarity matrix are then fused to obtain a spatiotemporal fusion similarity matrix. Using each customer as a node, the edge weights between nodes are determined based on the spatiotemporal fusion similarity between customers represented by the spatiotemporal fusion similarity matrix. A graph structure containing nodes and edges between nodes is constructed, and high-order features are extracted from each node based on the graph structure to obtain the customer fusion features corresponding to each customer. Retrieve knowledge information that matches the customer fusion characteristics of each customer from a preset knowledge base, wherein the knowledge information includes service introduction information and service success script information; Based on the knowledge information and the customer fusion characteristics, prompt words are constructed, and a large-scale model is generated to produce an information push strategy based on the prompt words.

2. The method according to claim 1, characterized in that, The process involves using each customer as a node, determining the edge weights between nodes based on the spatiotemporal fusion similarity between customers represented by the spatiotemporal fusion similarity matrix, constructing a graph structure containing nodes and edges between them, and extracting higher-order features from each node based on the graph structure to obtain the customer fusion features corresponding to each customer, including: Using each customer as a node, target elements with a value greater than the preset spatiotemporal fusion similarity threshold are identified in the spatiotemporal fusion similarity matrix according to the preset spatiotemporal fusion similarity threshold. Edges between nodes are constructed based on the target elements, and the edge weights between nodes are determined based on the values ​​of the target elements, thus constructing a graph structure containing nodes and edges between nodes. Based on the original behavioral data of each customer, construct the original characteristics corresponding to each customer; For each node in the graph structure, the betweenness centrality of the node is calculated as the structural importance of the node. The structural importance is concatenated with the original features of the node to obtain the node features. The topological distance between the node and other nodes is calculated using a network topological distance algorithm. Based on the topological distance, the feature fusion node corresponding to the node is determined. Based on the node features and edge weights corresponding to the feature fusion node, higher-order feature extraction is performed on the node features to obtain the customer fusion features corresponding to each customer.

3. The method according to claim 2, characterized in that, The step of extracting higher-order features from the node features of the fused nodes based on the node features and edge weights to obtain customer fusion features for each customer includes: The low-order fusion feature corresponding to the node is determined by the product of the edge weight between the node and its neighboring nodes and the node feature corresponding to the neighboring node. The low-order feature corresponding to the node is determined by the low-order fusion feature and the node feature corresponding to the node. The higher-order fusion feature corresponding to the node is determined by multiplying the edge weight between the node and its neighboring nodes with the lower-order feature corresponding to the neighboring node. The higher-order feature corresponding to the node is determined by combining the lower-order feature and the higher-order fusion feature. The higher-order feature is then used as the customer fusion feature of the customer corresponding to the node.

4. The method according to claim 2, characterized in that, After extracting high-order features from each node based on the graph structure to obtain the customer fusion features corresponding to each customer, the method further includes: Customer groups are clustered based on the customer fusion characteristics of each customer, and customer group characteristics of each customer group are determined based on the customer fusion characteristics of each customer in each customer group. The step of retrieving knowledge information matching the customer's fusion features from a preset knowledge base includes: Retrieve knowledge information that matches the characteristics of the customer group from a preset knowledge base; The process of constructing prompt words based on the knowledge information and the customer fusion features, and generating an information push strategy based on the prompt words through a strategy generation model, includes: For each customer in any customer group, prompt words are constructed based on the knowledge information and the customer group characteristics corresponding to the customer group. A strategy generation model is then used to generate an information push strategy corresponding to each customer in the customer group based on the prompt words.

5. The method according to claim 4, characterized in that, The method further includes: Based on the first time period, acquire the original behavior update data of each customer, and construct the original update features corresponding to each customer based on the original behavior update data; update the node features of each node in the graph structure based on the original update features, so as to update the customer fusion features and customer group features corresponding to each node; and / or, Based on the second time period, customer groups are re-clustered according to the customer fusion characteristics of each node in the graph structure to determine the updated customer group characteristics.

6. The method according to any one of claims 1 to 5, characterized in that, The prompts include structured strategy guidance information, which guides the strategy generation model to generate communication key point strategies, recommendation service and recommendation reason strategies, recommendation channel strategies, and recommendation timing strategies for the customer based on the knowledge information and the customer integration characteristics. After constructing prompt words based on the knowledge information and the customer fusion features, and generating an information push strategy based on the prompt words using a strategy generation model, the method further includes: The content generation agent generates content recommendation information based on the communication key point strategy, the recommendation service, and the recommendation reason strategy. The intelligent agent delivers content recommendation information to the customer based on the recommended channel strategy and the recommended timing strategy. When the customer interacts based on the received content recommendation information, the dialogue agent generates an interactive dialogue according to the communication key point strategy and the customer's interaction information, and updates the customer's original features in the graph structure based on the interactive dialogue with the customer.

7. The method according to claim 6, characterized in that, After delivering the content recommendation information to the customer, the method further includes: Obtain the reach effect information of the content recommendation information, score the content recommendation information based on the reach effect information, and if the score is higher than a preset score threshold, perform at least one of the following: New service success script information is extracted based on the content recommendation information, and the customer integration characteristics corresponding to the customer and the new service success script information are added to the preset knowledge base; Based on the prompt words and the content recommendation information, the strategy generation model is optimized and trained.

8. A device for generating information push strategies based on a large model, characterized in that, The device includes: The data acquisition module is used to acquire raw behavioral data from multiple customers and divide the raw behavioral data into time-dimensional data and spatial-dimensional data. The matrix construction module is used to construct a time-dimensional behavior similarity matrix based on the time-dimensional data of different customers, construct a spatial-dimensional behavior similarity matrix based on the spatial-dimensional data of different customers, and fuse the time-dimensional behavior similarity matrix and the spatial-dimensional behavior similarity matrix to obtain a spatiotemporal fusion similarity matrix. The feature fusion module is used to determine the edge weights between nodes based on the spatiotemporal fusion similarity between customers represented by the spatiotemporal fusion similarity matrix, with each customer as a node, construct a graph structure containing nodes and edges between nodes, and extract high-order features from each node based on the graph structure to obtain the customer fusion features corresponding to each customer. The knowledge retrieval module is used to retrieve knowledge information that matches the customer fusion characteristics of each customer from a preset knowledge base, wherein the knowledge information includes service introduction information and service success script information; The strategy generation module is used to construct prompt words based on the knowledge information and the customer fusion features, and generate information push strategies based on the prompt words through the strategy generation big model.

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

10. A computer device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

Images