E-commerce product recommendation method and system based on big data analysis

Abstract

The application provides an e-commerce product recommendation method and system based on big data analysis. It relates to the technical field of big data analysis. The method comprises the following steps: determining the user data to be analyzed of a target user, and determining category data of multiple categories based on user data mining; based on the category data of multiple categories, respectively performing cross-category indexing from the level of multiple categories of representation information to obtain multiple-category indexing portrait elements of the category data of multiple categories; based on the multiple-category indexing portrait elements of the category data of multiple categories, determining multiple undetermined portrait elements for the user data; performing portrait element reasoning based on the multiple undetermined portrait elements and the category data of multiple categories to obtain user portrait elements for the user data, and then generating a target user portrait. Finally, the personalized product recommendation for the target user is realized according to the user portrait. The method utilizes big data analysis technology, and effectively improves the accuracy of the recommendation system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of big data analysis, and particularly relates to an e-commerce product recommendation method and system based on big data analysis. BACKGROUND

[0002] In the e-commerce field, with the rapid development of big data technology and the increasing richness of user behavior data, personalized product recommendation has become an important means to improve user experience, enhance user stickiness and promote sales conversion. Traditional e-commerce product recommendation methods often analyze data based on a single dimension such as user purchase history, browsing records or search keywords, and then recommend related products to users. However, these methods ignore the diversity and complexity of user behavior and the potential correlation between different categories of data, resulting in limited accuracy and relevance of the recommended results.

[0003] In order to more comprehensively understand user needs and preferences and improve the accuracy of product recommendation, the industry has begun to explore multi-dimensional and cross-category e-commerce product recommendation methods based on big data analysis. Such methods attempt to mine and analyze various behavior data (such as clicks, views, adds to cart, purchases, etc.) and basic attribute data (such as age, gender, region, etc.) of users on e-commerce platforms to build more detailed and comprehensive user portraits. However, how to effectively index and integrate between multiple categories of data, and how to accurately infer user portrait elements based on these data to generate target user portraits, are still challenges faced by current technology. SUMMARY

[0004] The purpose of the present application is to provide an e-commerce product recommendation method and device based on big data analysis. The embodiments of the present application are implemented as follows:

[0005] In a first aspect, the embodiments of the present application provide an e-commerce product recommendation method based on big data analysis, which comprises:

[0006] determining user data to be analyzed for a target user, and mining and determining category data of multiple categories for the user data;

[0007] based on the category data of the multiple categories, respectively indexing across categories at the level of the representation information of the multiple categories to obtain multi-category index portrait elements of the category data of the multiple categories;

[0008] based on the multi-category index portrait elements of the category data of the multiple categories, determining multiple undetermined portrait elements for the user data;

[0009] perform portrait element reasoning based on the plurality of pending portrait elements and the category data of the plurality of categories to obtain a user portrait element for the user data, and generate a target user portrait of the target user based on the user portrait element;

[0010] recommend products to the target user according to the target user portrait.

[0011] In another aspect, the embodiment of the present application provides an e-commerce product recommendation system based on big data analysis, comprising a memory and a processor, the memory stores a computer program capable of running on the processor, and the processor implements the steps in the above method when executing the program.

[0012] The present application improves the accuracy and personalization level of e-commerce product recommendation through the method based on big data analysis. First, the user data of the target user is deeply mined to determine the category data of multiple categories, and cross-category indexing is performed on the representation information level of these category data, so as to obtain more rich and accurate multi-category index portrait elements, which helps to more comprehensively understand the user's interest preferences. Then, a plurality of pending portrait elements are determined according to the index portrait elements, and specific user portrait elements are obtained through the portrait element reasoning process combined with the category data, and finally the target user portrait is generated. This process can effectively improve the accuracy of user portrait construction, and then the e-commerce platform can provide personalized product recommendation services for the target user according to the more accurate user portrait, which not only improves the user's shopping experience, but also improves the efficiency and accuracy of the recommendation system. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the description of the embodiments of the present application will be briefly introduced.

[0014] Figure 1 An implementation process schematic diagram of an e-commerce product recommendation method based on big data analysis provided by the embodiment of the present application.

[0015] Figure 2 A hardware entity schematic diagram of an e-commerce product recommendation system based on big data analysis provided by the embodiment of the present application. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be further described in detail below in combination with the drawings and embodiments. The described embodiments should not be regarded as limiting the present application, and all other embodiments obtained by those skilled in the art without creative labor shall fall within the scope of protection of the present application.

[0017] Figure 1An implementation flowchart of an e-commerce product recommendation method based on big data analysis provided by an embodiment of the present application is shown in FIG. 1. As shown in FIG. 1, the method comprises the following steps. Figure 1

[0018] In step 100, the user data to be analyzed of a target user is determined, and category data of multiple categories is determined by mining the user data.

[0019] The recommendation system obtains the basic attribute data of the user through the user registration information, such as the name, age, gender, occupation, and other information filled in by the user when registering on the e-commerce platform. The recommendation system also obtains the behavior monitoring data by monitoring the user's behavior on the e-commerce platform. For example, when the user browses the goods on the platform, the recommendation system records the user's click, collection, add-to-cart, purchase, and other behaviors.

[0020] In order to more accurately analyze the user data, the collected raw data can be preprocessed. The steps of preprocessing include data cleaning, data conversion, and data normalization, etc. Data cleaning refers to removing noise, missing values, and outliers in the data to improve the quality of the data. The recommendation system uses statistical methods or machine learning algorithms to detect and handle missing values and outliers. Data conversion refers to converting the raw data into a format suitable for analysis. The recommendation system converts the text data into a numerical vector for numerical calculation and analysis. Data normalization refers to scaling the data to a specific range to eliminate the influence of the dimension of the data. The recommendation system uses methods such as min-max normalization or z-score normalization to normalize the data.

[0021] After completing the data preprocessing, the recommendation system mines the user data to determine category data of multiple categories. This is achieved through data mining algorithms such as clustering analysis, association rule mining, and classification algorithms, etc. Clustering analysis refers to dividing user data into different categories or clusters, and users in each cluster have similar features and behaviors. The recommendation system uses the K-means clustering algorithm to cluster users according to age, gender, consumption habits, and other features, obtaining different user groups. Association rule mining refers to discovering the association relationship between different variables in the data. The recommendation system uses the Apriori algorithm to discover the association rules between user purchases, such as users who have purchased a mobile phone are also likely to purchase a mobile phone case. Classification algorithm refers to classifying user data into different categories. The recommendation system uses the decision tree algorithm to classify users into high-value users, medium-value users, and low-value users according to the basic attribute data and behavior monitoring data of the users.

[0022] ​For the mining of basic attribute data and behavior monitoring data, the recommendation system adopts different strategies. For basic attribute data, the recommendation system constructs attribute feature vectors, taking each attribute as a dimension of the vector. For example, age is represented by a numerical value, and occupation is represented by a code. By analyzing these attribute feature vectors, the relationships and patterns between different attributes are discovered. For behavior monitoring data, the recommendation system constructs behavior sequences, recording the user's behavior at different time points. For example, the user's click behavior is represented as a time series, and by analyzing the behavior sequence, the user's behavior rules and trends are discovered.

[0023] In step 200, based on the category data of the multiple categories, cross-category indexing is performed on the level of the representation information of the multiple categories, respectively, to obtain multiple-category indexing portrait elements of the category data of the multiple categories.

[0024] The representation information of the multiple categories refers to information that can represent the characteristics of the category data. It is some attributes or feature vectors of the data or an abstract representation after processing. For the basic attribute data of users, the representation information is age interval, occupation type, etc. For behavior monitoring data, the representation information is click frequency, collection preference, etc. Cross-category indexing refers to associating and matching between different categories of representation information to discover potential connections and patterns.

[0025] Before cross-category indexing, the recommendation system performs feature extraction and conversion on the category data of the multiple categories, converting it into representation information suitable for indexing. The recommendation system uses natural language processing techniques to process the user's text description information, extracting keywords and topics; uses machine learning algorithms to analyze the user's behavior data, obtaining behavior feature vectors. These representation information will serve as the basis for cross-category indexing.

[0026] The recommendation system associates and matches the representation information of the multiple categories through cross-category indexing. This is achieved through various technical means, such as similarity calculation, association rule mining, etc. Similarity calculation refers to calculating the similarity between the representation information of different categories of data. The optional similarity measurement methods include Euclidean distance, cosine similarity, etc. The calculation formula of Euclidean distance is: where x and y are two representation information vectors, and n is the dimension of the vector. where is the dot product of the vector, and The similarity is the modulus of the vector respectively. The recommendation system finds the representation information of other category data most similar to the representation information of the current category data according to the result of similarity calculation, so as to realize cross-category indexing. The association rule mining is to find the potential rules and patterns by mining the association relationship between different category data. The recommendation system uses association rule mining algorithms such as Apriori algorithm to find the frequent item set and association rules between different category data, so as to realize cross-category indexing.

[0027] Through cross-category indexing, the recommendation system obtains the multi-category index portrait elements of the respective category data of the multiple categories. The multi-category index portrait element refers to an element obtained through cross-category indexing and capable of reflecting the characteristics of a user at the level of multiple category data. These elements are the user's interest preferences, consumption habits, behavior patterns, etc. A young office woman, whose basic attribute data representation information is between 25-30 years old, and whose behavior monitoring data representation information is that she often clicks on fashion and beauty products and has purchased high-end skincare products. Through cross-category indexing, the recommendation system obtains the multi-category index portrait elements of the user, such as "young office woman, high interest in fashion and beauty, and certain consumption ability".

[0028] Step 300, based on the multi-category index portrait elements of the respective category data of the multiple categories, determining a plurality of pending portrait elements for the user data.

[0029] The multi-category index portrait elements are obtained by cross-category indexing the representation information of the category data of multiple categories, and these elements reflect the characteristics of the user at the level of different category data. However, these elements are numerous and have certain redundancy and uncertainty, so the recommendation system needs to screen and determine them to obtain more targeted and representative pending portrait elements.

[0030] The recommendation system analyzes the multi-category index portrait elements of each category data of multiple categories when determining multiple pending portrait elements. This involves evaluating the importance, relevance, and reliability of the elements. Importance evaluation refers to judging the importance of a portrait element in describing user characteristics and preferences. For a user of an e-commerce platform, the preference for purchasing high-end electronic products is more important than occasionally browsing home goods, as the former better reflects the user's main consumption tendency. Relevance evaluation refers to judging the degree of association between the portrait element and the user data. If the user's behavior monitoring data is mainly concentrated on clothing goods, then the portrait elements related to clothing have higher relevance. Reliability evaluation refers to judging the credibility and stability of the portrait element. Some portrait elements based on a small amount of data or accidental behavior have low reliability and need to be filtered and excluded. One method is a statistical analysis-based method that evaluates the importance and reliability of portrait elements by calculating statistical indicators such as frequency of occurrence and distribution. Count the number of times each portrait element appears in the multi-category index portrait elements, and elements with higher occurrence frequencies are more important and reliable. Another method is a machine learning-based method that uses classification algorithms and clustering algorithms to classify and cluster portrait elements, thereby discovering relationships and patterns between different portrait elements. The recommendation system uses the K-means clustering algorithm to divide portrait elements into different clusters, with elements in each cluster having similar characteristics. Then, according to the size and characteristics of the clusters, the importance and relevance of the elements in each cluster are evaluated.

[0031] The recommendation system also considers the relevance and complementarity between portrait elements. Some portrait elements may not be important on their own, but when combined with other elements, they provide more comprehensive and accurate user information. The two portrait elements "likes outdoor sports" and "buys sports equipment" individually reflect only part of the user's characteristics, but when combined, they more clearly depict the user's sports preferences and consumption behavior. The recommendation system discovers these complementary element combinations by mining association rules between portrait elements, thereby better determining pending portrait elements.

[0032] Step 400, based on the multiple pending portrait elements and the multiple categories of category data, portrait element reasoning is performed to obtain a user portrait element for the user data, and a target user portrait of the target user is generated based on the user portrait element;

[0033] The multiple pending portrait elements are determined based on the multi-category index portrait elements of each category data of multiple categories in step 300, and they represent the preliminary extracted user-related characteristic elements from different angles. The multiple categories of category data contain raw information such as user basic attribute data and behavior monitoring data. The recommendation system combines these two and uses portrait element reasoning to deeply mine the user's potential characteristics and needs.

[0034] Image element inference employs algorithms and models to discover hidden patterns and relationships from pending image elements and category data. The recommendation system uses machine learning algorithms such as decision trees, neural networks, etc. to analyze and model these data. The decision tree algorithm constructs a decision tree by dividing and classifying data, and infers the user's image elements based on different feature conditions. The neural network simulates the structure of neurons in the human brain to perform nonlinear mapping and learning on data, thereby uncovering complex patterns and relationships in the data.

[0035] During image element inference, the recommendation system first needs to perform feature fusion on multiple pending image elements and multiple categories of category data. Feature fusion refers to the integration of data features from different sources and types to form a more comprehensive and representative feature set. For basic attribute data and pending image elements, the recommendation system concatenates or weights the feature vectors to obtain a new fused feature vector. Assuming that the pending image elements have "like outdoor sports" and the basic attribute data have "age 30" and "male", the recommendation system converts these information into feature vectors and then fuses them to obtain a new vector containing more information.

[0036] Next, the recommendation system uses the fused features for image element inference. Taking the decision tree algorithm as an example, the recommendation system judges and selects at the nodes of the decision tree based on the feature values in the fused features, and finally reaches the leaf node. The label corresponding to the leaf node is the inferred user image element. Assuming that a branch of the decision tree is divided according to "age" and "whether to like outdoor sports", when the age in the fused features is 30 and likes outdoor sports, the corresponding branch is followed to infer that the user is an "active outdoor enthusiast" image element.

[0037] For neural network algorithms, the recommendation system takes the fused features as input, and through the forward propagation process of the neural network, it performs nonlinear transformation through multiple hidden layers and finally outputs the inferred user image element. The training process of the neural network is to continuously adjust the weights and biases of the network through a large amount of historical data, so that the output of the network is as close as possible to the true user image element.

[0038] Through image element inference, the recommendation system obtains user image elements for user data. These user image elements are more accurate and detailed descriptions of user characteristics and needs, which take into account pending image elements and multiple categories of category data. User image elements include "high-consumption fashion enthusiasts" and "health-conscious housewives", etc. These elements can more comprehensively reflect users' consumption preferences, lifestyles, and values.

[0039] After obtaining the user portrait elements, the recommendation system generates a target user portrait for the target user based on these elements. The target user portrait is a comprehensive description of the target user, which integrates and organizes multiple user portrait elements into a structured and intuitive user feature model. The recommendation system displays the target user portrait in a graphical manner, such as creating a user portrait card that contains the user's basic information, interests, consumption habits, and other key features.

[0040] Step 500, according to the target user portrait, the target user is recommended products.

[0041] The target user portrait is a comprehensive description of the target user's characteristics, needs, interests, and consumption habits. The recommendation system selects the most attractive products for the target user from the vast number of products on the e-commerce platform based on this portrait. To achieve this goal, the recommendation system uses recommendation algorithms and strategies to match and associate the features in the target user portrait with the attributes of the products.

[0042] The recommendation system uses a content-based recommendation algorithm. This algorithm mainly matches the content features of the products with the interest preferences in the user portrait. If the target user portrait shows that the user is a "young tech enthusiast with high interest and consumption willingness for smartphones and digital accessories", the recommendation system will select products that meet this feature from the product library of the e-commerce platform, such as new smartphones, high-performance tablets, and fashionable digital earphones. The recommendation system will analyze the attributes of these products, such as brand, model, performance parameters, and price, and then compare them with the needs in the user portrait to select the most suitable products for recommendation. The content-based recommendation algorithm uses similarity calculation to measure the matching degree of products and user portraits. The optional similarity measurement methods include cosine similarity. The recommendation system will sort the products according to the cosine similarity value, and the products with higher similarity values will be recommended to the user.

[0043] The recommendation system also uses a collaborative filtering recommendation algorithm. Collaborative filtering algorithms are divided into user-based collaborative filtering and item-based collaborative filtering. User-based collaborative filtering algorithm finds other users with similar interests and behaviors to the target user, and then recommends the products that these similar users like to the target user. Assuming that the target user portrait shows that the user likes outdoor sports equipment, the recommendation system will find other users who also like outdoor sports equipment in the user database, analyze the products that these similar users have recently purchased or collected, such as professional hiking shoes and outdoor tents, and then recommend these products to the target user. Item-based collaborative filtering algorithm recommends products based on the similarity between products. If the target user purchases a sports watch, the recommendation system will analyze other products similar to this sports watch, such as other watches with the same functions or brands, and then recommend these similar products to the user.

[0044] In addition to the two optional recommendation algorithms mentioned above, the recommendation system also combines multiple algorithms for hybrid recommendation to improve the accuracy and diversity of recommendations. The recommendation system first uses a content-based recommendation algorithm to filter out a product set that matches the target user portrait, and then uses a collaborative filtering algorithm to further filter and sort this set, resulting in more accurate recommendation results.

[0045] When making product recommendations, the recommendation system also considers real-time information and dynamic changes of products. The information of product price, inventory, promotion activities on e-commerce platforms will be updated constantly, and the recommendation system will obtain these information in time and take it into consideration in the recommendation decision-making range. If a product is undergoing a time-limited discount activity, the recommendation system will preferentially recommend it to the target user to improve the user's purchase willingness.

[0046] The recommendation system also needs to make personalized recommendations according to different scenarios and usage habits of target users. For users browsing e-commerce platforms on computers, the recommendation system displays recommended products in a list format, providing detailed product information and pictures; for users browsing on mobile devices, the recommendation system displays recommended products in a card or waterfall style to adapt to the screen characteristics and user operation habits of mobile devices.

[0047] As an implementation manner, based on the category data of the plurality of categories, each cross-category index is performed on the level of the representation information of the plurality of categories, to obtain a plurality of category index portrait elements of the category data of the plurality of categories, including:

[0048] Step 210, determining a portrait element mapping relationship set; the portrait element mapping relationship set includes the contrast category representation information of the contrast user data under the plurality of categories and the contrast portrait element corresponding to the contrast category representation information respectively;

[0049] Step 220, based on the commonality measurement result between the category data of the plurality of categories and the contrast category representation information of the contrast user data under the plurality of categories, indexing in the portrait element mapping relationship set to obtain a plurality of category index portrait elements of the category data of the plurality of categories.

[0050] In step 210, the recommendation system first needs to determine the portrait element mapping relationship set, which is the basis for subsequent cross-category indexing. The portrait element mapping relationship set is a collection that stores the corresponding relationship between the reference user data and the reference portrait element, where the reference category representation information of the reference user data in multiple categories is regarded as a known feature pattern, and the reference portrait element is a user feature description corresponding to these feature patterns. In order to construct this mapping relationship set, the recommendation system collects a large amount of historical user data and classifies and labels these data. The recommendation system extracts the basic attribute data and behavior monitoring data of numerous users from the database of the e-commerce platform, divides these data according to different categories such as age, occupation, and consumption habit, then analyzes the data features in each category, groups the data with similar features, and labels each group of data with the corresponding reference portrait element. For example, for the user group whose age is between 20-30 years old, whose occupation is student or newly employed, and who often buys fashion trendy goods, the recommendation system labels the corresponding reference portrait element as "young fashion consumer group".

[0051] When constructing the portrait element mapping relationship set, the recommendation system needs to use data processing and analysis techniques. The recommendation system uses data mining algorithms to perform clustering analysis on historical user data, groups similar user data into different clusters, and each cluster represents a typical user feature pattern. The recommendation system uses the K-means clustering algorithm, which assigns data points to different clusters through iteration, so that the similarity of data points within a cluster is the highest, and the similarity of data points between clusters is the lowest. After clustering, the recommendation system labels each cluster with the corresponding reference portrait element, thereby forming the portrait element mapping relationship set.

[0052] In step 220, the commonality measure refers to the process of measuring the similarity between two kinds of data. The recommendation system calculates the similarity between the category data of multiple categories and the reference category representation information, finds the reference category representation information that is most similar to the current category data, and then determines the corresponding reference portrait element as the multi-category indexing portrait element. Optional methods include Euclidean distance, cosine similarity, etc.

[0053] The recommendation system obtains the similarity values between the category data of the multiple categories and the control category representation information after performing the commonality measurement. The recommendation system indexes the portrait element mapping relationship set according to the similarity values, finds the control category representation information with the highest similarity to the current category data, and then takes the control portrait element corresponding to the control category representation information as the multi-category index portrait element of the current category data. For example, the recommendation system analyzes the basic attribute data and the behavior monitoring data of a user, finds that the data of the user has the highest similarity to the control category representation information corresponding to the “young fashion consumer group” in the portrait element mapping relationship set by calculating the similarity, and then takes the “young fashion consumer group” as the multi-category index portrait element of the data of the user.

[0054] As an implementation, the indexing of the portrait element mapping relationship set based on the commonality measurement results between the category data of the multiple categories and the control category representation information of the control user data under the multiple categories to obtain the multi-category index portrait element of each category data of the multiple categories includes:

[0055] In step 221, the commonality measurement is performed on each category data of the multiple categories and the control category representation information under the multiple categories to obtain the commonality measurement result of each category data.

[0056] In step 222, the target measurement result meeting the commonality requirement is determined for each category data from the commonality measurement result of each category data.

[0057] In step 223, the multi-category index portrait element of each category data is respectively indexed in the portrait element mapping relationship set based on the control category representation information corresponding to the target measurement result of each category data to obtain the multi-category index portrait element of each category data.

[0058] In step 221, the category data of the multiple categories has been determined in the previous steps, for example, the basic attribute data and the behavior monitoring data of the user, and the control category representation information is stored in the portrait element mapping relationship set, which is the feature representation of the control user data under the multiple categories. The optional method includes Manhattan distance. Manhattan distance refers to the sum of the distances between two points in a multi-dimensional space along the coordinate axis direction, and the calculation formula is: The commonality measurement result reflects the similarity between each category data and each control category representation information by performing the commonality measurement on each category data.

[0059] In step 222, since each category data is measured for commonality with multiple comparison category representation information in step 221, multiple similarity values are obtained, and the recommendation system screens the similarity values meeting the commonality requirement as the target measurement result. The commonality requirement is set according to specific business requirements and data characteristics, and a similarity threshold is set. Only the measurement result with a similarity value greater than the threshold is considered to meet the commonality requirement.

[0060] In this way, the recommendation system screens the measurement result corresponding to the comparison category representation information most similar to the current category data from a large number of commonality measurement results, providing an accurate basis for subsequent indexing operations.

[0061] In step 223, the portrait element mapping relationship set stores the correspondence between the comparison category representation information and the comparison portrait element, and the recommendation system finds the corresponding comparison portrait element in this relationship set according to the comparison category representation information corresponding to the target measurement result. These comparison portrait elements are the multi-category index portrait elements of each category data.

[0062] As an implementation manner, the multiple categories include a monitoring data category and an attribute data category; the category data of the multiple categories includes behavior monitoring data and basic attribute data of the user data; the portrait element mapping relationship set includes behavior comparison category representation information of comparison user data in the monitoring data category, a behavior comparison portrait element corresponding to the behavior comparison category representation information, attribute comparison category representation information of the comparison user data in the attribute data category, and an attribute comparison portrait element corresponding to the attribute comparison category representation information;

[0063] In step 220, the commonality measurement result between the category data of the multiple categories and the comparison category representation information of the comparison user data in the multiple categories is indexed in the portrait element mapping relationship set to obtain the multi-category index portrait element of the category data of the multiple categories, including:

[0064] Step 10, respectively perform representation information extraction based on the behavior monitoring data and the basic attribute data to obtain behavior data representation and attribute data representation;

[0065] Step 20, respectively perform behavior representation commonality measurement on the behavior data representation and the behavior comparison category representation information and the attribute comparison category representation information, and obtain the multi-category index portrait element of the behavior monitoring data in the behavior comparison portrait element and the attribute comparison portrait element based on the behavior representation commonality measurement result;

[0066] Step 30, the attribute data representation is respectively with the behavior control category representation information and the attribute control category representation information, and the attribute representation commonality is measured, and the multi-category index portrait element of the basic attribute data is obtained in the behavior control portrait element and the attribute control portrait element based on the attribute representation commonality measurement result.

[0067] In step 10, the behavior monitoring data refers to the operation behavior records of users on the e-commerce platform, such as clicking, collecting, purchasing and the like of goods; the basic attribute data is some inherent characteristic information of the user, such as age, gender, occupation and the like. The representation information extraction is the process of converting the original data into more representative and computational feature representation, which can reflect the core features and internal laws of the data, and provide a basis for subsequent commonality measurement and portrait element index.

[0068] For the representation information extraction of behavior monitoring data, an optional method adopted by the recommendation system is to convert the behavior data into vector representation. The recommendation system counts the number of clicks, the number of collections and the number of purchases of different categories of goods by the user within a certain period of time, and forms a vector with these statistical values. Assuming that the user clicks on mobile phone class goods 5 times, computer class goods 3 times and home appliance class goods 2 times in a week, the behavior data representation is represented as vector (5, 3, 2). The recommendation system also considers the time sequence and frequency distribution of behavior and other factors to extract more complex features from the behavior data. For example, using time series analysis method, the user's behavior is arranged in time sequence, and the periodicity, trend and other features of the behavior are extracted.

[0069] For the representation information extraction of basic attribute data, the recommendation system encodes and quantizes the attribute data. For the age attribute, it is divided into different age groups, such as 18-25 years old, 26-35 years old, etc., and then the corresponding encoding is used to represent each age group; for the occupation attribute, an occupation classification system is established, and different occupations are encoded. Assuming that the user is 28 years old and the occupation is programmer, after encoding, the attribute data representation is represented as vector (2, 5), where 2 represents the encoding of 26-35 years old age group and 5 represents the encoding of programmer occupation.

[0070] In step 20, the behavior control category representation information is the control feature representation related to the behavior data in the portrait element mapping relationship set, and the attribute control category representation information is the control feature representation related to the attribute data. The behavior representation commonality measurement is a process of measuring the similarity between the behavior data representation and the control category representation information, and by calculating the similarity, the recommendation system finds out the control category representation information most similar to the behavior data representation, and then determines the corresponding control portrait element.

[0071] In step 30, the process of attribute characterization commonality measurement is similar to the process of behavior characterization commonality measurement, and the most similar control category characterization information to the attribute data characterization is found by calculating the similarity, and the corresponding control portrait element is determined.

[0072] As an implementation, the behavior data characterization is respectively behavior characterization commonality measured with the behavior control category characterization information and the attribute control category characterization information, and the multi-category index portrait element of the behavior monitoring data is obtained in the behavior control portrait element and the attribute control portrait element based on the behavior characterization commonality measurement result.

[0073] In step 21, the behavior data characterization is behavior characterization commonality measured with the behavior control category characterization information, and the first behavior characterization commonality measurement result is obtained.

[0074] In step 22, the first behavior category index portrait element is determined in the behavior control portrait element based on the behavior control category characterization information corresponding to the first behavior characterization commonality measurement result representing the commonality requirement.

[0075] In step 23, the behavior data characterization is behavior characterization commonality measured with the attribute control category characterization information, and the second behavior characterization commonality measurement result is obtained.

[0076] In step 24, the first attribute category index portrait element is determined in the attribute control portrait element based on the attribute control category characterization information corresponding to the second behavior characterization commonality measurement result representing the commonality requirement.

[0077] In step 25, the multi-category index portrait element of the behavior monitoring data is obtained based on the first behavior category index portrait element and the first attribute category index portrait element.

[0078] In step 21, the behavior data characterization is the result obtained after the characterization information extraction of the user behavior monitoring data, which reflects the behavior characteristics of the user in a form convenient for computer processing and analysis. The behavior control category characterization information is the behavior control category characterization information of the portrait element mapping relationship set representing a group of representative behavior patterns. The behavior characterization commonality measurement is an operation for measuring the similarity between the behavior data characterization and the behavior control category characterization information, and the most similar behavior pattern to the current behavior of the user is found through this measurement.

[0079] In step 22, meeting the commonality requirement means that the similarity score reaches a certain threshold, which is set according to business needs and data characteristics. The behavior-contrast portrait element is a portrait element in the portrait element mapping relationship set corresponding to the behavior-contrast category representation information, which describes the characteristics and preferences of users with corresponding behavior patterns.

[0080] In step 23, the attribute-contrast category representation information is the attribute-contrast category representation information in the portrait element mapping relationship set corresponding to the attribute data category of the user data, representing a set of representative user attribute characteristics. Although the behavior data representation mainly reflects the behavior information of the user, by measuring the commonality with the attribute-contrast category representation information, the potential association between user behavior and attributes is discovered. If the data distribution of the attribute-contrast category representation information is relatively discrete, a measurement method more suitable for processing discrete data is selected, such as Hamming distance. Hamming distance is used to measure the number of different characters at corresponding positions between two equal-length strings, and is more effective when processing attribute-contrast category representation information in the form of binary vectors.

[0081] In step 24, similar to step 22, meeting the commonality requirement means that the similarity score reaches a preset threshold. The attribute-contrast portrait element is a portrait element in the portrait element mapping relationship set corresponding to the attribute-contrast category representation information, which describes the characteristics of users with corresponding attribute characteristics.

[0082] In step 25, this step is to integrate the index portrait elements obtained from the behavior and attribute angles to form a more comprehensive and accurate multi-category index portrait element of the behavior monitoring data.

[0083] In the implementation of step 25, the recommendation system uses a splicing method to integrate the first behavior category index portrait element and the first attribute category index portrait element. However, in actual application, the integration method needs to be optimized according to business needs. For example, different weights are assigned according to the importance of different portrait elements, and then weighted combination is performed.

[0084] Steps 21-25 measure the commonality of the behavior data representation with the behavior-contrast category representation information and the attribute-contrast category representation information respectively, determine the corresponding index portrait elements, and finally integrate to obtain the multi-category index portrait element of the behavior monitoring data. The recommendation system can deeply analyze the behavior characteristics of users from multiple angles and mine the potential association between user behavior and attributes.

[0085] As an implementation, the attribute data representation is respectively compared with the behavior comparison category representation information and the attribute comparison category representation information for attribute representation commonality measurement, and a multi-category index image element of the basic attribute data is obtained in the behavior comparison image element and the attribute comparison image element based on the attribute representation commonality measurement result, including:

[0086] Step 31, attribute representation commonality measurement is performed on the attribute data representation and the behavior comparison category representation information to obtain a first attribute representation commonality measurement result;

[0087] Step 32, based on the behavior comparison category representation information corresponding to the first attribute representation commonality measurement result meeting the commonality requirement, a second behavior category index image element is determined in the behavior comparison image element;

[0088] Step 33, attribute representation commonality measurement is performed on the attribute data representation and the attribute comparison category representation information to obtain a second attribute representation commonality measurement result;

[0089] Step 34, based on the attribute comparison category representation information corresponding to the second attribute representation commonality measurement result meeting the commonality requirement, a second attribute category index image element is determined in the attribute comparison image element;

[0090] Step 35, based on the second behavior category index image element and the second attribute category index image element, a multi-category index image element of the basic attribute data is obtained.

[0091] In step 31, the attribute data representation is a representation form obtained after feature extraction and conversion of the basic attribute data of the user, which can reflect the core features of the basic attribute of the user. The behavior comparison category representation information is a comparison feature representation of the image element mapping relationship related to the user behavior, which contains typical features in different behavior modes. The purpose of attribute representation commonality measurement is to measure the similarity between the attribute data representation and the behavior comparison category representation information, and to find the potential relationship between the basic attribute of the user and some behavior modes through the measurement.

[0092] In step 32, the commonality requirement is a pre-set similarity threshold. Only when the similarity value in the first attribute representation commonality measurement result is greater than the threshold, it is considered that the corresponding behavior comparison category representation information has sufficient similarity with the attribute data representation, meeting the commonality requirement.

[0093] In step 33, the attribute comparison category representation information is a comparison feature representation related to the user basic attribute in the portrait element mapping relationship set, which covers the typical features under different attribute combinations. This step is similar to step 31, and the similarity measurement method is also used to measure the similarity between the attribute data representation and the attribute comparison category representation information. However, the comparison information here is related to the attribute.

[0094] In step 34, similar to step 32, a similarity threshold is set as the judgment standard of the commonality requirement.

[0095] In step 35, this step is the integration of the results of the previous steps. The features related to the basic attribute data extracted from the behavior and attribute are combined to obtain a more comprehensive and accurate multi-category index portrait element of the basic attribute data.

[0096] As another implementation, in step 10, the process of performing representation information extraction based on the behavior monitoring data and the basic attribute data respectively to obtain behavior data representation and attribute data representation is completed by a multi-category representation analysis network, and the multi-category representation analysis network is obtained by the following process:

[0097] Step a, obtaining training binary tuple data, the training binary tuple data including behavior monitoring training data and basic attribute training data;

[0098] Step b, performing representation information extraction on the behavior monitoring training data and the basic attribute training data respectively by the initialized initial multi-category representation analysis network to obtain behavior data representation of the behavior monitoring training data and attribute data representation of the basic attribute training data;

[0099] Step c, based on the behavior data representation and the attribute data representation, iteratively debugging the initial multi-category representation analysis network until a preset debugging stop condition is reached to obtain the multi-category representation analysis network.

[0100] In step a, the behavior monitoring training data refers to the relevant information collected from a large number of historical user behaviors, covering various operation behaviors of users on the e-commerce platform, such as clicking, collecting, purchasing, forwarding, etc. These behavior data can reflect the user's interest preferences and consumption habits. The basic attribute training data is some inherent feature information about the user, such as age, gender, occupation, region, etc. These attribute data provide a basis for understanding the basic background and characteristics of the user.

[0101] In step b, this process is regarded as the process of feature extraction and transformation of data by the network, and the network tries to mine the information most representing the characteristics of the data from the original data. The initial multi-class representation analysis network is a neural network model with a pre-defined structure, and its parameters are randomly initialized before the training starts.

[0102] In step c, the purpose of iterative debugging is to adjust the parameters of the network so that the behavioral data representation and attribute data representation output by the network can better reflect the characteristics and inherent laws of the input data.

[0103] The recommendation system uses a loss function to measure the difference between the representation information output by the network and the expected result. The optional loss function is the mean square error loss function (MSE), and its calculation formula is where n is the number of samples, is the expected result, is the output result of the network.

[0104] In each iteration, the recommendation system first calculates the loss value between the behavioral data representation and attribute data representation output by the current network and the expected result. The expected result is the ideal representation information obtained according to prior knowledge or expert annotation. Then, the recommendation system uses an optimization algorithm to adjust the parameters of the network to reduce the loss value. The optional optimization algorithm is stochastic gradient descent, adaptive moment estimation, etc. The update formula of the stochastic gradient descent algorithm (SGD) is where is the network parameter at the tthiteration, is the learning rate, is the gradient of the loss function with respect to the parameter .

[0105] The recommendation system repeatedly iterates this process, and each iteration updates the parameters of the network in the direction of reducing the loss value. When the loss value is reduced to below the preset threshold, or the number of iterations reaches the preset maximum number, the recommendation system considers that the debugging stop condition is reached, and the network obtained at this time is the trained multi-class representation analysis network.

[0106] As an implementation, the plurality of categories includes a monitoring data category and an attribute data category; the category data of the plurality of categories includes behavioral monitoring data and basic attribute data;

[0107] The portrait element reasoning based on the plurality of pending portrait elements and the category data of the plurality of categories to obtain a user portrait element for the user data, comprising:

[0108] Step 410, feature fusion of the basic attribute data and the plurality of pending portrait elements to obtain first fusion representation information; and feature fusion of the behavior monitoring data and the plurality of pending portrait elements to obtain second fusion representation information;

[0109] Step 420, portrait element reasoning based on the first fusion representation information and the second fusion representation information to obtain a user portrait element for the user data.

[0110] In step 410, feature fusion refers to integrating data features of different sources or types to form a more comprehensive and representative feature set, so as to more accurately reflect the user's characteristics and needs.

[0111] The recommendation system performs feature fusion on the basic attribute data and the plurality of pending portrait elements, aiming to combine the user's basic characteristics with the preliminary extracted feature elements to mine more in-depth and comprehensive user feature information. For feature fusion of the basic attribute data and the pending portrait elements, the recommendation system adopts multiple methods. One optional method is to extract the representation information respectively and then splice the representation information. The recommendation system converts the age, gender, occupation, etc. information in the basic attribute data into a feature vector, converts the pending portrait elements into a corresponding feature vector, and then splices the two vectors together to form a new vector, which is the first fusion representation information. Assuming that the user's age in the basic attribute data is 30 years old, the gender is male, and the occupation is engineer, the converted feature vector is (30, 1, 2) (where 1 represents male and 2 represents engineer occupation code); the pending portrait elements are "technology product enthusiasts" and "high consumption ability", and the converted feature vector is (1, 1) (assuming 1 represents having this feature), then the first fusion representation information is the spliced vector (30, 1, 2, 1, 1).

[0112] The recommendation system performs feature fusion on the behavior monitoring data and the plurality of pending portrait elements, aiming to combine the user's behavior characteristics with the preliminary extracted feature elements to further mine the user's interests and needs. For feature fusion of the behavior monitoring data and the pending portrait elements, the same splicing method is adopted. The recommendation system converts the click count, collection commodity list, purchase commodity details, etc. information in the behavior monitoring data into a feature vector, converts the pending portrait elements into a corresponding feature vector, and then splices the two vectors together to form the second fusion representation information.

[0113] In step 420, portrait element reasoning requires the recommendation system to use algorithms and models to discover hidden rules and relationships from the fused feature information, so as to infer user portrait elements that can accurately describe the user's characteristics and needs.

[0114] As an implementation form, the portrait element inference based on the first fusion feature information and the second fusion feature information obtains a user portrait element for the user data, including:

[0115] In step 421, the first fusion feature information is standardized and mapped by the cross-modal standardization mapping network to obtain first standardized and mapped fusion feature information.

[0116] In step 422, the second fusion feature information is standardized and mapped by the cross-modal standardization mapping network to obtain second standardized and mapped fusion feature information.

[0117] In step 423, the portrait element inference network is used to perform portrait element inference based on the first standardized and mapped fusion feature information and the second standardized and mapped fusion feature information to obtain a user portrait element for the user data.

[0118] In step 421, the first fusion feature information is obtained by feature fusion of the basic attribute data and the plurality of pending portrait elements in step 410, which integrates the basic features of the user and the preliminary extracted feature elements. However, since the data from different sources has different feature distributions and scales, directly using these information for portrait element inference will result in poor model performance. The role of the cross-modal standardization mapping network is to process the first fusion feature information to make its feature distribution and scale more uniform, which is convenient for subsequent network processing.

[0119] The cross-modal standardization mapping network is a trained neural network model that learns the mapping relationship between different modal data and converts the input first fusion feature information into a form more suitable for portrait element inference network processing. The recommendation system inputs the first fusion feature information into the cross-modal standardization mapping network, and the network will transform the input information according to its internal parameters. This transformation is regarded as a process of normalizing and aligning data, so that different features have comparability.

[0120] For example, suppose the first fusion feature information includes the user's age, occupation code, and feature vector corresponding to the pending portrait element. The age value range is between 18-80 years old, the occupation code is an integer between 1-100, and the value of the feature vector corresponding to the pending portrait element is between 0-1. The scales of these different features are quite different. If they are directly input into the portrait element inference network, the influence of some features will be too large, while the influence of other features will be ignored. The cross-modal standardization mapping network will process these features and map them to a unified feature space, so that each feature can play a reasonable role in the inference process.

[0121] The structure of the network is a multi-layer perceptron (MLP) including an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is determined according to the dimension of the first fused feature information, and the number of neurons in the output layer is determined according to the input requirements of the subsequent portrait element inference network. The recommendation system uses techniques such as batch normalization (Batch Normalization) to accelerate the training of the network and improve stability. The formula of batch normalization is where is the input data, is the mean of the small batch data, is the variance of the small batch data, is a small constant to avoid division by zero error. After batch normalization, the mean of the data is 0 and the variance is 1, which helps the network converge faster.

[0122] In step 422, the second fused feature information is obtained by feature fusion of the behavior monitoring data and the plurality of pending portrait elements, which reflects the user's behavior characteristics and the preliminary extracted feature elements. Similar to the first fused feature information, the second fused feature information also has the problem of different feature scales and inconsistent distribution, so it needs to be processed by the cross-modal canonical mapping network.

[0123] The recommendation system inputs the second fused feature information into the trained cross-modal canonical mapping network, which will transform it according to the same mapping rule and convert it into a representation in a unified feature space. For example, the second fused feature information includes the number of clicks, the number of collected goods, and the feature vector corresponding to the pending portrait elements. The value range of the number of clicks and the number of collected goods is very different, and through the processing of the cross-modal canonical mapping network, these features will be mapped to a scale, so that they can play a more fair role in the subsequent portrait element inference.

[0124] The training of the cross-modal canonical mapping network is based on a large amount of historical data. The recommendation system uses the first training behavior data representation of the first training user data, the first training attribute data representation, and the prior user portrait to debug and obtain the transition cross-modal canonical mapping network. In this process, the network learns the mapping relationship between different modal data and can effectively perform canonical mapping on the input information. For the canonical mapping of the second fused feature information, the recommendation system directly uses the trained cross-modal canonical mapping network to ensure the consistency and accuracy of the processing.

[0125] In step 423, the portrait element inference network is a neural network model specially used for inferring user portrait elements from the norm-mapped feature information. It receives the first norm-mapped fused representation information and the second norm-mapped fused representation information as inputs, and outputs user portrait elements that can accurately describe the user's characteristics and needs through a series of calculations and transformations.

[0126] In training the portrait element inference network, the recommendation system uses the second training behavior norm-mapped representation information, the second training attribute norm-mapped representation information, and the prior portrait elements of the second training user data. The prior portrait elements are portrait elements that accurately describe the user's characteristics and are used as labels for training. The recommendation system inputs the second training behavior norm-mapped representation information and the second training attribute norm-mapped representation information into the initialized initial portrait element inference network, and the network outputs the inferred portrait elements. Then, the recommendation system optimizes the transition cross-modal norm-mapping network and the initial portrait element inference network based on the inferred portrait elements and the prior portrait elements, and then performs collaborative debugging again until the preset debugging stopping condition is reached, obtaining the converged cross-modal norm-mapping network and portrait element inference network.

[0127] When the portrait element inference network is trained, the recommendation system inputs the first norm-mapped fused representation information and the second norm-mapped fused representation information of the new user into the trained network, and the network outputs the user portrait elements for the user data.

[0128] As another implementation, the cross-modal norm-mapping network and the portrait element inference network are trained by the following steps:

[0129] Step S10, based on the first training behavior data representation, the first training attribute data representation, and the prior user portrait of the first training user data, the transition cross-modal norm-mapping network is obtained by debugging, wherein the first training behavior data representation is obtained by representing information mining on the behavior monitoring data of the first training user data; and the first training attribute data representation is obtained by representing information mining on the basic attribute data of the first training user data.

[0130] The first training user data is a part of data selected by the recommendation system from a large amount of user data for training, which contains the user's behavior monitoring data and basic attribute data. The first training behavior data representation is obtained by representing information mining on the behavior monitoring data in the first training user data, which involves converting the user's behavior data into a more representative and computational feature vector.

[0131] The first training attribute data representation is obtained by performing representation information mining on the basic attribute data of the first training user data. The basic attribute data includes information such as age, gender, occupation, and region of the user. The recommendation system encodes and quantizes these attribute data and converts them into feature vectors. For the age attribute, different age ranges such as 18-25 years old, 26-35 years old, etc. are divided, and each age range is represented by a corresponding code; for the occupation attribute, an occupation classification system is established to encode different occupations. The prior user portrait is a priori known accurate description of the features of the user corresponding to the first training user data, which is obtained based on historical data, expert experience or other reliable sources. The prior user portrait contains information about the user's interests, consumption habits, values, etc.

[0132] The transition cross-modal canonical mapping network is a network that has not been fully trained, and its function is to perform canonical mapping on data of different modalities, so that these data have better alignment and comparability in the feature space, facilitating subsequent portrait element reasoning. When debugging the transition cross-modal canonical mapping network, the recommendation system continuously adjusts the parameters of the network so that the mapping results of the first training behavior data representation and the first training attribute data representation are closer to the features reflected by the prior user portrait.

[0133] When debugging the transition cross-modal canonical mapping network, the recommendation system inputs the first training behavior data representation and the first training attribute data representation into the initialized initial cross-modal canonical mapping network, and the network performs mapping processing on the input data to obtain the mapped representation information. Then, the recommendation system uses an initialized initial portrait element reasoning network to perform user portrait prediction on the mapped representation information to obtain a predicted user portrait for the first training user data. There is a certain difference between the predicted user portrait and the prior user portrait, and the recommendation system uses a loss function to measure this difference.

[0134] The recommendation system uses an optimization algorithm (such as stochastic gradient descent or adaptive moment estimation) to adjust the parameters of the initial cross-modal canonical mapping network according to the value of the loss function, so that the value of the loss function continuously decreases.

[0135] In step S20, the second training behavior data representation of the second training user data is obtained by performing representation information mining on the behavior monitoring data of the second training user data. The second training attribute data representation of the second training user data is obtained by performing representation information mining on the basic attribute data of the second training user data.

[0136] For obtaining the second training behavior data representation, the recommendation system processes and analyzes the user's behavior monitoring data using natural language processing techniques. For example, if the behavior monitoring data is the user's review of a product, the recommendation system uses word embedding techniques to convert each word in the review into a vector, and then processes these vectors through a deep learning model (such as a long short-term memory network LSTM) to obtain the representation of the entire review. In addition, the recommendation system also uses feature engineering methods to extract some statistical features from the behavior monitoring data, such as the number of browsed products, the frequency of clicking products, etc., and then combines these features into a vector as the second training behavior data representation. Suppose the user's behavior monitoring data is a list containing multiple product browsing records, each record including the name of the product, the browsing time, etc. The recommendation system calculates the number of products browsed by the user in different time periods, as well as the browsing proportion of different types of products, and combines these statistical information as features into a vector. The calculation formula is represented as: where represents the value of the i-th feature, for example is the number of products browsed in the past day, is the browsing proportion of electronic products, etc.

[0137] For obtaining the second training attribute data representation, the recommendation system processes the basic attribute data using one-hot encoding or embedding techniques. One-hot encoding is to convert each attribute value into a binary vector, with only one element being 1 and the rest being 0, representing the value of the attribute. For example, for the user's gender attribute, male is represented as [1, 0] and female is represented as [0, 1]. Embedding techniques map attribute values to a low-dimensional vector space, so that vectors with similar attribute values are closer in space. For example, for the user's age attribute, use the embedding layer to map the age value to a 10-dimensional vector space. The recommendation system combines these processed attribute vectors to obtain the second training attribute data representation. Suppose the user's basic attribute data includes age, gender and occupation, after one-hot encoding and embedding processing, vectors A, G and O are obtained respectively, then the second training attribute data representation P is represented as P = [A; G; O], where ";" represents the concatenation operation of vectors.

[0138] For obtaining prior image elements, the recommendation system refers to historical data or expert knowledge. Historical data refers to user data collected and processed by the recommendation system in the past. Through analysis and mining of these data, some user behavior patterns and preference rules are discovered. For example, by analyzing the shopping data of a large number of young male programmers, it is found that they generally have a high demand for electronic products, especially high-performance computers and mobile phones. Expert knowledge is the experience and judgment of professionals in the field. They make predictions and judgments on user image elements based on their professional knowledge and experience. The recommendation system combines historical data and expert knowledge to determine prior image elements. For example, the recommendation system uses Bayes' theorem to integrate historical data and expert knowledge. Assuming H represents a certain image element of the user (such as interest in mobile phones), D represents historical data, and E represents expert knowledge, then according to Bayes' theorem, the probability of the prior image element is: where P(H) is the prior probability, P(D, E|H) is the likelihood probability, and P(D, E) is the evidence probability. The recommendation system determines the prior image element by calculating this probability.

[0139] In step S30, the second training behavior data representation is standardized and mapped by the transition cross-modal standard mapping network to obtain second training behavior standard mapping representation information; and the second training attribute data representation is standardized and mapped by the transition cross-modal standard mapping network to obtain second training attribute standard mapping representation information.

[0140] When implementing step S30, the recommendation system uses a fully connected layer to perform linear transformation on the input representation information, and then introduces a non-linear factor through an activation function to map the input data to a new feature space. Assuming that the second training behavior data representation is a vector, the second training attribute data representation is a vector, and the fully connected layer in the transition cross-modal standard mapping network is represented as Y = WX + b, where W is a weight matrix, b is a bias vector, X is or Y is the output vector after standard mapping. The activation function selects the ReLU (Rectified Linear Unit) function, whose formula is By processing the output of the fully connected layer through the ReLU function, the expression ability of the network is enhanced. Another technical means is to perform batch normalization on the input representation information during the standard mapping process, so that the data has a similar distribution at each layer input.

[0141] The recommendation system also needs to consider the characteristics of different modal data when performing canonical mapping. For the second training behavior data representation, since it reflects the user's behavior information, it has high dynamics and diversity. For example, the user's behavior will change over time, promotion activities, and other factors. Therefore, in the canonical mapping process, the time factor is introduced, and the time series analysis method is used to process the behavior data representation. The behavior data representation is regarded as a time series, and a recurrent neural network (RNN) or its variants (such as LSTM, GRU) is used to model it to capture the temporal dependencies in the behavior data. When processing the second training attribute data representation, since the attribute data is relatively stable, such as the user's age, gender, and other information that will not change in the short term, the focus is on feature selection and dimension compression of the data. Principal component analysis (PCA) and other methods are used to reduce the dimensionality of the attribute data representation, remove redundant information in the data, and improve the computational efficiency of the model.

[0142] In practical applications, the recommendation system may encounter problems such as large data volume and limited computing resources. To solve these problems, the recommendation system uses distributed computing technology. The computing tasks of the transition cross-modal canonical mapping network are distributed to multiple computing nodes for parallel processing, thereby improving the computing speed. The recommendation system also uses model compression techniques such as pruning, quantization, and other methods to compress the transition cross-modal canonical mapping network, reducing the number of model parameters and lowering the computational complexity.

[0143] Step S40, infer the portrait element based on the second training behavior canonical mapping representation information and the second training attribute canonical mapping representation information through the initialized initial portrait element inference network, and obtain the inferred portrait element for the second training user data;

[0144] When implementing step S40, one optional technical means used by the recommendation system is to use a multi-layer perceptron. The multi-layer perceptron is a feedforward neural network composed of an input layer, a hidden layer, and an output layer. The recommendation system concatenates the second training behavior canonical mapping representation information and the second training attribute canonical mapping representation information into an input vector, which is input to the input layer of the multi-layer perceptron. Assuming that the second training behavior canonical mapping representation information is a vector with dimension m, and the second training attribute canonical mapping representation information is a vector with dimension n, then the input vector X = [ ; ] with dimension m + n. The hidden layer of the multi-layer perceptron processes the input vector through a series of linear transformations and nonlinear activation functions. The nonlinear activation function is selected, for example, the Sigmoid function or ReLU function, etc., to introduce non-linear factors and enhance the expression ability of the network. After processing through multiple hidden layers, the output layer outputs the estimated image element vector Y, whose dimension depends on the number of image elements that need to be inferred.

[0145] Another technical means is to use a convolutional neural network. If the second training behavior specification mapping representation information and the second training attribute specification mapping representation information are represented as a two-dimensional or multi-dimensional data structure, such as arranging behavior data in a matrix form according to a time sequence, the recommendation system uses a convolutional neural network for image element inference. The convolutional neural network processes the input data through convolutional layers, pooling layers, and fully connected layers. The convolutional layer uses a convolution kernel to perform sliding convolution operations on the input data to extract local features of the data. Assuming that the input data is X, the convolution kernel is K, and the convolution operation is represented as Y = X * K, where * represents the convolution operation. The pooling layer is used to down-sample the output of the convolutional layer to reduce the dimension of the data while retaining important feature information. The fully connected layer linearly transforms the output of the pooling layer to finally output the estimated image element.

[0146] When performing image element inference, the recommendation system also needs to consider the interaction between different features. An attention mechanism is used to automatically assign weights to different features, so that the network pays more attention to important features. The attention mechanism calculates the attention score of each feature and then performs weighted summation of the features according to the score. Assuming that the input features are , the attention score is calculated by an attention function, such as , where represents the attention score of the i-th feature. Then the attention score is normalized by the Softmax function to obtain the attention weight . Finally, the features are weighted and summed according to the attention weight to obtain the weighted feature representation .

[0147] In practical applications, the recommendation system will face problems of data noise and overfitting. To solve these problems, the recommendation system adopts regularization methods such as L1 and L2 regularization. L2 regularization constrains the complexity of the model by adding the sum of squares of weights in the loss function, and the loss function is represented as , where L is the original loss function, is the regularization parameter, is the weight of the model. The recommendation system also uses the Dropout technique to randomly discard some neurons during training to prevent over-reliance between neurons, thereby improving the generalization ability of the model. Specifically, in each training batch, the output of some neurons is set to 0 with a certain probability p, which forces the network to learn more robust features.

[0148] Step S50, based on the inferred image element and the prior image element, the transition cross-modal canonical mapping network and the initial image element inference network are optimized and co-optimized again until the preset debugging stop condition is reached, and the converged cross-modal canonical mapping network and image element inference network are obtained.

[0149] When implementing step S50, the recommendation system uses the following technical means to complete the optimization and co-optimization of the network. First, the recommendation system defines a loss function to measure the difference between the inferred image element and the prior image element. For example, the mean square error loss function is used to measure the average square error between two vectors, and the smaller the error, the closer the inferred image element is to the prior image element.

[0150] The recommendation system uses the gradient descent algorithm to minimize the loss function. Gradient descent algorithm is an iterative optimization algorithm, which calculates the gradient of the loss function with respect to the network parameters, and then updates the network parameters in the opposite direction of the gradient, so that the loss function gradually decreases.

[0151] During optimization, the recommendation system co-optimizes the transition cross-modal canonical mapping network and the initial image element inference network. This means that when updating the network parameters, the parameters of both networks need to be considered. Treat the two networks as a whole, calculate the gradient of the loss function with respect to the entire network parameters, and then update them together.

[0152] The recommendation system also uses the batch training method to improve the training efficiency. Batch training refers to selecting a small batch of data from the training data set for training each time, rather than using the entire training data set. Assuming that there are N samples in the training data set, and m samples are selected as a small batch each time, the training data set is divided into small batches. In each iteration, the recommendation system uses a small batch of data to calculate the loss function and gradient, and then updates the network parameters. This method reduces the amount of calculation and speeds up the training.

[0153] As an implementation, the first training behavior data representation, the first training attribute data representation, and the prior user portrait based on the first training user data are used to debug the transition cross-modal canonical mapping network, including:

[0154] Step S11, obtaining the first training behavior data representation, the first training attribute data representation, and the prior user portrait of the first training user data;

[0155] Step S12, performing canonical mapping on the first training behavior data representation by the initialized initial cross-modal canonical mapping network to obtain first training behavior canonical mapping representation information; performing canonical mapping on the first training attribute data representation by the initialized initial cross-modal canonical mapping network to obtain first training attribute canonical mapping representation information;

[0156] Step S13, performing user portrait prediction on the first training behavior canonical mapping representation information and the first training attribute canonical mapping representation information by the initialized initial portrait element inference network to obtain a predicted user portrait for the first training user data.

[0157] Step S14, iteratively debugging the initial cross-modal canonical mapping network based on the predicted user portrait and the prior user portrait until a preset debugging stop condition is reached to obtain a transition cross-modal canonical mapping network.

[0158] In step S11, the first training user data is a data set for initial training of the model, containing information of a large number of users. The first training behavior data representation is a result of representation information mining on behavior monitoring data in the first training user data, which converts the behavior information of the user into a form convenient for computer processing and analysis. The first training attribute data representation is a result of representation information mining on basic attribute data of the first training user data, reflecting the basic attribute characteristics of the user. The prior user portrait is a user portrait known in advance for the first training user data, which is obtained based on historical data, expert knowledge or other reliable sources, serving as a reference standard for subsequent model training.

[0159] In the implementation of step S11, the recommendation system uses various technical means to obtain these data. For the first training behavior data representation, the recommendation system uses deep learning models such as recurrent neural networks (RNN) or its variants long short-term memory networks (LSTM), gated recurrent units (GRU) to process behavior sequence data. Taking LSTM as an example, it can effectively process behavior data with time sequence and capture long-term dependencies between behaviors. Assuming that the behavior monitoring data is a time series of product browsing records, LSTM takes the browsing information at each time step as input, and through its internal memory unit and gating mechanism, outputs a vector that can represent the features of the entire behavior sequence, i.e. the first training behavior data representation. For the first training attribute data representation, the recommendation system uses one-hot encoding, embedding technology, etc. to process the basic attribute data. For example, for the gender attribute of the user, one-hot encoding is used to represent male as [1, 0] and female as [0, 1]; for the age attribute of the user, an embedding layer is used to map the age value to a low-dimensional vector space. Prior user portraits are obtained by collecting and analyzing historical sales data, user feedback information, etc. and are also determined with the help of experts' experience and judgment.

[0160] In step S12, the initial cross-modal canonical mapping network is a network in an initial state, and its parameters are randomly initialized. The purpose of canonical mapping is to align the features of data in different modalities (behavior data and attribute data) so that they have a unified format and scale, facilitating subsequent network processing.

[0161] In step S13, the initial portrait element inference network is also a network in an initialized state, which receives the behavior and attribute representation information after canonical mapping, and outputs the predicted user portrait through internal calculation and processing of the network.

[0162] In step S14, the core of this step is to compare the differences between the predicted user portrait and the prior user portrait, adjust the parameters of the initial cross-modal canonical mapping network, so that the prediction result is closer to the prior user portrait.

[0163] In order to measure the difference between the predicted user portrait and the prior user portrait, the recommendation system defines a loss function. For example, the mean square error loss function is used to calculate the average square error between two vectors, and the smaller the error, the closer the prediction result is to the prior user portrait.

[0164] As an implementation, the method further includes:

[0165] Step 310, element frequency analysis is performed on the multi-category index image elements of the category data of the multiple categories respectively, to obtain element frequency analysis results;

[0166] Step 320, the image element with the most occurrences in the element frequency analysis results is taken as the pending image element for the user data.

[0167] In step 310, the category data of the multiple categories contains different aspects of information about the user, such as basic attribute data and behavior monitoring data of the user, and the multi-category index image elements are obtained by indexing in the image element mapping relationship set. Element frequency analysis refers to counting the number of occurrences of each image element in all multi-category index image elements to understand the frequency of each image element.

[0168] In implementing step 310, the recommendation system uses a hash table data structure to implement element frequency analysis. The hash table efficiently stores and looks up elements, taking each image element as a key and the number of occurrences as a value stored in the hash table. When traversing the multi-category index image elements, if the element is already in the hash table, its corresponding value is incremented by 1; if the element is not in the hash table, it is inserted into the hash table and the value is initialized to 1. The time complexity is O(n), where n is the number of multi-category index image elements, which can quickly complete element frequency analysis.

[0169] In step 320, the pending image element is an intermediate result before the final user image element is determined, which is based on the results of element frequency analysis and selects the image element with the highest frequency, as these image elements better represent the user's characteristics and preferences.

[0170] In implementing step 320, the recommendation system traverses the hash table to find the key with the maximum value, i.e., the image element with the most occurrences. If there are multiple image elements with the same number of occurrences and the maximum value, the recommendation system selects according to the pre-set rules, such as selecting the first element in alphabetical order of the image element, or setting priorities for different image elements according to business needs and selecting the element with the highest priority.

[0171] From the overall process, steps 310-320 play a role in bridging the gap between the previous steps and the subsequent steps in the entire e-commerce product recommendation method. In the previous steps, the recommendation system obtains multi-category index image elements of multiple categories of category data through cross-category indexing, which reflect the user's characteristics from different angles. Steps 310-320 then filter and refine these image elements to find the most representative image elements as the pending image elements, providing more focused information for subsequent image element reasoning and user image generation.

[0172] Figure 2 A hardware entity schematic diagram of an e-commerce product recommendation system provided by an embodiment of the present application is shown in Figure 2 The hardware entity of the e-commerce product recommendation system P includes a processor P1 and a memory P2, wherein the memory P2 stores a computer program capable of running on the processor P1, and the processor P1 implements the steps in the method of any of the above embodiments when executing the program.

[0173] The above is only an embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. An e-commerce product recommendation method based on big data analysis, characterized in that, The method comprises: determining user data to be analyzed of a target user, determining category data of multiple categories for the user data; determining a set of portrait element mapping relationships; the set of portrait element mapping relationships comprises contrast category representation information of the contrast user data under the multiple categories and respective corresponding contrast portrait elements of the contrast category representation information; based on a commonality measurement result between the category data of the multiple categories and the contrast category representation information of the contrast user data under the multiple categories, indexing in the set of portrait element mapping relationships to obtain multiple-category index portrait elements of the category data of the multiple categories respectively; based on the multiple-category index portrait elements of the category data of the multiple categories respectively, determining multiple pending portrait elements for the user data; based on the multiple pending portrait elements and the category data of the multiple categories, performing portrait element reasoning to obtain a user portrait element for the user data, and generating a target user portrait of the target user based on the user portrait element; according to the target user portrait, recommending a product to the target user; wherein: the multiple categories comprise a monitoring data category and an attribute data category; the category data of the multiple categories comprises behavior monitoring data and basic attribute data of the user data; the set of portrait element mapping relationships comprises behavior contrast category representation information of the contrast user data under the monitoring data category and corresponding behavior contrast portrait elements of the behavior contrast category representation information, and attribute contrast category representation information of the contrast user data under the attribute data category and corresponding attribute contrast portrait elements of the attribute contrast category representation information; the based on a commonality measurement result between the category data of the multiple categories and the contrast category representation information of the contrast user data under the multiple categories, indexing in the set of portrait element mapping relationships to obtain multiple-category index portrait elements of the category data of the multiple categories respectively, comprises: based on the behavior monitoring data and the basic attribute data respectively, performing representation information extraction to obtain behavior data representation and attribute data representation; respectively performing behavior representation commonality measurement on the behavior data representation and the behavior contrast category representation information and the attribute contrast category representation information, and based on the behavior representation commonality measurement result, obtaining the multiple-category index portrait elements of the behavior monitoring data in the behavior contrast portrait elements and the attribute contrast portrait elements; respectively performing attribute representation commonality measurement on the attribute data representation and the behavior contrast category representation information and the attribute contrast category representation information, and based on the attribute representation commonality measurement result, obtaining the multiple-category index portrait elements of the basic attribute data in the behavior contrast portrait elements and the attribute contrast portrait elements.

2. The e-commerce product recommendation method based on big data analysis according to claim 1, wherein, the respectively performing behavior representation commonality measurement on the behavior data representation and the behavior contrast category representation information and the attribute contrast category representation information, and based on the behavior representation commonality measurement result, obtaining the multiple-category index portrait elements of the behavior monitoring data in the behavior contrast portrait elements and the attribute contrast portrait elements, comprises: The behavior data representation is subjected to behavior representation commonality measurement with the behavior control category representation information, to obtain a first behavior representation commonality measurement result; A first behavior category index image element is determined in the behavior control image element based on behavior control category representation information corresponding to the first behavior representation commonality measurement result representing satisfaction of the commonality requirement; The behavior data representation is subjected to behavior representation commonality measurement with the attribute control category representation information, to obtain a second behavior representation commonality measurement result; A first attribute category index image element is determined in the attribute control image element based on attribute control category representation information corresponding to the second behavior representation commonality measurement result representing satisfaction of the commonality requirement; The first behavior category index image element and the first attribute category index image element are used to obtain a multi-category index image element of the behavior monitoring data; Alternatively, the attribute data representation is subjected to attribute representation commonality measurement with the behavior control category representation information and the attribute control category representation information respectively, and a multi-category index image element of the basic attribute data is obtained in the behavior control image element and the attribute control image element based on attribute representation commonality measurement results, and the method comprises: The attribute data representation is subjected to attribute representation commonality measurement with the behavior control category representation information, to obtain a first attribute representation commonality measurement result; A second behavior category index image element is determined in the behavior control image element based on behavior control category representation information corresponding to the first attribute representation commonality measurement result representing satisfaction of the commonality requirement; The attribute data representation is subjected to attribute representation commonality measurement with the attribute control category representation information, to obtain a second attribute representation commonality measurement result; A second attribute category index image element is determined in the attribute control image element based on attribute control category representation information corresponding to the second attribute representation commonality measurement result representing satisfaction of the commonality requirement; The second behavior category index image element and the second attribute category index image element are used to obtain a multi-category index image element of the basic attribute data. 3.The e-commerce product recommendation method based on big data analysis of claim 1, wherein, The multiple categories include a monitoring data category and an attribute data category; and category data of the multiple categories includes behavior monitoring data and basic attribute data; The image element inference is performed based on the multiple pending image elements and the category data of the multiple categories to obtain a user image element for the user data, and the method comprises: Feature fusion is performed on the basic attribute data and the multiple pending image elements to obtain first fusion representation information, and feature fusion is performed on the behavior monitoring data and the multiple pending image elements to obtain second fusion representation information; The image element inference is performed based on the first fusion representation information and the second fusion representation information to obtain a user image element for the user data; The image element inference is performed based on the first fusion representation information and the second fusion representation information to obtain a user image element for the user data, and the method comprises: The first fusion feature information is norm-mapped by a cross-modal norm-mapping network to obtain first norm-mapped fusion feature information; The second fusion feature information is norm-mapped by the cross-modal norm-mapping network to obtain second norm-mapped fusion feature information; The portrait element inference network performs portrait element inference based on the first norm-mapped fusion feature information and the second norm-mapped fusion feature information to obtain a user portrait element for the user data.

4. The e-commerce product recommendation method based on big data analysis according to claim 3, wherein, The cross-modal norm-mapping network and the portrait element inference network are trained by the following steps: The first training behavior data feature of the first training user data, the first training attribute data feature, and the prior user portrait are obtained by debugging a transition cross-modal norm-mapping network, wherein the first training behavior data feature is obtained by performing feature mining on behavior monitoring data of the first training user data; The first training attribute data feature is obtained by performing feature mining on basic attribute data of the first training user data; The second training behavior data feature of the second training user data, the second training attribute data feature, and the prior portrait element are obtained; The second training behavior data feature is obtained by performing feature mining on behavior monitoring data of the second training user data; The second training attribute data feature is obtained by performing feature mining on basic attribute data of the second training user data; The second training behavior norm-mapped feature information is obtained by norm-mapping the second training behavior data feature by the transition cross-modal norm-mapping network, and the second training attribute norm-mapped feature information is obtained by norm-mapping the second training attribute data feature by the transition cross-modal norm-mapping network; The inferred portrait element for the second training user data is obtained by performing portrait element inference on the second training behavior norm-mapped feature information and the second training attribute norm-mapped feature information by the initialized initial portrait element inference network; The transition cross-modal norm-mapping network and the initial portrait element inference network are optimized based on the inferred portrait element and the prior portrait element, and then are cooperatively debugged again until a preset debugging stop condition is reached, to obtain the converged cross-modal norm-mapping network and portrait element inference network.

5. The e-commerce product recommendation method based on big data analysis according to claim 4, wherein, The first training behavior data feature of the first training user data, the first training attribute data feature, and the prior user portrait are obtained by debugging a transition cross-modal norm-mapping network, wherein the first training behavior data feature is obtained by performing feature mining on behavior monitoring data of the first training user data; The first training behavior data feature of the first training user data, the first training attribute data feature, and the prior user portrait are obtained by debugging a transition cross-modal norm-mapping network, wherein the first training behavior data feature is obtained by performing feature mining on behavior monitoring data of the first training user data; The first training behavior norm-mapped feature information is obtained by norm-mapping the first training behavior data feature by the initialized initial cross-modal norm-mapping network, and the first training attribute norm-mapped feature information is obtained by norm-mapping the first training attribute data feature by the initialized initial cross-modal norm-mapping network; perform user portrait prediction on the first training behavior specification mapping feature information and the first training attribute specification mapping feature information through the initialized initial portrait element inference network, to obtain a predicted user portrait for the first training user data; perform iterative debugging on the initial cross-modal specification mapping network based on the predicted user portrait and the prior user portrait until a preset debugging stop condition is reached, to obtain a transition cross-modal specification mapping network.

6. The e-commerce product recommendation method based on big data analysis according to claim 1, wherein, The determining of the multiple pending portrait elements for the user data based on the multiple-category index portrait elements of the category data of the multiple categories comprises: performing element frequency analysis on the multiple-category index portrait elements of the category data of the multiple categories, to obtain an element frequency analysis result; taking the portrait element with the highest occurrence frequency included in the element frequency analysis result as a pending portrait element for the user data.

7. An e-commerce product recommendation system based on big data analysis, characterized in that, A device comprises a memory and a processor, the memory storing a computer program capable of running on the processor, and the processor implements the steps in the method of any one of claims 1 to 6 when executing the program.

Images