Intelligent product recommendation method, system and equipment and storage medium

Abstract

The invention relates to the technical field of data processing, can be applied to the field of medical treatment and finance, and discloses an intelligent product recommendation method, system and device and a storage medium. And identifying a disease risk probability and a key risk factor of the customer based on a health risk identification model, determining an insurance recommendation scheme based on insurance scheme matching logic, generating consultation verbal skills and scheme presentation contents of the customer, and touching the customer through multiple communication modes. Based on the health risk identification model, the insurance scheme matching logic, the consultation verbal skill and the scheme presentation content generation technology, the problems of low customer matching degree, large agent burden and insufficient feedback closed loop in the existing recommendation mode are solved.

Description

Technical Field

[0001] This application relates to the field of data processing technology and can be applied to the medical and financial fields. Specifically, it relates to a product intelligent recommendation method, system, device, and storage medium. Background Technology

[0002] With the rapid development of the insurance market, customers' demand for personalized and precise insurance services is increasing. Currently, the marketing model for insurance products mainly relies on agent experience and traditional customer profiles. Although some insurance companies have tried to apply big data analysis and intelligent recommendation technologies, existing solutions still have many shortcomings and are difficult to meet market demands and business development requirements.

[0003] First, customer segmentation is primarily based on basic demographic tags (age, gender, income), lacking in-depth modeling of customers' health risk factors (such as medical examination data, health habits, and medical history). Traditional recommendation models, by segmenting customers solely through simple demographic characteristics, cannot accurately identify individual differences in health risk and protection needs, resulting in recommendations that lack personalization and scientific rigor, and are difficult to match with customers' actual risk protection requirements.

[0004] Secondly, most existing recommendation systems remain at the product recommendation level, failing to automatically generate personalized consultation scripts and solution introductions for customers. When communicating with customers, agents need to manually organize their language to explain product advantages and match customer needs. This is not only inefficient but also prone to inaccurate communication and information transmission errors, affecting customer understanding and willingness to purchase insurance.

[0005] In summary, existing insurance recommendations have significant shortcomings in terms of personalization, intelligence, and automation. There is an urgent need for an intelligent insurance product recommendation solution that can deeply explore customers' disease risks and achieve intelligent support across the entire process, in order to improve recommendation accuracy, marketing efficiency, and customer satisfaction. Summary of the Invention

[0006] To address the aforementioned issues, this application provides a product intelligent recommendation method, system, device, and storage medium. Based on a health risk identification model, insurance plan matching logic, consultation scripts, and plan presentation content generation technology, it solves the problems of low customer matching and heavy agent burden in existing recommendation models.

[0007] The embodiments of this application adopt the following technical solutions: Firstly, this application provides a product intelligent recommendation method, including: Collect multi-dimensional health data from customers, extract multi-modal features from the multi-dimensional health data, and fuse the multi-modal features into a unified health profile vector; Based on the unified health profile vector, the customer's disease risk probability and key risk factors are identified using a health risk identification model. Based on disease risk probability, key risk factors, insurance product information, and customer preference characteristics, insurance recommendation schemes are determined according to insurance scheme matching logic. Based on the unified health profile vector and insurance recommendation scheme, customer consultation scripts and scheme presentation content are generated, and the consultation scripts and scheme presentation content are delivered to customers through various communication methods.

[0008] Secondly, this application also provides a product intelligent recommendation system, including: The data acquisition unit is used to collect multi-dimensional health data from customers, extract multi-modal features from the multi-dimensional health data, and fuse the multi-modal features into a unified health profile vector. The risk identification unit is used to identify the customer's disease risk probability and key risk factors based on a health risk identification model using a unified health profile vector. The solution generation unit is used to determine the recommended insurance solution based on the insurance solution matching logic, according to the probability of disease risk, key risk factors, insurance product information and customer preference characteristics. The execution interaction unit is used to generate customer consultation scripts and plan presentation content based on the unified health profile vector and insurance recommendation scheme, and to reach customers with consultation scripts and plan presentation content through various communication methods.

[0009] Thirdly, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described intelligent product recommendation method.

[0010] Fourthly, this application also provides a computer-readable storage medium storing a computer program that, when instructed by a processor, implements the steps of the above-described intelligent product recommendation method.

[0011] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects: This application first collects multi-dimensional health data from customers, extracts multimodal features from the multi-dimensional health data, and merges the multimodal features into a unified health profile vector. Secondly, based on the unified health profile vector, it identifies the customer's disease risk probability and key risk factors using a health risk identification model. Next, based on the disease risk probability, key risk factors, insurance product information, and customer preference characteristics, it determines an insurance recommendation plan based on insurance plan matching logic. Finally, it generates customer consultation scripts and plan presentation content based on the unified health profile vector and the insurance recommendation plan, and delivers the consultation scripts and plan presentation content to customers through various communication methods.

[0012] This application, by collecting multi-dimensional health data and using multi-modal feature fusion to generate a unified health profile vector, combined with a health risk identification model, achieves a shift in recommendation mode from "demographic label" driven to "disease risk driven." The insurance recommendation scheme, based on the customer's precise disease risk probability and key risk factors, can accurately match the customer's individual health risk and protection needs, avoiding the extensive model of traditional recommendations based on demographic labels, and significantly improving the targeting and scientific rigor of the recommendations.

[0013] This application automatically generates consultation scripts and solution presentation content for clients, eliminating the need for agents to manually organize language and prepare solution materials, significantly reducing workload and improving communication efficiency. It also ensures the accuracy and professionalism of information delivery, enhancing client comprehension.

[0014] This application constructs a complete closed loop of data collection, risk identification, solution generation, and execution interaction to ensure continuous optimization and self-evolution, providing strong support for improving long-term marketing conversion efficiency.

[0015] Based on the technical solution proposed in this application, the accuracy, intelligence level and sustainable optimization capability of insurance recommendations are significantly improved, providing a guarantee for insurance companies to achieve differentiated competition and high-quality growth. Attached Figure Description

[0016] 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 A flowchart illustrating a product intelligent recommendation method according to an embodiment of this application is shown; Figure 2 A schematic diagram of the structure of a product intelligent recommendation system according to an embodiment of this application is shown; Figure 3 A schematic diagram of the resulting electronic device according to an embodiment of this application is shown. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] The purpose of this application is to provide a smart product recommendation method to address the problems of low customer matching accuracy and high agent burden in existing recommendation models. To achieve the above objective, Figure 1 This application illustrates a product intelligent recommendation method according to an embodiment of the present application, from... Figure 1 As can be seen, this embodiment includes steps S110 to S150: Step S110: Collect multi-dimensional health data of customers, extract multi-modal features from the multi-dimensional health data, and fuse the multi-modal features into a unified health profile vector.

[0019] Customer health information comes from a wide range of sources, and a single type of data cannot comprehensively and accurately reflect a customer's actual health status. Therefore, this embodiment collects multi-dimensional health data from customers and performs multi-modal feature fusion to generate a unified health profile vector.

[0020] In some optional implementations, the multi-dimensional health data includes: structured health data, unstructured health data, and time-series health data; step S110 involves collecting the customer's multi-dimensional health data, extracting multimodal features from the multi-dimensional health data, and fusing the multimodal features into a unified health profile vector, including: collecting the customer's structured health data, unstructured health data, and time-series health data; converting the structured health data into a structured feature vector through a feature embedding layer; extracting the semantic feature vector of the unstructured health data through a natural language processing model in the insurance field; extracting the dynamic health feature vector of the time-series health data through a temporal convolutional network or a Transformer temporal model; and using a cross-attention mechanism to weightedly fuse the structured feature vector, semantic feature vector, and dynamic health feature vector to generate a unified health profile vector.

[0021] During the data collection phase, customer health data is collected from the following three dimensions: structured health data, unstructured health data, and time-series health data.

[0022] Structured health data may include, but is not limited to, physical examination indicators (such as blood pressure, heart rate, blood sugar, etc.) and insurance records.

[0023] Unstructured data may include, but is not limited to: answers to open-ended questions in health questionnaires, agent interview records, and customer self-reported health texts.

[0024] Time-series health data may include, but is not limited to, exercise trajectory data and sleep monitoring data.

[0025] Differentiated feature extraction methods are used for different types of health data.

[0026] For structured health data, a feature embedding layer is used to convert the structured health data into structured feature vectors. The feature embedding layer can first normalize the structured health data to eliminate the influence of dimensions, and then map it to a fixed-dimensional vector representation through an embedding matrix.

[0027] For unstructured health data, semantic feature vectors are extracted using a natural language processing (NLP) model specific to the insurance industry. Unstructured health data, such as customer self-reported health texts, is input into a pre-trained NLP model specific to the insurance industry (such as the BERT model or a large-scale model specific to the insurance industry). Since the NLP model has learned from a large amount of insurance-related corpus, it can accurately identify health-related semantic information in unstructured health data.

[0028] For time-series health data, dynamic health feature vectors are extracted using temporal convolutional networks (TCNs) or Transformer temporal models. Time-series health data exhibits temporal dependencies. TCNs, through dilated convolutions, can capture the dependencies in long-series data, while Transformer temporal models use a self-attention mechanism to focus on health data changes at key time points.

[0029] A multimodal fusion layer, such as cross-attention, is used to weight and fuse structured feature vectors, semantic feature vectors, and dynamic health feature vectors to obtain a complete unified health profile vector for the customer. The cross-attention mechanism can adaptively learn the importance weights of different modal features. For example, for customers with abnormal physical examination indicators, the weight of the structured feature vector will be increased accordingly; for customers who report long-term discomfort, the weight of the semantic feature vector will be strengthened; and for customers whose recent exercise habits have changed significantly, the weight of the dynamic health feature vector will be increased. Through this weighted fusion method, the unique information of different modal data is preserved while achieving organic data integration. The resulting unified health profile vector can comprehensively and accurately depict the customer's health status, providing reliable data support for subsequent health risk identification.

[0030] Step S120: Based on the unified health profile vector, identify the customer's disease risk probability and key risk factors using the health risk identification model.

[0031] Traditional risk prediction models excel at pattern recognition, but tend to learn "correlation" rather than "causation." In other words, traditional risk prediction models often only output risk outcomes without explaining the causes of those risks. Consequently, traditional risk prediction models struggle to support customer trust and the explainability and compliance requirements of insurance business. This embodiment proposes a health risk identification model that can not only accurately predict the probability of a customer's disease risk but also identify key risk factors.

[0032] In some optional implementations, the health risk identification model includes a deep neural network and a causal inference framework; step S120, based on the unified health profile vector, identifies the customer's disease risk probability and key risk factors based on the health risk identification model, including: performing risk identification on the unified health profile vector through a deep neural network and outputting the customer's disease risk probability; performing attribution analysis on the disease risk probability through a causal inference framework and outputting key risk factors based on the causal relationship between the disease and the risk factors.

[0033] Deep neural networks (such as DNNs and Transformers) possess powerful feature learning and pattern recognition capabilities, enabling them to uncover complex health risk correlations from a unified health profile vector. The unified health profile vector is input into a pre-trained deep neural network, which then outputs the customer's disease risk probabilities. For example, a customer might have a 20% risk of diabetes, a 35% risk of cardiovascular disease, and an 8% risk of malignant tumors.

[0034] The pre-training process of deep neural networks can be based on existing model training methods, that is, based on a large amount of labeled sample data, the parameters of the deep neural network are continuously adjusted through the backpropagation algorithm to optimize the loss function to ensure prediction accuracy.

[0035] Causal inference frameworks (such as Do-Why and IV-based frameworks) enable precise explanations of the causes of risk. These frameworks identify causal relationships between key risk factors and disease by performing attribution analysis on the probability of disease risk, thereby pinpointing the key risk factors.

[0036] For example, the causal inference framework first constructs a causal graph to clarify the potential causal relationships between each feature in the unified health profile vector and the disease; then, it uses intervention analysis to determine the causal strength between the features and the disease; finally, it outputs key risk factors containing the names and degrees of influence of the risk factors. For instance, a high BMI and a family history of diabetes may lead to a 35% increased risk of diabetes.

[0037] The final output of the health risk identification model includes not only the probability of disease risk, but also key risk factors; it maintains the predictive accuracy of deep learning while introducing causal inference to enhance interpretability, meeting the transparency requirements of insurance scenarios; and it transforms the subsequent recommendation process from being driven by "demographic labels" to being driven by "disease risk".

[0038] Step S130: Based on the disease risk probability, key risk factors, insurance product information, and customer preference characteristics, determine the recommended insurance plan based on the insurance plan matching logic.

[0039] Insurance plan matching is a core step in connecting a customer's health risks with insurance products. Traditional recommendation models, with their single-mode recommendations, are prone to problems such as recommending only one type of product and insufficient matching with the customer's affordability. This embodiment uses a two-layer recommendation engine—a health-driven recommendation engine and a customer preference recommendation engine—to ensure that the insurance recommendation plan is both scientific and practical.

[0040] In some optional implementations, the insurance scheme matching logic includes: a health-driven recommendation engine and a customer preference recommendation engine; step S130, determining an insurance recommendation scheme based on the insurance scheme matching logic according to the disease risk probability, key risk factors, insurance product information, and customer preference characteristics, including: determining the protection gap based on the disease risk probability and key risk factors; matching candidate insurance products from the insurance product information based on the protection gap using the health-driven recommendation engine; and ranking the candidate insurance products based on the customer preference characteristics using the customer preference recommendation engine to determine the insurance recommendation scheme.

[0041] The customer's protection gap can be determined based on the probability of disease risk and key risk factors. For example, for a customer with a 20% probability of diabetes risk and a high BMI and family history of diabetes as key risk factors, the protection gap can be determined to be critical illness insurance + medical insurance; for a customer with a 35% risk of cardiovascular disease and age and a family history of cardiovascular disease as key risk factors, the protection gap can be determined to be medical insurance + long-term care insurance.

[0042] The insurance plan matching logic includes a two-layer recommendation engine: the first layer is a health-driven recommendation engine, and the second layer is a customer preference recommendation engine.

[0043] The first layer of the health-driven recommendation engine is used to match candidate insurance products from the insurance product database that can cover core protection gaps.

[0044] The insurance product database stores information on various insurance products, including product type, coverage, sum insured, premium, and claims terms. Based on the customer's protection gaps and insurance product information, the health-driven recommendation engine filters candidate insurance products covering core protection gaps from the database.

[0045] The second-layer customer preference recommendation engine is used to personalize and sort candidate insurance products based on customer preference characteristics, so as to prioritize recommend insurance products that are acceptable to the customer, thereby forming an insurance recommendation plan.

[0046] Customer preference characteristics may include, but are not limited to, historical insurance purchase types, insurance product browsing history, and premium payment ability. Based on these customer preference characteristics, the customer preference recommendation engine uses a learning ranking model (such as LambdaMART or DeepFM) to personalize the ranking of candidate insurance products. The ranking objective can be constrained by payment ability and conversion rate; that is, it considers both the customer's probability of accepting the candidate insurance product and their ability to pay for it, thus ranking candidate insurance products with a higher probability of acceptance and greater payment ability higher.

[0047] The ranked candidate insurance products form the final insurance recommendation. The output of the insurance recommendation can include not only a single insurance product, but also a combination of multiple products.

[0048] In some optional implementations, after step S130, which determines the insurance recommendation scheme based on the insurance scheme matching logic, the method further includes: constructing an insurance product knowledge graph; wherein the nodes of the insurance product knowledge graph include: diseases, risk factors, insurance terms, and claims cases, and the edges of the insurance product knowledge graph include: causal relationships between diseases and risk factors, coverage relationships between insurance terms and diseases, and applicability relationships between claims cases and insurance terms; and performing compliance and completeness checks on the insurance recommendation scheme based on the insurance product knowledge graph, and correcting the insurance recommendation scheme when the check results are abnormal.

[0049] Insurance terms are complex, and insurance recommendations need to be verified for compliance and coverage completeness; otherwise, recommendations may not comply with regulations or have inappropriate coverage. To ensure the compliance and coverage completeness of insurance recommendations, after the recommendations are generated based on a two-layer recommendation engine, they are further checked and corrected based on an insurance product knowledge graph.

[0050] The nodes of the insurance product knowledge graph include: diseases, risk factors, insurance terms, and claims cases. The edges between nodes include: causal relationships between diseases and risk factors, coverage relationships between insurance terms and diseases, and applicability relationships between claims cases and insurance terms. For example, diabetes (disease node) and high BMI (risk factor node) are connected by a causal relationship (edge); critical illness insurance terms (insurance terms node) and diabetes (disease node) are connected by a coverage relationship (edge); and diabetes claims cases (claims case node) and critical illness insurance terms (insurance terms node) are connected by an applicability relationship (edge). The construction of the insurance product knowledge graph is based on multi-source information such as insurance industry regulations, product terms text, and historical claims data to ensure the accuracy of nodes and edges.

[0051] The system uses an insurance product knowledge graph to check whether insurance recommendations comply with insurance regulatory rules and product terms, preventing non-compliant recommendations. If compliance issues are found, the insurance recommendations are automatically corrected. For example, if a recommendation contains non-compliant statements, the non-compliant statements will be deleted or adjusted.

[0052] Based on an insurance product knowledge graph, the system uses graph databases (such as Neo4j and GraphSAGE) to check whether recommended insurance plans comprehensively cover the customer's key risk factors. If any issues with coverage comprehensiveness are found, the recommended insurance plans are automatically corrected. For example, if a key risk factor includes hypertension, and there is a causal relationship between hypertension (risk factor node) and cardiovascular disease (disease node) in the insurance product knowledge graph, the system needs to check whether the recommended insurance plans include insurance products covering cardiovascular disease. If any are missing, the system automatically adds the relevant insurance products and corrects the recommended insurance plans.

[0053] By checking and correcting based on insurance product knowledge graphs, we ensure that insurance recommendations not only meet compliance requirements but also fully cover customers' disease risks, thereby improving the scientific rigor and reliability of insurance recommendations.

[0054] Step S140: Generate customer consultation scripts and plan presentation content based on the unified health profile vector and insurance recommendation scheme, and deliver the consultation scripts and plan presentation content to the customer through various communication methods.

[0055] In traditional referral models, communication between agents and clients primarily relies on text or voice. Text or voice communication requires agents to manually craft their language, potentially leading to low efficiency and inaccurate information delivery. This implementation automatically generates consultation scripts and solution presentations, reaching clients through multiple communication methods to provide communication support for agents and improve client comprehension accuracy.

[0056] In some optional implementations, step S140 involves generating customer consultation scripts and plan presentation content based on the unified health profile vector and insurance recommendation scheme. These scripts and content are then delivered to customers through various communication methods, including: generating consultation scripts based on the unified health profile vector and insurance recommendation scheme, using a large language model combined with enhanced retrieval; the consultation scripts include: disease risk reminders, explanations of coverage gaps, introductions to insurance recommendation schemes, and insurance application guidance; generating plan presentation content based on the unified health profile vector and insurance recommendation scheme using multi-role collaborative intelligent agents; these multi-role collaborative intelligent agents include: a risk explanation intelligent agent; a plan generation intelligent agent; and a compliance intelligent agent; the risk explanation intelligent agent explains the correspondence between key risk factors and insurance recommendation schemes, the plan generation intelligent agent generates textual descriptions and visual charts of the insurance recommendation scheme, and the compliance intelligent agent verifies the compliance of the plan presentation content; and delivering the consultation scripts and plan presentation content to users through agent applications, WeChat Work, online stores, and outbound telephone calls based on multimodal interaction components.

[0057] The consultation scripts are generated based on a Large Language Model (LLM) combined with Retrieval Augmentation (RAG). The LLM boasts powerful natural language generation capabilities, generating natural language scripts tailored to the specific communication scenario based on a unified health profile vector and insurance recommendation scheme. Retrieval Augmentation provides professional support through an insurance domain knowledge base, ensuring the personalization, professionalism, and accuracy of the consultation scripts.

[0058] The consultation script comprises four levels: introduction, explanation, recommendation, and action. "Introduction" includes disease risk reminders, which, based on a health risk identification framework, alerts the client to potential disease risks and key risk factors. "Explanation" includes explaining coverage gaps, clarifying the client's current coverage gaps in conjunction with the disease risk reminders, helping them understand the necessity of insurance. "Recommendation" includes introducing recommended insurance plans, detailing the product content, coverage scope, sum insured, premium, claims terms, and other key information. "Action" includes guiding the client through the application process with clear instructions.

[0059] The solution content is generated through multi-role collaborative agents to ensure its richness and intuitiveness. These agents include: a risk interpretation agent, a solution generation agent, and a compliance agent.

[0060] The risk explanation agent clarifies the correspondence between key risk factors and insurance recommendations, helping clients understand why they need an insurance recommendation.

[0061] The agent generates textual descriptions and visual charts for recommended insurance plans. These charts may include, but are not limited to: health risk radar charts (visually displaying the probability of a customer contracting various diseases), coverage gap bar charts (comparing the gap between current coverage and required coverage), and product coverage pie charts (showing the coverage ratio of the recommended insurance plan for disease risks). By transforming complex insurance information into intuitive and easy-to-understand graphics, the barrier to customer comprehension is lowered.

[0062] The compliance agent verifies the compliance of the content presented by the solution, filters out sensitive words and other violations, and ensures that the content presented by the solution complies with insurance regulatory requirements.

[0063] Based on multi-role collaborative intelligent agents, it enables comprehensive generation from product recommendations to intelligent scripts and solution introductions, significantly reducing the burden on agents.

[0064] Consultation scripts and solution presentations are delivered to users via a multimodal interactive component using various communication methods. These methods include, but are not limited to, agent apps, WeChat Work, online stores, and outbound phone calls. The consultation scripts and solution presentations remain consistent across all communication methods. The multimodal interactive component supports various presentation formats, including text, charts, and voice, and can deliver consultation scripts and solution presentations to customers in an appropriate modal format according to the scenario.

[0065] For example, agents can send the solution content to customers as images via WeChat and conduct text conversations with customers based on consultation scripts; agents can also make outbound calls, conduct voice communication with customers based on consultation scripts, and subsequently send the solution content to customers via SMS links.

[0066] In addition, during the interaction between the agent and the client through various communication methods, the conversation content can be captured in real time, and appropriate wording suggestions and auxiliary information can be pushed to the client in real time to help the agent respond to the client quickly and accurately, thereby improving communication efficiency and professionalism.

[0067] In some optional implementations, the above method further includes: collecting customer interaction feedback data, eliminating irrelevant data through causal reasoning to obtain valid feedback data; using the valid feedback data as a reward signal for reinforcement learning, updating the parameters of the health risk identification model and the rules of the insurance plan matching logic through a policy gradient method; and comparing and optimizing the updated health risk identification model and insurance plan matching logic based on an A / B testing mechanism.

[0068] Product recommendations in the insurance sector require long-term optimization; static rules alone are insufficient to adapt to customer needs and market changes. To achieve closed-loop optimization of the recommendation process, this embodiment collects customer interaction feedback data and continuously updates the health risk identification model and insurance plan matching logic through an optimization mechanism, achieving self-evolution.

[0069] Interactive feedback data encompasses customer behavior data regarding insurance recommendations, comprehensively reflecting customer feedback on these recommendations. This data may include, but is not limited to: click-through rate of insurance recommendations, page dwell time, reasons for customer rejection, types of inquiries, actual insurance enrollment rate, types of insurance products purchased, and long-term renewal rate.

[0070] Because interactive feedback data may contain irrelevant and distracting information, causal reasoning is used to filter the data, eliminating irrelevant data to obtain valid feedback. Causal reasoning identifies the feedback information that truly influences customer decisions by analyzing the causal relationship between interactive feedback data and recommendation effectiveness. For example, if a customer refuses to purchase insurance because of "high premiums," this interactive feedback data is directly related to the recommendation effect and is considered valid feedback data; conversely, if a customer accidentally clicks and immediately closes a promotional event, this interactive feedback data is irrelevant to the recommendation effect and is excluded. This filtering ensures that subsequent optimizations are based on genuine and valid customer feedback, thereby improving optimization effectiveness.

[0071] Valid feedback data is used as a reward signal for reinforcement learning, and policy gradient methods (such as PPO, DQN, etc.) are used to update the parameters of the health risk identification model and the rules of the insurance scheme matching logic.

[0072] The updated health risk identification model and insurance plan matching logic were compared and optimized. Customers were randomly divided into an experimental group and a control group. The experimental group used the updated model and logic, while the control group used the original model and logic. The optimization effect was verified by comparing the recommendation effects of the two groups. If the experimental group performed significantly better than the control group, the updated model and logic were officially put into use; if the effect did not meet expectations, the optimization direction was further adjusted based on the test results, and the process was continuously iterated.

[0073] Figure 2 This application illustrates a product intelligent recommendation system according to an embodiment of the present application, from... Figure 2 As can be seen, the product intelligent recommendation system 200 includes: The data acquisition unit 210 is used to collect multi-dimensional health data of customers, extract multi-modal features of multi-dimensional health data, and fuse multi-modal features into a unified health profile vector. Risk identification unit 220 is used to identify the customer’s disease risk probability and key risk factors based on a health risk identification model according to a unified health profile vector. The scheme generation unit 230 is used to determine the recommended insurance scheme based on the insurance scheme matching logic, according to the disease risk probability, key risk factors, insurance product information and customer preference characteristics. The interaction unit 240 is used to generate customer consultation scripts and plan presentation content based on the unified health profile vector and insurance recommendation scheme, and to reach customers with the consultation scripts and plan presentation content through various communication methods.

[0074] In some optional implementations, in the above system, multi-dimensional health data includes: structured health data, unstructured health data, and time-series health data; the data acquisition unit 210 is used to: collect the customer's structured health data, unstructured health data, and time-series health data; convert the structured health data into structured feature vectors through a feature embedding layer; extract the semantic feature vectors of the unstructured health data through an insurance domain natural language processing model; extract the dynamic health feature vectors of the time-series health data through a temporal convolutional network or a Transformer temporal model; and use a cross-attention mechanism to weightedly fuse the structured feature vectors, semantic feature vectors, and dynamic health feature vectors to generate a unified health profile vector.

[0075] In some optional implementations, in the above system, the health risk identification model includes a deep neural network and a causal inference framework; the risk identification unit 220 is used to: identify risks in a unified health profile vector through a deep neural network and output the customer's disease risk probability; and perform attribution analysis on the disease risk probability through the causal inference framework and output key risk factors based on the causal relationship between the disease and the risk factors.

[0076] In some optional implementations, the insurance scheme matching logic in the above system includes: a health-driven recommendation engine and a customer preference recommendation engine; the scheme generation unit 230 is used to: determine the protection gap based on the disease risk probability and key risk factors; match candidate insurance products from insurance product information based on the health-driven recommendation engine according to the protection gap; and sort the candidate insurance products based on the customer preference recommendation engine according to customer preference characteristics to determine the insurance recommendation scheme.

[0077] In some optional implementations, in the above system, the solution generation unit 230 is further used to: construct an insurance product knowledge graph; wherein the nodes of the insurance product knowledge graph include: diseases, risk factors, insurance clauses, and claims cases, and the edges of the insurance product knowledge graph include: causal relationships between diseases and risk factors, coverage relationships between insurance clauses and diseases, and application relationships between claims cases and insurance clauses; and to perform compliance and completeness checks on insurance recommendation solutions based on the insurance product knowledge graph, and to correct insurance recommendation solutions when the check results are abnormal.

[0078] In some optional implementations, in the above system, the interaction unit 240 is used to: generate consultation scripts based on a unified health profile vector and insurance recommendation schemes, using a large language model combined with retrieval enhancement; wherein the consultation scripts include: disease risk reminders, explanations of protection gaps, introductions to insurance recommendation schemes, and insurance application guidance; generate scheme presentation content based on a multi-role collaborative intelligent agent, wherein the multi-role collaborative intelligent agent includes: a risk explanation intelligent agent, a scheme generation intelligent agent, and a compliance intelligent agent; the risk explanation intelligent agent explains the correspondence between key risk factors and insurance recommendation schemes, the scheme generation intelligent agent generates textual descriptions and visual charts of insurance recommendation schemes, and the compliance intelligent agent verifies the compliance of the scheme presentation content; and reach users with the consultation scripts and scheme presentation content through agent applications, WeChat Work, online stores, and outbound telephone calls based on multimodal interaction components.

[0079] In some optional implementations, the system further includes a feedback optimization unit: used to collect customer interaction feedback data, eliminate irrelevant data through causal reasoning to obtain effective feedback data; use the effective feedback data as a reward signal for reinforcement learning, and update the parameters of the health risk identification model and the rules of the insurance plan matching logic through a policy gradient method; and compare and optimize the updated health risk identification model and insurance plan matching logic based on an A / B testing mechanism.

[0080] Specifically, the product intelligent recommendation system 200 can adopt a five-layer architecture: The data acquisition unit 210 integrates the customer's multimodal health data into a unified health profile vector, thereby consolidating the customer's health data; Risk identification unit 220 uses deep neural networks and a causal inference framework to identify and attribute disease risks to customers. Solution generation unit 230 generates differentiated guarantee combinations through a two-layer recommendation engine and knowledge graph checks; The execution interaction unit 240 generates personalized consultation scripts and visual solution introductions through the AI ​​Agent, and completes solution outreach based on multimodal interaction components; The feedback optimization unit performs reinforcement learning and causal reasoning based on customer behavior feedback to achieve model optimization and logic updates.

[0081] It should be noted that the aforementioned intelligent product recommendation system 200 can implement all of the aforementioned intelligent product recommendation methods, which will not be elaborated upon further.

[0082] Figure 3 This invention illustrates a schematic diagram of the structure of an electronic device according to an embodiment of the present application. Figure 3As shown, the electronic device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used for communication with external devices via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a product intelligent recommendation method.

[0083] In one embodiment, the electronic device provided in this application includes a memory and a processor. The memory stores a database and a computer program that can run on the processor. When the processor executes the computer program, it implements the steps of the aforementioned intelligent product recommendation method.

[0084] The above is as stated in this application. Figure 2 The product intelligent recommendation system method disclosed in the illustrated embodiments can be applied to a processor or implemented by a processor. During implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by software instructions. The processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The steps of the method disclosed in the embodiments of this application can be directly embodied as being executed by a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0085] In one embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the aforementioned intelligent product recommendation method.

[0086] It should be noted that the functions or steps that the above-mentioned electronic devices or computer-readable storage media can achieve can be referred to the relevant descriptions in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0087] Those skilled in the art will understand that all or part of the processes in the methods of 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 of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0088] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above.

[0089] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A product intelligent recommendation method, characterized in that, include: Collect multi-dimensional health data from customers, extract multi-modal features from the multi-dimensional health data, and fuse the multi-modal features into a unified health profile vector; Based on the unified health profile vector, the customer's disease risk probability and key risk factors are identified using a health risk identification model. Based on disease risk probability, key risk factors, insurance product information, and customer preference characteristics, insurance recommendation schemes are determined according to insurance scheme matching logic. Based on the unified health profile vector and insurance recommendation scheme, customer consultation scripts and scheme presentation content are generated, and the consultation scripts and scheme presentation content are delivered to customers through various communication methods.

2. The method according to claim 1, characterized in that, The multidimensional health data includes: structured health data, unstructured health data, and time-series health data; The process of collecting multi-dimensional health data from customers, extracting multi-modal features from the multi-dimensional health data, and fusing the multi-modal features into a unified health profile vector includes: Collect customers' structured health data, unstructured health data, and time-series health data; Structured health data is converted into structured feature vectors through a feature embedding layer; Extract semantic feature vectors from unstructured health data using natural language processing models in the insurance field; Dynamic health feature vectors of time-series health data can be extracted using temporal convolutional networks or Transformer temporal models. A cross-attention mechanism is used to weight and fuse structured feature vectors, semantic feature vectors, and dynamic health feature vectors to generate a unified health profile vector.

3. The method according to claim 1, characterized in that, The health risk identification model includes deep neural networks and a causal inference framework; The process of identifying the customer's disease risk probability and key risk factors based on a unified health profile vector and a health risk identification model includes: Risk identification is performed on the unified health profile vector using a deep neural network, and the probability of the customer's disease risk is output. Attribution analysis of disease risk probability is performed using a causal inference framework, and key risk factors are output based on the causal relationship between disease and risk factors.

4. The method according to claim 1, characterized in that, The insurance scheme matching logic includes: a health-driven recommendation engine and a customer preference recommendation engine; The process of determining recommended insurance plans based on disease risk probability, key risk factors, insurance product information, and customer preference characteristics, using insurance plan matching logic, includes: The coverage gap is determined based on the probability of disease risk and key risk factors; Based on the coverage gap, a health-driven recommendation engine is used to match candidate insurance products from insurance product information; Based on customer preference characteristics, candidate insurance products are ranked using a customer preference recommendation engine to determine the insurance recommendation scheme.

5. The method according to claim 4, characterized in that, Following the step of determining the insurance recommendation scheme based on the insurance scheme matching logic, the method further includes: Construct an insurance product knowledge graph; the nodes of the insurance product knowledge graph include: diseases, risk factors, insurance terms, and claims cases; the edges of the insurance product knowledge graph include: causal relationship between diseases and risk factors, coverage relationship between insurance terms and diseases, and application relationship between claims cases and insurance terms. The insurance recommendation scheme is checked for compliance and completeness based on the insurance product knowledge graph, and the insurance recommendation scheme is corrected when the check results are abnormal.

6. The method according to claim 1, characterized in that, The process involves generating customer consultation scripts and plan presentation content based on a unified health profile vector and insurance recommendation scheme, and then delivering these scripts and plan presentation content to customers through various communication methods, including: Based on a unified health profile vector and an insurance recommendation scheme, consultation scripts are generated using a large language model combined with retrieval enhancement. These scripts include: disease risk reminders, explanations of coverage gaps, introductions to recommended insurance schemes, and guidance on purchasing insurance. Based on the unified health profile vector and insurance recommendation scheme, the content is presented by a multi-role collaborative intelligent agent. The multi-role collaborative intelligent agent includes: risk interpretation intelligent agent, scheme generation intelligent agent, and compliance intelligent agent. The risk interpretation intelligent agent explains the correspondence between key risk factors and insurance recommendation schemes, the scheme generation intelligent agent generates textual descriptions and visual charts of insurance recommendation schemes, and the compliance intelligent agent verifies the compliance of the content presented by the scheme. Consultation scripts and solution presentations are based on multimodal interactive components and reach users through agent applications, WeChat Work, online stores, and outbound telephone calls.

7. The method according to claim 1, characterized in that, The method further includes: Collect customer interaction feedback data and eliminate irrelevant data through causal reasoning to obtain effective feedback data; Effective feedback data is used as a reward signal for reinforcement learning, and the parameters of the health risk identification model and the rules of the insurance scheme matching logic are updated through the policy gradient method. Based on the A / B testing mechanism, the updated health risk identification model and insurance plan matching logic are compared and optimized.

8. A product intelligent recommendation system, characterized in that, include: The data acquisition unit is used to collect multi-dimensional health data from customers, extract multi-modal features from the multi-dimensional health data, and fuse the multi-modal features into a unified health profile vector. The risk identification unit is used to identify the customer's disease risk probability and key risk factors based on a health risk identification model using a unified health profile vector. The solution generation unit is used to determine the recommended insurance solution based on the insurance solution matching logic, according to the probability of disease risk, key risk factors, insurance product information and customer preference characteristics. The execution interaction unit is used to generate customer consultation scripts and plan presentation content based on the unified health profile vector and insurance recommendation scheme, and to reach customers with consultation scripts and plan presentation content through various communication methods.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the product intelligent recommendation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is instructed by the processor, it implements the steps of the product intelligent recommendation method as described in any one of claims 1 to 7.

Images