Insurance product recommendation method based on picture fuzzy set and collaborative filtering

Through the insurance product recommendation method based on picture fuzzy set and collaborative filtering, the problem that the personalized needs of new users and historical users is difficult to meet, the accuracy and diversity of recommendations are improved, and low-cost and efficient personalized recommendations are achieved.

CN119006118APending Publication Date: 2024-11-22NANJING AUDIT UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411197210.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-29
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

The existing insurance product recommendation method has cold start problems, low user information utilization rate, difficult to meet personalized needs, low accuracy and diversity of recommendation results, and portability and Low reusability.

Method used

Adopt insurance product recommendation method based on picture fuzzy sets and collaborative filtering. By collecting user historical evaluation information and browsing data, using collaborative filtering recommendation algorithm to provide personalized recommendations to new users, and combining user similarity and insurance product similarity. Provide recommendations for historical users.

Benefits of technology

It improves user satisfaction and retention rate, achieves the satisfaction of personalized needs, improves the accuracy and diversity of recommended results, reduces costs, and improves the reusability of the algorithm.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119006118A_ABST
    Figure CN119006118A_ABST
Patent Text Reader

Abstract

The invention discloses an insurance product recommendation method based on a picture fuzzy set and collaborative filtering, and the method comprises the steps: collecting historical evaluation information of a user, and representing the historical evaluation information as a picture fuzzy set data set; collecting user browsing data, and representing the user browsing data as a picture fuzzy set data set; the similarity of the insurance products is calculated by using user browsing data, and a collaborative filtering recommendation algorithm based on the insurance products for new users is proposed accordingly; main features are extracted by utilizing historical evaluation information of users, the similarity between the users is calculated, and on this basis, the insurance product mixed recommendation method based on the user similarity and the insurance product similarity is provided for the historical users in combination with user browsing data. According to the invention, personalized insurance product recommendation methods are provided for new users and historical users respectively. According to the insurance product mixed recommendation method for the historical users, the recommendation accuracy and diversity can be further improved by adjusting the values of the parameters according to the insurance product mixed recommendation method for the historical users.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of product recommendation, and specifically relates to an insurance product recommendation method based on image fuzzy sets and collaborative filtering. Background Art

[0002] The insurance market is gradually maturing, but many people do not have a deep understanding of insurance and lack awareness of the importance of insurance, resulting in limited coverage and depth of insurance. At the same time, there are many types of insurance products on the market, but the homogeneity phenomenon is serious. There is a lack of personalized products and services for different consumer groups, which cannot meet the diverse insurance needs. In addition, although the development of science and technology has brought new opportunities to the insurance industry, the overall integration of the insurance industry and technology still needs to be improved to better utilize technologies such as big data and artificial intelligence to improve service efficiency and user experience. Therefore, in view of the characteristics of the insurance industry with many product types and different user needs, the insurance product recommendation method urgently needs to solve challenges and difficulties such as low user information utilization, difficulty in meeting user personalized needs, and difficulty in product classification.

[0003] At present, patent (CN117611285A) "A method, device, equipment and medium for recommending insurance products based on multiple platforms" recommends products by collecting user behavior information on different platforms; patent (CN106933992B) "A method, device, equipment and storage medium for recommending insurance products" and patent (CN115409630A) "A method for accurately recommending insurance products based on a hybrid recommendation algorithm" adopt a method for recommending products to users based on user similarity; patent (CN117974331A) "A method and system for recommending insurance products based on an e-commerce platform" enhances the accuracy of recommendations by feature grouping and grouping the user's insurance browsing history data set based on cursor speed change trend data.

[0004] Existing recommendation algorithms are mostly based on user similarity, and are not portable enough for new users and historical users’ new needs. The need to meet relevant preset conditions increases the cost of algorithm expansion and reuse. Although the scoring-based insurance product recommendation method is simple, it is mostly designed for special scenarios, has low reusability, and is difficult to meet the personalized needs of users. The insurance product recommendation method based on recommendation algorithm has high requirements for data acquisition and storage, and has low accuracy when facing new needs of new users and historical users. Summary of the invention

[0005] Purpose of the invention: In view of the above problems, the present invention proposes an insurance product recommendation method based on image fuzzy sets and collaborative filtering to solve the cold start problem when recommending insurance products to new users, the problem that it is difficult to meet user personalized needs due to low utilization of historical user information, the low accuracy and diversity of recommendation results, and the problem of low portability and reusability of existing insurance product recommendation methods.

[0006] Technical solution: To achieve the purpose of the present invention, the technical solution adopted by the present invention is:

[0007] An insurance product recommendation method based on image fuzzy sets and collaborative filtering includes the following steps:

[0008] Step 1: Collect user historical evaluation information and represent it as a set of image fuzzy set data;

[0009] Step 2: Collect user browsing data and represent it as a set of image fuzzy set data;

[0010] Step 3: Using the image fuzzy set data set obtained in Step 1 and Step 2, a recommendation list of insurance products is provided to new users through a collaborative filtering recommendation algorithm based on insurance products;

[0011] Step 4: Using the image fuzzy set data set obtained in steps 1 and 2, a hybrid algorithm based on user similarity and insurance product similarity is used to provide a list of insurance product recommendations for historical users.

[0012] Furthermore, step 1 collects user historical evaluation information, which is represented as a picture fuzzy set data set, as follows:

[0013] Step 1.1: Perform feature recognition on the user's historical text evaluation information to obtain a feature set;

[0014] Step 1.2: Use sentiment dictionary to perform sentiment analysis on user historical text evaluation information, and then normalize the obtained sentiment value;

[0015] Step 1.3: Convert the text evaluation information given by each user into positive sentiment value, neutral sentiment value and negative sentiment value under each feature;

[0016] Step 1.4: Convert the positive sentiment value, neutral sentiment value and negative sentiment value obtained in step 1.3 into the positive membership, neutral membership and negative membership of the image fuzzy number, and use the image fuzzy number to represent the user's evaluation value of a certain insurance product under each feature.

[0017] Furthermore, step 1.1 performs feature recognition on the user's historical text evaluation information to obtain a feature set, including:

[0018] (1.1.a) Based on the user historical text evaluation data set S, cross-screen the sentiment vocabulary data set to establish the initial feature word set F of the evaluation information, and calculate the i-th feature word f in the set F i TF-IDF i :

[0019]

[0020] Among them, tf i is a feature word f in set F i The word frequency, that is, f i The ratio of the number of occurrences in set S to the number of occurrences of all words in S; idf i is the inverse document frequency, N is the number of comments on insurance products by all users in the set S, and n i The set S contains the feature word f i Number of user reviews;

[0021] (1.1.b) Set threshold and filter TF-IDF i The insurance product evaluation feature words with values ​​higher than the threshold form a feature word set of the user's historical evaluation information;

[0022] (1.1.c) Sort the feature words in the feature word set obtained in (1.1.b) by their frequency of occurrence, and select the first G feature words with the highest frequency of occurrence that meet the requirements of the recommended user from the feature word set to form a feature word set F' = {f g '}, g∈{1,2,…,G}, where f g ' represents the g-th feature word.

[0023] Furthermore, step 1.2 uses a sentiment dictionary to perform sentiment analysis on the user's historical text evaluation information, and then normalizes the obtained sentiment value, including:

[0024] (1.2.a) Union the stuttering dictionary and the sentiment dictionary, and mark the sentiment words to get a new word segmentation dictionary;

[0025] (1.2.b) Find the sentiment words of the sentence, and record the word polarity, sentiment value and position according to the word segmentation dictionary, wherein the word polarity includes positive, negative and neutral;

[0026] (1.2.c) Search for a degree word before the position of the sentiment word, stop searching when it is found, set a weight value for the degree word, multiply the weight value by the sentiment value, and get a new sentiment value;

[0027] (1.2.d) Search for negative words before the position of the sentiment word, find all the negative words, if the number of negative words is odd, multiply the sentiment value by -1, if it is even, the sentiment value remains unchanged;

[0028] (1.2.e) Search for exclamation marks and question marks. If there is an exclamation mark at the end of a sentence, look for a sentiment word. If there is one, the corresponding sentiment value is +2. If there is a question mark at the end of a sentence, the sentence is judged as negative with a value of +2.

[0029] (1.2.f) Normalize the sentiment value after sentiment analysis. The normalization formula is as follows:

[0030] a ugk =a ugk ' / max{a max ,|a min |}

[0031] Among them, a ugk is the normalized sentiment value, a ugk ' is the sentiment value of the u-th user towards the k-th insurance product under the g-th feature obtained after sentiment analysis of the current sentence, g∈{1,2,…,G}, G is the total number of features, k∈{1,2,…,K}, K is the total number of insurance products, and a max and a min They are the maximum and minimum sentiment values ​​after sentiment analysis of all users' text evaluation information of all insurance products.

[0032] Furthermore, step 1.3 converts the text evaluation information given by each user into positive sentiment value, neutral sentiment value and negative sentiment value under each feature, as follows:

[0033] Assume that after sentiment analysis, the text evaluation given by the current user u is divided into S sentences. Then the positive sentiment value of the s-th sentence of the k-th insurance product under the g-th feature is Neutral sentiment and negative sentiment Calculated according to the following formula:

[0034]

[0035] Where u is an integer on [1,U], representing the u-th user among U users; s is an integer on [1,S], representing the s-th clause among S clauses; is the sentiment value of the s-th sentence of the k-th insurance product by the u-th user under the g-th feature,

[0036] Further, use the image fuzzy number p ugk It represents the evaluation value of user u on the k-th insurance product under the g-th feature. The formula is as follows:

[0037]

[0038] in, μ ugk , η ugk and v ugk They are p ugk The positive, neutral and negative affiliations of

[0039] Image fuzzy set P gk ={p ugk} represents the evaluation value of all users on the kth insurance product under the gth feature, u=1,2,…,U, U is the total number of users, g∈{1,2,…,G}, G is the total number of features.

[0040] Furthermore, step 2 collects user browsing data, which is represented as a picture fuzzy set data set, as follows:

[0041] Step 2.1: First, collect the user's browsing data and normalize the browsing data of user u browsing the k-th insurance product;

[0042] (2.1.a) The browsing time of user u browsing the kth insurance product is normalized as follows:

[0043]

[0044] Among them, T uk is the normalized browsing time, t uk is the actual browsing time of user u browsing the kth insurance product, T0≤t uk ≤T1, T0 and T1 are the upper and lower limits of normal browsing time respectively. Browsing time not within [T0, T1] is considered abnormal browsing and will not be normalized.

[0045] (2.1.b) The number of visits, clicks, and scroll bar drags for user u when browsing the kth insurance product are normalized as follows:

[0046]

[0047] Among them, I uk ,D uk ,E uk are the normalized visits, clicks, and scroll bar dragging times of user u browsing the kth insurance product, i uk d uk and e uk are the actual number of visits, clicks, and scroll bar drags for user u when browsing the kth insurance product, i umax d umax and e umax The actual total number of visits, clicks, and scroll bar drags for all insurance products visited by user u;

[0048] Step 2.2: Convert the normalized browsing data of user u browsing the kth insurance product into the corresponding positive sentiment value A uk ', Neutral sentiment value B uk ' and negative sentiment value C uk ', the formula is as follows:

[0049]

[0050] Among them, SA uk represents the collection / bookmarking behavior of user u when browsing the kth insurance product, SA uk =1 indicates that user u has this behavior when browsing the kth insurance product, otherwise SA uk =0; T max and T min is an adjustable parameter, representing the upper and lower limits of the normal browsing behavior of users, T uk >T max It means that user u stayed on the page for too long when browsing the kth insurance product, that is, the user has left. uk <T min It means that user u stayed on the page for too short a time when browsing the k-th insurance product and did not browse the information on the page normally;

[0051] Step 2.3: Use the image to blur the number q uk It represents the browsing data of user u browsing the k-th insurance product. The formula is as follows:

[0052]

[0053] Among them, N uk '=A uk '+B uk '+C uk ';μ uk '、η uk ' and v uk 'respectively q uk The positive, neutral and negative affiliations of

[0054] Picture fuzzy set Q u = {q uk} represents the browsing data of user u browsing all insurance products, k = 1, 2, ... K, K is the total number of insurance products.

[0055] Furthermore, step 3 uses a collaborative filtering recommendation algorithm based on insurance products to provide a recommended list of insurance products for new users, as follows:

[0056] Step 3.1: Calculate the similarity between the target user and other users;

[0057] Assume there are U users in total, and the similarity S(u * ,u) is calculated as follows:

[0058]

[0059] Among them, q u*k =(μ u*k ',η u*k ',ν u*k ') and q uk =(μ uk ',η uk ',ν uk ') are the image fuzzy sets and Q u The kth image fuzzy number represents the target user u * and the browsing data of the u-th user browsing the k-th insurance product, where K is the total number of insurance products; S(q u*k ,q uk ) is calculated as follows:

[0060]

[0061] Among them, μ u*k ',η u*k ',ν u*k ' and μ uk ',η uk ',ν uk 'respectively q u*k and q uk The positive, neutral and negative membership of u*k '=1-μ u*k '-η u*k '-v u*k ' and π uk '1-μ uk '-η uk '-v uk 'respectively q u*k and q uk The rejection degree of membership;

[0062] Step 3.2: Generate a recommended list of insurance products;

[0063] First, predict the target user u * The comment value given to the k-th insurance product The calculation is as follows:

[0064]

[0065] The target user u will be calculated * Predicted evaluation values ​​for K insurance products Sort from largest to smallest and select the top N insurance products to form a recommended list of insurance products.

[0066] Furthermore, step 4 uses a hybrid algorithm based on user similarity and insurance product similarity to provide a list of insurance product recommendations for historical users, including:

[0067] Step 4.1: Generate a prediction score for insurance products based on user similarity using the method in step 3

[0068] Step 4.2: Generate insurance product prediction scores based on insurance product similarity;

[0069] (4.2.a) Obtain the weight value of the feature;

[0070] Assuming there are K insurance products and G features, the evaluation value of user u on the kth insurance product under the gth feature is expressed as p ugk =(μ ugk ,η ugk ,ν ugk ), then the entropy of the comment information under the g-th feature is e(p ug ) is expressed as

[0071]

[0072] Among them, λ≥1 is the adjustment parameter, For p ugk =(μ ugk ,η ugk ,ν ugk )'s complement;

[0073] Then the weight w of the g-th feature is g The calculation is as follows:

[0074]

[0075] (4.2.b) Calculate the similarity S(k) between the historical evaluation of insurance product k* by the target user u* and other insurance products k * ,k), the calculation formula is as follows:

[0076]

[0077] Where k∈{1,2,…,K} and k≠k * , and p ugk =(μ ugk ,η ugk ,ν ugk ) are the image fuzzy sets and P gkThe fuzzy number of the u-th image represents the target user u * For k * insurance products and historical evaluation data of the k-th insurance product; The calculation formula is as follows:

[0078]

[0079] in, and μ ugk ,η ugk ,ν ugk They are and p ugk The positive membership, neutral membership and negative membership of and π ugk =1-μ ugk -η ugk -ν ugk They are and p ugk The rejection degree of membership;

[0080] (4.2.c) Predict the target user u * Evaluation value of the k-th insurance product The calculation is as follows:

[0081]

[0082] Step 4.3: Generate a mixed recommendation list of insurance products based on user similarity and insurance product similarity. The formula is as follows:

[0083]

[0084] Among them, 0≤α≤1 is the adjustment parameter;

[0085] The target user u will be calculated * Predicted mixed rating values ​​for K insurance products Sort from largest to smallest and select the top N insurance products to form a recommended list of insurance products.

[0086] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0087] The present invention discloses an insurance product method based on image fuzzy sets and collaborative filtering. The method first uses image fuzzy sets to represent user historical evaluation information and user browsing data processing, and then proposes suitable insurance products based on collaborative filtering for historical users and new users, respectively, to provide users with a personalized insurance product recommendation method, thereby improving user satisfaction and retention rate. The method has good portability for new user needs, and the algorithm is relatively simple, which ensures low cost and high reusability. The recommendation results of the method show good accuracy and diversity, and basically meet the personalized needs of users. The method can not only efficiently utilize user information, but also meet the personalized needs of users on the basis of realizing product classification, thereby ensuring that satisfactory insurance products are recommended to different users. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0089] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0090] The insurance product recommendation method based on image fuzzy sets and collaborative filtering described in the present invention captures user historical text evaluation information and user browsing data, and uses image fuzzy sets to represent them to form a data sampling set that can represent node data features. An insurance product recommendation method based on product similarity is proposed for user browsing data, and an insurance product hybrid recommendation method based on user similarity and insurance product similarity is proposed for user historical text evaluation information.

[0091] Process such as Figure 1 As shown, the specific steps include:

[0092] Step 1: Collect user historical evaluation information and represent it as a set of image fuzzy set data.

[0093] First, the user's historical evaluation information is feature identified and sentiment analyzed, and then the processed evaluation information is transformed. Since the user's sentiment is divided into positive, neutral and negative, and it is difficult for a user to evaluate the insurance product under each attribute, the image fuzzy number (corresponding to positive membership, neutral membership, negative membership and rejection membership) can be used to represent the above information for the subsequent calculation. The specific steps are as follows:

[0094] Step 1.1: Perform feature recognition on the user's historical text evaluation information to obtain a feature set. The details are as follows:

[0095] (1.1.a) Based on the user historical evaluation data set S, the initial feature word set F of the evaluation information is established through existing pre-processing methods such as word segmentation and stop word removal (refer to the "Sentiment Analysis Word Set" published by CNKI and the "Sentiment Vocabulary Ontology Library" published by Dalian University of Technology for cross-screening), and the feature word f in F is calculated. i TF-IDF i :

[0096]

[0097] Among them, tf i is a feature word f in set F i The word frequency, that is, f i The ratio of the number of occurrences in set S to the number of occurrences of all words in S; idf i is the inverse document frequency, N is the number of comments on insurance products by all users in the set S, and n i The set S contains the feature word f i The number of user comments.

[0098] (1.1.b) Set thresholds and filter TF-IDF according to specific application scenarios and data sets i The insurance product evaluation feature words with values ​​higher than the threshold form a feature word set of the user's historical evaluation information.

[0099] (1.1.c) Sort the feature words in the feature word set obtained in (1.1.b) by their frequency of occurrence, and select the first G feature words with the highest frequency of occurrence that meet the requirements of the recommended user (such as insurance product type, insurance product features, insurance product amount, etc.) from the feature word set to form a feature word set F' = {f g '}, g∈{1,2,…,G}, where f g ' represents the g-th feature word.

[0100] Step 1.2: Use sentiment dictionary to perform sentiment analysis on user historical text evaluation information, and then normalize the obtained sentiment values, including:

[0101] (1.2.a) Union the stuttering dictionary and the sentiment dictionary, label the sentiment words, and derive a new word segmentation dictionary.

[0102] (1.2.b) Find the sentiment words of the sentence, and record the word polarity, sentiment value and position according to the word segmentation dictionary, wherein the word polarity includes positive, negative and neutral.

[0103] (1.2.c) Search for a degree word before the position of the sentiment word, stop searching when it is found, set a weight value for the degree word, multiply the weight value by the sentiment value, and get a new sentiment value.

[0104] (1.2.d) Search for negative words before the sentiment word and find all the negative words. If the number of negative words is odd, multiply the sentiment value by -1. If it is even, the sentiment value remains unchanged.

[0105] (1.2.e) Search for exclamation marks and question marks. If there is an exclamation mark at the end of a sentence, look for a sentiment word. If there is one, the corresponding sentiment value is +2. If there is a question mark at the end of a sentence, the sentence is judged as negative with a value of +2.

[0106] (1.2.f) Normalize the sentiment value after sentiment analysis. The normalization formula is as follows:

[0107] a ugk =a ugk ' / max{a max ,|a min |}

[0108] Among them, a ugk is the normalized sentiment value, a ugk ' is the sentiment value of the u-th user towards the k-th insurance product under the g-th feature obtained after sentiment analysis of the current sentence, g∈{1,2,…,G}, G is the total number of features, k∈{1,2,…,K}, K is the total number of insurance products, and a max and a min They are the maximum and minimum sentiment values ​​after sentiment analysis of all users' text evaluation information of all insurance products.

[0109] Step 1.3: Convert the text evaluation information given by each user into positive sentiment value, neutral sentiment value, and negative sentiment value under each feature. Assume that after sentiment analysis, it is determined that the text evaluation given by the current user u (1≤u≤U) is divided into S sentences. Then the positive sentiment value of the s-th sentence of the k-th insurance product under the g-th feature is Neutral sentiment and negative sentiment Calculated according to the following formula:

[0110]

[0111] Where u is an integer on [1,U], representing the u-th user among U users; s is an integer on [1,S], representing the s-th clause among S clauses; is the sentiment value of the s-th sentence of the k-th insurance product by the u-th user under the g-th feature,

[0112] Step 1.4: Use the image to blur the number p ugk It represents the evaluation value of user u on the k-th insurance product under the g-th feature:

[0113]

[0114] in, μ ugk , η ugk and v ugk They are p ugk Positive affiliation, neutral affiliation and negative affiliation.

[0115] Image fuzzy set P gk ={p ugk} represents the evaluation value of all users on the kth insurance product under the gth feature, u=1,2,…,U, U is the total number of users, g∈{1,2,…,G}, G is the total number of features.

[0116] Step 2: Collect user browsing data and represent it as a set of image fuzzy set data.

[0117] First, we collect the user's browsing data, which mainly includes browsing time, number of visits (clicks), favorites / bookmarks, number of scroll bar drags, and other browsing behaviors of a certain insurance product page. Since there is no limit on the number and type of pages a user browses and the browsing behavior is not fixed, we use fuzzy numbers to represent the user's browsing behavior. The specific steps are as follows:

[0118] Step 2.1: Normalize the browsing data of user u browsing the k-th insurance product as follows:

[0119] (2.1.a) The browsing time of user u browsing the kth insurance product is normalized as follows:

[0120]

[0121] Among them, T uk is the normalized browsing time, t uk is the actual browsing time of user u browsing the kth insurance product, T0≤t uk ≤T1, T0 and T1 are the upper and lower limits of the normal browsing time. The browsing time not within [T0, T1] is abnormal browsing and will not be normalized.

[0122] (2.1.b) The number of visits, clicks, and scroll bar drags for user u when browsing the kth insurance product are normalized as follows:

[0123]

[0124] Among them, I uk ,D uk ,E uk are the normalized visits, clicks, and scroll bar dragging times of user u browsing the kth insurance product, iuk d uk and e uk are the actual number of visits, clicks, and scroll bar drags for user u when browsing the kth insurance product, i umax d umax and e umax The actual total number of visits, total number of clicks, and total number of scroll bar drags for user u to access all insurance products.

[0125] Step 2.2: Convert the normalized browsing data of user u browsing the kth insurance product into the corresponding positive sentiment value A uk ', Neutral sentiment value B uk ' and negative sentiment value C uk ', the formula is as follows:

[0126]

[0127] Among them, SA uk represents the collection / bookmarking behavior of user u when browsing the kth insurance product, SA uk =1 indicates that user u has this behavior when browsing the kth insurance product, otherwise SA uk =0; T max and T min is an adjustable parameter, representing the upper and lower limits of the normal browsing behavior of users, T uk >T max It means that user u stayed on the page for too long when browsing the kth insurance product, that is, the user has left. uk <T min It means that user u stayed on the page for too short a time when browsing the k-th insurance product and did not browse the information on the page normally.

[0128] Step 2.3: Use the image to blur the number q uk It represents the browsing data of user u browsing the k-th insurance product. The formula is as follows:

[0129]

[0130] Among them, N uk '=A uk '+B uk '+C uk ';μ uk '、η uk ' and v uk 'respectively q uk Positive affiliation, neutral affiliation and negative affiliation.

[0131] Picture fuzzy set Q u = {q uk} represents the browsing data of user u browsing all insurance products, k = 1, 2, ... K, K is the total number of insurance products.

[0132] Step 3: Propose a collaborative filtering recommendation algorithm based on insurance products for new users to provide new users with a personalized list of insurance product recommendations.

[0133] Step 3.1: Calculate the similarity between the target user and other users.

[0134] Assume there are U users in total, and the similarity S(u * ,u) is calculated as follows:

[0135]

[0136] Among them, q u*k =(μ u*k ',η u*k ',ν u*k ') and q uk =(μ uk ',η uk ',ν uk ') are the image fuzzy sets and Q u The kth image fuzzy number represents the target user u * and the browsing data of the u-th user browsing the k-th insurance product, where K is the total number of insurance products; S(q u*k ,q uk ) is calculated as follows:

[0137]

[0138] Among them, μ u*k ',η u*k ',ν u*k ' and μ uk ',η uk ',ν uk 'respectively q u*k and q uk The positive, neutral and negative membership of u*k '=1-μ u*k '-η u*k '-v u*k ' and π uk '1-μ uk '-η uk '-v uk 'respectively q u*k and q uk The rejection degree of membership.

[0139] Step 3.2: Generate a recommended list of insurance products.

[0140] First, predict the target user u * The comment value given to the k-th insurance product The calculation is as follows:

[0141]

[0142] The target user u will be calculated * Predicted evaluation values ​​for K insurance products Sort from largest to smallest and select the top N insurance products to form a recommended list of insurance products.

[0143] Step 4: Use a hybrid algorithm based on user similarity and insurance product similarity to provide a list of recommended insurance products for historical users, including:

[0144] Step 4.1: Generate a prediction score for insurance products based on user similarity using the method in step 3

[0145] Step 4.2: Generate insurance product prediction scores based on insurance product similarities.

[0146] (4.2.a) Obtain the weight value of the feature;

[0147] Assuming there are K insurance products and G features, the evaluation value of user u on the kth insurance product under the gth feature is expressed as p ugk =(μ ugk ,η ugk ,ν ugk ), then the entropy of the comment information under the g-th feature is e(p ug ) is expressed as

[0148]

[0149] Among them, λ≥1 is the adjustment parameter, For p ugk =(μ ugk ,η ugk ,ν ugk )'s complement;

[0150] Then the weight w of the g-th feature is g The calculation is as follows:

[0151]

[0152] (4.2.b) Calculate the similarity S(k) between the historical evaluation of insurance product k* by the target user u* and other insurance products k * ,k), the calculation formula is as follows:

[0153]

[0154] Where k∈{1,2,…,K} and k≠k * , and p ugk =(μ ugk ,η ugk ,ν ugk ) are the image fuzzy sets and P gk The fuzzy number of the u-th image represents the target user u * For k * insurance products and historical evaluation data of the k-th insurance product; The calculation formula is as follows:

[0155]

[0156] in, and μ ugk ,η ugk ,ν ugk They are and p ugk The positive membership, neutral membership and negative membership of and π ugk =1-μ ugk -η ugk -ν ugk They are and p ugk The rejection degree of membership;

[0157] (4.2.c) Predict the target user u * Evaluation value of the k-th insurance product The calculation is as follows:

[0158]

[0159] Step 4.3: Generate a mixed recommendation list of insurance products based on user similarity and insurance product similarity. The formula is as follows:

[0160]

[0161] Among them, 0≤α≤1 is the adjustment parameter;

[0162] The target user u will be calculated * Predicted mixed rating values ​​for K insurance products Sort from largest to smallest and select the top N insurance products to form a recommended list of insurance products.

[0163] It can be found that the present invention proposes a personalized insurance product recommendation method for new users and historical users, respectively. It has good portability and meets the personalized needs of users. The mixed recommendation method of insurance products based on user similarity and insurance product similarity can further improve the recommendation accuracy and diversity by adjusting the value of parameter α.

Claims

1. An insurance product recommendation method based on image fuzzy sets and collaborative filtering, characterized in that: The following steps are involved: Step 1: Collect user historical evaluation information and represent it as a set of image fuzzy set data; Step 2: Collect user browsing data and represent it as a set of image fuzzy set data; Step 3: Using the image fuzzy set data set obtained in Step 1 and Step 2, a recommendation list of insurance products is provided to new users through a collaborative filtering recommendation algorithm based on insurance products; Step 4: Using the image fuzzy set data set obtained in steps 1 and 2, a hybrid algorithm based on user similarity and insurance product similarity is used to provide a list of insurance product recommendations for historical users.

2. The insurance product recommendation method based on image fuzzy sets and collaborative filtering according to claim 1, characterized in that: Step 1 collects user historical evaluation information and represents it as a picture fuzzy set data set. The method is as follows: Step 1.1: Perform feature recognition on the user's historical text evaluation information to obtain a feature set; Step 1.2: Use sentiment dictionary to perform sentiment analysis on user historical text evaluation information, and then normalize the obtained sentiment value; Step 1.3: Convert the text evaluation information given by each user into positive sentiment value, neutral sentiment value and negative sentiment value under each feature; Step 1.4: Convert the positive sentiment value, neutral sentiment value and negative sentiment value obtained in step 1.3 into the positive membership, neutral membership and negative membership of the image fuzzy number, and use the image fuzzy number to represent the user's evaluation value of a certain insurance product under each feature.

3. The insurance product recommendation method based on image fuzzy sets and collaborative filtering according to claim 2 is characterized in that: Step 1.1 performs feature recognition on the user's historical text evaluation information to obtain a feature set, including: (1.1.a) Based on the user historical text evaluation data set S, cross-screen the sentiment vocabulary data set to establish the initial feature word set F of the evaluation information, and calculate the i-th feature word f in the set F i TF-IDF i : TF-IDF i =tf i ×idf i , Among them, tf i is a feature word f in set F i The word frequency, that is, f i The ratio of the number of occurrences in set S to the number of occurrences of all words in S; idf i is the inverse document frequency, N is the number of comments on insurance products by all users in the set S, and n i The set S contains the feature word f i Number of user reviews; (1.1.b) Set threshold and filter TF-IDF i The insurance product evaluation feature words with values ​​higher than the threshold form a feature word set of the user's historical evaluation information; (1.1.c) Sort the feature words in the feature word set obtained in (1.1.b) by their frequency of occurrence, and select the first G feature words with the highest frequency of occurrence that meet the requirements of the recommended user from the feature word set to form a feature word set F' = {f g '}, g∈{1,2,…,G}, where f g ' represents the g-th feature word.

4. The insurance product recommendation method based on image fuzzy sets and collaborative filtering according to claim 2, characterized in that: Step 1.2 uses a sentiment dictionary to perform sentiment analysis on the user's historical text evaluation information, and then normalizes the obtained sentiment values, including: (1.2.a) Union the stuttering dictionary and the sentiment dictionary, and mark the sentiment words to get a new word segmentation dictionary; (1.2.b) Find the sentiment words of the sentence, and record the word polarity, sentiment value and position according to the word segmentation dictionary, wherein the word polarity includes positive, negative and neutral; (1.2.c) Search for a degree word before the position of the sentiment word, stop searching when it is found, set a weight value for the degree word, multiply the weight value by the sentiment value, and get a new sentiment value; (1.2.d) Search for negative words before the position of the sentiment word, find all the negative words, if the number of negative words is odd, multiply the sentiment value by -1, if it is even, the sentiment value remains unchanged; (1.2.e) Search for exclamation marks and question marks. If there is an exclamation mark at the end of a sentence, look for a sentiment word. If there is one, the corresponding sentiment value is +2. If there is a question mark at the end of a sentence, the sentence is judged as negative with a value of +2. (1.2.f) Normalize the sentiment value after sentiment analysis. The normalization formula is as follows: a ugk =a ugk ' / max{a max ,|a min |} Among them, a ugk is the normalized sentiment value, a ugk ' is the sentiment value of the u-th user towards the k-th insurance product under the g-th feature obtained after sentiment analysis of the current sentence, g∈{1,2,…,G}, G is the total number of features, k∈{1,2,…,K}, K is the total number of insurance products, and a max and a min They are the maximum and minimum sentiment values ​​after sentiment analysis of all users' text evaluation information of all insurance products.

5. The insurance product recommendation method based on image fuzzy sets and collaborative filtering according to claim 2, characterized in that: Step 1.3 converts the text evaluation information given by each user into positive sentiment value, neutral sentiment value and negative sentiment value under each feature as follows: Assume that after sentiment analysis, the text evaluation given by the current user u is divided into S sentences. Then the positive sentiment value of the s-th sentence of the k-th insurance product under the g-th feature is Neutral sentiment and negative sentiment Calculated according to the following formula: Where u is an integer on [1,U], representing the u-th user among U users; s is an integer on [1,S], representing the s-th clause among S clauses; is the sentiment value of the s-th sentence of the k-th insurance product by the u-th user under the g-th feature, 6. The insurance product recommendation method based on image fuzzy sets and collaborative filtering according to claim 5, characterized in that: Fuzzy number p with image ugk It represents the evaluation value of user u on the k-th insurance product under the g-th feature. The formula is as follows: in, μ ugk , η ugk and v ugk They are p ugk The positive, neutral and negative affiliations of Image fuzzy set P gk ={p ugk } represents the evaluation value of all users on the kth insurance product under the gth feature, u=1,2,…,U, U is the total number of users, g∈{1,2,…,G}, G is the total number of features.

7. The insurance product recommendation method based on image fuzzy sets and collaborative filtering according to claim 1, characterized in that: Step 2 collects user browsing data, which is represented as a set of image fuzzy set data, as follows: Step 2.1: First, collect the user's browsing data and normalize the browsing data of user u browsing the k-th insurance product; (2.1.a) The browsing time of user u browsing the kth insurance product is normalized as follows: Among them, T uk is the normalized browsing time, t uk is the actual browsing time of user u browsing the kth insurance product, T0≤t uk ≤T1, T0 and T1 are the upper and lower limits of normal browsing time respectively. Browsing time not within [T0, T1] is considered abnormal browsing and will not be normalized. (2.1.b) The number of visits, clicks, and scroll bar drags for user u when browsing the kth insurance product are normalized as follows: Among them, I uk ,D uk ,E uk are the normalized number of visits, clicks, and scroll bar drags for user u browsing the kth insurance product, i uk d uk and e uk are the actual number of visits, clicks, and scroll bar drags for user u when browsing the kth insurance product, i umax d umax and e umax The actual total number of visits, clicks, and scroll bar drags for all insurance products visited by user u; Step 2.2: Convert the normalized browsing data of user u browsing the kth insurance product into the corresponding positive sentiment value A uk ', Neutral sentiment value B uk ' and negative sentiment value C uk ', the formula is as follows: Among them, SA uk represents the collection / bookmarking behavior of user u when browsing the kth insurance product, SA uk =1 indicates that user u has this behavior when browsing the kth insurance product, otherwise SA uk =0; T max and T min is an adjustable parameter, representing the upper and lower limits of the normal browsing behavior of users, T uk >T max It means that user u stayed on the page for too long when browsing the kth insurance product, that is, the user has left. uk <T min It means that user u stayed on the page for too short a time when browsing the k-th insurance product and did not browse the information on the page normally; Step 2.3: Use the image to blur the number q uk It represents the browsing data of user u browsing the k-th insurance product. The formula is as follows: Among them, N uk '=A uk '+B uk '+C uk ';μ uk '、η uk ' and v uk 'respectively q uk The positive, neutral and negative affiliations of Picture fuzzy set Q u = {q uk } represents the browsing data of user u browsing all insurance products, k = 1, 2, ... K, K is the total number of insurance products.

8. The insurance product recommendation method based on image fuzzy sets and collaborative filtering according to any one of claims 1 to 7, characterized in that: Step 3 uses a collaborative filtering recommendation algorithm based on insurance products to provide a recommended list of insurance products for new users. The method is as follows: Step 3.1: Calculate the similarity between the target user and other users; Assume there are U users in total, and the similarity S(u * ,u) is calculated as follows: Among them, q u*k =(μ u*k ',η u*k ',ν u*k ') and q uk =(μ uk ',η uk ',ν uk ') are the image fuzzy sets Q u* and Q u The kth image fuzzy number represents the target user u * and the browsing data of the u-th user browsing the k-th insurance product, where K is the total number of insurance products; S(q u*k ,q uk ) is calculated as follows: Among them, μ u*k ',η u*k ',ν u*k ' and μ uk ',η uk ',ν uk 'respectively q u*k and q uk The positive, neutral and negative membership of u*k '=1-μ u*k '-η u*k '-v u*k ' and π uk '1-μ uk '-η uk '-v uk 'respectively q u*k and q uk The rejection degree of membership; Step 3.2: Generate a recommended list of insurance products; First, predict the target user u * The comment value given to the k-th insurance product The calculation is as follows: The target user u will be calculated * Predicted evaluation values ​​for K insurance products Sort from largest to smallest and select the top N insurance products to form a recommended list of insurance products.

9. The insurance product recommendation method based on image fuzzy sets and collaborative filtering according to claim 8, characterized in that: Step 4 uses a hybrid algorithm based on user similarity and insurance product similarity to provide a list of insurance product recommendations for historical users, including: Step 4.1: Generate a prediction score for insurance products based on user similarity using the method in step 3 Step 4.2: Generate insurance product prediction scores based on insurance product similarity; (4.2.a) Obtain the weight value of the feature; Assuming there are K insurance products and G features, the evaluation value of user u on the kth insurance product under the gth feature is expressed as p ugk =(μ ugk ,η ugk ,ν ugk ), then the entropy of the comment information under the g-th feature is e(p ug ) is expressed as Among them, λ≥1 is the adjustment parameter, For p ugk =(μ ugk ,η ugk ,ν ugk )'s complement; Then the weight w of the g-th feature is g The calculation is as follows: (4.2.b) Calculate the similarity S(k) between the historical evaluation of insurance product k* by the target user u* and other insurance products k * ,k), the calculation formula is as follows: Where k∈{1,2,…,K} and k≠k * , p ugk* =(μ ugk* ,η ugk* ,ν ugk* ) and p ugk =(μ ugk ,η ugk ,ν ugk ) are the image fuzzy sets P gk* and P gk The fuzzy number of the u-th image represents the target user u * For k * insurance products and historical evaluation data of the kth insurance product; S(p ugk* ,q ugk ) is calculated as follows: Among them, μ ugk* ,η ugk* ,ν ugk* and μ ugk ,η ugk ,ν ugk They are p ugk* and p ugk The positive, neutral and negative membership of ugk* =1-μ ugk* -η ugk* -ν ugk* and π ugk =1-μ ugk -η ugk -ν ugk They are p ugk* and p ugk The rejection degree of membership; (4.2.c) Predict the target user u * Evaluation value of the k-th insurance product The calculation is as follows: Step 4.3: Generate a mixed recommendation list of insurance products based on user similarity and insurance product similarity. The formula is as follows: Among them, 0≤α≤1 is the adjustment parameter; The target user u will be calculated * Predicted mixed rating values ​​for K insurance products Sort from largest to smallest and select the top N insurance products to form a recommended list of insurance products.

Citation Information

Cited By

  • Client guarantee demand analysis and insurance product recommendation method based on data elements

    CN121120271A