Commodity recommendation method and system

Abstract

The invention relates to a commodity recommendation method and system, and belongs to the technical field of commodity recommendation. The commodity recommendation method comprises the following steps: collecting user information data and commodity information data, and desensitizing the user information data; constructing user features through the obtained data features to form a user portrait vector; extracting theme features of the commodity information by using a theme modeling algorithm through the commodity information data, and calculating the distribution condition of each theme feature; mining a commodity relationship between the commodity information data and the user behavior data, and establishing a time function of a change trend based on a complementary substitution relationship of commodities; and based on the user portrait, the commodity information data and the commodity relationship, establishing a recommendation score function of the recommended commodity, and generating a recommendation list of the commodity information according to recommendation scores. According to the method, matching is completed by respectively extracting the features of the user and the commodity, and the comprehensive recommendation score is given based on the relationship between the commodities, so that the accuracy and richness of commodity recommendation are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of commodity recommendation, and specifically relates to a commodity recommendation method and system. Background Art

[0002] Existing product recommendation systems are often based on collecting users' historical purchase records, mining information centered on user characteristics, and recommending to users products that have been recently browsed by users with similar interests. In this process, the collected data is often desensitized in real time according to set rules.

[0003] However, since the rules set in the data desensitization process cannot fully meet the needs of different data application scenarios, some user information has been leaked. At the same time, the matching degree between recommended products and user characteristics is not considered, and the subject features of products are less explored, resulting in insufficient accuracy of recommendations. The correlation between user-purchased products and recommended products is not considered, resulting in algorithm recommendations that often frequently recommend products that users no longer need, or even always recommend similar products that have already been purchased. Therefore, the effectiveness and richness of recommended products are insufficient. Summary of the invention

[0004] In order to solve the above problems existing in the prior art, the present invention provides a product recommendation method and system, which completes the matching by extracting the features of users and products respectively, and then gives a comprehensive recommendation score based on the relationship between products, thereby improving the accuracy and richness of product recommendations.

[0005] The purpose of the present invention can be achieved through the following technical solutions: A first aspect of the present disclosure provides a product recommendation method, comprising the following steps: Data collection and preprocessing: Collect user information and product information from users with their permission, clean and encode the collected data to make it consistent in format, and desensitize the data; User portrait construction: construct user features through preprocessed data features, convert all user features into vector form, form user portrait vector, and add labels to user portraits; Obtain the topic distribution of product information: Use the topic modeling algorithm to extract the topic features of product information through the product information data, and calculate the distribution of each topic feature; Constructing a time function of commodity relationships: mining commodity relationships between commodity information data and user behavior data, and establishing a time function of changing trends based on the complementary and substitutive relationship of commodities; Generate recommendation list: Establish recommendation scoring function for recommended products based on user portrait, product information data and product relationship, and generate recommendation list of product information according to recommendation score; The method of constructing a commodity relationship time function comprises the following steps: Use association rule mining algorithms to mine recommended product combinations from all product purchase records, and then filter association rules based on user behavior data to establish a set of complementary products; Using the similarity calculation method, a threshold is set to calculate the similarity between the recommended products and the purchased products in the user behavior data to filter out a set of alternative products; Based on the established complementary commodity set and substitute commodity set, a time function based on commodity relationship is established according to the time variation characteristics of the complementary and substitute commodity relationship.

[0006] Furthermore, the establishment of a complementary product set comprises the steps of: Use association rule mining algorithms to find frequent item sets about recommended products in all transaction records; generate association rules from frequent item sets, and calculate the support and confidence of the rules; Set the minimum support threshold and confidence threshold to filter the association rules, and then remove the rules that do not contain the user's purchase of goods based on the user behavior data; The association rules finally obtained by screening are used to obtain the included product combinations and establish a complementary product set for recommended products.

[0007] Furthermore, the data collection and preprocessing includes the following steps: Data segmentation: Identify the sensitive status of the collected data and divide the data into sensitive fields and regular fields. De-identify the sensitive fields and use the data engine to extract features for regular fields. De-identification: Desensitize sensitive fields of user information, including: Delete directly identifying information: Delete fields in the data that directly identify individuals, including name, phone number, account number, etc. Replace with anonymous identifiers: Assign a randomly generated anonymous identifier to each user ID and use a hash algorithm to convert sensitive fields into a fixed-length hash value; Geographic location fuzzification: fuzzy geographic location data, blurring precise latitude and longitude information to the city or region level; Time fuzzification: Convert precise time data into months or quarters; Extract features: By defining the rule features of the data engine, scan the user data one by one and generate corresponding data features, including: Rule feature construction: Build rules through multiple dimensions, including user basic data rules, consumption data rules, behavior data rules, social data rules, and feedback data rules. Perform logical definition for each type of data rule, including scope definition, condition combination, and dynamic addition; Data feature generation: Based on the constructed data rule features, the data features of each user data are generated through logical definition.

[0008] Furthermore, the user information data includes user basic data, user behavior data and user feedback data, wherein the user basic data includes user ID, gender, age, geographic location and registration time; the user behavior data includes purchase records, browsing history and search keywords; the user feedback data includes the user's click rate, purchase rate and evaluation feedback on recommended products.

[0009] Furthermore, obtaining the subject distribution of product information includes the following steps: Constructing a document-word matrix: Convert the preprocessed text data of the product information into a document-word matrix. Suppose there are M documents and V words. The document-word matrix D is expressed as: ; Where, d MV represents the frequency of occurrence of word V in document M; Generate topic distribution vector: Use the LDA model to train the document-word matrix, set the number of topics N, and generate an N-dimensional topic distribution vector I for each document D, expressed as: ; In the formula, Indicates that given a document D, the Nth topic R N Probability of occurrence.

[0010] Furthermore, the time function based on commodity relationship is expressed as: K m,c (Δt)=N(Δt|0,σ c ); K n,s (Δt)=-N(Δt|0,σ s1 )+N(Δt|u s , σ s2 ); In the formula, K m,c (Δt) represents the degree of influence between commodity m and complementary commodity c, N(Δt|0, σ c ) means the mean is 0 and the standard deviation is σ c Normal distribution, Δt represents the time interval between the last purchase record of complementary product c; K n,s (Δt) represents the degree of influence between commodity n and substitute commodity s, N(Δt∣0,σ s1 ) means the mean is 0 and the standard deviation is σ s1 The normal distribution of N(Δt|us , σ s2 ) indicates that the mean is u s , standard deviation is σ s2 The normal distribution of u indicates the positive impact of alternative products on recommendations in the long term; s represents the mean of the substitution relationship, σ s1 and σ s2 The subscripts S1 and S2 are used to distinguish the standard deviation of negative impact from the standard deviation of positive impact.

[0011] Furthermore, the recommendation scoring function F is expressed as follows: ; In the formula, S(A, B) represents the similarity between user portraits A and B, M(H, I) represents the matching degree between user behavior data H and topic distribution vector I, α, β, γ are weight coefficients, Indicates the degree of influence between commodity m and complementary commodity c, Indicates the degree of influence between product n and substitute product c, represents the user's rating of product m in the complementary product set C, Represents the user's rating of product n in the set of alternative products S.

[0012] Furthermore, generating a recommendation list further includes: Set the minimum recommendation score according to the recommendation requirements, filter out the product information that meets the recommendation conditions and push it to users.

[0013] A second aspect of the present disclosure provides a product recommendation system for implementing a product recommendation method as described above, including a user portrait module, a product information modeling module and a product recommendation module; The user portrait module is used to collect user information data and construct a user portrait by extracting user features, including the following steps: Clean and preprocess the collected raw data, including deduplication, filling missing values, data format conversion, and outlier detection and processing; Identify the sensitive status of the collected data, divide the data into sensitive fields and regular fields, de-identify the sensitive fields, and use the data engine to extract features for regular fields; Extract features for building user profiles from preprocessed data, including extracting features from the following dimensions: demographic characteristics, behavioral characteristics, interests and hobbies, lifestyle and values, and sentiment analysis; Features of different dimensions are combined into a comprehensive feature and a unified vector representation is established for each user.

[0014] As a preferred technical solution of the present invention, the commodity information modeling module is used to extract topic features according to commodity information data and calculate the distribution of topic features, including the following steps: Convert the preprocessed text data of product information into a document-word matrix; use the LDA model to train the document-word matrix, set the number of topics, and generate a topic distribution vector with a dimension corresponding to the number of topics for each document; The product recommendation module is used to match the user with the user group of interest preference according to the user portrait, obtain the matching degree with the current user through the distribution of product theme features, and then perform weighted correction on the recommended product score based on the time function considering the product relationship, and finally push the product information according to the comprehensive score of the recommended product; The time function of commodity relations includes a time function established according to the temporal variation trend of commodities in a complementary relationship and a time function established according to the temporal variation trend of commodities in a substitution relationship.

[0015] The beneficial effects of the present invention are: The present invention first collects user data, and to protect data privacy, the collected data is de-identified and data engine feature extracted only for product recommendation. De-identification desensitizes sensitive fields of user data, and data engine extraction directly performs rule-based extraction on the data and only retains data features. Developers only receive de-identified and extracted data features and cannot directly access the original data, thereby reducing the risk of user privacy data being leaked due to defects in desensitization rules and protecting user privacy data.

[0016] The present invention first processes the user information data and the product information data separately, establishes user portraits and product information topic distribution for users and products respectively, then performs group matching on users based on user portraits, recommends group interest products to users, and then completes matching with users based on topic distribution, ensuring that the recommendation results meet both the user's interest preferences and the attribute requirements of product information; when making recommendations, the impact of products that have a complementary and alternative relationship with purchased products on the recommendation of target items over time cannot be ignored, so on the basis of the first two steps, a time function of the product relationship is established, and it is included in the recommended product score in the form of a weighted item, so as to explore the user's potential needs from the relationship between products, thereby improving the accuracy of recommended products and the richness of categories. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to facilitate understanding by those skilled in the art, the present invention is further described below with reference to the accompanying drawings.

[0018] Figure 1 A flowchart of a product recommendation method provided by an embodiment of the present invention; Figure 2 A schematic diagram of a process for constructing a commodity relationship time function provided by an embodiment of the present invention; Figure 3 A schematic diagram of the structure of a product recommendation system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation method, structure, characteristics and effects of the present invention are described in detail below in combination with the accompanying drawings and preferred embodiments.

[0020] This embodiment provides a product recommendation method, such as Figure 1 As shown, the following steps are included: S1. Data collection and preprocessing: Collect user information and product information from users by obtaining their permission, clean and encode the collected data to make the data format consistent, and desensitize the data.

[0021] It is understandable that user information data includes basic user data, user behavior data and user feedback data, among which basic user data includes user ID, gender, age, geographic location, registration time, etc., which are used to build basic information of user portraits and help understand the basic characteristics and potential needs of users; user behavior data includes purchase records (product ID, purchase time, purchase quantity, payment amount), browsing history (product ID, browsing time, dwell time), search keywords (search time, search content), etc., which are used to analyze users' interests, preferences and consumption habits, form user behavior models, and support personalized recommendations; product information data includes product information (product ID, name, category, price, inventory), promotional activities (activity ID, activity type, start time, end time), user reviews (product ID, user ID, rating, comment content), etc., which are used to evaluate marketing effectiveness, analyze the attractiveness of different products and activities to users, and provide basic data for personalized recommendations; user feedback data includes users' click-through rate, purchase rate, evaluation feedback, etc. on recommended products, which are used to evaluate the effectiveness of the recommendation system and help optimize the recommendation model.

[0022] S11. Data division: Identify the sensitive status of the collected data and divide the data into sensitive fields and regular fields. De-identify the sensitive fields and use the data engine to extract features for the regular fields. S12. De-identification: Desensitize sensitive fields of user information, including: Delete directly identifying information: Delete fields in the data that directly identify individuals, including name, phone number, account number, etc. Replace with anonymous identifiers: Assign a randomly generated anonymous identifier to each user ID, and use a hash algorithm (such as SHA-256, MD5, etc.) to convert sensitive fields into a fixed-length hash value; Geographic location fuzzification: fuzzify geographic location data (such as GPS coordinates) to blur precise latitude and longitude information to the city or region level; Time fuzzification: Convert precise time data (such as registration time, access time, etc.) into months or quarters.

[0023] S13, feature extraction: By defining the rule features of the data engine, the user data is scanned one by one, and corresponding data features are generated, including: Rule feature construction: Build rules through multiple dimensions, including user basic data rules, consumption data rules, behavior data rules, social data rules and feedback data rules, and logically define each type of data rule, including range definition (such as age, consumption amount, etc.), condition combination (such as judging user activity) and dynamic addition (data features of users that have increased significantly, such as purchase frequency); Data feature generation: Based on the constructed data rule features, the data features of each user data are generated through logical definition.

[0024] It should be noted that we collect users’ basic data, behavioral data, and feedback data, and build user portraits through feature extraction and data de-identification. At the same time, we collect product information data from e-commerce platforms to make personalized recommendations.

[0025] S2. User portrait construction: Construct user features through preprocessed data features, convert all user features into vector form, form a user portrait vector, and add labels to the user portrait.

[0026] It should be noted that user characteristics include basic user characteristics (such as age group, geographic location, etc.), consumption characteristics (such as consumption frequency, average consumption amount, popular purchase categories, etc.), behavioral characteristics (such as browsing time, diversity of search keywords, etc.), social characteristics (such as the number of likes and comments, etc.) and feedback characteristics (such as the average rating of purchased goods). Among them, adding labels to user portraits is to add labels to user portraits based on user characteristics and behaviors, such as "high-spending users", "electronic product enthusiasts", "socially active people", etc.

[0027] S3. Obtaining the topic distribution of product information: extracting the topic features of product information using the topic modeling algorithm through the product information data, and calculating the distribution of each topic feature, including the following steps: S31, constructing a document-word matrix: converting the preprocessed text data of the commodity information data into a document-word matrix. Suppose there are M documents and V words, the document-word matrix D can be expressed as: ; Where, d MV Represents the frequency of occurrence of word V in document M.

[0028] S32, generate topic distribution vector: use LDA model to train the document-word matrix, set the number of topics N, and for each document D, generate an N-dimensional topic distribution vector I, expressed as: ; In the formula, Indicates that given a document D, the Nth topic R N Probability of occurrence.

[0029] It is understandable that commonly used topic modeling algorithms include Latent Dirichlet Allocation (LDA), Non-negative Matrix Factorization (NMF), etc. In this embodiment, LDA is used for topic modeling. By extracting the topic distribution in the product information data of the e-commerce platform, it is convenient to match users based on topics in subsequent personalized recommendations.

[0030] S4. Constructing the time function of commodity relationship: mining the commodity relationship between commodity information data and user behavior data, and establishing the time function of the change trend based on the complementary substitution relationship of commodities, such as Figure 2 As shown, the following steps are included: S41, using an association rule mining algorithm to mine a combination of recommended products from all product purchase records, and then filtering association rules based on user behavior data to establish a complementary product set, including: Use association rule mining algorithms (such as Apriori or FP-Growth) to find frequent item sets about recommended products in all transaction records; for example, if recommended product E and product F are often purchased together, then {E, F} is a frequent item set.

[0031] Generate association rules from frequent item sets, calculate the support and confidence of the rules, set the minimum support threshold and confidence threshold to filter the association rules, and then remove the rules that do not contain the user's purchase of goods based on user behavior data; The association rules finally obtained by screening are used to obtain the included product combinations and establish a complementary product set for recommended products.

[0032] It should be noted that in obtaining a set of complementary products, an association rule mining algorithm is first used to mine frequent item sets for all user purchase records of recommended products in order to obtain a more comprehensive complementary combination of products. Then, the obtained complementary combination of products is screened based on the purchase records of the current user's behavioral data, thereby establishing a complementary combination of products exclusive to the current user, improving the personalization of the set and improving the accuracy of subsequent product recommendations.

[0033] S42: Using a similarity calculation method, setting a threshold to calculate the similarity between the recommended products and the purchased products in the user behavior data, and filtering out a set of alternative products.

[0034] In this embodiment, the cosine distance between the commodities is calculated to screen out a set of commodities with a substitution relationship, and by analyzing the purchase history of the user, it is identified which commodities are often purchased together (complementary relationship) or often purchased as substitutes (substitution relationship), thereby obtaining a set of complementary commodities and a set of substitute commodities, which will not be described in detail here.

[0035] S43. Based on the established complementary commodity set and substitute commodity set, a time function based on commodity relationship is established according to the time variation characteristics of the complementary and substitute commodity relationship: K m,c (Δt)=N(Δt|0,σ c ); K n,s (Δt)=-N(Δt|0,σ s1 )+N(Δt|u s , σ s2 ); In the formula, K m,c (Δt) represents the degree of influence between commodity m and complementary commodity c, N(Δt|0, σ c ) means the mean is 0 and the standard deviation is σ c Normal distribution, Δt represents the time interval between the last purchase record of complementary product c; K n,s (Δt) represents the degree of influence between commodity n and substitute commodity s, N(Δt∣0,σ s1 ) means the mean is 0 and the standard deviation is σ s1 The normal distribution of N(Δt|u s , σ s2 ) indicates that the mean is u s , standard deviation is σ s2 The normal distribution of u indicates the positive impact of alternative products on recommendations in the long term; s It represents the mean of the substitution relationship, reflecting the time point when users may want to change the product after using it for a period of time. The subscripts S1 and S2 are used to distinguish the standard deviation of negative impact and the standard deviation of positive impact.

[0036] It should be noted that the purchase relationship of goods changes over time. For example, the purchase expectation of complementary goods may be strong in the short term, while the impact of substitute goods may be negative in the short term and may turn positive over time. Therefore, the time function of the complementary relationship is designed to have a positive initial value and decay faster, using a normal distribution with a mean of zero as its time function; the time function of the substitute relationship is designed to be two opposite normal distributions as its time function. Therefore, by introducing the time factor, the relationship between goods that changes over time is taken into account, which improves the accuracy of the recommendation.

[0037] S5. Generate a recommendation list: Establish a recommendation scoring function for recommended products based on user portraits, product information data, and product relationships, and generate a recommendation list of product information based on the recommendation score.

[0038] Among them, the formula of the recommendation scoring function F is as follows: ; In the formula, S(A, B) represents the similarity between user portraits A and B, M(H, I) represents the matching degree between user behavior data H and topic distribution vector I, α, β, γ are weight coefficients, Indicates the degree of influence between commodity m and complementary commodity c, Indicates the degree of influence between product n and substitute product c, represents the user's rating of product m in the complementary product set C, Represents the user's rating of product n in the set of alternative products S.

[0039] It should be noted that the minimum recommendation score is set according to the recommendation requirements, and the product information that meets the recommendation conditions is screened and pushed to the user. In the recommendation scoring function, the user's score for product m is calculated by obtaining the lift in the process of mining the association rules of the complementary product set to represent the user's preference for product m, and the user's score for product n is represented by obtaining the cosine distance of the alternative product set to represent the user's preference for product n.

[0040] This embodiment also provides a product recommendation system. Figure 3 As shown, it includes a user portrait module, a product information modeling module and a product recommendation module.

[0041] The user portrait module is used to collect user information data and construct a user portrait by extracting user features, including the following steps: Clean and preprocess the collected raw data, including: Deduplication: Remove duplicate records to avoid misleading analysis results caused by duplicate data; Filling missing values: handling missing items in user information, using methods such as mean filling, median filling, most frequent value filling or predicted filling; Data format conversion: converting data into a standard format, such as converting a time field into a unified date format, or converting a categorical variable into a numerical variable; Outlier detection and processing: Identify and process outliers, such as purchase amounts that are too high or too low, to avoid negative impacts on subsequent analysis; Identify the sensitive status of the collected data, divide the data into sensitive fields and regular fields, de-identify the sensitive fields, and use the data engine to extract features for regular fields; Extract features for building user portraits from the preprocessed data, including extracting features from the following dimensions: Demographic characteristics: age, gender, region, occupation, etc.; Behavioral characteristics: purchase history, browsing history, search keywords, etc. Interests and hobbies: such as favorite product categories, brand preferences, content consumption habits, etc.; Lifestyle and values: user’s values, consumption level, risk preference, etc. (can be obtained through indirect analysis); Sentiment analysis: Extract users’ sentiments (positive, negative, or neutral) toward a product or service by analyzing text data such as users’ comments and feedback.

[0042] Features of different dimensions are combined into a comprehensive feature and a unified vector representation is established for each user.

[0043] The commodity information modeling module is used to extract topic features according to commodity information data and calculate the distribution of topic features, including the following steps: The preprocessed text data of the product information data is converted into a document-word matrix; the document-word matrix is ​​trained using the LDA model, the number of topics is set, and a topic distribution vector with a dimension corresponding to the number of topics is generated for each document.

[0044] The product recommendation module is used to match user groups with interest preferences for users based on user portraits, and obtain the matching degree with the current user through the distribution of product theme features, and then perform weighted correction on the recommended product scores based on a time function that considers product relationships, and finally push product information based on the comprehensive scores of recommended products.

[0045] The time function of commodity relations includes a time function established according to the temporal variation trend of commodities in a complementary relationship and a time function established according to the temporal variation trend of commodities in a substitution relationship.

[0046] The present invention first collects user data, and to protect data privacy, the collected data is de-identified and data engine feature extracted only for product recommendation. De-identification desensitizes sensitive fields of user data, and data engine extraction directly performs rule-based extraction on the data and only retains data features. Developers only receive de-identified and extracted data features and cannot directly access the original data, thereby reducing the risk of user privacy data being leaked due to defects in desensitization rules and protecting user privacy data.

[0047] The present invention first processes the user information data and the product information data separately, establishes user portraits and product information topic distribution for users and products respectively, then performs group matching on users based on user portraits, recommends group interest products to users, and then completes matching with users based on topic distribution, ensuring that the recommendation results meet both the user's interest preferences and the attribute requirements of product information; when making recommendations, the impact of products that have a complementary and alternative relationship with purchased products on the recommendation of target items over time cannot be ignored, so on the basis of the first two steps, a time function of the product relationship is established, and it is included in the recommended product score in the form of a weighted item, so as to explore the user's potential needs from the relationship between products, thereby improving the accuracy of recommended products and the richness of categories.

[0048] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technical personnel in this field can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A commodity recommendation method, characterized in that: The following steps are involved: Data collection and preprocessing: Collect user information and product information from users with their permission, clean and encode the collected data to make it consistent in format, and desensitize the data; User portrait construction: construct user features through preprocessed data features, convert all user features into vector form, form user portrait vector, and add labels to user portraits; Obtain the topic distribution of product information: Use the topic modeling algorithm to extract the topic features of product information through the product information data, and calculate the distribution of each topic feature; Constructing a time function of commodity relationships: mining commodity relationships between commodity information data and user behavior data, and establishing a time function of changing trends based on the complementary and substitutive relationship of commodities; Generate recommendation list: Establish recommendation scoring function for recommended products based on user portrait, product information data and product relationship, and generate recommendation list of product information according to recommendation score; The method of constructing a commodity relationship time function comprises the following steps: Use association rule mining algorithms to mine recommended product combinations from all product purchase records, and then filter association rules based on user behavior data to establish a set of complementary products; Using the similarity calculation method, a threshold is set to calculate the similarity between the recommended products and the purchased products in the user behavior data to filter out a set of alternative products; Based on the established complementary commodity set and substitute commodity set, a time function based on commodity relationship is established according to the time variation characteristics of the complementary and substitute commodity relationship.

2. A commodity recommendation method according to claim 1, characterized in that: The data collection and preprocessing includes the following steps: Data segmentation: Identify the sensitive status of the collected data and divide the data into sensitive fields and regular fields. De-identify the sensitive fields and use the data engine to extract features for regular fields. De-identification: Desensitize sensitive fields of user information, including: Deleting directly identifying information: Deleting fields in the data that directly identify individuals, including name, phone number, and account information; Replace with anonymous identifiers: Assign a randomly generated anonymous identifier to each user ID and use a hash algorithm to convert sensitive fields into a fixed-length hash value; Geographic location fuzzification: fuzzy geographic location data, blurring precise latitude and longitude information to the city or region level; Time fuzzification: Convert precise time data into months or quarters; Extract features: By defining the rule features of the data engine, scan the user data one by one and generate corresponding data features, including: Rule feature construction: Build rules through multiple dimensions, including user basic data rules, consumption data rules, behavior data rules, social data rules, and feedback data rules. Perform logical definition for each type of data rule, including scope definition, condition combination, and dynamic addition. Data feature generation: Based on the constructed data rule features, the data features of each user data are generated through logical definition.

3. A commodity recommendation method according to claim 2, characterized in that: The user information data includes user basic data, user behavior data and user feedback data, wherein the user basic data includes user ID, gender, age, geographic location and registration time; the user behavior data includes purchase records, browsing history and search keywords; the user feedback data includes the user's click rate, purchase rate and evaluation feedback on recommended products.

4. A commodity recommendation method according to claim 1, characterized in that: The step of establishing a complementary commodity set comprises the following steps: Use association rule mining algorithms to find frequent item sets about recommended products in all transaction records; generate association rules from frequent item sets, and calculate the support and confidence of the rules; Set the minimum support threshold and confidence threshold to filter the association rules, and then remove the rules that do not contain the user's purchase of goods based on the user behavior data; The association rules finally obtained by screening are used to obtain the included product combinations and establish a complementary product set for recommended products.

5. A commodity recommendation method according to claim 1, characterized in that: The method of obtaining the subject distribution of commodity information includes the following steps: Constructing a document-word matrix: Convert the preprocessed text data of the product information into a document-word matrix. Suppose there are M documents and V words. The document-word matrix D is expressed as: ； Where, d MV Represents the frequency of occurrence of word V in document M; Generate topic distribution vector: Use the LDA model to train the document-word matrix, set the number of topics N, and for each document D, generate an N-dimensional topic distribution vector I, expressed as: ； In the formula, Indicates that given a document D, the Nth topic R N Probability of occurrence.

6. A commodity recommendation method according to claim 1, characterized in that: The time function based on commodity relationship is expressed as: K m,c (Δt)=N(Δt∣0,σ c ); K n,s (Δt)=-N(Δt∣0,σ s1 )+N(Δt∣u s ,s s2 ); In the formula, K m,c (Δt) represents the degree of influence between commodity m and complementary commodity c, N(Δt|0, σ c ) means the mean is 0 and the standard deviation is σ c The normal distribution of Δt represents the time interval between the last purchase record of complementary product c; K n,s (Δt) represents the degree of influence between commodity n and substitute commodity s, N(Δt|0, σ s1 ) means the mean is 0 and the standard deviation is σ s1 The normal distribution of N(Δt|u s , σ s2 ) indicates that the mean is u s , standard deviation is σ s2 The normal distribution of indicates the positive impact of alternative products on recommendations in the long term; s represents the mean of the substitution relationship, σ s1 and σ s2 The subscripts S1 and S2 are used to distinguish the standard deviation of negative impact from the standard deviation of positive impact.

7. A commodity recommendation method according to claim 1, characterized in that: The recommendation scoring function F is expressed as follows: ； In the formula, S(A, B) represents the similarity between user portraits A and B, M(H, I) represents the matching degree between user behavior data H and topic distribution vector I, α, β, γ are weight coefficients, Indicates the degree of influence between commodity m and complementary commodity c, Indicates the degree of influence between product n and substitute product c, represents the user's rating of product m in the complementary product set C, Represents the user's rating of product n in the set of alternative products S.

8. A commodity recommendation method according to claim 1, characterized in that: The generating of the recommendation list further includes: Set the minimum recommendation score based on the recommendation requirements, filter out product information that meets the recommendation conditions and push it to users.

9. A product recommendation system, used to implement a product recommendation method according to any one of claims 1 to 8, characterized in that: Includes user portrait module, product information modeling module and product recommendation module; The user portrait module is used to collect user information data and construct a user portrait by extracting user features, including the following steps: Clean and preprocess the collected raw data, including deduplication, filling missing values, data format conversion, and outlier detection and processing; Identify the sensitive status of the collected data, divide the data into sensitive fields and regular fields, de-identify the sensitive fields, and use the data engine to extract features for regular fields; Extract features for building user profiles from preprocessed data, including extracting features from the following dimensions: demographic characteristics, behavioral characteristics, interests, lifestyles and values, and sentiment analysis; Features of different dimensions are combined into a comprehensive feature and a unified vector representation is established for each user.

10. A commodity recommendation system according to claim 9, characterized in that: The commodity information modeling module is used to extract topic features according to commodity information data and calculate the distribution of topic features, including the following steps: Convert the preprocessed text data of product information into a document-word matrix; use the LDA model to train the document-word matrix, set the number of topics, and generate a topic distribution vector with a dimension corresponding to the number of topics for each document; The product recommendation module is used to match the user with the user group of interest preference according to the user portrait, obtain the matching degree with the current user through the distribution of product theme features, and then perform weighted correction on the recommended product score based on the time function considering the product relationship, and finally push the product information according to the comprehensive score of the recommended product; The time function of commodity relations includes a time function established according to the temporal variation trend of commodities in a complementary relationship and a time function established according to the temporal variation trend of commodities in a substitution relationship.

Images