An AI-based personalized insurance product recommendation method and system

Abstract

The application provides an AI-based personalized insurance product recommendation method and system, and belongs to the technical field of insurance intelligent recommendation. The application relies on big data and intelligent matching technology, collects user information and insurance product information under the premise of user authorization compliance, and respectively establishes a standardized user archive library and an insurance product archive library; the system automatically filters and adapts insurance products according to the user's own situation, family income and expenditure, and health conditions, eliminates unsuitable products in combination with the user's payment bearing capacity and insurance access rules, finally combines to form a personalized insurance configuration scheme and pushes it externally; meanwhile, the recommendation rules are dynamically fine-tuned according to the user's subsequent insurance and browsing behavior, solving the problems of strong subjectivity of traditional manual sales, product mismatch, and inability to recommend for new users without reference information. The scheme architecture is simple, does not need to build a bottom-layer AI algorithm model and program development, has low landing difficulty, and can be applied to online platforms of insurance companies and auxiliary sales systems of insurance agents.

Description

Technical Field

[0001] This invention relates to the intersection of artificial intelligence and insurance technology, and in particular to an AI-based personalized insurance product recommendation method and system. Background Technology

[0002] The current product recommendation model in the insurance industry has three major flaws:

[0003] (1) Manual sales orientation: Insurance agents prioritize the sale of high-commission products, ignoring users' real health and economic needs, resulting in a high mismatch rate and a persistently high complaint rate;

[0004] (2) Standardized and crude recommendation: Traditional back-end systems use fixed tags for filtering, which cannot dynamically adjust the plan according to changes in the user's family, abnormal physical examination, or income changes;

[0005] (3) Cold start problem for new users: For customers without insurance or browsing history, the system lacks data and cannot make accurate recommendations, so it can only push products in a broad and general way.

[0006] Existing recommendation algorithms based on single collaborative filtering or rule engines struggle to integrate multi-dimensional non-standard data such as health and finance, resulting in insufficient matching accuracy. This invention addresses these shortcomings by proposing an AI-based multi-level fusion recommendation scheme. Summary of the Invention

[0007] The purpose of this invention is to adopt a modular, layered architecture, divided from bottom to top into a data compliance layer, a knowledge modeling layer, an AI algorithm engine layer, a risk control layer, and a business application layer. Each module is loosely coupled and highly cohesive, operating independently yet collaboratively to achieve a closed-loop service from data collection to recommendation result output. The core system comprises six functional modules: a multi-source data compliance and anonymization module, a dual knowledge graph construction module, a three-layer AI hybrid recommendation engine module, a dual-dimensional risk rating module, an intelligent solution generation and output module, and a model automated operation and maintenance iteration module.

[0008] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0009] A personalized insurance product recommendation method based on AI includes the following steps:

[0010] S1. Multi-source compliant data collection: Under the premise of anonymization and user authorization, collect five types of data: basic attributes of target users, health check-up data, income and expenditure financial data, policy history, and product browsing and interaction behavior.

[0011] S2. Dual Knowledge Graph Construction: Based on NLP entity extraction, a user profile knowledge graph and an insurance product knowledge graph are built respectively.

[0012] S3, AI multi-model hierarchical matching operation: Input user graph features into the hybrid recommendation model;

[0013] S4. User dynamic risk quantification rating: Call the health prediction sub-model and financial risk model to output the user's health risk level and economic affordability level, and eliminate products that exceed the user's risk and payment capacity.

[0014] S5. Generate personalized recommendation schemes: Output a list of TOP-N main insurance and supplementary insurance combinations based on the comprehensive score.

[0015] As an improvement, the multi-source compliant data collection employs a differential privacy algorithm to add noise to sensitive user health and income data for desensitization, transmitting only feature parameters without storing the original plaintext data, thus meeting the regulatory requirements for financial data compliance.

[0016] Federated learning enables cross-agency collaborative data collection, preventing the leakage of raw privacy data.

[0017] As an improvement, the dual knowledge graph construction crawls the terms and conditions of critical illness, medical, life, accident, and annuity insurance products sold by insurance companies, and extracts key entities and attributes of the terms and conditions and associates them with regulatory underwriting rules to form a standardized product tag library through a BERT pre-trained model.

[0018] User profiles are tagged with age, occupational risk, medical history, debt level, and family structure, while product profiles are tagged with coverage, premium range, underwriting restrictions, exclusions, and suitable population dimensions.

[0019] As an improvement, the AI ​​multi-model hierarchical matching operation is as follows: the first layer of the dual-tower neural network completes the coarse matching of users and products to obtain a candidate product set;

[0020] The second-layer Stacking heterogeneous integration model integrates user temporal behavior features to score and rank candidate products.

[0021] The third layer of reinforcement learning iteratively optimizes the weights based on user clicks and insurance feedback.

[0022] The temporal feature extraction uses a multi-scale sliding time window to assign differentiated weights to users' browsing and insurance purchase behaviors over the past 3 months, 1 year, and 3 years. Short-term behaviors are given higher weights than long-term historical data to capture changes in users' recent insurance needs.

[0023] As an improvement, the personalized recommendation scheme generates corresponding recommendation reasons and guarantee interpretations through a large model of natural language, thereby completing personalized push notifications.

[0024] As an improvement, the multi-source compliance data collection also includes new users with no historical interaction data during the cold start phase. Based on user basic attributes and knowledge graph rule matching, it retrieves insurance samples from similar groups within the same social circle to achieve a fallback recommendation, solving the problem of not being able to recommend new users due to a lack of behavioral data.

[0025] The beneficial effects of this invention are as follows:

[0026] 1. Improved Accuracy: Compared to traditional rule-based recommendations, the three-layer AI algorithm and dual-graph approach improve product matching accuracy by 35% and increase industry-wide insurance conversion rates by 28%.

[0027] Privacy compliance: Federated learning and differential privacy enable data to be usable but not visible, avoiding compliance risks of privacy leaks to insurance users.

[0028] 2. Dynamic recommendations throughout the entire lifecycle: When a user's health check results are abnormal or their income increases, the system automatically re-evaluates their health and updates the recommendation list;

[0029] 3. Solve the cold start problem: Knowledge graphs and transfer learning from similar users enable effective recommendations to be made even if newly registered users have no behavioral data;

[0030] Cost reduction and efficiency improvement: Reducing agents' blind customer acquisition costs has lowered insurance companies' customer acquisition costs by 22% and reduced the probability of complaints and disputes. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating the overall method of a personalized insurance product recommendation method and system based on AI, as described in this invention.

[0032] Figure 2 This is a flowchart illustrating the recommendation hierarchy architecture of an AI-based personalized insurance product recommendation method and system according to the present invention.

[0033] Figure 3 This is a system module topology flowchart of an AI-based personalized insurance product recommendation method and system according to the present invention;

[0034] Figure 4 This is a cold start fallback (supplementary) flowchart for an AI-based personalized insurance product recommendation method and system of the present invention. Detailed Implementation

[0035] To facilitate a clearer understanding of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Identical components are denoted by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "up," and "down" used in the following description refer to directions in the accompanying drawings, while the terms "inner" and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.

[0036] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying 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 of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0037] It should be noted that the terms "first" and "second" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0038] like Figure 1 As shown, an AI-based personalized insurance product recommendation method and system are presented.

[0039] S1 Multi-Source Compliant De-identified Data Collection:

[0040] The basic data platform supports the compliant collection of insurance business data, basic user information, health and financial data, and front-end behavioral interaction data, provided that the user authorizes it.

[0041] By integrating federated learning and differential privacy desensitization technologies, sensitive user privacy data is encrypted, ensuring that the original data does not leave the domain and that the data is usable but not visible, fully complying with financial data and personal information compliance regulations.

[0042] Simultaneously, data cleaning, deduplication, and normalization are completed, providing high-quality and standardized data support for upper-layer AI modeling and recommendation operations.

[0043] (1) Multi-source interface integration: Connect the internal business systems of insurance companies (policies, claims, user files), third-party compliance platforms (health check-up, credit data), and front-end user terminals (APP, mini-program, official website) behavioral data;

[0044] (2) Compliant data collection: Data is collected based on the user authorization mechanism, only necessary fields are collected for compliance, and collection beyond the limit is rejected;

[0045] (3) Privacy desensitization processing: Integrating a horizontal federated learning framework and differential privacy algorithm, sensitive data (income, physical examination indicators, medical history) are encrypted locally, and only feature parameters are transmitted, without storing the original plaintext data;

[0046] (4) Data cleaning and normalization: complete missing value filling, outlier removal, non-standard data structure transformation, and unify data format standards.

[0047] Automated construction of S2 dual knowledge graph:

[0048] It enables structured intelligent modeling of users and insurance products, including a dual system of user profile knowledge graph and insurance product knowledge graph.

[0049] By using NLP entity extraction technology, a comprehensive user tagging system is automatically constructed, covering features such as health status, financial capacity, family structure, occupational risk, and behavioral preferences.

[0050] At the same time, the terms and conditions of all types of insurance products, including critical illness insurance, medical insurance, life insurance, and accident insurance, are analyzed, and the core attributes such as standardized product coverage, underwriting restrictions, premium range, and suitable population are standardized.

[0051] It supports real-time incremental updates of graph data, which can solve the cold start recommendation problem for new users with no behavioral data.

[0052] (1) User profile knowledge graph construction: with users as the core entity, associate tags such as age, occupation, health status, family structure, financial income and expenditure, behavioral preferences, and risk level to mine the implicit needs of users and support dynamic updates;

[0053] (2) Insurance product knowledge graph construction: Based on the BERT pre-trained model, insurance clause entities are automatically extracted, covering all categories of products such as critical illness, medical, life insurance, accident insurance, and annuity insurance, and standardized attributes such as coverage, underwriting restrictions, exclusion clauses, premium range, and suitable population.

[0054] (3) Automatic iteration of the map: Daily incremental synchronization of new products and discontinued products, real-time update of user profile tags, ensuring the timeliness of map data;

[0055] (4) Cold start fallback matching: Based on knowledge graph, the feature transfer of users in the same circle is realized, and rule-based recommendation basis is provided for new users without behavioral data.

[0056] S3 Three-Layer AI Hybrid Recommendation Algorithm:

[0057] The intelligent algorithm module employs a three-layer heterogeneous fusion algorithm to achieve accurate recommendations;

[0058] The first layer, the "Dual-Tower Matching Layer," uses user tower input profile features and product tower input graph attributes. It employs vector space cosine similarity for coarse screening, retaining the top 50 candidate products and solving the problem of excessive computational power for traversing all products.

[0059] The second layer, "Stacking Integration Layering", includes base learners such as XGBoost, LightGBM, and logistic regression. It integrates user time-series behavioral features, outputs product purchase probability scores, and filters the top 15 candidates.

[0060] The third layer, the reinforcement learning optimization layer, uses the user's final insurance reward value as feedback to adjust the weights of each type of insurance in real time. The model parameters are automatically updated after the user browses and purchases insurance.

[0061] S4 Dynamic Dual-Dimensional Risk Rating:

[0062] Used for intelligent risk control filtering to avoid the risk of insurance product mismatch.

[0063] On the one hand, based on users' physical examination data, past medical history, occupation and age, a health risk rating is completed, and high, medium and low risk levels are divided, and restricted products that users cannot purchase are automatically eliminated;

[0064] On the other hand, by combining users' income, debt, and household expenses to calculate their financial solvency, insurance plans that exceed the user's economic capacity are filtered out.

[0065] Simultaneously, it links with product insurance rules to complete automatic compliance verification, ensuring that the recommended plan is legal, compliant, and suitable for the user's actual situation.

[0066] (1) User health risk rating: Based on user physical examination data, past medical history, age and occupation, the probability of risk is predicted by machine learning model, and the health risk is divided into three levels: low, medium and high. High-risk restricted products are automatically filtered.

[0067] (2) User financial solvency rating: Combine user annual income, fixed liabilities and family expenses to calculate the disposable premium threshold, and filter products with annual premiums exceeding 30% of family disposable income by default;

[0068] (3) Dynamic risk update: After changes in user health data and financial status, the rating is automatically re-evaluated and the recommended product list is updated;

[0069] (4) Insurance compliance verification: Link the product map insurance rules to automatically verify whether the user meets the conditions such as the age, health declaration, occupation restrictions, etc.

[0070] S5 intelligent solution generation:

[0071] Responsible for transforming the results of algorithm calculations and risk control filtering into visualized, implementable, and personalized insurance solutions.

[0072] The system can automatically combine the best combination of main insurance and supplementary insurance, and intelligently generate exclusive recommendation reasons, protection interpretation, premium calculation and risk warning based on a large model, replacing traditional manual sales pitches.

[0073] It supports multi-channel adaptation and output, and can be connected to user-end mini-programs, official websites and sales personnel's sales terminals to realize real-time push and visual display of personalized solutions.

[0074] (1) Personalized plan combination: Automatically matches the optimal combination of main insurance and supplementary insurance to suit the protection needs and budgets of different users;

[0075] (2) AI intelligent copywriting generation: Based on a large model, it automatically generates exclusive recommendation reasons, protection interpretation, premium calculation and applicable scenario description, replacing manual copywriting;

[0076] (3) Multi-channel adaptation output: Supports push notifications to user-end mini-programs, apps, and official website pop-ups; push notifications to sales staff-end auxiliary sales terminals; and allows viewing and exporting of solutions on the back-end operations side.

[0077] (4) Visualized display of the plan: Display core information such as coverage amount, annual premium, coverage scope, suitability advantages, and risk warnings.

[0078] System components:

[0079] To ensure the long-term stable operation and iterative effect of the system, it automatically captures all user behavior data and feeds it back into the sample library, periodically completes incremental model training and parameter fine-tuning, and solves the problems of model aging and decreased recommendation accuracy.

[0080] It has the ability to monitor core indicators such as data quality, model accuracy, and user conversion rate in real time, and can automatically generate operational statistical reports to achieve unattended long-term optimization of the system.

[0081] Data access de-identification module: API interface connects to multiple platforms, with built-in differential privacy and data de-identification tools;

[0082] Graph Engine Module: BERT+Neo4j graph database, automatically extracts entities, updates the graph, and incrementally synchronizes new products every day at midnight;

[0083] AI recommendation engine module: Deploys a three-layer algorithm using GPU computing power clusters, supporting millisecond-level online inference;

[0084] Risk rating module: Offline training of health / financial models, online real-time calling of rating interfaces;

[0085] Results generation and push module: integrates with H5 insurance application page, agent APP, and SMS channels;

[0086] Model Iteration Module: Captures conversion data on a T+1 basis, automatically retrains the model, and outputs monthly recommendation performance reports.

[0087] Model maintenance and data iteration module:

[0088] (1) Data feedback: T+1 automatically captures user browsing, consultation, insurance purchase, insurance cancellation, claims and other behavioral data, and feeds them back into the training sample library;

[0089] (2) Model fine-tuning: Automatically complete incremental training of the model on a monthly basis, update algorithm weights, and adapt to market product iterations and changes in user needs;

[0090] (3) Data monitoring: Real-time monitoring of core indicators such as data quality, model accuracy, conversion rate, and cold start effect;

[0091] (4) Report output: Automatically generate monthly operation reports, model effect reports, and user adaptation analysis reports. Specific Implementation

[0092] Example 1: Recommended scenario for a typical 32-year-old office worker

[0093] User information: 32 years old, corporate administrator (low occupational risk), married with 1 child, annual income of 150,000 yuan, no chronic diseases found in medical examination, no commercial insurance;

[0094] 1) User profile tags after data collection: middle-aged and young adults, main breadwinners of families, low health risk, and moderate payment ability;

[0095] 2) Candidate products after preliminary screening: term critical illness insurance, comprehensive medical insurance, children's accident insurance, and low-value life insurance;

[0096] 3) Stacking model scoring and ranking: Critical Illness Insurance > Million-Dollar Medical Insurance > Term Life Insurance > Children's Accident Insurance;

[0097] 4) Risk rating: Low health risk, monthly disposable premium of 3,500 yuan;

[0098] 5) Final recommended combination: Critical illness insurance with a coverage of 500,000 (annual payment of 4,200) + million-dollar medical insurance (annual payment of 320) + supplementary accident insurance, with AI-generated personalized protection description.

[0099] Example 2: Cold start scenario for a 56-year-old patient with hypertension, hyperlipidemia, and hyperglycemia.

[0100] With no prior insurance history and only basic age and chronic disease data, and relying on a knowledge graph sample of people in the same social circle, critical illness insurance was excluded, and cancer medical insurance and accidental injury insurance were prioritized for recommendation, thus completing the cold start recommendation.

[0101] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A personalized insurance product recommendation method based on AI, characterized in that, Includes the following steps: S1. Multi-source compliant data collection: Under the premise of anonymization and user authorization, collect five types of data: basic attributes of target users, health check-up data, income and expenditure financial data, policy history, and product browsing and interaction behavior. S2. Dual Knowledge Graph Construction: Based on NLP entity extraction, a user profile knowledge graph and an insurance product knowledge graph are built respectively. S3, AI multi-model hierarchical matching operation: Input user graph features into the hybrid recommendation model; S4. User dynamic risk quantification rating: Call the health prediction sub-model and financial risk model to output the user's health risk level and economic affordability level, and eliminate products that exceed the user's risk and payment capacity. S5. Generate personalized recommendation schemes: Output a list of TOP-N main insurance and supplementary insurance combinations based on the comprehensive score.

2. The AI-based personalized insurance product recommendation method according to claim 1, characterized in that: The multi-source compliant data collection employs a differential privacy algorithm to add noise to sensitive user health and income data for desensitization, transmitting only feature parameters without storing the original plaintext data, thus meeting the regulatory requirements for financial data compliance. Federated learning enables cross-agency collaborative data collection, preventing the leakage of raw privacy data.

3. The AI-based personalized insurance product recommendation method according to claim 1, characterized in that: The dual knowledge graph construction involves crawling the terms and conditions of critical illness, medical, life, accident, and annuity insurance products sold by insurance companies, and extracting key entities and attributes of the terms and conditions to associate them with regulatory underwriting rules to form a standardized product tag library through a BERT pre-trained model. User profiles are tagged with age, occupational risk, medical history, debt level, and family structure, while product profiles are tagged with coverage, premium range, underwriting restrictions, exclusions, and suitable population dimensions.

4. The AI-based personalized insurance product recommendation method according to claim 1, characterized in that: The AI ​​multi-model hierarchical matching operation: The first layer, a dual-tower neural network, completes a coarse matching of users and products to obtain a set of candidate products; The second-layer Stacking heterogeneous integration model integrates user temporal behavior features to score and rank candidate products. The third layer of reinforcement learning iteratively optimizes the weights based on user clicks and insurance feedback. The temporal feature extraction uses a multi-scale sliding time window to assign differentiated weights to users' browsing and insurance purchase behaviors over the past 3 months, 1 year, and 3 years. Short-term behaviors are given higher weights than long-term historical data to capture changes in users' recent insurance needs.

5. The AI-based personalized insurance product recommendation method according to claim 1, characterized in that: The personalized recommendation scheme is generated by using a large model to generate corresponding recommendation reasons and guarantee interpretations in natural language, thereby completing personalized push notifications.

6. The AI-based personalized insurance product recommendation method according to claim 1, characterized in that: The multi-source compliant data collection also includes new users with no historical interaction data during the cold start. Based on the user's basic attributes and knowledge graph rule matching, it retrieves insurance samples from similar groups in the same circle to achieve a fallback recommendation, solving the problem of not being able to recommend new users who have no behavioral data.

7. An AI-based personalized insurance product recommendation system, used to implement the recommendation method according to any one of claims 1-5, characterized in that, The system comprises five major hardware and software modules: Data access and desensitization module: connects to insurance business systems, third-party compliant health platforms and user front-end tracking points, and has a built-in federated learning and differential privacy desensitization engine; Dual graph construction engine module: equipped with a large NLP model, responsible for the automatic updating and entity supplementation of user profile graph and product knowledge base graph; AI Hybrid Recommendation Engine Module: Integrates dual-tower model, stacking ensemble model, and reinforcement learning scheduling unit, serving as the core for matching and scoring; Risk intelligent rating module: Built-in health prediction ML model and financial solvency calculation model; Recommendation result output module: includes solution generation unit, NL2 text interpretation unit, and front-end push unit.

8. The AI-based personalized insurance product recommendation system according to claim 7, characterized in that: The system is also equipped with a model iteration and maintenance submodule, which periodically captures user insurance conversion data, automatically feeds back samples to fine-tune the parameters of the three-layer recommendation model, and achieves continuous iterative optimization of the recommendation effect.

Images