A method and apparatus for generating personalized insurance business data

Abstract

This application discloses a method and apparatus for generating personalized insurance business data, relating to the fields of big data and artificial intelligence technology, and addresses the technical problem that existing recommendation systems lack the ability to deeply understand and dynamically model multi-source heterogeneous data. The method includes: collecting multi-source heterogeneous user data; performing preprocessing and feature extraction to construct user feature vectors; using AI algorithms for dynamic modeling to generate user profiles and risk scores; generating a recommendation list based on the dynamic profile using a recommendation algorithm incorporating reinforcement learning; calling a generative AI model to generate natural language recommendation reasons; and updating model parameters based on user interaction behavior data feedback. Through multi-source data fusion, AI dynamic modeling, reinforcement learning to optimize recommendations, generative AI to enhance explanations, and closed-loop feedback updates, accurate, personalized, and interpretable intelligent recommendations for insurance products are achieved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method and apparatus for generating personalized insurance business data based on big data and artificial intelligence analysis. Background Technology

[0002] Existing insurance product recommendation systems generally rely on static customer questionnaires or limited historical insurance data, primarily employing rule-based matching for product recommendations. These systems are characterized by low implementation costs but offer limited personalization and struggle to cover diverse customer risk needs. In practice, customers spend significant time browsing product terms, and the recommendations often do not align with their risk profiles, impacting customer satisfaction and conversion rates.

[0003] Meanwhile, traditional recommendation methods mostly employ content-based filtering or collaborative filtering algorithms, lacking the ability to deeply model user behavior. Furthermore, the data sources used by recommendation systems are limited to insurance purchase behavior, partial health data, or sales history, failing to form a complete customer profile, resulting in limited accuracy and adaptability of the recommendation model.

[0004] Existing insurance recommendation systems suffer from several shortcomings in multi-source heterogeneous data environments. These shortcomings include a lack of cross-modal semantic understanding capabilities, a lack of ability to perceive the dynamic evolution of user risk status, and a lack of ability to model the causal relationship between recommendation strategies and risk changes. Specifically, these shortcomings manifest in the following technical deficiencies:

[0005] Data semantic understanding dimension: Existing systems cannot understand the deep semantic relationships between heterogeneous data. Existing systems treat wearable device step counts, medical examination indicators, and social emotional text as independent feature fields; lacking cross-modal semantic understanding capabilities, they cannot recognize the deep semantic meaning of "increased health risk" jointly pointed to by "decreased step count + abnormal heart rate + negative social emotions", resulting in the dimensionality reduction of high-dimensional data and serious information loss.

[0006] Dynamic modeling dimension: Existing model parameters are fixed and cannot capture the implicit evolution of user status. Traditional models (logistic regression, collaborative filtering) have fixed parameters after training. When user behavior patterns change gradually (such as a decrease in exercise volume for three consecutive months), the model cannot perceive this implicit risk evolution. It can only update the record when the user actively purchases insurance or suffers an accident. The recommendations are seriously lagging behind the user's true status.

[0007] Strategic Collaboration Dimension: Existing recommendation mechanisms cannot establish a dynamic balance between "utilizing known preferences" and "exploring potential risks." Collaborative filtering based on historical behavior can only recommend products that users have already shown interest in, and cannot proactively identify risks that users are not yet aware of (such as the risk of serious illness in a sub-healthy state), resulting in high-risk users being recommended low-risk products, creating a "risk mismatch."

[0008] Causal reasoning dimension: Existing systems can only discover correlations, but cannot identify the reasons for changes in risk. Existing recommendation systems are based on feature correlations (e.g., older age → recommending critical illness insurance), lacking causal inference capabilities, and cannot answer "why the user's risk status has changed." This results in recommendation reasons that can only describe "what" but cannot explain "why," affecting user trust.

[0009] From an objective evaluation of existing technologies, the following shortcomings exist: the recommendation data dimensions are limited, making it difficult to accurately depict customer profiles; the analysis technology is relatively lagging, failing to fully utilize AI technologies such as machine learning and deep learning for dynamic modeling; the recommendation algorithm rules are simple, with weak generalization ability, making it unable to adapt to changes in user needs; and there is a lack of high-performance data processing platforms, which cannot support the real-time analysis of large-scale, multi-source heterogeneous data. Summary of the Invention

[0010] This application proposes a method and apparatus for generating personalized insurance business data, which solves the technical problem that existing insurance recommendation systems lack the ability to deeply understand and dynamically model multi-source heterogeneous user data, resulting in the system's inability to automatically generate insurance business data based on changes in user status data.

[0011] In a first aspect, embodiments of this application propose a method for generating personalized insurance business data, comprising the following steps:

[0012] The system collects multi-source heterogeneous data from users through a data access cluster. This multi-source heterogeneous data includes wearable device data, medical examination data, consumer behavior data, and social text data.

[0013] The collected multi-source heterogeneous data is cleaned, standardized, and feature-extracted to construct user feature vectors;

[0014] AI algorithms are used to model user feature vectors to generate dynamic user profiles and risk scores. The AI ​​algorithms include clustering algorithms, classification algorithms, and sequence modeling algorithms.

[0015] Based on dynamic user profiles and risk scores, a recommendation algorithm incorporating reinforcement learning is used to generate a personalized insurance product recommendation list.

[0016] The generative AI model is invoked to generate natural language recommendation reasons based on the recommendation list and user profile;

[0017] In response to interaction data on the recommendation list and / or the recommendation reasons, the model parameters of the AI ​​algorithm and the recommendation algorithm are updated.

[0018] Optionally, the acquisition of multi-source heterogeneous data further includes: converting different formats of data from different data sources into a unified internal standard format through a protocol conversion module, and achieving high-concurrency data access through a message queue module of the data access cluster.

[0019] Optionally, preprocessing of the collected multi-source heterogeneous data may further include: using the KNN algorithm to fill in missing values, using the Z-score method to standardize the data, and / or using the Local Outlier Factor (LOF) algorithm to detect and remove outliers.

[0020] Optionally, modeling user feature vectors using AI algorithms may further include one or more of the following: using K-means clustering algorithm to classify users into different health risk levels; using logistic regression algorithm to predict the probability of users experiencing future risks; and using RNN sequence model to perform time-series modeling of user behavior sequences.

[0021] Optionally, generating a personalized insurance product recommendation list using a recommendation algorithm that integrates reinforcement learning further includes: employing collaborative filtering and content recommendation for multi-path recall, and introducing a reinforcement learning DQN algorithm to dynamically optimize the recommendation list.

[0022] Optionally, the process of calling a generative AI model to generate natural language recommendation reasons further includes: using a BERT fine-tuning model to identify the user's intent in the natural language input, and calling a GPT generative pre-trained model to generate personalized semantic recommendation explanations based on the user profile and recommendation results.

[0023] Optionally, updating the model parameters of the AI ​​algorithm and recommendation algorithm further includes: collecting user operation behavior data in real time and using a reinforcement learning model to update the strategy parameters online based on user feedback; and / or, adding the behavior data from the previous collection period to the training set in an offline state and periodically updating the model.

[0024] Secondly, embodiments of this application also provide a personalized insurance business data generation apparatus for implementing the method described in any of the above-mentioned embodiments, comprising:

[0025] The data acquisition module is used to collect multi-source heterogeneous data from users, including wearable device data, medical examination data, consumer behavior data, and social text data.

[0026] The data processing module is used to clean, standardize, and extract features from the collected multi-source heterogeneous data to construct user feature vectors;

[0027] The analysis engine module is used to model user feature vectors using AI algorithms to generate dynamic user profiles and risk scores. The AI ​​algorithms include clustering algorithms, classification algorithms, and sequence modeling algorithms.

[0028] The recommendation engine module is used to generate a personalized insurance product recommendation list based on dynamic user profiles and risk scores, using a recommendation algorithm that integrates reinforcement learning.

[0029] The user interface module is used to display recommendation results and receive user interactions, including calling a generative AI model to generate natural language recommendation reasons based on the recommendation list and user profile, and collecting user interaction behavior data on the recommendation results;

[0030] The feedback update module is used to update the model parameters of the analysis engine module and the recommendation engine module in response to the interaction behavior data of the recommendation list and / or the recommendation reasons.

[0031] Thirdly, embodiments of this application also provide a personalized insurance business data generation system based on big data and artificial intelligence analysis, including:

[0032] The database layer is used to store heterogeneous data from multiple sources, including distributed file systems, columnar databases, relational databases, graph databases, vector databases, and cache databases.

[0033] The server layer is equipped with a personalized insurance business data generation device as described in the second aspect embodiment, or is configured to perform the steps of the personalized insurance business data generation method as described in the first aspect embodiment. The server layer is functionally divided into a data access cluster, a data processing cluster, an AI modeling cluster, a recommendation engine cluster, a natural language cluster, and a business service cluster. Each cluster makes service calls through an API gateway.

[0034] The terminal layer includes user terminals, agent terminals, management terminals, and / or development terminals. Each terminal interacts with the server layer through an API gateway to perform one or more of the following tasks: receiving recommendation results, displaying recommendation reasons, collecting user feedback, and implementing manual intervention and operational management.

[0035] Fourthly, embodiments of this application also provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the personalized insurance business data generation method as described in any of the preceding claims.

[0036] Fifthly, embodiments of this application also provide a processor for running a computer program, wherein the computer program, when running, performs the steps of the personalized insurance business data generation method as described in any of the preceding claims.

[0037] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects:

[0038] This invention achieves end-to-end intelligence in the recommendation process by introducing multi-source heterogeneous data, combining AI technology for feature modeling, dynamic recommendation, and semantic interpretation, significantly improving the following aspects:

[0039] Recommendation accuracy: High-dimensional profiling and risk assessment models effectively improve the relevance of insurance matching;

[0040] Personalized recommendations: AI models learn individual behaviors and interests to make customized recommendations;

[0041] Fast response time: Utilizes a big data distributed framework to quickly process terabytes of data;

[0042] High interpretability: Generative models and graph neural networks enhance the understandability of recommended content;

[0043] Privacy protection: We employ federated learning and data encryption strategies to enhance modeling capabilities while ensuring security.

[0044] Experiments revealed a significant improvement in conversion rates after using AI-enhanced recommendations. A / B testing showed that GPT explanations increased user click-through rates and conversion rates.

[0045] This invention achieves end-to-end intelligent recommendation processes by introducing multi-source heterogeneous data, combining AI technology for feature modeling, dynamic recommendation, and semantic interpretation. Regarding accuracy, high-dimensional profiling and risk assessment models effectively improve the relevance of insurance matching; for personalization, AI models learn individual behaviors and interests to provide customized recommendations; it boasts fast response speed, rapidly processing terabytes of data using a big data distributed framework; it offers strong interpretability, with generative models and graph neural networks enhancing the understandability of recommended content; and for privacy protection, federated learning and data encryption strategies are employed to improve modeling capabilities while ensuring security. Attached Figure Description

[0046] 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:

[0047] Figure 1 A flowchart illustrating an embodiment of the personalized insurance business data generation method of this application;

[0048] Figure 2 This is a flowchart of another embodiment of the method of this application, which includes data acquisition and preprocessing steps;

[0049] Figure 3 This is a flowchart of another embodiment of the method of this application, which includes dynamic modeling and recommendation generation steps;

[0050] Figure 4 This is a flowchart of another embodiment of the method of this application, which includes feedback update and closed-loop optimization steps;

[0051] Figure 5 This is a structural diagram of an embodiment of the personalized insurance business data generation device of this application;

[0052] Figure 6 This is an example architecture diagram of the personalized insurance business data generation system of this application. Detailed Implementation

[0053] 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.

[0054] Through embodiments of these methods, apparatuses, systems, processors, and storage devices, the present invention provides a system for intelligent recommendation of personalized insurance products based on big data and artificial intelligence analysis. It adopts a distributed microservice architecture and achieves end-to-end personalized intelligent recommendation of insurance products through multi-source heterogeneous data fusion collection, AI-driven dynamic user profile modeling, intelligent recommendation algorithm engine, and generative AI-enhanced interpretation.

[0055] System overall concept: The system structure is divided into five major modules. The bottom layer relies on distributed computing and storage architecture (such as Apache Hadoop+HBase, Spark Streaming), the middle layer relies on AI modeling capabilities, and the upper layer provides recommendation API interfaces with a microservice architecture.

[0056] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0057] Example 1: Personalized Insurance Business Data Generation Method

[0058] Figure 1 This is a flowchart illustrating an embodiment of the personalized insurance business data generation method of this application. Figure 1 As shown, the method includes the following steps 110-160.

[0059] Step 110: Collect multi-source heterogeneous data from users through the data access cluster.

[0060] This step involves the collection of multi-source heterogeneous data. This multi-source heterogeneous data includes wearable device data, medical examination data, consumer behavior data, and social text data. For example, the collection dimensions include:

[0061] User demographic attributes: age, gender, income, family structure;

[0062] Lifestyle behaviors: Daily average steps, sleep duration, and heart rate fluctuations are collected via smart bracelets;

[0063] Medical data: such as BMI, past medical history;

[0064] Consumer preferences: frequency of product categories purchased and amount spent on e-commerce platforms;

[0065] Social data: keyword emotion recognition and health attitude modeling.

[0066] Real-world case studies show that:

[0067] Based on a sample of 300,000 users, 85% of users are active on WeChat and Alipay, and more than 65% of users have linked their accounts to health-related apps;

[0068] The system collects more than 100GB of data per day and stably supports data access with a QPS of over 2000.

[0069] In one embodiment, step 110 further includes sub-steps 111-112, such as... Figure 2 As shown.

[0070] Figure 2 This is a flowchart of another embodiment of the method of this application, which includes data acquisition and preprocessing steps, including:

[0071] Step 111: Use the protocol conversion module to convert data of different formats from different data sources into the system's internal standard format.

[0072] For example, this step achieves protocol conversion through the protocol conversion module.

[0073] Specifically, the API interface is used to connect to the following data sources:

[0074] Wearable devices (such as Apple Watch, Huawei Band): collect step count, heart rate, and sleep quality;

[0075] Health apps (such as Keep and Codoon): collect exercise frequency and duration;

[0076] Medical examination data: provided by the hospital or a third-party testing center, in PDF / HL7 format;

[0077] Data update frequency: Automatically synchronized once a day, and collected in real time after user authorization.

[0078] Step 112: Implement high-concurrency data access through the message queue module.

[0079] This step enables message queue access. Preferably, message queue technologies such as Kafka are used to buffer and smooth data spikes, ensuring no data loss. It supports parallel processing by multiple consumers, processing over 100GB of data per day and supporting high-concurrency access with QPS of 2000+.

[0080] Step 120: Clean, standardize, and extract features from the collected multi-source heterogeneous data to construct user feature vectors.

[0081] This step involves data preprocessing and feature engineering. In the preferred embodiment of this application, SparkSQL+MLlib is specifically used to perform data preprocessing.

[0082] This step 120 further includes sub-steps 121-123, such as... Figure 2 As shown.

[0083] Step 121: Missing value handling. Preferably, the KNN algorithm is used to fill in missing values.

[0084] Data integrity is improved by finding the K nearest neighbors that are most similar to the missing sample and using their mean or median to fill in the missing data.

[0085] Step 122: Data standardization. Preferably, the Z-score method is used for data standardization.

[0086] The calculation formula is: This converts features of different dimensions into a unified scale, making model training easier.

[0087] Step 123: Outlier Detection. Preferably, the Local Outlier Factor (LOF) algorithm is used to detect and remove outliers.

[0088] The calculation formula is:

[0089]

[0090] Used to identify potentially erroneous data collection values ​​(such as steps exceeding 100,000 steps / day).

[0091] The cleaning process also includes removing records with a missing rate of more than 20%.

[0092] Step 130: Use AI algorithms to model user feature vectors and generate dynamic user profiles and risk scores.

[0093] This step enables dynamic modeling of user risk status. Preferably, the AI ​​algorithm includes clustering algorithms, classification algorithms, and sequence modeling algorithms.

[0094] This step 130 further includes sub-steps 131-133, such as... Figure 3 As shown.

[0095] Figure 3 This is a flowchart of another embodiment of the method of this application, which includes dynamic modeling and recommendation generation steps. Figure 3 As shown, step 130 further includes:

[0096] Step 131, Cluster Modeling

[0097] Preferably, the K-means clustering algorithm is used to classify users into different health risk levels.

[0098] The objective function of K-means is:

[0099] Specifically, customers are clustered into three categories: high health and activity level; medium health risk; and high risk requiring protection.

[0100] The model was trained using 3 million historical physical examination data points, and the SSE error convergence criterion was used as the convergence condition.

[0101] Step 132, Classification Modeling

[0102] Preferably, a logistic regression algorithm is used to predict the probability of a user experiencing an accident in the future.

[0103] The scoring model uses logistic regression to predict the probability of an accident in the next cycle:

[0104]

[0105] Feature dimensions: 15, including number of accidents, vehicle age, customer loyalty score, etc.

[0106] Based on the probability of claim prediction using regression analysis, users are categorized as follows:

[0107] High risk: P > 0.6;

[0108] Medium risk: 0.3 < P < 0.6;

[0109] Low risk: P < 0.3.

[0110] Step 133, Sequence Modeling

[0111] Preferably, an RNN sequence model is used to perform time-series modeling of user behavior sequences.

[0112] Sequence modeling of customer semantic behavior using RNN networks: Capture the time-dependent relationships of user behavior and perceive the evolution of hidden risks.

[0113] Step 140: Based on dynamic user profiles and risk scores, a personalized insurance product recommendation list is generated using a recommendation algorithm that integrates reinforcement learning.

[0114] This step implements the intelligent recommendation strategy generation. Specifically, the recommendation engine module comprehensively employs multiple recommendation algorithms:

[0115] Collaborative filtering (CF);

[0116] Content-based recommendations;

[0117] A hybrid recommendation model is used, which combines reinforcement learning (DQN) to optimize the recommendation sequence.

[0118] This step 140 further includes sub-steps 141-142, such as... Figure 3 As shown.

[0119] Step 141, Multi-channel recall

[0120] Preferably, this application employs collaborative filtering and content recommendation for multi-channel recall.

[0121] The collaborative filtering algorithm is represented as:

[0122] Content recommendation similarity is calculated as follows: This leads to a hybrid recommendation model.

[0123] Step 142, Reinforcement Learning Optimization

[0124] Preferably, this application introduces the reinforcement learning DQN algorithm to dynamically optimize the recommendation list.

[0125] By combining reinforcement learning (DQN) to optimize recommendation sequences and adaptively adjusting recommendation strategies based on real-time user feedback, the co-evolution of user potential risk perception and recommendation strategies can be achieved.

[0126] Recommended strategies vary depending on the risk level:

[0127] Recommended for high-risk users: Comprehensive insurance package: Compulsory traffic accident liability insurance + vehicle damage insurance + third-party liability insurance (coverage of 1 million RMB) + no deductible;

[0128] Recommended for medium-risk users: Standard combination: Compulsory traffic accident liability insurance + vehicle damage insurance + third-party liability insurance (coverage of 500,000 yuan);

[0129] Recommended for low-risk users: Simplified combination: Compulsory traffic accident liability insurance + third-party liability insurance (coverage of 300,000 yuan).

[0130] Figure 3 This demonstrates a hybrid workflow structure with parallel processing preceding serial processing:

[0131] Phase 1 (Parallel Processing): Steps 131-133 are executed in parallel. The three modeling steps occur simultaneously and independently, each characterizing user features from different dimensions: clustering modeling outputs risk level labels, classification modeling outputs risk probability scores, and sequence modeling outputs behavioral temporal features. Together, they constitute a complete dynamic user profile.

[0132] The second stage (serial processing): Steps 141 and 142 are executed sequentially. Step 141 (multi-path recall) receives the output results of the three parallel steps and fuses multi-dimensional features to generate a candidate product list; Step 142 (reinforcement learning optimization) based on Step 141, dynamically optimizes and sorts the candidate list to generate the final recommendation result.

[0133] Step 150: Call the generative AI model to generate natural language recommendation reasons based on the recommendation list and user profile.

[0134] This step enables the generation of interpretable recommendations. More preferably, step 150 includes sub-steps 151-152.

[0135] Step 151, Intent Recognition

[0136] Preferably, this application uses the BERT fine-tuning model to identify the intent in the user's natural language input.

[0137] For example: Fine-tune the intent recognition module using the pre-trained language model BERT to recognize the following types of customer intent:

[0138] "I want to know about retirement security";

[0139] "I just had a baby and I'd like to buy some insurance."

[0140] Output structured intent information and match product categories: pension insurance, education savings insurance, whole life insurance, etc.

[0141] Step 152, Text Generation

[0142] Preferably, this application calls the GPT generative pre-trained model to generate personalized semantic recommendation explanations based on user profiles and recommendation results.

[0143] Generative AI (such as GPT) can be introduced to generate recommended explanations for users, for example:

[0144] "Since your average number of steps in the past 30 days is 3400 and your BMI is high, the system suggests that you focus on the XX critical illness insurance and chronic disease management combination product."

[0145] This transforms recommendation results from "black box output" to "understandable expression".

[0146] Step 160: In response to the interaction behavior data of the recommendation list and / or the recommendation reasons, update the model parameters of the AI ​​algorithm and the recommendation algorithm.

[0147] This step implements feedback updates and closed-loop optimization. Preferably, this step 160 further includes sub-steps 161-163, such as... Figure 4 As shown.

[0148] Figure 4 This is a flowchart of another embodiment of the method of this application, which includes feedback update and closed-loop optimization steps. It includes:

[0149] Step 161: Real-time feedback collection

[0150] By collecting user action data in real time, user clicks, browsing, and order placement behaviors are reported to the message queue in real time, forming a behavior feedback data stream.

[0151] Step 162, Online Update

[0152] The reinforcement learning model updates policy parameters online based on user feedback.

[0153] The reinforcement learning model calculates reward values ​​based on user feedback, updates the policy network, and achieves dynamic optimization of the recommendation policy.

[0154] Step 163, Offline Update

[0155] In offline mode, behavioral data from the previous collection period is added to the training set, and the model is updated periodically.

[0156] For example, if the period is daily, then the offline training on the next day will add the previous day's behavioral data to the training set, forming a closed-loop optimization of data-model-recommendation.

[0157] Example 2: Application Scenario of Personalized Recommendations for Health Insurance Products

[0158] In a specific application scenario, the method of this application is applied to a personalized recommendation system for health insurance products. The target users are urban professionals aged 25 to 45, and the goal is to recommend health insurance products that are suitable for their health conditions and lifestyles.

[0159] Step 210: Multi-source data acquisition (corresponding to steps 110-112)

[0160] Connect to the following data source using the API interface:

[0161] Wearable devices (such as Apple Watch, Huawei Band): collect step count, heart rate, and sleep quality;

[0162] Health apps (such as Keep and Codoon): collect exercise frequency and duration;

[0163] Medical examination data: provided by the hospital or a third-party testing center, in PDF / HL7 format;

[0164] Data update frequency: Automatically synchronized once a day, and collected in real time after user authorization.

[0165] Step 220: Data preprocessing (corresponding to steps 120-123)

[0166] Cleaning: Remove records with a missing rate exceeding 20%;

[0167] Standardization: Use Z-score standardization;

[0168] Outlier detection: The Local Outlier Factor (LOF) algorithm is introduced to identify potentially erroneous collected values ​​(such as steps exceeding 100,000 steps / day).

[0169] Step 230: Customer health profile modeling (corresponding to steps 130-131)

[0170] The K-means algorithm was used to cluster customers into 3 classes:

[0171] High level of health and activity;

[0172] Medium-term health risks;

[0173] High risks require protection.

[0174] The model was trained using 3 million historical physical examination data points, and the SSE error convergence criterion was used as the convergence condition.

[0175] Step 240: Recommendation generation and display (corresponding to steps 140-142 and 150-152)

[0176] The system displays recommendation results, scoring them based on factors such as sum insured, coverage, and premium.

[0177] It supports simulated premium calculation and one-click customer service consultation.

[0178] Example 3: Intelligent Recommendation Application Scenarios for Car Insurance Renewal

[0179] In another specific application scenario, the method of this application is applied to an intelligent recommendation system for car insurance renewal, serving existing customers with a history of purchasing car insurance for more than 3 consecutive years, with the goal of optimizing renewal conversion rates and improving user satisfaction.

[0180] Step 310, Data Acquisition (corresponding to steps 110-111)

[0181] Data sources include:

[0182] Vehicle accident records over the years (accident time, claim amount, reason for claim);

[0183] Vehicle status data (vehicle brand, vehicle age, mileage).

[0184] Customer behavior data (insurance purchase time, number of interactions, claims satisfaction);

[0185] Step 320: Risk scoring model construction (corresponding to steps 130 and 132)

[0186] The scoring model uses logistic regression to predict the probability of an accident in the next cycle:

[0187]

[0188] Feature dimensions: 15, including number of accidents, vehicle age, customer loyalty score, etc.

[0189] Step 330, Personalized Recommendation Model (corresponding to steps 140-142)

[0190] Based on the probability of claim prediction using regression analysis, users are categorized as follows:

[0191] High risk: P > 0.6;

[0192] Medium risk: 0.3 < P < 0.6;

[0193] Low risk: P < 0.3.

[0194] Recommended strategy:

[0195] Recommended for high-risk users: Comprehensive insurance package: Compulsory traffic accident liability insurance + vehicle damage insurance + third-party liability insurance (coverage of 1 million RMB) + no deductible;

[0196] Recommended for medium-risk users: Standard combination: Compulsory traffic accident liability insurance + vehicle damage insurance + third-party liability insurance (coverage of 500,000 yuan);

[0197] Recommended for low-risk users: Simplified combination: Compulsory traffic accident liability insurance + third-party liability insurance (coverage of 300,000 yuan).

[0198] Step 340: Optimize the user interface experience (corresponding to steps 150 and 152)

[0199] Dynamic recommendation reasons are displayed, such as: "You have no claims record in the past two years, so we recommend a moderate reduction in your third-party liability insurance coverage." One-click renewal and cost comparison are supported.

[0200] Example 4: Application Scenario of Joint Recommendation for Life Insurance and Pension Planning

[0201] In another specific application scenario, the method proposed in this application is applied to a joint recommendation system for life insurance and pension planning. Combined with AI intent recognition, the target group is people aged 35 and above with family burdens and long-term protection needs, especially customers who are new to insurance.

[0202] Step 410: Natural Language Understanding + Semantic Recognition (corresponding to steps 150-151)

[0203] For example, using the pre-trained language model BERT to fine-tune the intent recognition module, the following types of customer intent can be identified:

[0204] "I want to know about retirement security";

[0205] "I just had a baby and I'd like to buy some insurance."

[0206] Output structured intent information and match product categories: pension insurance, education savings insurance, whole life insurance, etc.

[0207] Step 420: Data fusion modeling (corresponding to steps 120 and 130)

[0208] Customer input: income level, expected retirement age, age of family members, etc.

[0209] Modeling the pension gap (based on retirement at age 60 and survival at age 85):

[0210] Pension gap = (Monthly expenditure × 12) × (Life expectancy - Retirement age) - Existing savings

[0211] Example: Monthly expenses are 8,000 yuan, and there are already 200,000 yuan in savings. The shortfall after retirement is: (8,000 × 12) × (85 - 60) - 200,000 = 2,680,000 yuan.

[0212] Step 430: Construction of the joint recommendation model (corresponding to steps 130, 133, and 140)

[0213] Sequence modeling of customer semantic behavior using RNN networks:

[0214] The model outputs a combination of joint safeguard recommendations, for example:

[0215] Whole life insurance (coverage of 500,000);

[0216] Education insurance (annual premium of 10,000 yuan, 10-year term).

[0217] Annuity insurance (returned annually).

[0218] Step 440: Visualization of Recommendation Results (corresponding to steps 150 and 152)

[0219] Use charts to illustrate the proportion of coverage gaps filled by different types of insurance;

[0220] Recommendation reason: Natural language generation: Combined with the ChatGPT-style text generation module, it enhances user trust and acceptance.

[0221] Experiments show that this method can improve underwriting efficiency and accuracy compared to traditional underwriting methods. At the same time, the analysis of underwriting reports generated by AIGC can help underwriters understand the basis for decision-making and learn professional knowledge, thereby improving their professional level and ability, and has significant commercial application value.

[0222] Example 5: Personalized Insurance Business Data Generation Device

[0223] Figure 5 This is a structural diagram of an embodiment of the personalized insurance business data generation device of this application. Figure 5 As shown, the device 500 includes the following modules:

[0224] The data acquisition module 51 is used to collect multi-source heterogeneous data from users, including wearable device data, medical examination data, consumer behavior data, and social text data.

[0225] This module implements steps 110, 111~112 in the aforementioned method embodiments.

[0226] Furthermore, the data acquisition module 51 includes:

[0227] The protocol conversion unit is used to convert data of different formats from different data sources into the system's internal standard format.

[0228] The message queue unit is used to enable high-concurrency data access.

[0229] The data processing module 52 is used to clean, standardize, and extract features from the collected multi-source heterogeneous data to construct user feature vectors.

[0230] This module implements steps 120, 121-123 in the aforementioned method embodiments.

[0231] Furthermore, the data processing module 52 includes:

[0232] The missing value processing unit uses the KNN algorithm to fill in missing values.

[0233] Standardized units are established using the Z-score method for data standardization.

[0234] The outlier detection unit uses the Local Outlier Factor (LOF) algorithm to detect and remove outliers.

[0235] The analysis engine module 53 is used to model user feature vectors using AI algorithms to generate dynamic user profiles and risk scores. The AI ​​algorithms include clustering algorithms, classification algorithms, and sequence modeling algorithms.

[0236] This module corresponds to steps 130, 131-133 in the aforementioned method embodiments.

[0237] Furthermore, the analysis engine module 53 includes:

[0238] The clustering modeling unit uses the K-means clustering algorithm to classify users into different health risk levels;

[0239] The classification modeling unit uses a logistic regression algorithm to predict the probability of a user's future claims.

[0240] The sequence modeling unit uses an RNN sequence model to perform time-series modeling of user behavior sequences.

[0241] The recommendation engine module 54 is used to generate a personalized insurance product recommendation list based on dynamic user profiles and risk scores, using a recommendation algorithm that integrates reinforcement learning.

[0242] This module implements steps 140, 141-142 in the aforementioned method embodiments.

[0243] Furthermore, the recommendation engine module 54 includes:

[0244] The multi-path recall unit is used to perform multi-path recall using collaborative filtering and content recommendation.

[0245] Reinforcement Learning Optimization Unit: Used to dynamically optimize the recommendation sequence using the reinforcement learning DQN algorithm, and adaptively adjust the recommendation strategy based on real-time user feedback.

[0246] The user interface module 55 is used to display recommendation results and receive user interactions, including calling a generative AI model to generate natural language recommendation reasons based on the recommendation list and user profile, and collecting user interaction behavior data on the recommendation results.

[0247] This module implements steps 150, 151-152 in the aforementioned method embodiments.

[0248] Furthermore, the user interface module 55 includes:

[0249] The intent recognition unit is used to recognize the intent in the user's natural language input using a BERT fine-tuning model.

[0250] The text generation unit is used to call the GPT generative pre-trained model to generate personalized semantic explanations based on user profiles and recommendation results.

[0251] The feedback update module 56 is used to update the model parameters of the analysis engine module and the recommendation engine module in response to the interaction behavior data of the recommendation list and / or the recommendation reasons.

[0252] This module implements steps 160-163 in the aforementioned method embodiments.

[0253] Furthermore, the feedback update module 56 includes:

[0254] The real-time feedback unit is used to collect user click, browsing, and order placement behavior data through the real-time computing module.

[0255] The online update unit is used to update the policy parameters of the reinforcement learning model online based on user feedback;

[0256] The offline update unit is used to add behavioral data from the previous collection period to the training set in an offline state, and periodically update the model.

[0257] Example 6: Computer System

[0258] Figure 6 This is an architecture diagram of an embodiment of the personalized insurance business data generation system of this application. The system 600 adopts a distributed microservice architecture, achieving end-to-end personalized intelligent recommendations for insurance products through multi-source heterogeneous data fusion and collection, AI-driven dynamic user profile modeling, intelligent recommendation algorithm engine, and generative AI-enhanced interpretation. The system consists of a three-layer architecture: a database layer 610, a server layer 620, and a terminal layer 630. Data interaction and business collaboration between the layers are achieved through an API gateway 627.

[0259] Database layer 610 is used to store multi-source heterogeneous data, including:

[0260] Distributed File System 611: Employs HDFS technology to store raw collected data, including massive amounts of heterogeneous data from multiple sources such as wearable device logs, physical examination report PDFs, and social text data;

[0261] Columnar Database 612: Employs HBase technology to store user profile data and behavioral feature matrices, supporting high-concurrency read / write and real-time querying;

[0262] Relational database 613: Employs MySQL technology to store transactional data such as product information, insurance records, and user accounts, ensuring data consistency;

[0263] Graph Database 614: Employs Neo4j technology to store the user-product-risk tag relationship network, supporting multi-hop relationship queries;

[0264] Vector Database 615: Employs Milvus or FAISS technology to store user semantic vectors and product feature vectors, supporting similarity retrieval and semantic matching;

[0265] Cache Database 616: Employs Redis technology to store popular recommendation results and user session information, reducing response latency.

[0266] Preferably, the database layer 610 adopts a hybrid storage architecture, selecting an appropriate storage engine based on different data types and access characteristics.

[0267] Server layer 620 is deployed with the personalized insurance business data generation device 500 as described above, or configured to perform the steps of the personalized insurance business data generation method as described above.

[0268] The server layer 620 is divided into the following clusters according to their functions, and each cluster makes service calls through the API gateway 627:

[0269] Data access cluster 621: responsible for the collection and access of multi-source data, corresponding to the data acquisition module 51 in device 500, and corresponding to steps 110~112.

[0270] Preferably, the data access cluster further includes:

[0271] Data Acquisition and Adaptation Unit: Connects to various data source APIs, supports different protocols such as HTTP / MQTT / HL7, and acquires raw data from wearable devices, hospital physical examination systems, e-commerce platforms, etc.

[0272] Protocol conversion unit: Converts data of different formats into the system's internal standard JSON format;

[0273] Message queue unit: Employs Kafka technology to buffer and smooth data spikes, ensuring no data loss and supporting parallel processing by multiple consumers.

[0274] Data processing cluster 622: responsible for data cleaning, processing and feature engineering, corresponding to data processing module 52 in device 500, and corresponding to steps 120-123.

[0275] Preferably, the data processing cluster further includes:

[0276] Offline ETL unit: Employs Spark technology to process all data in batches every night, generating training samples and feature wide tables;

[0277] Real-time computing unit: Employing Spark Streaming or Flink technology, it processes real-time user behavior data in seconds and updates online features;

[0278] Feature engineering unit: performs feature derivation, feature selection and feature standardization, and outputs the model input feature vector.

[0279] AI modeling cluster 623: responsible for model training and inference services, corresponding to the analysis engine module 53 in device 500, and corresponding to implementation steps 130~133.

[0280] Preferably, the AI ​​modeling cluster further includes:

[0281] Model training unit: Using frameworks such as TensorFlow, PyTorch, and XGBoost, K-means clustering models, logistic regression models, and RNN sequence models are trained on a timed basis;

[0282] Online inference unit: Provides REST / gRPC interface to output model prediction results in real time, including risk score, click-through rate prediction, etc.

[0283] Model Version Management Unit: Responsible for model registration, version control, and canary releases, and supports A / B testing distribution.

[0284] Recommendation engine cluster 624: This is the core decision-making unit of the system, responsible for the execution and optimization of recommendation strategies. It corresponds to recommendation engine module 54 in device 500 and implements steps 140-142.

[0285] Preferably, the recommendation engine cluster further includes:

[0286] Multi-path recall unit: Obtains a list of candidate product IDs from multiple strategies such as collaborative filtering, vector similarity, and popularity supplementation;

[0287] Ranking Prediction Unit: Calls the online inference unit to predict CTR / CVR on the candidate set and outputs a list of products with predicted scores;

[0288] Reordering strategy unit: Performs diversity control, business rule filtering, inserts manual intervention results, and generates the final recommendation list;

[0289] Experimental traffic distribution unit: Distributes traffic to different experimental groups based on user ID, supporting A / B testing comparisons of algorithm strategies.

[0290] Natural Language Cluster 625: Responsible for semantic understanding and text generation, corresponding to some functions of the user interface module 55 in device 500 (intent recognition and text generation), and corresponding to implementation steps 150-152.

[0291] Preferably, the natural language cluster further includes:

[0292] Intent recognition unit: Employs a fine-tuned BERT model to recognize the intent in the user's natural language input, such as "consult about pension" or "learn about critical illness insurance".

[0293] Semantic vector unit: Converts user behavior text into feature vectors and stores them in the vector database 615;

[0294] Text generation unit: Calls the GPT model to generate personalized recommendation reasons based on recommendation results and user profiles.

[0295] Business Service Cluster 626: Responsible for business logic processing and external interface exposure.

[0296] Furthermore, the service cluster further includes:

[0297] User Service Unit: Handles user registration, login, authorization management, and other related business;

[0298] Product Service Unit: Provides functions such as product information query and product listing / delisting;

[0299] Order Service Unit: Handles transaction processes such as insurance order placement and policy inquiry.

[0300] API Gateway 627: Using Nginx or Spring Cloud Gateway technology, it is responsible for request routing, rate limiting, authentication, and load balancing. It is the unified entry point for terminal layer 630 to access server layer 620.

[0301] The terminal layer 630 includes a user terminal 631, an agent terminal 632, a management terminal 633, and a development terminal 634. Each terminal interacts with the server layer 620 through an API gateway 627 to receive recommendation results, display recommendation reasons, collect user feedback, and implement manual intervention and operation management.

[0302] User-side 631: Targeting insurance consumers, using WeChat mini-program / APP / H5 technology.

[0303] Preferably, the user terminal further includes:

[0304] Registration and login module: Supports one-click login via mobile phone number / WeChat, and user registration;

[0305] Authorization Management Module: Authorizes the collection of data from wearable devices and health apps;

[0306] Personal Information Module: Maintains basic information such as age, occupation, and income;

[0307] Policy management module: View purchased policies and receive renewal reminders;

[0308] Homepage Recommendation Module: Displays a personalized list of recommended products;

[0309] Recommendation Reasons Module: Displays personalized explanation statements generated by GPT;

[0310] Product details module: View product terms, premiums, and coverage;

[0311] Text consultation module: Supports natural language input for insurance questions;

[0312] Premium calculation module: Input the sum insured / term and calculate the premium in real time;

[0313] Online insurance application module: Fill in the insurance information and complete the online payment;

[0314] Electronic policy module: policy download and policy verification.

[0315] Seat Terminal 632: Designed for insurance advisors and customer service personnel, using a web-based management backend technology.

[0316] Preferably, the seat end further includes:

[0317] Customer list module: Displays assigned customers and supports filtering by criteria;

[0318] Customer profiling module: View customer health profiles and risk scores;

[0319] Behavior tracking module: View customer browsing and click history;

[0320] Policy view module: View the customer's purchased products and renewal dates;

[0321] Recommendation intervention module: Manually adjust the priority of customer recommendation strategies;

[0322] Product push module: Pushes specified products to the client;

[0323] Script suggestion module: AI generates communication scripts tailored to this customer;

[0324] Follow-up record module: Records phone / WeChat communication content;

[0325] Task reminder module: Set up customer follow-up reminders.

[0326] Management Terminal 633: Designed for product managers and operations personnel, it adopts a web-based management backend technology.

[0327] Preferably, the management terminal further includes:

[0328] Core metrics module: Displays DAU, conversion rate, and average number of recommendations per user;

[0329] Real-time monitoring module: real-time QPS, response latency, and error rate;

[0330] Experiment configuration module: Create a new AB experiment and set the flow rate ratio;

[0331] Version management module: Manages different algorithm versions;

[0332] Effect Comparison Module: Comparison of indicators between the experimental group and the control group;

[0333] Recall strategy configuration module: Enable / disable multiple recall channels;

[0334] Ranking weight configuration module: Adjust feature weights and business rules;

[0335] Product listing module: Add new insurance products and enter terms and conditions.

[0336] Development side 634: Targeting algorithm engineers and operations personnel, it adopts an internal tool platform technology.

[0337] Preferably, the development end further includes:

[0338] Model upload module: Upload the trained model file;

[0339] Version control module: Manages model versions and supports rollback;

[0340] Gray-scale release module: Sets the gray-scale traffic ratio for the model;

[0341] Service monitoring module: View CPU / memory / latency for each service;

[0342] Log query module: Retrieve logs by user ID / time;

[0343] Call chain tracing module: Traces the entire chain of a single request;

[0344] Data Quality Module: View field missing rate and outliers;

[0345] Configuration management module: Modify service configurations and the changes take effect dynamically.

[0346] System working process:

[0347] The system's process for providing personalized insurance recommendations includes the following steps:

[0348] Step S1: Data Acquisition and Access

[0349] After the user completes authorization through the user terminal 631, the data acquisition and adaptation unit of the data access cluster 621 connects to multiple data sources such as wearable devices, hospital physical examination systems, and e-commerce platforms via API interfaces. The raw data is converted into a standard format by the protocol conversion unit and written to the message queue unit.

[0350] Step S2: Data Processing and Profile Building

[0351] The real-time computing unit of the data processing cluster 622 consumes data from the message queue unit and updates real-time user features in seconds; the offline ETL unit processes all data in batches every night to generate training samples. The feature engineering unit outputs feature vectors, user profiles are written to columnar database 612, graph relationships are updated to graph database 614, and semantic vectors are stored in vector database 615.

[0352] Step S3: Model Training and Update

[0353] The model training unit of the AI ​​modeling cluster 623 periodically reads training data from the distributed file system 611 to train K-means clustering models, logistic regression models, and RNN sequence models. After the models pass evaluation, they are released to the online inference unit via the model version management unit.

[0354] Step S4: Real-time Recommendation Service

[0355] A user initiates a recommendation request through client 631, which is routed to business service cluster 626 via API gateway 627. The multi-path recall unit of recommendation engine cluster 624 obtains a candidate set from collaborative filtering, vector similarity, and popularity supplementation. The ranking prediction unit calls the online inference unit to perform CTR prediction, and the re-ranking strategy unit performs diversity control to generate the final recommendation list. The text generation unit of natural language cluster 625 calls the GPT model to generate personalized recommendation reasons, and the results are returned to client 631 for rendering and display via API gateway 627.

[0356] Step S5: Interactive Feedback and Closed-Loop Optimization

[0357] The user terminal (631) reports user clicks, browsing, and order placement behaviors to the message queue unit in real time. The real-time computing unit updates user real-time features, uses a reinforcement learning model to calculate reward values ​​based on user feedback, and updates the policy network. The next day, offline training adds the previous day's behavior data to the training set, forming a closed-loop optimization of data-model-recommendation.

[0358] Step S6: Human Intervention and Operational Management

[0359] The recommendation intervention unit on the agent side 632 allows manual adjustment of customer recommendation strategies; the experiment configuration unit on the management side 633 allows configuration of A / B test traffic; and the development side 634 completes the deployment of new models through the canary release unit.

[0360] Example 7: Computer-readable storage medium

[0361] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the personalized insurance business data generation method as described in any of the preceding embodiments.

[0362] The computer-readable storage medium may be the medium included in the system or apparatus described in any of the foregoing embodiments, or it may be a medium that exists independently and is not assembled into the device.

[0363] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0364] Example 8: Processor

[0365] This application also provides a processor for running a computer program, which, when running, executes the steps of the personalized insurance business data generation method as described in any of the preceding embodiments.

[0366] The processor can be deployed in server-layer data access clusters, data processing clusters, AI modeling clusters, recommendation engine clusters, natural language clusters, or business service clusters to perform parallel processing of computational tasks such as multi-source heterogeneous data acquisition, feature engineering, model inference, recommendation generation, and feedback updates.

[0367] Preferably, the processor adopts a heterogeneous multi-core architecture design, and achieves efficient and optimized execution of the method through hardware-level functional partitioning and dedicated computing units.

[0368] The processor integrates multiple data interface controllers, including a wearable device interface controller supporting Bluetooth 5.3 / BLE and ANT+ protocols, a medical data interface controller integrating an HL7 / FHIR protocol stack hardware acceleration unit, a consumer data interface controller integrating a PCIe 4.0 / DMA controller, and a social text data interface controller integrating a network protocol offloading engine. These hardware-level interface controllers can achieve zero-copy acquisition of multi-source data and reduce CPU interrupt load.

[0369] In terms of core architecture, the processor adopts a heterogeneous multi-core design, divided into multiple core clusters with specialized functions. The first core cluster is the AI ​​inference core, dedicated to accelerating AI model inference, supporting efficient parallel computation of K-means clustering, logistic regression, and RNN sequence models, corresponding to step 130. The second core cluster is the reinforcement learning core, dedicated to updating reinforcement learning policy networks, integrating a DQN algorithm hardware acceleration unit, supporting real-time reward computation and online policy parameter updates, corresponding to steps 140 and 160. The third core cluster is the generative AI core, dedicated to generative AI text generation, integrating a Transformer model hardware acceleration engine, supporting low-latency inference of GPT-like models, corresponding to step 150. The fourth core cluster is the data processing core, dedicated to data preprocessing, integrating SIMD instruction set acceleration for KNN imputation, Z-score normalization, and LOF anomaly detection, corresponding to step 120. The fifth core cluster is the feedback update core, dedicated to the feedback update mechanism, supporting real-time behavioral data acquisition, offline training data aggregation, and model parameter version management, corresponding to step 160.

[0370] To further improve processing efficiency, the processor also integrates multiple dedicated hardware acceleration units. The AI ​​model inference acceleration unit includes a tensor computation engine composed of 2048 MAC units, supporting FP16 / INT8 mixed-precision computation with a peak computing power of 100 TOPS. It also integrates a sparse inference engine and a 32MB on-chip SRAM model cache. The reinforcement learning policy network acceleration unit implements the forward propagation and backward update of the DQN network in a hardware pipeline, supporting 100,000 policy updates per second, and integrates an experience replay buffer controller and a reward computation unit. The generative AI text generation acceleration unit includes a Transformer engine supporting hardware parallel computation of multi-head attention mechanisms, with a single inference latency of less than 5ms. It integrates 512MB of embedded storage for high-frequency word vector caching and is equipped with a beam search hardware accelerator. The data preprocessing acceleration unit integrates a KNN accelerator implemented in hardware using KD-trees, a batch data parallel normalization Z-score computation unit, and a LOF anomaly detection engine that implements fast computation of locally reachable density in hardware.

[0371] At the storage level, the processor employs a multi-level on-chip memory architecture, including a 64KB L1 instruction / data cache per core, a 2MB L2 cache per core cluster, a 32MB on-chip L3 shared cache, 512MB on-chip embedded DRAM, and a high-bandwidth memory controller connecting to off-chip HBM2e memory. This storage hierarchy ensures that frequently accessed model parameters and user profile data can be quickly accessed on-chip, reducing memory access latency. The on-chip interconnect network adopts an 8×8 mesh topology, supporting QoS priority routing and broadcast multicast functions to ensure that real-time recommendation requests obtain low-latency paths.

[0372] The processor can be deployed in the server layer of the personalized insurance business data generation system described in Embodiment Six. Specifically, it includes a multi-source data acquisition interface unit and a fourth core cluster deployed in the data access cluster, a fourth core cluster and a data preprocessing acceleration unit deployed in the data processing cluster, a first core cluster and an AI inference acceleration unit deployed in the AI ​​modeling cluster, a second core cluster and a reinforcement learning acceleration unit deployed in the recommendation engine cluster, a third core cluster and a generative AI acceleration unit deployed in the natural language cluster, and a fifth core cluster and a feedback update acceleration unit deployed in the business service cluster.

[0373] When the processor executes the method described in this application embodiment, its workflow is as follows: The multi-source data acquisition interface unit acquires multi-source data from wearable devices, medical systems, social platforms, etc., in parallel through a dedicated controller, and writes it directly to HBM2e memory via DMA to achieve zero-copy data acquisition; after the fourth core cluster is woken up, the data preprocessing acceleration unit reads the raw data from HBM2e, and performs KNN imputation, Z-score standardization, and LOF anomaly detection in parallel in hardware, and stores the processing results in embedded DRAM; the first core cluster reads the feature vector from the embedded DRAM, and the AI ​​inference acceleration unit performs K-means clustering, logistic regression prediction, etc. in parallel. The RNN sequence modeling generates user profiles and risk scores, which are stored in the L3 shared cache. The second core cluster reads profile data from the L3 shared cache, and the reinforcement learning acceleration unit performs multi-path recall and DQN policy optimization, generating a recommendation list which is also stored in the L3 shared cache. The third core cluster reads the recommendation list from the L3 shared cache, and the generative AI acceleration unit calls the Transformer hardware engine to generate personalized recommendation reasons, with the results stored in embedded DRAM. The fifth core cluster collects user interaction behavior, and the reinforcement learning acceleration unit updates the policy network parameters online according to the reward, while simultaneously writing the behavioral data into the experience replay buffer for offline training.

[0374] The processor provided in this embodiment, through a heterogeneous multi-core architecture and a dedicated hardware acceleration unit, achieves hardware-level optimized execution of the method described in this application embodiment. Compared with general-purpose CPUs, it has the following advantages: hardware pipelined processing significantly reduces end-to-end recommendation latency from hundreds of milliseconds to less than 10 milliseconds; the dedicated acceleration unit supports processing tens of thousands of recommendation requests per second, achieving high throughput processing; hardware acceleration can reduce power consumption by more than 80% compared to general-purpose computing; the hardware-based reinforcement learning policy network supports true real-time online updates; a single chip integrates the entire process of data acquisition, processing, modeling, recommendation, generation, and feedback, making it suitable for edge deployment and end-user applications.

[0375] Other embodiments

[0376] Furthermore, this application also proposes an electronic device (or computing 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 method described in any embodiment of this application.

[0377] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory. Memory may include non-persistent storage in computer-readable media, random access memory (RAM), and / or non-volatile memory such as read-only memory (ROM) or flash RAM.

[0378] General instructions

[0379] It should be noted that the execution subject of each step of the method provided in the embodiments can be the same device, or the method can be executed by different devices. For example, the execution subject of steps 110 and 120 can be device 1, and the execution subject of step 130 can be device 2; or the execution subject of step 110 can be device 1, and the execution subject of steps 120 and 130 can be device 2; and so on.

[0380] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0381] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0382] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0383] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0384] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0385] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0386] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for generating personalized insurance business data, characterized in that, Includes the following steps: The system collects multi-source heterogeneous data from users through a data access cluster. This multi-source heterogeneous data includes wearable device data, medical examination data, consumer behavior data, and social text data. The collected multi-source heterogeneous data is cleaned, standardized, and feature-extracted to construct user feature vectors; AI algorithms are used to model user feature vectors to generate dynamic user profiles and risk scores. The AI ​​algorithms include clustering algorithms, classification algorithms, and sequence modeling algorithms. Based on dynamic user profiles and risk scores, a recommendation algorithm incorporating reinforcement learning is used to generate a personalized insurance product recommendation list. The generative AI model is invoked to generate natural language recommendation reasons based on the recommendation list and user profile; In response to interaction data on the recommendation list and / or the recommendation reasons, the model parameters of the AI ​​algorithm and the recommendation algorithm are updated.

2. The method for generating personalized insurance business data according to claim 1, characterized in that, The acquisition of multi-source heterogeneous data further includes: converting different formats of data from different data sources into the system's internal standard format through a protocol conversion module, and achieving high-concurrency data access through the message queue module of the data access cluster.

3. The method for generating personalized insurance business data according to claim 1, characterized in that, For the collected multi-source heterogeneous data: KNN algorithm is used to fill in missing values, Z-score method is used for data standardization and / or Local Outlier Factor (LOF) algorithm is used to detect and remove outliers.

4. The method for generating personalized insurance business data according to claim 1, characterized in that, Includes one or more of the following: Users are divided into different health risk levels using the K-means clustering algorithm; Logistic regression algorithm is used to predict the probability of a user's future claims. An RNN sequence model is used to perform time-series modeling of user behavior sequences.

5. The method for generating personalized insurance business data according to claim 1, characterized in that, Collaborative filtering and content recommendation are used for multi-path recall, and the reinforcement learning DQN algorithm is introduced to dynamically optimize the recommendation list.

6. The method for generating personalized insurance business data according to claim 1, characterized in that, The BERT fine-tuning model is used to identify the intent in the user's natural language input, and the GPT generative pre-trained model is called to generate personalized semantic explanations based on user profiles and recommendation results.

7. The method for generating personalized insurance business data according to claim 1, characterized in that, By collecting user operation behavior data in real time, a reinforcement learning model is used to update the strategy parameters online based on user feedback; and / or, in offline mode, the behavior data from the previous collection period is added to the training set to periodically update the model.

8. A personalized insurance business data generation device, used to implement the method described in any one of claims 1 to 7, characterized in that, include: The data acquisition module is used to collect multi-source heterogeneous data from users, including wearable device data, medical examination data, consumer behavior data, and social text data. The data processing module is used to clean, standardize, and extract features from the collected multi-source heterogeneous data to construct user feature vectors; The analysis engine module is used to model user feature vectors using AI algorithms to generate dynamic user profiles and risk scores. The AI ​​algorithms include clustering algorithms, classification algorithms, and sequence modeling algorithms. The recommendation engine module is used to generate a personalized insurance product recommendation list based on dynamic user profiles and risk scores, using a recommendation algorithm that integrates reinforcement learning. The user interface module is used to display recommendation results and receive user interactions, including calling a generative AI model to generate natural language recommendation reasons based on the recommendation list and user profile, and collecting user interaction behavior data on the recommendation results; The feedback update module is used to update the model parameters of the analysis engine module and the recommendation engine module in response to the interaction behavior data of the recommendation list and / or the recommendation reasons.

9. A personalized insurance business data generation system based on big data and artificial intelligence analysis, characterized in that, include: The database layer is used to store heterogeneous data from multiple sources, including distributed file systems, columnar databases, relational databases, graph databases, vector databases, and cache databases. The server layer is equipped with the personalized insurance business data generation device as described in claim 8, or is configured to perform the steps of the personalized insurance business data generation method as described in any one of claims 1 to 7. The server layer is functionally divided into a data access cluster, a data processing cluster, an AI modeling cluster, a recommendation engine cluster, a natural language cluster, and a business service cluster. Each cluster makes service calls through an API gateway. The terminal layer includes user terminals, agent terminals, management terminals, and / or development terminals. Each terminal interacts with the server layer through an API gateway to perform one or more of the following tasks: receiving recommendation results, displaying recommendation reasons, collecting user feedback, and implementing manual intervention and operational management.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the personalized insurance business data generation method as described in any one of claims 1 to 7.

Images