Product recommendation method and device based on artificial intelligence, equipment and storage medium

Abstract

The present application relates to the field of natural language processing and artificial intelligence, and is suitable for the internet financial scene, and particularly relates to a product recommendation method, device and equipment based on artificial intelligence and a storage medium, the product recommendation method based on artificial intelligence comprises the following steps: obtaining user data of a target car insurance user, obtaining a non-car insurance product purchase probability of the target car insurance user according to the user data of the target user, and marking the target car insurance user as one of a first type of user, a second type of user and a third type of user according to the non-car insurance product purchase probability; predicting the demand for non-car insurance products of the target car insurance user according to the user data to obtain a non-car insurance product demand prediction list, and recommending products to the target car insurance user according to the non-car insurance product demand prediction list if the target car insurance user is a first type of user. Through multi-dimensional data modeling and purchase probability quantification, high-intention first type of users are accurately screened, blind recommendation is avoided, the conversion efficiency of non-car insurance products is greatly improved, and fine operation is realized.

Description

Technical Field

[0001] This invention relates to the fields of finance and artificial intelligence, and is applicable to internet finance scenarios. In particular, it relates to a product recommendation method, apparatus, device, and storage medium based on artificial intelligence. Background Technology

[0002] In the mainstream business scenario of cross-selling auto insurance and non-auto insurance in the insurance industry, promoting non-auto insurance products has become a core direction for enhancing customer value and increasing business volume. However, existing recommendation methods generally have technical shortcomings: on the one hand, they rely on manual experience or single static rules for customer screening and segmentation, resulting in coarse-grained segmentation and a lack of comprehensive consideration of multi-dimensional data such as user dynamic behavior and family attributes, leading to low accuracy in identifying high-intent potential customers; on the other hand, a systematic user demand prediction mechanism has not been established, and recommendation scripts are mostly standardized templates that cannot adapt to the personalized needs of different users, resulting in a lack of precise guidance in agent communication. These problems directly lead to low conversion efficiency of non-auto insurance products and a lack of refined operation. Summary of the Invention

[0003] This invention provides a product recommendation method, apparatus, device, and storage medium based on artificial intelligence to solve the technical problems of low conversion efficiency and insufficient refined operation of non-motor insurance products when cross-selling motor insurance and non-motor insurance in the prior art.

[0004] Firstly, an artificial intelligence-based product recommendation method is provided, the method comprising: Obtain user data of the target auto insurance user, including personal information data, insurance interaction data, and historical browsing data; Based on the user data and the pre-trained product type prediction model, the non-auto insurance product demand prediction of the target auto insurance user is performed, and a non-auto insurance product demand prediction list of the target auto insurance user is generated. Based on the user data and the pre-trained purchase probability prediction model, the purchase probability of non-auto insurance products for the target auto insurance user is obtained. If the probability of purchasing the non-motor insurance product falls within the first range, a first personalized recommendation scheme is generated for the target motor insurance user based on the non-motor insurance product demand prediction list, and products are recommended to the target motor insurance user according to the first personalized recommendation scheme.

[0005] Secondly, an artificial intelligence-based product recommendation device is provided, the artificial intelligence-based product recommendation device comprising: The data acquisition module is used to acquire user data of the target auto insurance user, including personal information data, insurance interaction data, and historical browsing data. The recommendation list acquisition module is used to predict the non-auto insurance product demand of the target auto insurance user based on the user data and a pre-trained product type prediction model, and generate a non-auto insurance product demand prediction list for the target auto insurance user. The purchase probability acquisition module is used to obtain the purchase probability of non-auto insurance products of the target auto insurance user based on the user data and the pre-trained purchase probability prediction model. The product recommendation module is used to generate a first personalized recommendation scheme for the target car insurance user based on the non-car insurance product demand prediction list if the purchase probability of the non-car insurance product is in the first range, and to recommend products to the target car insurance user according to the first personalized recommendation scheme.

[0006] Thirdly, a computer device is provided, 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 artificial intelligence-based product recommendation method.

[0007] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the aforementioned artificial intelligence-based product recommendation method.

[0008] The aforementioned AI-based product recommendation method, device, equipment, and storage medium first acquires user data of target auto insurance users, including personal information data, insurance interaction data, and historical browsing data, thus constructing a comprehensive data foundation. Secondly, based on multi-dimensional user data, a demand prediction model analyzes users' potential protection needs, generating a non-auto insurance product demand prediction list containing core and secondary demand products. Then, based on the collected user data, the prediction model calculates the purchase probability of non-auto insurance products. When the purchase probability falls within a first range, a first personalized recommendation scheme is generated based on the demand prediction list, and suitable products are pushed to the target user group, achieving precise conversion. Through multi-dimensional data modeling and purchase probability quantification, users with high-intent non-auto insurance product purchase probabilities within the first range are accurately screened, avoiding blind recommendations and significantly improving the conversion efficiency of non-auto insurance products, thus achieving refined operation. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1This is a schematic diagram of an application environment for an artificial intelligence-based product recommendation method provided in an embodiment of the present invention.

[0011] Figure 2 This is a schematic flowchart of the first embodiment of the product recommendation method based on artificial intelligence provided in this invention.

[0012] Figure 3 This is a schematic flowchart of a second embodiment of the product recommendation method based on artificial intelligence provided in an embodiment of the present invention.

[0013] Figure 4 yes Figure 2 A schematic diagram of a specific implementation method for step S30.

[0014] Figure 5 yes Figure 2 A schematic diagram of a specific implementation method for step S50.

[0015] Figure 6 This is a schematic flowchart of the fourth embodiment of the product recommendation method based on artificial intelligence provided in an embodiment of the present invention.

[0016] Figure 7 This is a schematic diagram of an artificial intelligence-based product recommendation device according to an embodiment of the present invention. Figure 8 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The AI-based product recommendation method provided in this invention can be applied to, for example... Figure 1 In the application environment, Figure 1 This is a schematic diagram of an application environment for an artificial intelligence-based product recommendation method provided in an embodiment of the present invention, such as... Figure 1As shown, this application environment includes a terminal 10 and a server 20. Terminal 10 and server 20 jointly execute an AI-based product recommendation method. It should be noted that terminal 10 and server 20 can be smartphones, tablets, laptops, desktop computers, etc., but are not limited to these. Terminal 10 and server 20 can be connected via Bluetooth, USB (Universal Serial Bus), or other communication methods; this invention does not impose any limitations on these connections.

[0019] Figure 2 This is a schematic flowchart of the first embodiment of the product recommendation method based on artificial intelligence provided in this invention, as follows: Figure 2 As shown, this AI-based product recommendation method may specifically include the following steps: S20. Obtain user data of the target auto insurance user, wherein the user data includes personal information data, insurance interaction data, and historical browsing data.

[0020] In this step, personal information data, including age, gender, family attributes (whether there are children / elderly), assets (whether there is a house / car), and occupation type, is extracted in batches from the customer information database and updated quarterly. Insurance interaction data, including car insurance policy years, renewal records, non-car insurance historical policy records, claims records for the past two years (car insurance / non-car insurance), and complimentary insurance receipt records, is synchronized in real time from the policy / claims system. Historical browsing data, including the app's non-car product click trajectory for the past 90 days, article / live stream viewing time, community participation records, and function service usage logs (such as chauffeur / car wash service call records), is collected daily from the behavior log system.

[0021] S30. Based on the user data and the pre-trained product type prediction model, predict the non-auto insurance product demand of the target auto insurance user and generate a list of predicted non-auto insurance product demand of the target auto insurance user.

[0022] In this step, the pre-trained product type prediction model can use an XGBoost multi-class classification model, targeting the non-auto insurance categories that users are most likely to purchase, including health insurance, home insurance, pet insurance, accident insurance, and travel insurance. The product purchase probability for each user is sorted from high to low, and the top 3 products are selected to form a demand prediction list. This list clearly identifies the core demand product (Top 1) and secondary demand products (Top 2-3), and labels the reasons for product suitability. For example, core demand: comprehensive medical insurance; suitability reasons: families with children + health insurance click records.

[0023] S40. Based on the user data and the pre-trained purchase probability prediction model, obtain the purchase probability of non-auto insurance products for the target auto insurance user.

[0024] In this step, the pre-trained purchase probability prediction model can use the XGBoost binary classification model, with the target label being whether non-auto insurance has been purchased in the past 12 months (1=purchase, 0=not purchased). The user data after S20 preprocessing is split into a training set and a test set in a 7:3 ratio. The model outputs the non-auto insurance purchase probability in the range of 0-1. When the non-auto insurance purchase probability is in the first range, the target auto insurance user can be marked as a high-intent user. The specific range can be set according to the actual situation. In a preferred implementation, the first range is [0.7, 1].

[0025] Furthermore, in practice, the probability of non-auto insurance purchases can be mapped to a grade score using a linear scaling formula, making it easier for staff to intuitively obtain customer information. In a specific embodiment, the grade score = purchase probability × 5. The first grade score range is 3.5-5.0, corresponding to a purchase probability ≥ 70%, marked as high-intent users and the core conversion target; the second grade score range is 2.0-3.4, corresponding to purchase probabilities [30%, 70%], marked as second-tier users, i.e., medium-intent users, requiring focused follow-up and cultivation; the third grade score range is 0-1.9, corresponding to a purchase probability < 30%, marked as third-tier users, i.e., low-intent users, using a lightweight cultivation strategy.

[0026] S50. If the probability of purchasing the non-motor insurance product is in the first range, then a first personalized recommendation scheme is generated for the target motor insurance user based on the non-motor insurance product demand prediction list, and products are recommended to the target motor insurance user according to the first personalized recommendation scheme.

[0027] In this step, if the purchase probability of the non-auto insurance product falls within the first range, it indicates that the target auto insurance user is a high-intent user. At this point, based on the target auto insurance user's user data and the core and secondary products in the demand prediction list, historical demonstration cases from the non-auto insurance product recommendation knowledge base are invoked. The sales pitch logic from these demonstration cases is extracted and combined with the target auto insurance user's product needs and characteristic data to form a first personalized recommendation scheme. The first personalized recommendation scheme includes a complete solution encompassing demand matching, scheme design, sales pitch generation, and execution. A GPT-like large model can be guided by PromptEngineering, inputting user data, the demand prediction list, and historical demonstration cases to generate the first personalized recommendation scheme. Product recommendations are then made based on the user's reach preferences in their historical browsing data, such as prioritizing WeChat private message recommendations for active WeChat users and prioritizing app pop-up recommendations for users who frequently open apps.

[0028] In one specific implementation, within an internet finance scenario, an insurance company needs to improve the conversion efficiency of non-motor insurance users for auto insurance, focusing on four core non-motor insurance categories: health insurance, home insurance, pet insurance, and travel insurance.

[0029] Step 1: Extract target user's personal information data, insurance interaction data, and historical browsing data from the core customer information database. Personal information data includes basic attributes: 35 years old, female, married, has 2 children (8-year-old boy + 5-year-old girl), parents alive (62 years old / 59 years old); asset attributes: owns a house (a residential property in a first-tier city), no pets, daily commuting relies on a private car (3-year-old, joint-venture brand SUV). Insurance interaction data includes insurance records: has purchased auto insurance from Ping An for 3 consecutive years, no historical non-auto insurance records, and received a "7-day trial insurance for million-dollar medical insurance" through the app in March 2024; claims records: no auto insurance / non-auto insurance claims in the past 2 years, and no complaint records. Historical browsing data includes product interactions: clicked the "million-dollar medical insurance" product details page 3 times, watched the "family protection planning" live stream for 15 minutes, and saved the "home property insurance purchase guide" article; service usage: used the Ping An Good Car Owner App parking payment function 5 times, clicked the "chauffeur service" entry once (without placing an order), and was active on the app for 12 days within the past 30 days. The acquired user data is preprocessed, including deduplication and completion, and then all data is standardized into a structured format.

[0030] Step Two: Based on the user data obtained in Step One, call the pre-trained XGBoost multi-class demand prediction model. Input demand-related features from the user data, such as "family attributes (children + elderly)," "asset attributes (homeownership)," and "behavioral characteristics (clicking on million-dollar medical insurance + watching family protection live streams)," and the model outputs the purchase probabilities of various products: Million-dollar medical insurance (family version): 72%; Home insurance (upgraded version): 58%; Travel accident insurance (annual version): 31%; Pet insurance: 0%. Sort by probability from high to low, and select the top 3 to form a demand prediction list. The list includes demand priority, recommended products, and reasons for suitability.

[0031] Step 3: Use the well-trained XGBoost binary classification model. The training samples for this model are data from 1 million car insurance users over the past 2 years, labeled as "whether non-car insurance has been purchased". Input the user data preprocessed in Step 1 into the model. The model input features include 28 core features such as "owning a house = 1", "receiving free insurance = 1", "million-dollar medical insurance clicks = 3", and "30-day active = 12". The model outputs a non-car insurance purchase probability of 78%.

[0032] Step 4: Assuming the preset first threshold is 0.7, and the probability of purchasing non-car insurance products obtained in Step 3 is 78%, which is in the first interval, mark the user as a high-intent user. At this point, successful non-auto insurance recommendation examples from the non-auto insurance product recommendation knowledge base are used to match similar cases of a user group with children, homeownership, and high activity levels, extracting the sales pitch logic. A customized recommendation script is generated through a large-scale model, combining the demand prediction list with user characteristics to generate a first personalized recommendation plan. This first personalized recommendation plan includes: demand matching, based on the user's characteristics of "family with children + focus on family medical care + high activity," matching core demand (million-dollar medical insurance for families) and secondary demand (upgraded home insurance); plan design, with the core recommendation being the million-dollar medical insurance for families (highlighting shared deductible and premium discounts), paired with home insurance (bundled discounts), and configured with a limited-time "free family health check" benefit; script generation, based on similar cases from the knowledge graph, generating personalized scripts through a large-scale model, including user name, auto insurance loyal customer status, and reasons for demand matching; and execution, using both App pop-up and WeChat private message channels for outreach, with 24-hour tracking and response, and supplementary follow-up scripts for "click-but-not-paid" cases.

[0033] The first personalized recommendation scheme described above will be implemented to make recommendations to the target auto insurance users.

[0034] Figure 3 This is a flowchart illustrating a second embodiment of the artificial intelligence-based product recommendation method provided in one embodiment of the present invention, as shown below. Figure 3 As shown, this AI-based product recommendation method may specifically include the following steps: S60. Obtain the interactive voice data between the agent and the target car insurance user when recommending products to the target car insurance user, and determine whether the product recommendation is successful based on the interactive voice data.

[0035] In this step, the entire voice recording of the call between the agent and the target auto insurance user is automatically recorded through the insurance company's call center system, and can be uploaded to the audio storage server in real time. The voice file is input into the ASR model, and after speaker recognition, timestamp marking, and keyword extraction, structured interactive text is output. If the interactive text contains statements from the user explicitly agreeing to the insurance purchase, such as "I want to purchase insurance," "Apply now," or "Place an order for me," or if the agent subsequently enters "User has submitted an insurance application" or "Premium has been paid" into the business system, the recommendation is considered successful. If the interactive text contains statements from the user explicitly rejecting the purchase (such as "I don't want to buy," "I don't need it," or "We'll talk about it later") or not explicitly agreeing or rejecting (such as "I'll think about it," or "I need to discuss it with my family"), the recommendation is considered unsuccessful, and the process enters the S70 secondary recommendation process.

[0036] S70. If the product recommendation fails, the agent's execution status of the first personalized recommendation scheme and the target car insurance user's product needs are obtained based on the interactive voice data. A second personalized recommendation scheme is generated based on the execution status and the product needs. The target car insurance user is then recommended a product based on the second personalized recommendation scheme.

[0037] In this step, the agent's execution of the first personalized recommendation plan includes whether the agent fully conveyed the product's core selling points, whether the reasons for matching the user's situation were appropriate, whether the agent responded to the user's initial questions, and whether the product was recommended according to the priority of the recommendation plan. This part can be done by using a large model to analyze the interaction text sentence by sentence. For example, if the recommendation plan requires "prioritizing the million-dollar medical family version, highlighting the family shared deductible," but the agent does not mention "family shared deductible" in the interaction text, it is marked as "core selling point omission"; if the agent does not recommend the home insurance upgrade version (a secondary need), it is marked as "product priority execution deviation"; combining all marked content, an "Agent Execution Report" is generated, clarifying the highlights and defects of the execution, such as the omission of 2 core selling points and the failure to guide the family to purchase insurance advantages.

[0038] Secondly, core demand / objection keywords are extracted from the user's speech-to-text using NLP keyword extraction models (such as TF-IDF+TextRank). Based on the insurance industry demand intent database, user needs are categorized using multi-classification models (such as CNN-BiLSTM) into: price-sensitive (keywords: too expensive, limited budget, discount); protection-demand (keywords: coverage, sum insured, claim conditions); product-suitable (keywords: parents, children, single-person protection); and service-requirement (keywords: payment method, after-sales service, consultation channels).

[0039] Then, we supplemented and corrected the shortcomings in our agent operations, and precisely adapted to the needs of users. We added missing selling points, such as a detailed explanation of the "family shared deductible." We adapted to user needs, such as highlighting the "monthly payment method" and "family insurance discount" to price-sensitive users; and explaining to users who want to insure their parents that the product explicitly supports supplementary insurance for seniors over 60 years old, with no medical examination required. We adjusted our sales pitch logic, such as first addressing user objections before emphasizing the product's value.

[0040] Finally, prioritize channels with high user acceptance. For example, the first recommendation can be made through the App and WeChat Work, and the second recommendation can be made by combining WeChat Work private messages and telephone follow-ups. Agents can obtain the full text of the second recommendation plan through the work platform.

[0041] In some implementations, this AI-based product recommendation method may specifically include the following steps: S80. If the product recommendation is successful, extract the step-by-step path from the first personalized recommendation scheme. The step-by-step path includes the recommendation script template and the complete execution steps. Based on the step-by-step path, the non-auto insurance product demand prediction list, and the user data acquisition demonstration case, add the demonstration case to the product recommendation knowledge base, and update the index of the product recommendation knowledge base.

[0042] In this step, a precise step-by-step path is extracted from the first personalized recommendation plan. The complete execution steps are broken down according to chronological order and action logic, clarifying whether the operation at each stage is automated or conducted manually, the completion time limit, and the input / output data (e.g., pre-construction preparation → channel selection → initial contact → response tracking → objection handling → conversion loop). Recommended script templates are extracted and categorized by scenario, including initial contact, objection handling, and insurance guidance. After removing personalized variables, placeholders and filling rules are added to ensure reusability.

[0043] The extracted step-by-step paths, non-motor insurance product demand prediction lists, target motor insurance user personal information data, insurance interaction data, and historical browsing data are integrated to form demonstration cases. These demonstration cases can also be supplemented with basic information such as case number, scenario, and success time, as well as conversion results such as insurance product, premium, and conversion time. The demonstration cases are then categorized and added to the product recommendation knowledge base according to user characteristic tags and business scenario tags; simultaneously, the knowledge base index is updated, adding relevant search dimensions to ensure rapid retrieval and matching in the future.

[0044] Figure 4 yes Figure 2 A schematic diagram of a specific implementation method for step S30 is shown below. Figure 4 As shown, step S30 includes: S31. Based on the user data and the pre-trained product type prediction model, obtain the probability of the target car insurance user's demand for multiple non-car insurance products.

[0045] In this step, the product type prediction model can adopt the XGBoost multi-class classification model. This model is suitable for the core task of predicting the probability of multi-product demand. It can simultaneously output the probability of a user's demand for multiple non-auto insurance product categories and effectively capture the complex relationship between user characteristics and different product demands, achieving better prediction accuracy than traditional multi-class classification models. The model is pre-trained based on historical data from auto insurance users over the past two years. The target variable is the non-auto insurance product category that the user is most likely to purchase, covering core categories such as health insurance, home insurance, accident insurance, travel insurance, and pet insurance. User data is input into the pre-trained non-auto insurance product demand prediction model. The model outputs the probability of the target auto insurance user's demand for each core non-auto insurance product category, with a value range of [0,1]. The closer the probability is to 1, the stronger the user's demand for that product.

[0046] S32. Sort the multiple non-motor insurance products from high to low according to the probability of demand to obtain the product recommendation list for the target motor insurance user.

[0047] In this implementation, the output demand probabilities for multiple non-motor insurance products are sorted from highest to lowest. Products with a demand probability ≥10% are prioritized for retention, while products with extremely low demand are eliminated to avoid redundant recommendations. The sorted list forms a product recommendation list containing three core elements: demand priority, product category, and matching reason. The matching reason is generated by combining user characteristics and behavioral tags, such as "Home insurance demand probability 65%, matching reason: user owns a house + recently viewed home insurance application guide." Based on the non-motor insurance product sorting results, core and secondary demand levels are defined. Typically, the top 3 products are selected as the focus of recommendations, with the top 1 being the core demand product and the top 2-3 being secondary demand products.

[0048] In some implementations, step S40 further includes: S41. Extract user features from the user data and generate user behavior labels using an unsupervised clustering algorithm. Based on the user behavior labels and a pre-trained purchase probability prediction model, obtain the non-auto insurance product purchase probability of the target auto insurance user.

[0049] In this step, static features that are not easily changed are extracted from personal information data and insurance interaction data, including age, gender, family attributes, asset status, occupation type, years of car insurance coverage, renewal records, and historical claims / complaint records. Dynamic features reflecting recent user behavior preferences are extracted from historical browsing data, including the number of clicks on non-car insurance products in the past 90 days, the duration of live stream / article viewing, service usage frequency, and the number of active days on the App in the past 30 days. Redundant features are eliminated through variance analysis and correlation analysis, retaining several core features to form a standardized feature matrix. After preprocessing the standardized feature matrix, user features are obtained.

[0050] The K-Means unsupervised clustering algorithm is employed. This algorithm is highly adaptable to structured data, computationally efficient, and can quickly segment user behavior into groups, meeting the processing needs of massive user volumes in insurance scenarios. Clustering error is analyzed using the Elbow Method, and the number of clusters is set to 5-8 based on the requirements of the insurance business scenario (e.g., 5 categories: high-frequency active users, family protection-focused users, price-sensitive users, service-dependent users, and low-frequency browsing users). After clustering, the behavioral characteristics of each user group are summarized, generating concrete behavioral tags (e.g., characteristics corresponding to "family protection-focused users": frequent clicking on family medical products, watching family protection-related live streams). These behavioral tags are then added as new features and fused with user characteristics to form an enhanced feature matrix.

[0051] A pre-trained XGBoost binary classification model is selected. The enhanced feature matrix is ​​input into the trained XGBoost model, and the output is the probability of the target car insurance user purchasing non-car insurance products, with a value range of [0,1]. The closer the probability is to 1, the stronger the user's purchase intention. The XGBoost binary classification model can be trained using historical data of car insurance users from the past two years as training samples, with "whether to purchase non-car insurance products" as the target variable.

[0052] Figure 5 yes Figure 2 A schematic diagram of a specific implementation method for step S50 is shown below. Figure 5 As shown, in some embodiments, step S50 may include: S51. If the purchase probability of the non-motor insurance product is in the second interval, then generate a third personalized recommendation scheme for the target motor insurance user based on the non-motor insurance product with the highest demand probability in the non-motor insurance product demand prediction list, and recommend products to the target motor insurance user based on the third personalized recommendation scheme.

[0053] In this step, if the probability of purchasing non-auto insurance products falls within the second range, it indicates that the target auto insurance user is a moderately interested user with a moderate need to purchase non-auto insurance products. In a preferred implementation, the second range is [0.3, 0.7]. From the list of predicted non-auto insurance product demand, the product with the highest demand probability is extracted as the core recommendation, while secondary demand products are eliminated to avoid distracting users with multiple product recommendations, thus aligning with the moderately interested and cautious decision-making characteristics of the second type of user. The third personalized recommendation scheme focuses on the core demand product, including a simplified recommendation script, the product's core selling points, reasons for suitability, and lightweight benefits; the script should highlight core advantages such as "high cost-effectiveness" and "comprehensive basic coverage," aligning with the moderately interested user's need for a balance between risk and cost.

[0054] S52. If the probability of purchasing non-motor insurance products is in the third interval, then based on the target motor insurance user's non-motor insurance product demand prediction list, push recommended advertisements for non-motor insurance products to the target motor insurance user.

[0055] In this step, if the purchase probability of the non-auto insurance product falls within the third interval, it indicates that the target auto insurance user is a low-intent user and currently has no need to purchase non-auto insurance products. In a preferred implementation, the third interval is [0, 0.3]. These users have a low purchase probability, and advertising recommendations are the primary approach. The advertisements mainly lower the user's cognitive threshold by introducing low-barrier products, gradually building awareness of non-auto insurance products among the three types of users. Specifically, one or two non-auto insurance products with a demand probability ≥ 15% can be selected as the target for push notifications, prioritizing low-barrier, low-premium products to reduce user resistance.

[0056] Figure 6 This is a schematic flowchart of the fourth embodiment of the product recommendation method based on artificial intelligence provided in one embodiment of the present invention, as shown below. Figure 7 As shown, in some embodiments, before step S20, the following steps are further included: S10. Obtain user data for all car insurance users, predict the user retention probability of each car insurance user based on a pre-trained user retention probability prediction model, and divide the car insurance users into a first priority user group, a second priority user group, and a third priority user group according to the user retention probability.

[0057] In this step, user data for all auto insurance users is retrieved through the insurance company's data platform, including personal information data, insurance interaction data, and historical browsing data. An XGBoost binary classification model is used as the user retention probability prediction model. This model effectively captures the complex correlation between user characteristics and retention behavior, and its prediction accuracy is superior to traditional models. The model is pre-trained based on historical data of auto insurance users over the past three years, with the target variable being whether the user will remain in the market over a relatively long period, typically 12 months. After feature preprocessing, the acquired user data is input into the pre-trained retention probability prediction model, which outputs the retention probability for each auto insurance user over the next 12 months, with values ​​ranging from [0,1]. The closer the probability is to 1, the stronger the user's retention potential. Based on the retention targets of the auto insurance business and the capacity of operational resources, a retention probability threshold is set. Users with a retention probability ≥80% are classified as the first priority user group, which has high retention potential and is considered core high-quality users. Users with a retention probability of 40%-79% are classified as the second priority user group, which has moderate retention potential and is considered key users for cultivation. Users with a retention probability <40% are classified as the third priority user group, which has low retention potential and is considered users for lightweight maintenance.

[0058] S11. Select target car insurance users sequentially from the first priority user group, the second priority user group, and the third priority user group.

[0059] In this step, following the principle of progressive priority, target users are first selected from the first priority user group. After completing the recommendation operation for this group, users are then selected from the second and third priority user groups in turn. At the same time, the selection ratio of each priority user is set in conjunction with business objectives. For example, the selection ratio of first priority users is not less than 60%, ensuring that core resources are focused on high-quality users.

[0060] The AI-based product recommendation method provided in this invention can be constructed based on AI, acquiring and processing relevant data using AI technology to achieve unattended AI-based product recommendations. Artificial intelligence (AI) is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results. Basic AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing technology, operating / interactive systems, and mechatronics. AI software technologies mainly include computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0061] In one embodiment, an AI-based product recommendation device is provided, which corresponds one-to-one with the AI-based product recommendation method in the above embodiments. For example... Figure 7 As shown, the AI-based product recommendation device includes: The data acquisition module 200 is used to acquire user data of the target auto insurance user, including personal information data, insurance interaction data, and historical browsing data. The recommendation list acquisition module 300 is used to predict the non-auto insurance product demand of the target auto insurance user based on the user data and the pre-trained product type prediction model, and generate a non-auto insurance product demand prediction list of the target auto insurance user. The purchase probability acquisition module 400 is used to acquire the non-auto insurance product purchase probability of the target auto insurance user based on the user data and the pre-trained purchase probability prediction model. The product recommendation module 500 is used to generate a first personalized recommendation scheme for the target car insurance user based on the non-car insurance product demand prediction list if the purchase probability of the non-car insurance product is greater than a first preset threshold, and to recommend products to the target car insurance user according to the first personalized recommendation scheme.

[0062] Specifically, the device also includes a recommendation judgment module and a secondary recommendation module.

[0063] The recommendation judgment module is used to acquire the interactive voice data between the agent and the target car insurance user when recommending products to the target car insurance user, and to determine whether the product recommendation is successful based on the interactive voice data.

[0064] The secondary recommendation module is used to, if the product recommendation fails, obtain the agent's execution status of the first personalized recommendation scheme and the product needs of the target car insurance user based on the interactive voice data, generate a second personalized recommendation scheme based on the execution status and the product needs, and recommend products to the target car insurance user based on the second personalized recommendation scheme.

[0065] Specifically, the device also includes a case extraction module.

[0066] The case extraction module is used to extract the step-by-step path in the first personalized recommendation scheme if the product recommendation is successful. The step-by-step path includes the recommendation script template and the complete execution steps. Based on the step-by-step path, the non-auto insurance product demand prediction list and the user data to obtain demonstration cases, the demonstration cases are added to the product recommendation knowledge base, and the index of the product recommendation knowledge base is updated.

[0067] Specifically, the recommendation list acquisition module 300 includes a product type prediction unit and a recommendation list acquisition subunit.

[0068] The product type prediction unit is used to obtain the probability of demand for multiple non-auto insurance products by the target auto insurance user based on the user data and the pre-trained product type prediction model.

[0069] The recommendation list acquisition sub-unit is used to sort multiple non-auto insurance products from high to low according to the probability of demand to obtain the product recommendation list for the target auto insurance user.

[0070] Specifically, the purchase probability acquisition module 400 includes a purchase probability acquisition subunit.

[0071] The purchase probability acquisition subunit is used to extract user features from the user data, generate user behavior labels through an unsupervised clustering algorithm, and obtain the non-auto insurance product purchase probability of the target auto insurance user based on the user behavior labels and a pre-trained purchase probability prediction model.

[0072] The product recommendation module 500 is further configured to generate a third personalized recommendation scheme for the target car insurance user based on the non-car insurance product with the highest demand probability in the non-car insurance product demand prediction list if the purchase probability of the non-car insurance product is in the second interval, and to recommend products to the target car insurance user based on the third personalized recommendation scheme.

[0073] The product recommendation module 500 is also used to push recommended advertisements for non-auto insurance products to the target auto insurance user based on the target auto insurance user's non-auto insurance product demand prediction list if the purchase probability of the non-auto insurance product is in the third interval.

[0074] Specifically, the device also includes a user grouping module and a user selection module.

[0075] The user segmentation module is used to acquire user data of all car insurance users, predict the user retention probability of each car insurance user based on a pre-trained user retention probability prediction model, and divide the car insurance users into a first priority user group, a second priority user group, and a third priority user group according to the user retention probability.

[0076] The user selection module is used to select target auto insurance users sequentially from the first priority user group, the second priority user group, and the third priority user group.

[0077] Figure 8 This is a schematic diagram of the structure of one embodiment of the computer device according to this application, as shown below. Figure 8 As shown, this application also provides a computer device, including: The memory and processor contain computer-readable instructions that, when executed by the processor, cause the processor to perform any step of the aforementioned AI-based product recommendation method.

[0078] This application also provides a computer-readable storage medium in which computer-readable instructions, when executed by one or more processors, cause the one or more processors to perform any step in an artificial intelligence-based product recommendation method. It is understood that the readable storage medium in this embodiment may be a volatile readable storage medium or a non-volatile readable storage medium.

[0079] 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 recommendation method based on artificial intelligence, characterized in that, include: Obtain user data of the target auto insurance user, including personal information data, insurance interaction data, and historical browsing data; Based on the user data and the pre-trained product type prediction model, the non-auto insurance product demand prediction of the target auto insurance user is performed, and a list of non-auto insurance product demand predictions for the target auto insurance user is generated. Based on the user data and the pre-trained purchase probability prediction model, the purchase probability of non-auto insurance products for the target auto insurance user is obtained. If the probability of purchasing the non-motor insurance product falls within the first range, a first personalized recommendation scheme is generated for the target motor insurance user based on the non-motor insurance product demand prediction list, and products are recommended to the target motor insurance user according to the first personalized recommendation scheme.

2. The product recommendation method based on artificial intelligence as described in claim 1, characterized in that, After recommending products to the target auto insurance user based on the first personalized recommendation scheme, the method further includes: Acquire the voice interaction data between the agent and the target car insurance user when recommending products to the target car insurance user, and determine whether the product recommendation is successful based on the voice interaction data. If the product recommendation fails, the agent's execution status of the first personalized recommendation scheme and the target car insurance user's product needs are obtained based on the interactive voice data. A second personalized recommendation scheme is generated based on the execution status and the product needs, and a product recommendation is made to the target car insurance user based on the second personalized recommendation scheme.

3. The product recommendation method based on artificial intelligence as described in claim 2, characterized in that, After determining whether the product recommendation was successful based on the interactive voice data, the process further includes: If the product recommendation is successful, the step-by-step path in the first personalized recommendation scheme is extracted. The step-by-step path includes the recommendation script template and the complete execution steps. Based on the step-by-step path, the non-auto insurance product demand prediction list, and the user data acquisition demonstration case, the demonstration case is added to the product recommendation knowledge base, and the index of the product recommendation knowledge base is updated.

4. The product recommendation method based on artificial intelligence as described in claim 1, characterized in that, The step of predicting the non-auto insurance product needs of the target auto insurance user based on the user data and a pre-trained product type prediction model, and generating a list of predicted non-auto insurance product needs for the target auto insurance user, includes: Based on the user data and the pre-trained product type prediction model, the probability of the target auto insurance user's demand for multiple non-auto insurance products is obtained. The product recommendation list for the target auto insurance user is obtained by sorting the multiple non-auto insurance products from high to low according to the probability of demand.

5. The product recommendation method based on artificial intelligence as described in claim 1, characterized in that, The step of obtaining the non-auto insurance product purchase probability of the target auto insurance user based on the user data and a pre-trained purchase probability prediction model includes: User features are extracted from the user data, and user behavior labels are generated using an unsupervised clustering algorithm. Based on the user behavior labels and a pre-trained purchase probability prediction model, the purchase probability of non-auto insurance products for the target auto insurance user is obtained.

6. The product recommendation method based on artificial intelligence as described in claim 1, characterized in that, After obtaining the non-auto insurance product purchase probability of the target auto insurance user based on the user data and the pre-trained purchase probability prediction model, the method further includes: If the purchase probability of the non-motor insurance product is in the second range, then a third personalized recommendation scheme is generated for the target motor insurance user based on the non-motor insurance product with the highest demand probability in the non-motor insurance product demand prediction list, and products are recommended to the target motor insurance user based on the third personalized recommendation scheme. If the probability of purchasing non-motor insurance products falls within the third range, then based on the target motor insurance user's predicted list of non-motor insurance product needs, recommended advertisements for non-motor insurance products will be pushed to the target motor insurance user.

7. The product recommendation method based on artificial intelligence as described in claim 1, characterized in that, Prior to obtaining the user data of the target auto insurance user, the process also included: Obtain user data for all car insurance users, predict the user retention probability of each car insurance user based on a pre-trained user retention probability prediction model, and divide the car insurance users into a first priority user group, a second priority user group, and a third priority user group according to the user retention probability. Target car insurance users are selected sequentially from the first priority user group, the second priority user group, and the third priority user group.

8. A product recommendation device based on artificial intelligence, characterized in that, include: The data acquisition module is used to acquire user data of the target auto insurance user, including personal information data, insurance interaction data, and historical browsing data. The recommendation list acquisition module is used to predict the non-auto insurance product demand of the target auto insurance user based on the user data and a pre-trained product type prediction model, and generate a non-auto insurance product demand prediction list for the target auto insurance user. The purchase probability acquisition module is used to obtain the purchase probability of non-auto insurance products of the target auto insurance user based on the user data and the pre-trained purchase probability prediction model. The product recommendation module is used to generate a first personalized recommendation scheme for the target car insurance user based on the non-car insurance product demand prediction list if the purchase probability of the non-car insurance product is greater than a first preset threshold, and to recommend products to the target car insurance user according to the first personalized recommendation scheme.

9. A computer 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 artificial intelligence-based product 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 executed by a processor, it implements the steps of the artificial intelligence-based product recommendation method as described in any one of claims 1 to 7.

Images