CN115131101A - A Personalized Intelligent Recommendation System for Insurance Products - Google Patents

Abstract

The invention discloses a personalized intelligent recommendation system for insurance products. First, the training data features need to be integrated. The integrated content is divided into two categories: customer agent feature vector and insurance product feature vector; customer agent feature vector: this part It is divided into two categories, customer information and information about the agent to which the customer belongs; customer information can be subdivided into customer attribute information and customer behavior information; the technical key points of the present invention include adding the feature of the agent dimension to the insurance product recommendation system , adding the factor representing the agent's sales habit into the prediction of the customer's possible potential interest points, so that the agent's marketing habit can be considered, and the recommendation effect can be comprehensively improved.

Description

A Personalized Intelligent Recommendation System for Insurance Products

The invention relates to the technical field of insurance product recommendation systems, in particular to a personalized intelligent recommendation system for insurance products.

Background technique

In the context of the Internet era, the traditional insurance marketing model has been continuously impacted by the Internet and has been greatly challenged. Online, intelligent and customized insurance product sales have become a trend. People have higher expectations for the products and services provided by the insurance business development system. While insurtech is constantly facing challenges, it also brings new opportunities to the development of the smart exhibition industry. How to make better use of insurtech, contextualize the marketing environment, build business development tools to mobilize consumers’ potential needs, reduce the cost of insurance agents’ marketing and business development, comprehensively improve efficiency and improve policy conversion rate, etc. New issues are waiting to be solved .

In recent years, major insurance companies have successively launched a variety of agent development platform tools, which have aggregated multiple functions such as customer management, intelligent double recording, and micro-mall. People exhibition industry empowerment.

On the other hand, with the rise of the Internet industry, intelligent recommendation systems have been widely used. People's demand for personalized information is growing, mining customers' personalized needs and providing more and more accurate services have become a common topic in all walks of life. Traditional recommendation systems are mainly based on user purchase history, product-related attributes and other data, with the help of association rules, content filtering and other algorithms, through the combination of rules and algorithms, to mine potential connections between users and products, and recommend more closely related products to users. To achieve the purpose of intelligent recommendation.

The insurance industry pays attention to providing competitive services. Through the intelligent recommendation system, based on the integration of customer behavior, insurance product information, and agent information, it is an important part of the enterprise to explore the potential needs of customers and recommend insurance products for customers more targetedly. core competitiveness.

The insurance industry is different from the ToC business model of the general Internet industry. The way of business development is usually through ToAtoC, the agent and then the customer. Combined with our own sales habits, we can make customized product recommendations for customers. Therefore, how to effectively integrate the uncertainty factor of the agent into the recommendation system and effectively complete the recommendation task has become a common topic in the industry.

Technical solutions of the prior art:

A recommended method for an insurance product, comprising the following steps:

S1. If the user currently uses the mobile marketing scenario, continue to S2~S3; if the user currently uses the Internet marketing scenario, then continue to S4~S6;

S2. Analyze and evaluate the security gap of the user according to the user's wealth and life cycle; S3. Recommend corresponding insurance products to the user according to the priority of the user's security responsibility and security gap; Tagging processing; S5. Perform clustering, association, and collaborative filtering processing on the tagged user data; S6. Recommend corresponding insurance products to users according to the processed user data and in combination with user scenarios.

In the S2 described in 2, specifically comprise the following steps:

S21. Calculate the user's premium gap according to the user's age, annual income, occupation, and historical protection; S22. According to the user's age, through the ideal package standard, package star ranking, premium expenditure level, and protection level

Evaluate the needs, calculate the user's insurance gap, and divide the user's insurance level requirements.

In the S21 described, the premium gap refers to the premium gap of the ideal package, and the specific calculation method is: premium gap = (the ideal package guaranteed payment ratio - the existing guaranteed payment ratio) × annual income; among which, the ideal package guarantee The payment expenditure ratio refers to: according to the current life cycle of the user, the annual premium expenditure under ideal conditions accounts for the ratio of the annual income; the existing insurance payment expenditure ratio refers to: the insurance products that the user has purchased, the annual premium expenditure accounts for the annual percentage of income.

In the described S22, the hierarchical division of the user's security level requirements needs to be evaluated according to the coverage of the user's security responsibility, the insured amount sufficiency of the personal liability protection, the premium expenditure ratio of the savings type liability protection, and the ideal package protection payment ratio; The adequacy of the insured amount of the personal liability guarantee refers to the proportion of the total insured amount of the user's historical personal liability guarantee to the total insured amount of the ideal package standard guarantee.

In the S22 described, the insured amount gap includes the insured amount gap of personal liability and the insured amount gap of savings liability, and the specific calculation method is: the insured amount gap of personal liability = the total insured amount of ideal package standard guarantee - historical policy coverage The insured amount of the protection liability; the insured amount gap of the savings liability = (the ratio of the savings protection expenditure in the ideal package - the existing savings protection expenditure ratio) × annual income; among which, the total insured amount of the ideal package standard protection refers to: according to In the current life cycle of the user, the sum insured of life insurance coverage, critical illness medical coverage, accident coverage, general medical coverage, and hospitalization subsidy coverage under ideal conditions; the coverage of coverage responsibilities covered by historical policies refers to: the user The sum insured of the purchased life insurance coverage, critical illness medical coverage, accident coverage, general medical coverage, and hospitalization subsidy coverage; the ratio of savings coverage in the ideal package refers to: according to the current life cycle of the user, under ideal circumstances The ratio of the required savings security expenditure to the annual income; the existing savings security expenditure ratio refers to the ratio of the user's existing savings security expenditure to the annual income.

In the described S3, specifically comprise the following steps: S31, according to the user's security responsibility priority, recommend the insurance product corresponding to the high-priority security responsibility to the user, and this insurance product is not purchased by the user before; Described user security Responsibility priority is set and adjusted according to the user's life cycle. S32. If the premium of the recommended high-priority insurance product has filled the premium gap of the user, no other insurance products are recommended; otherwise, the personal liability insurance is recommended to the user according to the insured gap of the user's personal liability Corresponding insurance products; S33. If the premiums of the recommended high-priority insurance products and the insurance products corresponding to personal liability protection have filled the user's premium gap, no other insurance products are recommended; otherwise, according to the user's savings The insured amount gap of similar liability, and recommends insurance products corresponding to savings-type liability protection to users.

In described S4, specifically comprise the following steps:

S41, predefine a user model, the user model is composed of a plurality of labels, each label describes an attribute of the user; S42, collect user data, including: the user's basic information data, the user's historical purchase data, the user's historical behavior data ;S43. Match the user data with various labels in the user model at regular intervals, and place labels on the user data.

In the described S5, specifically comprise the following steps:

S51, using K-means clustering algorithm to perform clustering processing on user data to obtain classified user data;

S52. Use the Apriori algorithm to analyze the historical purchase data in the classified user data, mine the relationship between the insurance product and the historical purchase data, and establish an association rule;

S53 , using the data embedding method to analyze the historical behavior data in the classified user data, mining the collaborative filtering relationship between the insurance product and the historical behavior data, and establishing collaborative filtering rules.

For implementing the recommended method as described in any one of claims 1 to 8;

It includes: the protection gap evaluation module, which is suitable for mobile marketing scenarios, analyzes and evaluates the user's protection gap according to the user's wealth and life cycle; the protection gap analysis module recommends the corresponding insurance to the user according to the priority of the user's protection responsibility and the protection gap product;

It also includes: user tagging module, which tags user data based on a preset user model; data processing module, which performs clustering, association, and collaborative filtering processing on tagged user data; strategy fusion The module recommends corresponding insurance products to users according to the processed user data and in combination with user scenarios.

The said protection gap includes the premium gap and the insured amount gap;

The security gap evaluation module includes: a wealth gap calculation module, which calculates the user's premium gap according to the user's age, annual income, occupation, and historical security; a life cycle gap calculation module, which is based on the user's age. The star ranking of the package, the level of premium expenditure, and the level of protection are evaluated, and the user's insurance gap is calculated.

Disadvantages of the prior art:

First, the information on the sales habits of insurance agents is not considered. Because the insurance industry often conducts business through agents, the sales habits of agents have a decisive impact on customer underwriting behavior. If this information is not considered, the prediction effect will be unsatisfactory and the effect of the recommendation system will be affected.

Second, when the number of customers and insurance products in the recommender system continues to increase, traditional collaborative filtering faces the problem of decreased effectiveness when dealing with data sparseness. When the amount of data continues to increase, the data gradually becomes sparse, the amount of information decreases, and it is difficult to calculate the similarity of customers and insurance products, resulting in unsatisfactory recommendation results.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to provide a personalized intelligent recommendation system for insurance products. The technical key points of the present invention include adding the feature of the agent dimension to the insurance product recommendation system, adding the feature of the agent dimension to the prediction of the possible potential points of interest of customers. The factors that represent the sales habits of the agents make the marketing habits of the agents be considered and comprehensively improve the recommendation effect. In the present invention, the method of matrix data expansion can still be applied under the circumstance that the customer's browsing and purchasing data are relatively coefficients, so as to ensure the quality of recommendation. In view of the above two points, the present invention helps to improve the recommendation effect of the intelligent recommendation scene of insurance products.

In order to solve the technical problem of appeal, the present invention adopts following technical scheme:

A personalized intelligent recommendation system for insurance products:

First, it is necessary to integrate the training data features. The integration content is divided into two categories: customer agent feature vector and insurance product feature vector;

Customer agent feature vector: This part is divided into two categories, customer information and information about the agent to which the customer belongs;

Customer information can be subdivided into customer attribute information and customer behavior information;

The reason for integrating customer agents into feature vectors is that each customer corresponds to a specific agent; when the same customer corresponds to multiple agents, the agent is selected by certain rules or used as a combination of different customer agents;

Insurance product feature vector: This part mainly includes the features related to insurance products, including the structural features related to insurance terms extracted from unstructured data through NLP in insurance terms, and also includes information about insurance product types, starting sales stored in the core system. Daily structured features, which can be queried in the database through the insurance product ID;

After having the customer agent feature vector and the insurance product feature vector, it is necessary to define the customer's score matrix for the insurance product. The score is weighted according to the customer's purchase and browsing behavior of the product; Therefore, the customer purchase weight can be set to 1, and the customer browsing weight can be set to 0.6. Calculate the customer's rating for the product based on the customer's historical purchases and browsing;

The scoring matrix needs to be decomposed into two hidden matrices with k hidden factors in the user and item dimensions by means of singular value decomposition; by multiplying the two hidden matrices, a full rank customer and product matrix can be obtained, thus Complete the matrix expansion to solve the problem that the rating matrix data is too sparse.

Through the above operations, the customer agent characteristic vector, the insurance product characteristic vector and the expanded customer and product rating matrix are obtained;

The next step is to use the above information and use a hybrid collaborative filtering method to obtain user predictions on product ratings;

First, perform a preliminary cluster analysis on the customer on the customer agent feature vector through the KMeans algorithm to form clusters of different customers, and calculate the customer's product score based on the mean of the product scores of all customers who have scored the product in the customer cluster. Obtain score a; similarly, cluster on the product feature vector to obtain product clusters with similar product characteristics, and calculate the average product score based on the product scores in the product cluster by customers, and obtain score b;

For the scoring of the above four cases, a weighting coefficient is introduced, and the weighted summation of the four cases is carried out to obtain the final score; the sum of the weighting coefficients of the four cases is 1; the value range of the coefficient of each case is 0 to 1; the acquisition of the weight The rule is that the final result obtained by the combination can better predict the customer's rating of the product; the weighting coefficient can be obtained through ABTest experiments to find the optimal combination;

The recommendation results calculated by the algorithm given by the algorithm are sorted in descending order according to the product score, and the top k products are screened as the candidate recommendation list; the last step is to add a product filtering module, which is used to prevent the agent from being available to the agent before the recommendation list is provided to the customer. Filter products for sale, products purchased by customers, and products recommended but not clicked by customers. The reason is that different insurance products are often only available to agents of a certain channel, and repeatedly recommending products purchased by customers to customers will cause customers’ disgust. Through the filtering module, the effect of the recommendation system can be further improved, and customer clicks can be increased. Rate.

Through the above steps, the recommendation list is finally generated. According to the business requirements, the top k products are screened as the final recommendation list. The offline model runs batches every day, and the recommendation list corresponding to each customer is calculated. The results are stored in the Redis database. After receiving the front-end request, the online service reads the recommendation list from memory and returns it to the user.

The above-mentioned personalized intelligent recommendation system for insurance products, wherein:

The customer attribute information includes the basic information of the customer's gender, age, and address obtained from the customer's insurance application process; such information is stored in the insurance company's core system database, and the category features are obtained from the database according to the customer ID and added to the feature vector;

Matching between customer behavior data and customer attribute data Matches the personal information left by the customer when they purchase the policy to identify the same person, and then associate;

Another important piece of information of the customer agent feature vector is the relevant characteristics of the agent to which the customer belongs; the agent to which the customer belongs refers to which agent the customer buys the policy from, then the customer belongs to the agent; the agent characteristics include the basic characteristics of the agent. Information, this information is linked to the agent ID through the customer's insurance policy, and can be queried in the core system; at the same time, the agent information also includes the agent's historical order information, including the types of products that the agent is good at selling. The ordering situation in the time dimension such as year, including the statistical information of the customer dimension that the agent has issued orders, including the average age of the customer, the distribution of customer gender, and the customer surrender rate.

The above-mentioned personalized intelligent recommendation system for insurance products, wherein:

The customer behavior data comes from the front-end embedding point of the former app; the online mall of the insurance company and the app client of the WeChat Store will have a user interface. The user interface displays the insurance products, insurance information, insurance services and other items of the insurance company to the user; the customer browses, purchases, and makes an appointment by clicking on the items of interest; through the front-end embedding, the insurance company collects the behavior information generated by the user during the browsing process , including clicking, searching, browsing duration, favorites, forwarding, and sharing; front-end tracking information is uniformly collected through tracking tools and stored in a designated database; the recommendation system accesses user behavior data in the database through user IDs; another part of customer behavior The information comes from the customer's history of purchasing, insuring, and underwriting insurance products. Such information needs to be aggregated in a certain time dimension, such as one month, three months, half a year, and one year, to form customer behavior labels; customer behavior data includes customer historical preservation and claims information.

The above-mentioned personalized intelligent recommendation system for insurance products, wherein:

For the scoring matrix, it needs to be decomposed into two hidden matrices with k hidden factors in the user and item dimensions by means of singular value decomposition. The vector is multiplied by the nth product implicit vector. The two latent matrices can be obtained by iterative optimization by gradient descent. Make the product of the two closer to the true rating matrix.

The above-mentioned personalized intelligent recommendation system for insurance products, wherein:

Based on the customer-product scoring matrix, the H-KNN hybrid collaborative filtering method is used, combined with the itembased collaborative filtering and customer based collaborative filtering algorithms, to calculate and search for the most similar customers to the customer and the most similar products to the product. The customer's score c for the product is predicted by the similar customer's score for the product, and the customer's score d for the product can also be inferred from the customer's score for the similar product.

The above-mentioned personalized intelligent recommendation system for insurance products, wherein:

The recommendation results calculated by the algorithm given by the algorithm are sorted in descending order according to the product scores, and the top k products are screened as the candidate recommendation list; this is not the final result, because insurance companies often have different recommended products at different times. Featured products need to be displayed in preference to algorithmically recommended products. Regardless of whether the result package given by the algorithm does not contain such products, such products need to occupy the first position in the list. At the same time, the situation of product replacement should also be considered. Often a hot-selling product needs to be updated after it has been launched for a period of time, which may be to update some terms. When the product is updated, it is necessary to integrate the new and old product data to prevent the situation that the new product is not recommended.

Compared with the prior art, the beneficial effects of the present invention are:

It can mainly solve two problems. The first one is to solve the problem of ignoring the sales habits of agents in the insurance product recommendation system. Traditional insurance product recommendation systems often only consider customer information and product information, but not agent information. Because the insurance industry often conducts business through agents, the sales habits of agents have a decisive impact on customer underwriting behavior. If this information is not considered, the prediction effect will be unsatisfactory and the effect of the recommendation system will be affected. The invention combines the agent information to which the client belongs and the client information, and comprehensively considers the influence of the agent on the client's choice.

The second problem solved by the present invention is to solve the problem of inaccurate recommendation caused by sparse data by adopting the method of hybrid collaborative filtering, firstly adopting the method of matrix decomposition. The reason for the sparse data may be that some customers purchased a small number of products, or some products were only purchased by a small number of customers. Hybrid collaborative filtering also comprehensively integrates the recommendation results of products based on similar attributes, the recommendation results based on customers with similar attributes, the prediction of the customer's rating of the product based on the rating of the product by similar customers, and the rating of similar products by the customer. Guess the customer's rating for the product. Through the above four-way collaborative filtering algorithm, the final result is obtained by weighting and mixing, so that the recommendation result is closer to the hidden needs of customers.

Instruction drawings

FIG. 1 is a schematic diagram of a personalized intelligent recommendation system for insurance products.

Detailed ways

The embodiments of the present invention will be described in detail below with reference to the examples, but those skilled in the art will understand that the following examples are only used to illustrate the present invention and should not be regarded as limiting the scope of the present invention. If the specific conditions are not indicated in the examples, it is carried out according to the conventional conditions or the conditions suggested by the manufacturer.

Definitions of acronyms and key terms

ToC: to customer

ToAtoC: to the agent to the client

NLP (natural language processing): natural language processing

KNN (K Nearest Neighbor) : K Nearest Neighbor Algorithm

item based collaborate filtering: item based collaborative filtering

Customer based collaborate filtering: User based collaborative filtering

This scheme first needs to integrate the training data features. The content of integration can be roughly divided into two categories: customer agent feature vector and insurance product feature vector. The following two types of feature vectors are introduced.

Customer agent feature vector: This part can be subdivided into two categories, customer information and information about the agent to which the customer belongs.

Customer information can be further subdivided into customer attribute information and customer behavior information.

The customer attribute information includes basic information such as the customer's gender, age, and address obtained from the customer's insurance application process. Such information is stored in the insurance company's core system database, and some category features are similar to customer addresses. Due to too many categories, preprocessing such as binning of discrete features is required. These basic information can be obtained from the database according to the customer ID and added to the feature vector.

Customer behavior data comes from the front-end embedded point of the app. App clients such as insurance companies’ online malls and WeChat Stores will have a user interface. The user interface displays the insurance products, insurance information, insurance services and other items of the insurance company for the user. Customers can browse, purchase, make reservations, etc. by clicking on the items they are interested in. Through front-end embedding, insurance companies can collect behavioral information generated by users during the browsing process, including clicks, searches, browsing duration, favorites, forwarding, and sharing. Front-end buried point information is uniformly collected and stored in the specified database through the tracking tool. The recommender system can access user behavior data in the database through the user ID. Another part of customer behavior information comes from the customer's history of purchasing, insuring, and underwriting insurance products. Such information needs to be aggregated in a certain time dimension, such as one month, three months, half a year, and one year, to form customer behavior labels. Generally speaking, the more recent the customer's behavior is, the more important it is, and the more recent the behavior is less important. If the same customer browses an item multiple times, it means that the customer is more interested in the item. Customer behavior data can also include information such as customer history preservation and claims settlement.

The matching between customer behavior data and customer attribute data can be done by matching the personal information left by the customer when purchasing the policy to identify the same person and then associate.

Another important piece of information in the client-agent feature vector is the relevant features of the client's agent. The agent to which the customer belongs refers to the agent from which the customer purchases the policy, then the customer belongs to the agent. The characteristics of the agent include the basic information of the agent, such as the agent's gender, age, and the length of time in the agency. This information is linked to the agent ID through the customer's policy and is queried in the core system. At the same time, the agent information also includes the historical order information of the agent, including the types of products that the agent is good at selling, and the order situation of the agent in the past three months, one year and other time dimensions. It can also include statistical information of the customer dimension that the agent has issued orders, including the average age of customers, customer gender distribution, customer surrender rate, etc.

Adding agent-related information is essential for insurance product recommendations. Because the sales model of insurance products is often to reach customers through agents. Therefore, the sales habits of agents are crucial to the underwriting behavior of customers. Because agents tend to recommend products that they are more familiar with and are better at selling to customers. And you may not recommend it too much for the types of insurance that you are not very good at selling. Therefore, it can be said that the agent's choice can affect the customer's choice to a certain extent.

The reason for integrating customer agents into feature vectors is that each customer corresponds to a specific agent. When there are multiple agents corresponding to the same client, the agent can be selected by certain rules (such as the most recently contracted agent). It can also be used as a combination of different client agents.

Insurance product feature vector: This part mainly contains features related to insurance products. Contains structured features related to insurance clauses extracted from unstructured data through NLP (natural language processing). It also includes structural features such as insurance product types, sales start dates, etc. stored in the core system. The insurance product ID can be queried in the database.

After having the customer agent feature vector and the insurance product feature vector, it is necessary to define the customer's score matrix for the insurance product. The score can be weighted according to the customer's purchase and browsing behavior of the product. For example, the weight of customer purchase will be higher than the weight of customer browsing, so the customer purchase weight can be set to 1, and the customer browsing weight can be set to 0.6. Calculate the customer's rating for the product based on the customer's historical purchases and browsing.

The scoring matrix needs to be decomposed into two hidden matrices with k hidden factors in the user and item dimensions by means of singular value decomposition. For example, the rating of the mth customer for the nth product can be obtained by multiplying the mth customer's latent vector and the nth product latent vector. The two latent matrices can be obtained by iterative optimization by gradient descent. Make the product of the two closer to the true rating matrix. By multiplying two implicit matrices, a full-rank customer and product matrix can be obtained, so as to complete the matrix expansion and solve the problem that the rating matrix data is too sparse.

Through the above operations, the customer agent feature vector, the insurance product feature vector and the expanded customer and product rating matrix are obtained.

The next step is to use the above information and use a hybrid collaborative filtering method to obtain user predictions on product ratings.

First, perform preliminary cluster analysis on customers on the customer agent feature vector by KMeans algorithm to form different customer clusters. rate a. In the same way, cluster on the product feature vector to obtain product clusters with similar product features, and calculate the average product score based on the product scores in the product cluster by customers, and obtain the score b.

Based on the customer-product scoring matrix, the H-KNN hybrid collaborative filtering method is used, combined with the itembased collaborative filtering and customer based collaborative filtering algorithms, to calculate and search for the most similar customers to the customer and the most similar products to the product. The customer's score c for the product is predicted by the similar customer's score for the product, and the customer's score d for the product can also be inferred from the customer's score for the similar product.

For the scoring of the above four cases, a weighting coefficient is introduced, and the weighted summation of the four cases is performed to obtain the final score. The sum of the weighting coefficients for the four cases is 1. The coefficients for each case range from 0 to 1. The weighted acquisition rule can better predict the customer's rating of the product for the final result obtained by the combination. This method can effectively integrate multi-channel scoring and obtain more accurate results. The weighting coefficient can be obtained by experimenting with ABTest to find the optimal combination.

The recommendation results calculated by the algorithm given by the algorithm are sorted in descending order according to the product scores, and the top k products are screened as the candidate recommendation list. But this is not the end result. Because insurance companies often have different main products at different times. Featured products need to be displayed in preference to algorithmically recommended products. Regardless of whether the result package given by the algorithm does not contain such products, such products need to occupy the first position in the list. At the same time, the situation of product replacement should also be considered. Often a hot-selling product needs to be updated after it has been launched for a period of time, which may be to update some terms. When the product is updated, it is necessary to integrate the new and old product data to prevent the occurrence of new products not being recommended.

The last step is to add the product filtering module. The function of this module is to filter the products not available for sale by the agent, the products purchased by the customer, and the products recommended but not clicked by the customer before the recommendation list is provided to the customer. The reason is that different insurance products may only be available for sale by agents of a certain channel, and repeatedly recommending products purchased by customers to customers will cause customers' disgust. Through the filtering module, the effect of the recommendation system can be further improved and the click rate of customers can be increased.

Through the above steps, a recommendation list is finally generated, and the top k items are filtered as the final recommendation list according to business requirements. The offline model runs batches every day, calculates the recommendation list corresponding to each customer, and stores the results in the Redis database. After receiving the front-end request, the online service reads the recommendation list from memory and returns it to the user.

The beneficial effects brought by the technical solution of the present invention:

The invention can comprehensively consider the influence of the agent on the customer's purchasing behavior, add the characteristics of the agent for model training, retain the sales habit data of the agent, and improve the recommendation effect of the insurance recommendation system. At the same time, even in the case of sparse data, through the method of matrix decomposition, the matrix is expanded to complement the customer product rating matrix, so as to improve the customer's click-through rate on recommended products and the policy conversion rate.

Claims (6)

1. The personalized intelligent recommendation system for insurance products is characterized in that training data characteristics need to be integrated firstly, and the integrated contents are divided into two types: the client agent feature vector and the insurance product feature vector; client agent feature vector: the part is divided into two categories, namely client information and information related to an agent to which the client belongs; the client information can be subdivided into client attribute information and client behavior information; the reason for integrating customer agents into feature vectors is that each customer will correspond to a particular agent; when the same client corresponds to a plurality of agents, the agents are selected according to a certain rule or different client agent combinations are selected; insurance product feature vector: the part mainly comprises insurance product related features, insurance clause related structured features extracted from unstructured data through NLP in insurance clauses, and also comprises insurance product risk types and sale date related structured features stored in a core system, and the insurance product related features are inquired in a database through an insurance product ID; after possessing the characteristic vector of the client agent and the characteristic vector of the insurance product, the scoring matrix of the client for the insurance product is defined, and the scoring is calculated according to the weighting of the client for the purchasing and browsing behavior of the product; the weight of the purchase of the customer is higher than the weight of the browse of the customer, so that the weight of the purchase of the customer can be set to be 1, and the weight of the browse of the customer can be set to be 0.6; decomposing a scoring matrix into two hidden matrixes of user and project dimensions with k hidden factors by adopting a singular value decomposition mode; through the operation, the characteristic vector of a client agent, the characteristic vector of an insurance product and the grading matrix of the expanded client and product are obtained; the next step is to use the above information and adopt a mixed collaborative filtering mode to obtain the product score prediction of the user; firstly, performing primary clustering analysis on customers on a customer agent characteristic vector through a KMeans algorithm to form different customer clusters, and calculating the product scores of the customers based on the average value of the product scores of all the customers scoring the products in the customer clusters to obtain a score a; similarly, clustering is carried out on the product feature vectors to obtain product clusters with similar product features, and the product scoring mean value is calculated based on the product scoring condition of the products in the product clusters by customers to obtain a score b; introducing a weighting coefficient for the scoring of the four conditions, and carrying out weighted summation on the four conditions to obtain a final score; the sum of the weighting coefficients of the four cases is 1; the coefficient value range in each case is 0 to 1; the weight obtaining rule is that the final result obtained by combination can better predict the product scoring condition of the customer; the weighting coefficients can be obtained by performing an experiment through ABTest to find an optimal combination; the recommendation results calculated by the algorithm are sorted in a descending order according to the product scores, and top k commodities are screened as a candidate recommendation list; the last step is to add a product filtering module which is used for filtering products which can not be sold by agents, products purchased by customers and products which are recommended but not clicked by customers before the recommendation list is provided for the customers, because different insurance products are usually sold only by agents in a certain channel and the products purchased by the customers are repeatedly recommended to the customers, which can cause the counter feeling of the customers, the effect of the recommendation system can be further improved and the click rate of the customers is increased through the filtering module, through the steps, the recommendation list is finally generated, top k products are screened as the final recommendation list according to the service requirements, the offline model carries out batch running every day, the recommendation list corresponding to each customer is calculated, the result is stored in a Redis database, the online service reads the recommendation list from the memory after receiving a front-end request, and returned to the user.

2. The system of claim 1, wherein the system further comprises: the client attribute information comprises basic information of client gender, age and address acquired from the client insurance application process; the information is stored in a core system database of the insurance company, and the category characteristics are obtained from the database according to the customer ID and added into the characteristic vector; the customer behavior data and the customer attribute data are matched, and the personal information left when the customer purchases the policy is matched to identify the same person, so that the correlation is carried out; another piece of important information of the characteristic vector of the client agent is related characteristics of the agent to which the client belongs; the agent to which the customer belongs means the agent from which the customer purchases the policy, and the customer belongs to the agent; the agent characteristics comprise basic information of the agent, and the information is related to the ID of the agent through the policy of the client and is inquired in the core system; meanwhile, the agent information also comprises historical order information of the agent, including the product type which the agent is good at selling, the order output condition of the agent in the time dimension of three months, one year and the like, and statistical information of the client dimension which the agent outputs the order, including the average age of the client, the gender distribution of the client and the refund rate of the client.

3. An insurance product personalized intelligent recommendation system according to claim 2, characterized in that: the customer behavior data come from a front app front end buried point; the user interface is like the items such as insurance products, insurance information, insurance service and the like of the insurance company displayed by the user; a client browses, purchases and makes an appointment by clicking an interested item; through the front-end embedded point, the insurance company collects behavior information generated by a user in the browsing process, including clicking, searching, browsing duration, collecting, forwarding and sharing; the front-end embedded point information is uniformly collected through an embedded point tool and stored in a designated database; the recommendation system accesses user behavior data in a database through the user ID; the other part of the customer behavior information comes from the purchasing, insurance application and underwriting behaviors of the insurance products in the customer history, and the information needs to be aggregated in a certain time dimension to form a customer behavior label in a similar dimension aggregation of one month, three months, half year, one year and the like; the customer behavior data comprises customer historical preservation and claim settlement information.

4. An insurance product personalized intelligent recommendation system according to claim 3, characterized in that: the scoring matrix is decomposed into two hidden matrixes with k hidden factors in user and project dimensionality by adopting a singular value decomposition mode, the scoring of the mth customer on the nth product can be obtained by multiplying the hidden vector of the mth customer and the hidden vector of the nth product, the two hidden matrixes can be obtained by iterative optimization through a gradient descent method, and the product of the two hidden matrixes is closer to a real scoring matrix.

5. An insurance product personalized intelligent recommendation system according to claim 4, characterized in that: on the basis of a customer-product scoring matrix, an H-KNN mixed collaborative filtering method is adopted, an item based marketing filtering algorithm and a customer based marketing filtering algorithm are combined, a customer most similar to the customer and a product most similar to the product are calculated and searched, the score c of the customer on the product is predicted through the score of the similar customer on the product, and the score d of the customer on the product can also be predicted through the score of the customer on the similar product.

6. An insurance product personalized intelligent recommendation system according to claim 5, characterized in that: the recommendation results calculated by the algorithm are sorted in a descending order according to the product scores, and top k commodities are screened as a candidate recommendation list; however, the method is not a final result, as insurance companies often have different main products at different times, the main products need to be displayed in preference to the recommended products of the algorithm, whether the result packet given by the algorithm does not contain the products, the products need to occupy the position with the priority of the list, meanwhile, the situation of product updating and updating is also considered, a hot product needs to be updated after being online for a period of time, some items may be updated, and when the product is updated, new and old product data need to be fused to prevent the new product from being generated in the situation that the new product is not recommended.

Images