Shop commodity recommendation method and device, equipment and medium

Abstract

The invention discloses a shop commodity recommendation method and device, equipment and a medium, and the method comprises the steps: obtaining a search text inputted by a user in response to a search request of the user; extracting a keyword of the search text, and determining a corresponding category according to the search text; in the first stage, shop level recall operation is executed on the shop feature library, and part of shops with shop features matched with the keywords and the category to which the shops belong are recalled to form a shop recommendation list; and in the second stage, performing a commodity level recall operation on the commodity feature data of the commodities of the shops in the shop recommendation list in the commodity feature library, and recalling a part of commodities of which the commodity features are matched with the keywords and the categories to which the commodities belong, so as to form a commodity recommendation list. According to the method, the system pressure is effectively reduced through two stages of level recall operations.

Description

Technical Field

[0001] The present application includes the field of e-commerce technology, and in particular includes a store product recommendation method and its device, equipment, and medium. Background Art

[0002] In the current e-commerce environment, store product recommendations are the key to improving user experience and increasing transaction volume. When users are looking for specific products, they often need to use the search function to quickly locate stores that provide related products and related products in the stores. However, existing store product recommendation methods usually search directly in a huge product database through the user's search text. This method consumes a lot of computing resources when processing large-scale data, causing excessive system pressure and affecting search efficiency and user experience.

[0003] Specifically, when processing user queries, existing store search methods usually directly perform a full search in a database containing millions or even tens of millions of products. Although this full search method can theoretically cover all search results, in actual applications, due to the large amount of data, the search process consumes a lot of computing resources and time, especially in high-concurrency situations, the system's response speed and stability are difficult to guarantee.

[0004] In order to deal with these problems, how to effectively reduce system pressure while ensuring search efficiency and user experience has become an urgent problem to be solved by current e-commerce platforms. Summary of the invention

[0005] The primary purpose of the present application is to solve at least one of the above problems and provide a store product recommendation method and its device, equipment, and medium.

[0006] According to one aspect of the present application, a store product recommendation method is provided, comprising the following steps:

[0007] Respond to user search requests and obtain search text entered by the user;

[0008] Extracting keywords from the search text, and determining corresponding categories according to the search text;

[0009] In the first stage, a store-level recall operation is performed on the store feature library to recall some stores whose store features match the keywords and the categories to which they belong, thereby forming a store recommendation list;

[0010] In the second stage, a product level recall operation is performed on the product feature data of the products owned by the stores in the store recommendation list in the product feature library, and some products whose product features match the keywords and the categories to which they belong are recalled to form a product recommendation list.

[0011] According to another aspect of the present application, a store product recommendation device is provided, comprising:

[0012] A search text acquisition module, configured to respond to a user search request and acquire a search text input by the user;

[0013] A search text processing module, configured to extract keywords from the search text and determine corresponding categories according to the search text;

[0014] The first-stage recall module is configured to perform a store-level recall operation on the store feature library in the first stage, recalling some stores whose store features match the keywords and the categories to which they belong, and forming a store recommendation list;

[0015] The second-stage recall module is configured to perform a product-level recall operation on the product feature data of the products of the stores in the store recommendation list in the product feature library in the second stage, and recall some products whose product features match the keywords and the categories to which they belong, to form a product recommendation list.

[0016] According to another aspect of the present application, a store product recommendation device is provided, comprising a central processing unit and a memory, wherein the central processing unit is used to call and run a computer program stored in the memory to execute the steps of the store product recommendation method described in the present application.

[0017] According to another aspect of the present application, a non-volatile readable storage medium is provided, which stores a computer program implemented according to the store product recommendation method in the form of computer-readable instructions, and when the computer program is called and executed by a computer, the steps included in the method are executed.

[0018] The technical solution of this application has many advantages, including but not limited to the following aspects:

[0019] This application effectively reduces the amount of data for a single retrieval by dividing the recall process into two stages - store-level recall and product-level recall, which not only speeds up the retrieval speed but also reduces the pressure on the system. In the first stage, this application only searches the store feature library instead of the entire product library. Since the number of stores is usually much smaller than the number of products, this step greatly reduces the scale of data that needs to be processed, can quickly respond to user search requests, and generate a preliminary store recommendation list. In the second stage, the products of the stores in the store recommendation list generated in the first stage are further retrieved. This step performs a more detailed screening of the products of each store to ensure that the products recommended to the user are not only related to the store, but also more in line with the user's search intent. In addition, the recall operation in the second stage allows for a more reasonable allocation of computing resources, that is, this application can centrally process product data that is highly relevant to user search requests. This resource allocation helps to improve the efficiency and response speed of the entire system. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A flowchart of an embodiment of a store product recommendation method of the present application;

[0021] Figure 2 This is a functional block diagram of the store merchandise recommendation device of this application;

[0022] Figure 3 This is a schematic diagram of the structure of a store product recommendation device used in this application. DETAILED DESCRIPTION

[0023] A store product recommendation method of the present application can be programmed as a computer program product, deployed in a client or server for operation and implementation. For example, in the exemplary application scenario of the present application, it can be deployed in the server of an e-commerce platform, wherein the e-commerce platform can be an e-commerce platform that opens independent station services. An independent station refers to a new type of official website (website) established on the SaaS technology platform with an independent domain name, private content, data, and rights and interests, independent business sovereignty and business entity responsibilities, supported by social cloud computing capabilities, and can independently and freely connect to third-party software tools, publicity and promotion media and channels. Usually, the seller of goods builds and operates an online store on an independent station, so that the online store can display its goods on the independent station, and provide e-commerce services such as searching for goods, adding shopping carts, settlement, logistics delivery, placing orders, and after-sales service on the online store to the buyers of goods (that is, users). The buyers of goods (users) perform various interactive operations with the e-commerce platform on the terminal device.

[0024] The detailed information of all commodities on the e-commerce platform, including commodity name, description, price, brand, specification, inventory, sales volume, etc., will be stored in the commodity database. Before the implementation of the typical embodiment of the present application, the e-commerce platform first extracts the commodity information of each commodity from the commodity database, and then calls the large model to extract keywords and predict the categories of these commodity information to obtain the commodity feature data of each commodity. In addition, in order to improve the internationalization capability of the search, the e-commerce platform will also translate the commodity feature data into multiple languages, and store the translated data in the commodity feature library. Furthermore, the e-commerce platform will also obtain the categories and keywords corresponding to the respective commodity information of all commodities in each store. According to the preset screening rules, a preset number of categories and keywords are screened out as the categories and keywords corresponding to the store information of the store, and the associated store unique identifier is stored in the store feature library.

[0025] Users can interact with the e-commerce platform by inputting questions related to searching for stores or products on the e-commerce platform, and the e-commerce platform obtains the search text input by the user by monitoring the interactive behavior, such as "Is there a children's bicycle store?" In a typical embodiment of the present application, the e-commerce platform obtains the search text input by the user by responding to the user's search request, and then extracts keywords from the search text and predicts the category to which it belongs, wherein this function can be achieved by deploying a large model in the store platform. Specifically, the user enters a search request on the e-commerce platform, such as "Is there a children's bicycle store?" The e-commerce platform obtains the search text input by the user by monitoring the user's interactive behavior. The search text is analyzed using a large model to extract keywords, such as "children's bicycle", and based on the extracted keywords, the large model further predicts the commodity category to which the keyword belongs, such as "children's bicycle" belongs to the "children's products" category. In one embodiment, the "children's products" category belongs to the broader "maternal and child products" category. That is, when predicting the category to which it belongs, the large model not only considers the primary category, but also deeply analyzes the multi-level categories to ensure the accuracy and relevance of the recommendation. For example, “children’s bicycle” is further subdivided into secondary categories such as “children’s bicycle > balance bike” or “children’s bicycle > tricycle”, and even more detailed tertiary categories such as “children’s bicycle > balance bike > suitable for 1-3 years old”.

[0026] Then, based on the extracted keywords and predicted categories, the e-commerce platform performs the first-stage recall from the store feature library, and selects some stores to form a store recommendation list. In one embodiment, the e-commerce platform can significantly improve the retrieval efficiency and accuracy by converting the store feature library into a graph database for retrieval. The graph database can take advantage of the graph structure to quickly perform association queries between nodes, thereby efficiently screening out qualified stores in the huge store feature library; and the e-commerce platform can more accurately match the keywords and categories entered by the user by calling the graph matching algorithm to ensure the relevance and accuracy of the recommendation results; in addition, the flexibility and scalability of the graph database enable the e-commerce platform to easily cope with the growing store and product data, maintain the efficient operation of the system and continuously improve the user experience. Next, the e-commerce platform retrieves the product information corresponding to each product in the store included in the store recommendation list from the product feature library, and performs the second-stage recall of these product information based on the keywords and categories, and finally obtains the products corresponding to the corresponding recalled product feature data to form a product recommendation list. In this way, the e-commerce platform not only improves the search efficiency and reduces the system pressure, but also improves the user experience, allowing users to find the required products and stores faster.

[0027] The store product recommendation method of the present application can be programmed as a computer program product and deployed in a client or server for execution. For example, in the exemplary application scenario of the present application, it can be deployed and implemented in the server of an e-commerce platform, thereby accessing the interface opened after the computer program product is running and performing human-computer interaction with the process of the computer program product through a graphical user interface to execute the method.

[0028] See also Figure 1 According to a store product recommendation method provided by the present application, in one embodiment, the method comprises the following steps:

[0029] Step S5100: respond to the user search request and obtain the search text input by the user;

[0030] In the e-commerce environment, user search is one of the main ways to obtain product information and store information. Users express their needs by entering search text, and e-commerce platforms need to respond to users' search requests quickly and accurately to provide relevant product and store recommendations. The search text entered by the user can be one or more keywords, that is, the user enters the name of the product they want to buy or related terms, or it can be a descriptive text, that is, the user enters a complete sentence to describe the desired product; the search text can include queries for specific attributes, that is, the user will specify the brand, color, size, price range, etc.; the search text can use natural language queries, that is, the user will enter in the form of questions, such as "Which mobile phone is suitable for taking pictures?"; the search text supports multi-language input, that is, the user can search in different languages.

[0031] Users' interactive behaviors on e-commerce platforms will be monitored. E-commerce platforms provide search box components for users to enter search text and monitor users' input behaviors in real time. Once users enter text in the search box and submit a search request, the e-commerce platform will immediately capture the search text for subsequent keyword extraction and category prediction.

[0032] In one embodiment, after capturing the search text input by the user, the e-commerce platform will pre-process it, such as removing unnecessary spaces, punctuation marks, and stop words, to improve the efficiency and accuracy of subsequent processing steps.

[0033] Step S5200: extract keywords of the search text, and determine corresponding categories according to the search text;

[0034] Keywords are extracted from the search text input by the user extracted in the previous step, and the corresponding category is determined according to the search text. In one embodiment, the e-commerce platform trains a keyword extraction model and a category classification model to implement this step. Specifically, the e-commerce platform collects a large amount of search text and corresponding keywords and category labels as training data. The e-commerce platform will design the architecture of the keyword extraction model and the category classification model. The keyword extraction model can use sequence labeling tasks, using models such as BiLSTM-CRF (bidirectional long short-term memory network-conditional random field) to identify keywords in the search text. The category classification model uses multi-classification tasks, and uses pre-trained language models such as BERT (Bidirectional Encoder Representations from Transformers) for fine-tuning to predict the corresponding category of the search text. After the architecture design of the keyword extraction model and the category classification model is completed, the e-commerce platform will use the prepared training data to train the keyword extraction model and the category classification model. The training process includes steps such as data input, model forward propagation, loss calculation and back propagation. E-commerce platforms will set appropriate loss functions and optimization algorithms, such as the cross entropy loss function and the Adam optimization algorithm, to improve the training effect and convergence speed of the model. The training process will be repeated for multiple rounds until the performance of the model reaches the preset indicators or the training data is fully learned.

[0035] After the model training is completed, the e-commerce platform will deploy the trained keyword extraction model and category classification model to the e-commerce platform environment. When the user enters the search text, the e-commerce platform will call the trained keyword extraction model and category classification model to process the search text. The keyword extraction model will identify the keywords in the search text, and the category classification model will predict the category to which the search text belongs. These keywords and category information will be used in subsequent store and product recall operations, thereby improving the accuracy and relevance of search results.

[0036] In another embodiment, the e-commerce platform extracts keywords and determines the category of the search text obtained in the previous step by calling a large model (such as the ChatGPT-4o model). A large model refers to a machine learning model with large-scale parameters and complex computing structures, such as BERT, GPT, etc. The large model learns complex patterns and features by training massive data, has a stronger generalization ability, and can make accurate predictions on unseen data. Specifically, the large model constructs a corresponding Prompt. Among them, "Prompt" refers to text or instructions that provide input to the large model to guide it to generate specific outputs. It is a text paragraph provided by the user when interacting with the large model, which is used to describe the information, answers, text, etc. that the user wants to obtain from the large model. The purpose of the Prompt is to guide the large model to produce the required response in order to better control the generated output. The prompt is based on the search text entered by the user and the category system of the e-commerce platform. For example, if the user enters "Do you have an Apple iPhone 13Pro with 256GB memory and graphite color?", the e-commerce platform generates a prompt corresponding to "Please extract keywords from the search text 'Do you have an Apple iPhone 13Pro with 256GB memory and graphite color?' and determine the category to which the search text belongs in the category system of the e-commerce platform". The prompt is input into the big model, and the keywords of the search text are "mobile phone", "Apple", "iPhone13Pro", "256GB", and "graphite color", and the category to which it belongs can be "electronic products>mobile phone>Apple>iPhone 13 series". The output of the big model can be controlled by changing the prompt to obtain the required output content.

[0037] In another embodiment, the e-commerce platform can adopt a rule- and statistics-based method to extract keywords and determine the category to which they belong. This method relies on predefined rules and historical data statistics to identify and classify the search text input by users. First, the e-commerce platform needs to define a series of rules to identify and extract keywords from the search text. These rules can be based on natural language processing techniques such as part-of-speech tagging, word frequency statistics, and stop word filtering. For example, the e-commerce platform can define to extract nouns, verbs, and adjectives as potential keywords and filter out common stop words (such as "的", "是", "吗", etc.). In addition, according to word frequency statistics, retain the words with higher frequencies. For the category matching rules, the e-commerce platform can define the mapping relationship between keywords and categories. For example, "手机" corresponds to the category "电子产品>手机", and "Apple" corresponds to the category "电子产品>手机>Apple". Determine the final category according to the combination of keywords. For example, "Apple iPhone13Pro" corresponds to the category "电子产品>手机>Apple>iPhone 13系列".

[0038] Before extracting keywords and determining the category to which they belong, it is necessary to preprocess the search text input by users. The preprocessing steps include word segmentation and part-of-speech tagging. For example, "请问有Apple iPhone 13Pro、内存为256GB、颜色为石墨色的手机吗?" can be segmented into "请问", "有", "Apple", "iPhone 13Pro", "内存", "为", "256GB", "颜色", "为", "石墨色", "的", "手机", "吗". Tag the part of speech for each lexical unit for subsequent application of the rules. For example, "Apple" is tagged as a noun, "iPhone 13Pro" is tagged as a noun phrase, and "256GB" is tagged as a quantifier, etc.

[0039] According to the predefined rules, the preprocessed text is subjected to keyword extraction. Specifically, potential keywords are extracted based on part-of-speech tagging and word frequency statistics. For example, "Apple", "iPhone13Pro", "256GB", "graphite", and "mobile phone" are extracted from the above segmentation results. Words that do not conform to the rules are removed, such as "please ask", "have", "for", "of", "may", etc. After the keywords are extracted, the category to which the search text belongs is determined according to the predefined category matching rules. Specifically, the corresponding categories are matched according to the combination of keywords. For example, "Apple iPhone 13Pro" matches the category "Electronic Products>Mobile Phones>Apple>iPhone 13 Series". The final category is determined according to the hierarchical relationship of the keywords. For example, after the combination of "Apple iPhone 13Pro", "256GB", and "graphite", it is finally determined to be "Electronic Products>Mobile Phones>Apple>iPhone 13 Series". Finally, the extracted keywords and the determined categories are output for subsequent use. For example, the output results are: Keywords: "mobile phone", "Apple", "iPhone 13Pro", "256GB", "graphite color"; category: "Electronic Products>Mobile Phones>Apple>iPhone 13 Series".

[0040] In order to improve the accuracy of keyword extraction and category determination, e-commerce platforms can optimize rules regularly. Optimization methods include data analysis and rule adjustment. Analyze user search behavior and historical data to identify common keyword combinations and category matching patterns. According to the data analysis results, adjust the rules for keyword extraction and category matching to improve the accuracy and efficiency of the system. Through rule-based and statistical methods, e-commerce platforms can achieve keyword extraction and category determination for search texts without relying on large-scale pre-trained models. This method has higher efficiency and accuracy when processing search texts in specific fields or specific types.

[0041] Step S5300: In the first stage, a store-level recall operation is performed on the store feature library to recall some stores whose store features match the keywords and the categories to which they belong, thereby forming a store recommendation list.

[0042] This application uses a two-stage recall operation to progressively and finely screen out the products and stores that best match the user's search intent. In the first stage of the store-level recall operation, it is necessary to recall some stores whose features match the keywords and categories in the user's search text from the store feature library to form a store recommendation list.

[0043] It should be noted that before the implementation of the typical embodiment of the present application, a store feature library is provided in the e-commerce platform. The store feature library is a database containing the corresponding features of all stores. The store feature library stores the store unique identification, categories and keywords of all stores in the e-commerce platform. Among them, the determination of store-related features can be provided by the e-commerce platform to the user when the user registers a store. The user sets the keywords and categories of the registered store through the input control, and then the e-commerce platform monitors the input of the store keywords and categories and stores them in the store feature library. The keywords and categories of the store entered by the user reflect the main sales categories and features of the store. The determination of store related can also be based on the preset screening rules. The categories and keywords of all the goods in the store are filtered out to a preset number of categories and keywords, as the categories and keywords of the corresponding store information of the store and stored in the store feature library. Among them, please refer to the subsequent specific implementation method for determining the corresponding keywords and categories of the store according to the screening rules, which will not be repeated here.

[0044] In one embodiment, after determining the keywords and categories of the search text entered by the user in the previous step, the e-commerce platform will match the keywords and categories. The keywords entered by the user are matched with the keywords in the store feature library. For example, if the keywords entered by the user are "red", "ladies" and "sports shoes", the e-commerce platform will search for stores matching these keywords in the store feature library. At the same time, the categories entered by the user are matched with the categories of the stores in the store feature library. In order to improve the accuracy and recall rate of matching, the e-commerce platform adopts a multi-dimensional matching strategy, including exact matching, fuzzy matching and extended matching. First, an exact match is performed, that is, a store in the store feature library is searched for stores that completely contain the keywords and categories entered by the user. If the exact match fails to find enough stores, a fuzzy match is performed, allowing a certain keyword or category deviation. On the basis of fuzzy matching, an extended match is also performed to find other stores related to the keywords and categories entered by the user. According to the above matching strategy, some stores that meet the conditions are recalled from the store feature library. The number of recalled stores can be controlled according to a preset threshold, for example, the top 100 most relevant stores are recalled. The recalled stores are sorted and filtered according to multiple dimensions to ensure the quality of the recommendation list, including relevance sorting, store rating sorting, sales data sorting, and personalized filtering. Based on the sorting and filtering results, a store recommendation list is generated, which contains the stores most relevant to the user's search request and is displayed on the user interface for users to browse and select.

[0045] Step S5400, in the second stage, a product level recall operation is performed on the product feature data of the products owned by the stores in the store recommendation list in the product feature library, and some products whose product features match the keywords and the categories to which they belong are recalled to form a product recommendation list.

[0046] The e-commerce platform will extract the product feature data of the products in the store in the store recommendation list from the product feature library. The extracted product feature data includes the keywords and categories of the product information of the store products in the store recommendation list. The e-commerce platform will perform product-level recall operations in the product feature library based on the keywords and categories entered by the user. The goal of the product-level recall operation is to filter out the products that best match the user's search intent and form a product recommendation list. Specifically, first, the e-commerce platform will match the keywords entered by the user with the keywords in the product feature data. Keyword matching can be achieved through a variety of methods, such as exact matching, fuzzy matching, and extended matching. Secondly, the e-commerce platform will match the categories entered by the user with the categories in the product feature data. Category matching is based on the hierarchical structure and semantic relationship of the categories. The e-commerce platform will search for products in the product feature data that belong to the category entered by the user and its subcategories. In order to improve the accuracy and recall rate of recall, the e-commerce platform will comprehensively consider matching strategies in multiple dimensions. In addition to matching keywords and categories, e-commerce platforms can also consider factors such as the price, brand, specifications, inventory status, sales data, and user reviews of products. During the matching process, e-commerce platforms will sort and filter the recalled products according to multiple dimensions to ensure the quality of the recommendation list. Sorting and filtering strategies can include relevance sorting, price sorting, sales volume sorting, score sorting, and personalized filtering. For example, e-commerce platforms can sort products according to their relevance to user search intent, and the higher the relevance, the higher the ranking. At the same time, e-commerce platforms can filter products according to factors such as price, sales volume, and user reviews to ensure that the products in the recommendation list have high quality and user satisfaction. Based on the sorting and filtering results, the e-commerce platform will generate a product recommendation list. This list contains the products that are most relevant to the user's search text, and is arranged in descending order according to factors such as relevance, price, sales volume, and score. When users browse the recommendation list, they can quickly find products that meet their search intent, thereby improving user experience and platform operation efficiency. Through this step, the e-commerce platform can effectively perform product-level recall operations on the product feature data of the products owned by the stores in the store recommendation list in the product feature library in the second phase, and recall some products whose features match the keywords entered by the user and the categories they belong to, to form an accurate product recommendation list, thereby improving the user experience and the platform's operational efficiency.

[0047] It can be known from the typical embodiments of the present application that the technical solution of the present application has many advantages, including but not limited to the following aspects:

[0048] This application effectively reduces the amount of data for a single retrieval by dividing the recall process into two stages - store-level recall and product-level recall, which not only speeds up the retrieval speed but also reduces the pressure on the system. In the first stage, this application only searches the store feature library instead of the entire product library. Since the number of stores is usually much smaller than the number of products, this step greatly reduces the scale of data that needs to be processed, can quickly respond to user search requests, and generate a preliminary store recommendation list. In the second stage, the products of the stores in the store recommendation list generated in the first stage are further retrieved. This step performs a more detailed screening of the products of each store to ensure that the products recommended to the user are not only related to the store, but also more in line with the user's search intent. In addition, the recall operation in the second stage allows for a more reasonable allocation of computing resources, that is, this application can centrally process product data that is highly relevant to user search requests. This resource allocation helps to improve the efficiency and response speed of the entire system.

[0049] The method of the present application takes advantage of the structure of the graph database and quickly performs association queries between nodes through a graph matching algorithm. This graph-based method can efficiently filter out qualified stores and products from a huge store and product feature library, further improving search efficiency.

[0050] Based on any embodiment of the method of the present application, before responding to the user's search request, the method includes:

[0051] Step S6100: extracting commodity information of each commodity from the commodity database;

[0052] Before the implementation of the typical embodiments of this application, the e-commerce platform first extracts the product information of each product from the product database. The product database contains detailed information about all products on the e-commerce platform, which is the basis for the e-commerce platform to manage, display, search, and recommend products. The product information in the product database includes multiple dimensions such as the name, description, price, brand, specifications, inventory status, sales data, and user reviews of the product. Specifically, after a product seller (i.e., a merchant) registers and opens a store on the platform, they need to upload the product information of the products they sell. This product information includes the basic attributes of the product (such as name, description, price, brand, specifications, etc.) and multimedia content such as product pictures and videos. The product information uploaded by the merchant is the main source of information in the product database. In one embodiment, after a user purchases a product, they can evaluate and rate the product, and this user feedback information (such as evaluation content, score, purchase quantity, etc.) will also be recorded in the product database as part of the product information. In another embodiment, the e-commerce platform will automatically generate some product-related data based on the sales situation and inventory changes of the product, such as sales volume and inventory quantity. These generated product-related data will also be stored in the product database for product management and display. To ensure the accuracy and integrity of the extracted product information, the e-commerce platform will regularly update the data in the product database and perform data cleaning and verification to remove duplicate, incorrect, or outdated product information.

[0053] Step S6200: Invoke the large model to extract keywords and predict the category to which each product belongs from the product information of each product, as the corresponding product features in the product feature data of the product.

[0054] After converting the product information of each product extracted from the product database into a format acceptable to the large model, it is input into the large model. First, the large model will preprocess the text information in the product information. The preprocessing steps include text cleaning, word segmentation, and stop word removal. Text cleaning mainly removes irrelevant information such as special characters, HTML tags, and spaces in the text to ensure the accuracy of subsequent processing. Word segmentation is to split the text into individual words, which is a basic step in natural language processing. Stop word removal is to filter out those words that appear frequently in the text but have little meaning for understanding the content, such as "de", "shi", "zai", etc. Then the large model will use word embedding technology to convert the segmented words into vector representations. Word embedding technology (such as Word2Vec, GloVe, BERT, etc.) can capture the semantic relationships between words and map the words into a high-dimensional vector space. In this way, each word can be represented by a vector.

[0055] Based on word embedding, just like using the big model to extract keywords and categories of search text, the corresponding prompt is constructed to guide the big model to extract keywords and predict the categories of product information to obtain the required output content. Keyword extraction can be achieved through a variety of methods, such as TF-IDF (term frequency-inverse document frequency), TextRank, and attention mechanism based on deep learning. The TF-IDF method evaluates the importance of words by calculating the frequency of words in the text and the inverse document frequency in the entire corpus. The TextRank method is a graph-based sorting algorithm that calculates the importance of words by constructing a co-occurrence relationship graph between words. The attention mechanism based on deep learning can automatically learn which words in the text are most critical to understanding the content. At the same time, the big model will predict the category. The category prediction can be based on multi-classification tasks. The big model will use the text information of the product and the existing product classification system to predict the category of the product by training the classification model. The classification model can be a traditional machine learning model (such as logistic regression, support vector machine) or a deep learning model (such as convolutional neural network CNN, recurrent neural network RNN, Transformer, etc.). The input of the classification model is the text vector representation of the product, and the output is the category to which the product belongs. In order to improve the accuracy of the category prediction, the large model can perform multimodal fusion by combining other information of the product (such as price, brand, sales data, etc.). Finally, the extracted keywords and the predicted categories will be integrated into the product feature data of the product. These feature data include not only keywords and categories, but also information such as the weight of the keywords and the confidence of the categories. These feature data will be used for subsequent product search recommendations.

[0056] Step S6300: Translate the product feature data into multiple languages ​​and store them in a product feature library, so that when the recall operation is performed, feature matching corresponding to the keyword and the category to which it belongs can be achieved based on the same language.

[0057] E-commerce platforms will call multilingual translation models to translate keywords and categories in product feature data. Multilingual translation models can be based on neural network architectures, such as Transformer, and can handle translation tasks between multiple languages. The input of the translation model is the keywords and categories in the product feature data, and the output is the corresponding expression of these features in the target language. In order to ensure the accuracy and consistency of the translation, e-commerce platforms will use high-quality translation models and combine contextual information for translation to avoid semantic deviation and ambiguity. After the translation is completed, the e-commerce platform will store the translated product feature data in the product feature library. The product feature library is a database that centrally stores product feature data, including product feature data and translated versions of this information in different languages. The design of the product feature library needs to consider efficient storage and fast retrieval of data to support large-scale product management and user search requests.

[0058] In order to achieve feature matching based on the same language in the recall operation, the e-commerce platform will extract product feature data in the corresponding language from the product feature library based on the user's language preference or the language of the search input when the user initiates a search request. For example, if the user searches in Chinese, the e-commerce platform will extract Chinese keywords and categories from the product feature library, and match and recall products based on these features. In this way, the e-commerce platform can ensure the language consistency between the user's search request and the product feature data, thereby improving the accuracy and relevance of the search results.

[0059] In one embodiment, in order to further improve the effect of the recall operation, the e-commerce platform can also store multilingual synonyms and near synonyms of keywords and their categories in the product feature library, so that semantic similarity can be more comprehensively considered during feature matching. For example, for the keyword "mobile phone", the product feature library stores multiple synonyms such as "mobile phone", "mobile phone", and "smartphone", and the translation versions of these synonyms in different languages ​​will also be stored, so that the user's search intent can be more flexibly matched in the recall operation.

[0060] Through this step, the e-commerce platform can effectively process multilingual product feature data and achieve feature matching based on the same language in recall operations, thereby improving user experience and platform operational efficiency.

[0061] In this embodiment, the commodity information of each commodity is extracted from the commodity database, and the large model is used to extract keywords and predict the categories of the commodity information as the corresponding commodity features in the commodity feature data. The large model flexibly constructs appropriate prompts to quickly and accurately extract keywords from commodity information and predict the categories according to different needs, greatly improving the efficiency and quality of generating commodity feature data. Then, the keywords and categories in the commodity feature data are translated through a multilingual translation model, and the translated commodity feature data is stored in the commodity feature library. This step ensures the flexible application of commodity feature data in different language environments, providing a better search experience for multilingual users.

[0062] Based on any embodiment of the method of the present application, before responding to the user's search request, the method includes:

[0063] Step S7100: Obtain the categories and keywords of all commodities in each store in the commodity feature database;

[0064] This step includes a comprehensive scan and data extraction of the product feature library to ensure that the information obtained is accurate and complete. First, the e-commerce platform will start the data extraction process and traverse each product record in the product feature library. The product feature library is a database that centrally stores product feature data, including product keywords, categories, and other feature data, as well as translations of this information in different languages. During the traversal process, the e-commerce platform will extract the categories and keywords of each product.

[0065] The category is the classification label of the product in the e-commerce platform, which is used to classify the product into a specific category. These categories are predefined and cover various attributes of the product, such as electronic products, clothing, household items, etc. When each product is added to the product feature library, it will be assigned one or more categories so that users can quickly find the required products through the category navigation. Keywords are words that describe the core attributes and characteristics of the product, which can be extracted from the text information such as the name, description, brand, specifications, etc. of the product. The keyword extraction process has been described in detail in the previous step, including text preprocessing, word embedding, keyword extraction and other steps. These keywords can reflect important information such as the function, purpose, material, color, etc. of the product. Through this step, the e-commerce platform can obtain the categories and keywords of all products in each store.

[0066] Step S7200: Filter out a preset number of categories and keywords according to a preset filtering rule as the categories and keywords of the store information corresponding to the store, and store the associated store unique identifier as the corresponding store feature in the store feature data in the store feature library;

[0067] The e-commerce platform will filter the obtained categories and keywords according to the preset filtering rules. The setting of filtering rules is based on indicators such as the popularity, sales volume, user evaluation, and keyword frequency of the product. For example, the e-commerce platform can select the categories and keywords of the top N products with the highest sales volume in each store, or select the categories and keywords of the top N products with the best user evaluation. A comprehensive score can also be calculated by assigning different weights to these indicators. Specifically, the e-commerce platform will first set a weight for each indicator, which reflects the importance of different indicators in determining the representativeness of the product. For example, sales volume may be assigned a higher weight because sales volume directly reflects the market acceptance of the product. Then, the e-commerce platform will calculate the comprehensive score of each product based on these weighted indicators. This comprehensive score represents multiple dimensions of the product and can more comprehensively reflect the popularity and popularity of the product. The e-commerce platform will set a preset threshold, and only those products with a comprehensive score exceeding this threshold will be selected to represent the main products of the store. This preset threshold can be determined based on historical data analysis, market trend forecasting, or business strategy. After selecting the products that represent the main products of the store, the e-commerce platform will further extract the categories and keywords of these products to label the store. These categories and keywords not only reflect the main sales categories of the store, but also capture the core characteristics of the store's products and the focus of users.

[0068] The filtering rules are designed to ensure that the extracted categories and keywords can represent the store’s core business and main products, thereby providing valuable information for subsequent store management and user searches.

[0069] In one embodiment, the e-commerce platform aggregates the categories of each commodity from the commodities on the shelves in the store, and selects the top three categories with the highest proportion as the category tags of the store. This step includes counting and sorting the categories of all commodities on the shelves in the store to determine the main business categories of the store. Specifically, first, the e-commerce platform will traverse all the commodities on the shelves in the store and extract the categories to which each commodity belongs. The e-commerce platform will count the extracted categories and calculate the proportion of each category in the store. The calculation of the proportion can be based on the number or sales of commodities. The e-commerce platform will count the number or sales of commodities under each category and calculate its proportion in the total number of commodities or total sales in the store. After the statistics are completed, the e-commerce platform will sort according to the proportion of the categories and select the top three categories with the highest proportion as the category tags of the store. The selected categories represent the main business categories of the store and can reflect the core business and main commodities of the store. These category tags will be stored as feature data of the store, which will be used for personalized recommendations of the store, store search optimization, store ranking improvement and other aspects. Through this embodiment, the e-commerce platform can effectively extract the main business categories of the store and use them as category tags for the store, thereby providing valuable information for store management and user search.

[0070] In another embodiment, the e-commerce platform aggregates the keywords of each product from the products on the shelves in the store, and performs word segmentation and semantic normalization on these keywords, and selects the top 10 keywords as the keywords of the store according to the word frequency. This process includes processing and screening the keywords of all the products on the shelves in the store to determine the main keywords of the store. First, the e-commerce platform traverses all the products on the shelves in the store and extracts the keywords of each product. Keywords are words that describe the core attributes and characteristics of the product, which can be extracted from the name, description, brand, specifications and other text information of the product. These keywords can reflect important information such as the function, use, material, color, etc. of the product. Then the e-commerce platform will perform word segmentation on the extracted keywords. Word segmentation is to split the keywords into independent words, which is a basic step in natural language processing. Word segmentation can decompose complex keywords into smaller semantic units, which is convenient for subsequent semantic normalization and word frequency statistics. After the word segmentation is completed, the e-commerce platform will perform semantic normalization on the words after word segmentation. Semantic normalization is to merge words with the same or similar semantics to reduce the redundancy and ambiguity of words. For example, for the words "mobile phone", "mobile phone" and "smart phone", the e-commerce platform will normalize these words into the same semantic unit "mobile phone". Semantic normalization can improve the accuracy and consistency of keywords and facilitate subsequent word frequency statistics and screening.

[0071] After the semantic normalization is completed, the e-commerce platform will perform word frequency statistics on the normalized words and calculate the frequency of occurrence of each word in the store. Word frequency statistics can reflect the importance and representativeness of words in the store, and high-frequency words represent the main commodities and core businesses of the store. After the word frequency statistics are completed, the e-commerce platform will sort the words according to their word frequencies and select the top 10 words with the highest word frequencies as the keywords of the store. The top 10 keywords represent the main commodities and core businesses of the store, and can reflect the main business direction and product characteristics of the store. These keywords will be stored as feature data of the store and used for personalized recommendations of the store, store search optimization, store ranking improvement and other aspects. Through this embodiment, the e-commerce platform can effectively extract the main keywords of the store and use them as keywords of the store, thereby providing valuable information for store management and user search.

[0072] After the screening is completed, the e-commerce platform will associate the selected categories and keywords with the unique identifier of the store. The unique identifier of the store is a unique identification code assigned by the e-commerce platform to each store to distinguish different stores. The e-commerce platform will store the associated store feature data in the store feature library.

[0073] In this embodiment, by obtaining the categories and keywords of all products in each store in the product feature library, and filtering out a preset number of categories and keywords according to preset filtering rules, the e-commerce platform can effectively extract the core business and main product features of the store. These feature data are associated with the store's unique identifier and stored in the store feature library, which is used to more accurately define the main business scope of each store and provide support for the store's personalized marketing and user-targeted recommendations.

[0074] Based on any embodiment of the method of the present application, in the first stage, a store-level recall operation is performed on the store feature library to recall some stores whose store features match the keywords and the categories to which they belong, to form a store recommendation list, including:

[0075] Step S5310: store corresponding store features in the store feature data as nodes of the first graph database, and establish corresponding edge connections between the nodes according to predefined business logic;

[0076] The first graph database is a graph-based storage system for storing and querying complex related data. Each store feature, including the category, keywords, and store unique identifier, is converted into a node and stored in the graph database. These nodes represent the core features of the store and are the basis for subsequent graph matching operations. The e-commerce platform then establishes edge connections between nodes in the graph database based on predefined business logic. These edge connections represent the associations between store features. For example, if two stores belong to the same category, the e-commerce platform will establish an edge between the nodes representing the two stores, indicating that the nodes of the two stores have similarities in business. Similarly, if the keywords of the two stores overlap, the e-commerce platform will also establish an edge connection between the nodes of the two stores, indicating that the two stores have similarities in product attributes. In this way, the e-commerce platform can build a complex graph structure that reflects the associations between stores.

[0077] Through this step, the e-commerce platform can effectively convert store feature data into nodes in the graph database and establish edge connections between nodes according to predefined business logic, thereby building a powerful graph structure to provide basic data support for subsequent store-level recall operations.

[0078] Step S5320: Based on the keywords and categories of the search text input by the user, a specific graph matching algorithm is used to match the keywords in the first graph database, and some stores whose characteristics match the keywords and categories are screened out to form a store recommendation list.

[0079] The e-commerce platform will extract the keywords and categories from the search text entered by the user. These keywords and categories will be used as search conditions for matching operations in the first graph database. The e-commerce platform will use a specific graph matching algorithm to match in the first graph database. The graph matching algorithm can be based on graph traversal and search technology, and can efficiently find nodes that meet the conditions in a complex graph structure. Common graph matching algorithms include graph-based depth-first search (DFS), breadth-first search (BFS), graph neural network (GNN), etc. In the matching process, the e-commerce platform will search for nodes that match these conditions in the graph database based on the keywords and categories entered by the user. Specifically, the e-commerce platform will first search for nodes that match the keywords entered by the user. These nodes represent the keywords in the product feature data, and the e-commerce platform will match them in the graph database based on the semantic similarity of the keywords. For example, if the keyword entered by the user is "mobile phone", the e-commerce platform will search for nodes containing keywords such as "mobile phone", "mobile phone", and "smartphone" in the graph database. At the same time, the e-commerce platform will search for nodes that match these categories in the graph database based on the categories entered by the user. These nodes represent the categories in the product feature data, and the e-commerce platform will match them in the graph database based on the hierarchy and semantic relationship of the categories. For example, if the category entered by the user is "electronic products", the e-commerce platform will search for nodes belonging to "electronic products" and its subcategories (such as "mobile phones", "tablets", "laptops", etc.) in the graph database.

[0080] During the search process, the e-commerce platform will comprehensively consider the matching results of keywords and categories, and filter out some stores whose characteristics match the keywords and categories entered by the user. The nodes of these stores are directly or indirectly associated with the search conditions in the graph database, representing stores that match the user's search intent. The e-commerce platform will extract these matching store nodes to form a store recommendation list.

[0081] Through this step, the e-commerce platform can leverage the powerful query capabilities of the First Graph Database to efficiently screen out stores that match the user's search intent and form an accurate store recommendation list, thereby improving the user experience and the platform's operational efficiency.

[0082] Based on any embodiment of the method of the present application, after selecting some stores whose characteristics match the keywords and the categories to which they belong, the method includes:

[0083] Step S5321: Calculate the text relevance score between the search text input by the user and the store feature data of each store in the store by using an information retrieval algorithm;

[0084] Information retrieval algorithm is a technology used to measure the similarity between texts. Common algorithms include TF-IDF (term frequency-inverse document frequency), BM25, cosine similarity, etc. First, the e-commerce platform will pre-process the search text entered by the user, including text cleaning, word segmentation, and removal of stop words. Then the e-commerce platform will use word embedding technology to convert the segmented words into vector representations. Word embedding technology (such as Word2Vec, GloVe, BERT, etc.) can capture the semantic relationship between words and map the words to a high-dimensional vector space. In this way, each word can be represented by a vector. Then, the e-commerce platform will compare the vector representation of the search text entered by the user with the text vector representation in the store feature data to calculate the similarity. Common similarity calculation methods include cosine similarity, Euclidean distance, Manhattan distance, etc.

[0085] When calculating similarity, the e-commerce platform will comprehensively consider the similarity between each word in the search text entered by the user and each word in the store feature data, and weight them according to the importance of the words. For example, the e-commerce platform can use the TF-IDF method to calculate the weight of each word. The TF-IDF method evaluates the importance of words by calculating the frequency of the word in the text and the inverse document frequency in the entire corpus. Words with high weights will occupy a larger proportion in the similarity calculation, thereby more accurately reflecting the user's search intent. Finally, the e-commerce platform will use the calculated similarity score as a text relevance score to measure the text relevance between the search text entered by the user and the store feature data. These text relevance scores will be used in the subsequent calculation of the total relevance score to rank and recommend stores. Through this step, the e-commerce platform can effectively calculate the text relevance score between the search text entered by the user and the store feature data, providing basic data support for accurate store recommendations.

[0086] Step S5322: determine whether the category of the search text input by the user matches the category in the store characteristic data of each store; if so, the category relevance score is recorded as 1, otherwise, it is recorded as 0;

[0087] The e-commerce platform will extract the category information from the search text entered by the user. Then the extracted category of the user search text is compared with the category in the store feature data. The category in the store feature data is the statistical result of the categories of all the products on the shelves in the store, representing the main business category of the store. The e-commerce platform will traverse each store in the store feature data and compare the category of the user search text with the category of the store one by one. During the comparison process, the e-commerce platform will determine whether the category of the user search text completely matches the category of the store. If the category of the user search text is exactly the same as the category of the store, the e-commerce platform will record the category relevance score as 1, indicating that the category is completely matched. For example, if the category of the user search text is "electronic products" and the category of the store is also "electronic products", the category relevance score is recorded as 1. If the category of the user search text does not completely match the category of the store, the e-commerce platform will record the category relevance score as 0, indicating that the category does not match. For example, if the category of the user's search text is "electronic products" and the category of the store is "clothing", the category relevance score is recorded as 0.

[0088] Through this step, the e-commerce platform can quickly determine whether the category of the user's search text matches the category of the store, and generate a category relevance score based on the matching results. These category relevance scores will be used in the subsequent calculation of the total relevance score to rank and recommend the stores. In this way, the e-commerce platform can ensure that the stores in the store recommendation list are consistent with the user's search intent in terms of category, and improve the accuracy and relevance of the search results.

[0089] Step S5323: determine whether the keywords in the search text input by the user match the keywords in the store feature data of each store. If they match, the keyword relevance score is recorded as 1, otherwise it is recorded as 0.

[0090] The e-commerce platform will extract keyword information from the search text entered by the user. Then the keywords of the extracted user search text are compared with the keywords in the store feature data. The keywords in the store feature data are the keyword statistics of all the products on the shelves in the store, representing the main products and core business of the store. The e-commerce platform will traverse each store in the store feature data and compare the keywords of the user search text with the keywords of the store one by one. During the comparison process, the e-commerce platform will determine whether the keywords of the user search text completely match the keywords of the store. If the keywords of the user search text are exactly the same as the keywords of the store, the e-commerce platform will record the keyword relevance score as 1, indicating that the keywords are completely matched. For example, if the keyword of the user search text is "smartphone" and the keywords of the store include "smartphone", the keyword relevance score is recorded as 1. If the keywords of the user search text do not completely match the keywords of the store, the e-commerce platform will record the keyword relevance score as 0, indicating that the keywords do not match. For example, if the keywords of the user search text are "smartphone" and the keywords of the store include "clothing" and "tops", the keyword relevance score is recorded as 0.

[0091] Through this step, the e-commerce platform can quickly determine whether the keywords in the user's search text match the store's keywords, and generate keyword relevance scores based on the matching results. These keyword relevance scores will be used in the subsequent calculation of the total relevance score to rank and recommend stores. In this way, the e-commerce platform can ensure that the stores in the store recommendation list are consistent with the user's search intent in terms of keywords, and improve the accuracy and relevance of search results.

[0092] Step S5324: Calculate the total relevance score of each store based on the corresponding text relevance score, category relevance score and keyword relevance score of each store, and sort the stores based on the total relevance score to form a store recommendation list.

[0093] In one embodiment, the e-commerce platform assigns weights to the text relevance score, category relevance score, and keyword relevance score of each store. These weights reflect the importance of different relevance scores in the calculation of the total score. For example, the text relevance score may be assigned a higher weight because the text relevance score can capture the semantic similarity between the user's search text and the store's feature data; while the category relevance score and the keyword relevance score may be assigned a lower weight because the category relevance score and the keyword relevance score are mainly used to determine the direct match between the category and the keyword. The e-commerce platform will then perform a weighted summation of the text relevance score, category relevance score, and keyword relevance score of each store based on the assigned weights to calculate the total relevance score of each store.

[0094] Through this formula, the e-commerce platform can comprehensively consider the degree of match between the user's search text and the store's feature data in three dimensions: text, category, and keywords, and calculate the comprehensive relevance score for each store. After calculating the total relevance score for each store, the e-commerce platform will sort the stores according to these total scores. The sorting process is carried out in order from high to low relevance total scores, and the stores with higher relevance total scores are ranked higher. In this way, the e-commerce platform can ensure that the stores that best match the user's search intent are displayed first in the recommendation list. Finally, the e-commerce platform will form a store recommendation list based on the sorting results. This list contains the stores that are most relevant to the user's search text, arranged in order from high to low relevance total scores. When users browse the recommendation list, they can quickly find stores that meet their search intent, thereby improving user experience and platform operational efficiency.

[0095] In this embodiment, by comprehensively utilizing information retrieval algorithms, category matching, and keyword matching, the total correlation score between the user search text and the store feature data is accurately calculated, thereby achieving accurate store recommendations. This embodiment comprehensively captures the degree of match between the user's search intent and the store feature data through multi-dimensional correlation calculation, dynamic weight allocation, and efficient sorting, ensuring that the store that best matches the user's search intent is displayed first, significantly improving the accuracy of store recommendations and user experience.

[0096] Based on any embodiment of the method of the present application, after selecting some stores whose characteristics match the keywords and the categories to which they belong, the method includes:

[0097] Step S5330: Calculate the similarity between the store features of each store in the store and the keywords and categories of the search text input by the user;

[0098] The e-commerce platform will extract keywords and category information from the search text entered by the user. Then the keywords and categories of the extracted user search text will be compared with the keywords and categories in the store feature data. The keywords and categories in the store feature data are the statistical results of the keywords and categories of all the products on the shelves in the store, representing the main products and core business of the store. The e-commerce platform will traverse each store in the store feature data and compare the keywords and categories of the user search text with the keywords and categories of the store one by one.

[0099] In one embodiment, the e-commerce platform directly uses a similarity calculation method to measure the similarity between the keywords and categories of the user's search text and the keywords and categories in the store feature data. Common similarity calculation methods include cosine similarity, Jaccard similarity, Euclidean distance, Manhattan distance, etc. Cosine similarity is a commonly used similarity calculation method that measures the similarity between two vectors by calculating the cosine value of the angle between the two vectors. The closer the cosine value is to 1, the more similar the two vectors are; the closer the cosine value is to 0, the less similar the two vectors are.

[0100] When calculating similarity, the e-commerce platform will comprehensively consider the similarity between each keyword and category in the user's search text and each keyword and category in the store feature data, and weight them according to their importance. For example, the e-commerce platform can use the TF-IDF method to calculate the weight of each word and category. The TF-IDF method evaluates the importance of words and categories by calculating their frequency in the text and their inverse document frequency in the entire corpus. Words and categories with high weights will account for a larger proportion in the similarity calculation, thereby more accurately reflecting the user's search intent.

[0101] Finally, the e-commerce platform will use the calculated similarity score as the similarity score between the store features and the user's search text to measure the degree of match between the store features and the user's search intent. These similarity scores will be used for subsequent store ranking and recommendation, thereby improving user experience and platform operation efficiency. Through this step, the e-commerce platform can effectively calculate the similarity between store features and user search text, providing basic data support for accurate store recommendations.

[0102] Step S5340: Sort the stores according to the calculated similarities to form a store recommendation list.

[0103] The e-commerce platform will aggregate and organize the similarity scores of each store. These similarity scores are calculated based on the similarity between the keywords and categories of the search text entered by the user and the keywords and categories in the store feature data. The higher the similarity score, the more closely the store features match the user's search intent. The e-commerce platform will then sort the stores based on the similarity scores. The sorting process is carried out in order from high to low similarity scores, and the stores with higher similarity scores are ranked higher. In this way, the e-commerce platform can ensure that the stores that best match the user's search intent are displayed first in the recommendation list. The sorting algorithm can be a simple score-based sorting or a more complex multi-factor-based sorting algorithm, such as weighted sorting, multi-level sorting, etc.

[0104] During the sorting process, the e-commerce platform can also consider other factors to further optimize the store recommendation list. For example, the e-commerce platform can weight the similarity score based on factors such as the store's historical performance, user reviews, sales, and inventory status. Specifically, the e-commerce platform can calculate a comprehensive score for each store, which is the weighted sum of the similarity score and other factors. For example, the e-commerce platform can give higher weights to stores with good historical performance and higher weights to stores with sufficient inventory, thereby improving the quality of the recommendation list and user experience.

[0105] Finally, the e-commerce platform will form a store recommendation list based on the sorting results. This list contains the stores that are most relevant to the user's search text, arranged in descending order of comprehensive scores. When users browse the recommendation list, they can quickly find stores that meet their search intent, thereby improving user experience and platform operational efficiency.

[0106] In this embodiment, by accurately calculating the similarity between store features and user search texts, and sorting stores according to the similarity, an accurate store recommendation list is formed. This embodiment comprehensively captures the degree of match between user search intent and store features by comprehensively considering the similarity between the keywords and categories in the user search text and the keywords and categories in the store feature data, thereby improving the accuracy and relevance of the recommendation results. At the same time, based on the calculated similarity scores, the e-commerce platform can dynamically adjust the ranking of stores to ensure that the stores that best match the user's search intent are displayed first. Through similarity calculation and dynamic sorting, the e-commerce platform can quickly generate an accurate store recommendation list to improve the user's search experience.

[0107] Based on any embodiment of the method of the present application, calculating the similarity between the store features of each store in the store and the keywords and categories of the search text input by the user includes:

[0108] Step S5331, obtaining the keywords and categories of the search text input by the user;

[0109] The e-commerce platform will obtain the keywords and categories of the search text entered by the user. Then, the natural language processing (NLP) technology is used to pre-process the search text entered by the user. Through this step, the e-commerce platform can effectively extract keywords and category information from the search text entered by the user, providing basic data support for subsequent similarity calculations. These keywords and category information will be used to construct new nodes in the second graph database and input into the graph neural network model for similarity prediction, thereby improving the accuracy and relevance of the store recommendation list.

[0110] Step S5332: corresponding to the second graph database constructed based on the store feature data of each store in the store feature library, construct the keywords and the categories to which the search text belongs as new nodes of the second graph database to update the second graph database;

[0111] The e-commerce platform will extract the store feature data of each store from the store feature library. The store feature library is a database that centrally stores store feature data, including store features such as the store's category, keywords, and store unique identifier. Then the e-commerce platform will build a second graph database based on the store feature data. The second graph database is a storage system based on a graph structure, which is used to store and query complex related data. Each store feature, including the category, keyword, and store unique identifier, will be converted into a node and stored in the second graph database. After building the second graph database, the e-commerce platform will construct the keywords and categories of the search text entered by the user as new nodes of the second graph database. Specifically, the e-commerce platform will create a new node for each store's keyword and category, and store it in the second graph database. These new nodes represent the core features of the user's search text and are the basis for subsequent similarity calculations.

[0112] In order to ensure the accuracy and consistency of the newly added nodes, the e-commerce platform will use high-quality word embedding technology to convert keywords and categories into vector representations. Word embedding technology (such as Word2Vec, GloVe, BERT, etc.) can capture the semantic relationship between words and map words to a high-dimensional vector space. In this way, each keyword and category can be represented by a vector. The e-commerce platform will store these vector representations as features of the newly added nodes in the second graph database. After the new nodes are stored, the e-commerce platform will update the second graph database and associate the new nodes with the original store feature nodes. Specifically, the e-commerce platform will establish edge connections between the new nodes and the store feature nodes according to the edge connection rules of the original second graph database.

[0113] Through this step, the e-commerce platform can effectively construct the keywords and categories of the search text entered by the user as new nodes of the second graph database, and update the graph database, thereby providing basic data support for subsequent similarity calculations. These new nodes and edge connections will be used for similarity prediction in the graph neural network model, thereby improving the accuracy and relevance of the store recommendation list.

[0114] Step S5333: input the updated second graph database into the graph neural network model pre-trained using the store feature data to predict the similarity between the keywords and categories of the search text and the feature data of each store.

[0115] The e-commerce platform will prepare an updated second graph database. The updated second graph database contains the original store feature nodes and the newly added user search text keywords and category nodes, as well as the edge connections between the nodes. These nodes and edge connections form a complex graph structure that reflects the relationship between store features and user search intent.

[0116] The e-commerce platform will input the updated second graph database into the pre-trained graph neural network model. Graph Neural Network (GNN) is a deep learning model specifically designed to process graph structured data, which can capture the complex relationships between nodes and the global structure of the graph. The graph neural network model consists of multiple graph convolutional layers, each of which can aggregate the neighbor information of the node and gradually propagate and update the feature representation of the node through multi-layer convolution operations.

[0117] The e-commerce platform will input the updated second graph database into the graph neural network model. The graph neural network model will gradually update the feature representation of the node through multi-layer graph convolution operations based on the structure of the graph and the node features. Specifically, the graph neural network model will first aggregate the neighbor information of each node, and then update the feature representation of the node through nonlinear transformation. This process will be repeated many times until the feature representation of the node reaches a stable state.

[0118] In the last layer of the graph neural network model, the e-commerce platform uses a fully connected layer (Fully ConnectedLayer) to map the feature representation of the node to the similarity score. Specifically, the e-commerce platform compares the keywords and category nodes of the user's search text with the feature nodes of each store to calculate the similarity score. Finally, the e-commerce platform uses the calculated similarity score as the similarity score between the store features and the user's search text to measure the degree of match between the store features and the user's search intent. These similarity scores will be used for subsequent store sorting and recommendations, thereby improving user experience and the platform's operational efficiency. Through this step, the e-commerce platform can effectively use the graph neural network model to predict the similarity between store features and user search texts, providing basic data support for accurate store recommendations.

[0119] It should be noted that before training the graph neural network model, the e-commerce platform first needs to prepare training data. That is, store feature data and corresponding labels. Store feature data is extracted from the store feature library, which includes store features such as the store's category, keywords, and store unique identifier. The label represents the similarity between store feature data, and the e-commerce platform will convert this data into a graph structure. Then the e-commerce platform will design the architecture of the graph neural network model. The graph neural network model consists of multiple graph convolution layers, each of which can aggregate the neighbor information of the node and gradually propagate and update the feature representation of the node through multi-layer convolution operations. The e-commerce platform will select the appropriate graph convolution layer type and number of layers according to the specific task requirements and data characteristics. After the model architecture design is completed, the e-commerce platform will use the prepared training data to train the graph neural network model. The training process includes the following steps: First, the e-commerce platform will input the training data into the graph neural network model and gradually update the feature representation of the node through multi-layer graph convolution operations. Specifically, the graph neural network model will first aggregate the neighbor information of each node, and then update the feature representation of the node through nonlinear transformation. This process will be repeated many times until the feature representation of the node reaches a stable state.

[0120] In each round of training, the e-commerce platform calculates the error between the similarity score output by the model and the true label, and uses the back-propagation algorithm to update the model parameters. The back-propagation algorithm calculates the gradient of the error to the model parameters and updates the model parameters using the gradient descent method or other optimization algorithms to minimize the error. The e-commerce platform sets appropriate loss functions and optimization algorithms, such as the mean square error (MSE) loss function and the Adam optimization algorithm, to improve the training effect and convergence speed of the model. The training process is repeated for multiple rounds until the performance of the model reaches the preset indicators or the training data is fully learned. The e-commerce platform will evaluate the model using the validation data set to monitor the training effect and generalization ability of the model. If the model performs poorly on the validation data set, the e-commerce platform will adjust the model architecture, hyperparameters or training strategy and retrain.

[0121] After the model training is completed, the e-commerce platform will use the trained graph neural network model to predict the similarity between the keywords and categories of the user's search text and the feature data of each store.

[0122] In this embodiment, by obtaining the keywords and categories of the search text input by the user, a second graph database is constructed and updated, and a pre-trained graph neural network model is used to predict the similarity between store features and the user search text, thereby achieving accurate store recommendations. This embodiment constructs the keywords and categories of the user search text as new nodes of the second graph database, dynamically updates the graph structure, captures the complex relationship between the user's search intent and the store features, and provides basic data support for similarity calculation. At the same time, using the pre-trained graph neural network model, the e-commerce platform can efficiently process complex graph structure data, capture the complex relationship between nodes and the global structure of the graph, thereby improving the accuracy and relevance of similarity prediction. Through the multi-layer graph convolution operation and fully connected layer mapping of the graph neural network model, the e-commerce platform can dynamically calculate the similarity between store features and user search texts, ensure that the recommendation results are highly matched with the user's search intent, and improve the user experience and the platform's operational efficiency.

[0123] See also Figure 2 According to one aspect of the present application, a store product recommendation device is provided, including a search text acquisition module 5100, a search text processing module 5200, a first-stage recall module 5300, and a second-stage recall module 5400, wherein the search text acquisition module 5100 is configured to respond to a user search request and obtain a search text input by a user; the search text processing module 5200 is configured to extract keywords of the search text and determine the corresponding category according to the search text; the first-stage recall module 5300 is configured to perform a store-level recall operation on the store feature library in the first stage, recall some stores whose store features match the keywords and the categories to which they belong, and form a store recommendation list; the second-stage recall module 5400 is configured to perform a product-level recall operation on the product feature data of the products of the stores in the store recommendation list in the product feature library in the second stage, recall some products whose product features match the keywords and the categories to which they belong, and form a product recommendation list.

[0124] On the basis of any embodiment of the device of the present application, the search text acquisition module 5100 includes: a product information acquisition submodule, which is configured to extract product information of each product from a product database; a product feature generation submodule, which is configured to call a large model to extract keywords and predict the categories to which the product information of each product belongs, as the corresponding product features in the product feature data of the product; a multilingual translation submodule, which is configured to translate the product feature data into multiple languages ​​and store them in a product feature library, so as to achieve feature matching corresponding to the keywords and the categories to which they belong based on the same language when executing the recall operation.

[0125] On the basis of any embodiment of the device of the present application, before the search text acquisition module 5100, it includes: a product feature acquisition submodule, which is configured to obtain the categories and keywords of all products in each store in the product feature library; a store feature composition submodule, which is configured to filter out a preset number of categories and keywords according to preset filtering rules as the categories and keywords of the store information corresponding to the store, and the associated store unique identifier is stored in the store feature library as the corresponding store feature in the store feature data.

[0126] Based on any embodiment of the device of the present application, the first-stage recall module 5300 includes: a graph database generation submodule, configured to store the corresponding store features in the store feature data as nodes of the first graph database in the first graph database, and establish corresponding edge connections between the nodes according to a predefined business logic; a feature matching submodule, configured to use a specific graph matching algorithm to match in the first graph database according to the keywords and categories of the search text input by the user, and screen out some stores whose store features match the keywords and categories to form a store recommendation list.

[0127] On the basis of any embodiment of the device of the present application, after the feature matching submodule, it includes: a text relevance score calculation submodule, which is configured to use an information retrieval algorithm to calculate the text relevance score between the search text input by the user and the store feature data of each store in the store; a category relevance score calculation submodule, which is configured to determine whether the category to which the search text input by the user belongs matches the category in the store feature data of each store. If so, the category relevance score is recorded as 1, otherwise it is recorded as 0; a keyword relevance score calculation submodule, which is configured to determine whether the keywords of the search text input by the user match the keywords in the store feature data of each store. If so, the keyword relevance score is recorded as 1, otherwise it is recorded as 0; a total relevance score calculation submodule, which is configured to calculate the total relevance score of each store based on the corresponding text relevance score, category relevance score and keyword relevance score of each store, and sort the stores according to the total relevance score to form a store recommendation list.

[0128] On the basis of any embodiment of the device of the present application, after the feature matching submodule, it includes: a similarity calculation submodule, which is configured to calculate the similarity between the store characteristics of each store in the store and the keywords and categories of the search text input by the user; a store sorting submodule, which is configured to sort the various stores according to the calculated similarities to form a store recommendation list.

[0129] Based on any embodiment of the device of the present application, the similarity calculation submodule includes: a text feature acquisition submodule, which is configured to obtain the keywords and categories of the search text input by the user; a second graph database update submodule, which is configured to correspond to the second graph database constructed based on the store feature data of each store in the store feature library, and construct the keywords and categories of the search text as new nodes of the second graph database to update the second graph database; a similarity prediction submodule, which is configured to input the updated second graph database into a graph neural network model pre-trained using the store feature data, to predict the similarity between the keywords and categories of the search text and the feature data of each store.

[0130] In order to solve the above technical problems, the present application also provides a computer device. Figure 3 As shown, a schematic diagram of the internal structure of a computer device. The computer device includes a processor, a computer-readable storage medium, a memory, and a network interface connected via a system bus. Among them, the computer-readable storage medium of the computer device stores an operating system, a database, and computer-readable instructions. The database may store a control information sequence. When the computer-readable instructions are executed by the processor, the processor can implement a store product recommendation method. The processor of the computer device is used to provide computing and control capabilities to support the operation of the entire computer device. The memory of the computer device may store computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor can execute the store product recommendation method of the present application. The network interface of the computer device is used to connect and communicate with a terminal. Those skilled in the art can understand that Figure 3 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0131] In this embodiment, the processor is used to execute Figure 2 The memory stores the program code and data required to execute the above modules or submodules. The network interface is used to transmit data between user terminals or servers. The memory in this embodiment stores the program code and data required to execute all modules / submodules in the store product recommendation device of this application, and the server can call the program code and data of the server to execute the functions of all submodules.

[0132] The present application also provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the store product recommendation method of any embodiment of the present application.

[0133] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments of the present application can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the aforementioned storage medium can be a computer-readable storage medium such as a disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0134] It will be understood by those skilled in the art that the various operations, methods, steps, measures, and schemes in the processes discussed in this application may be alternated, changed, combined, or deleted. Furthermore, other steps, measures, and schemes in the various operations, methods, and processes discussed in this application may also be alternated, changed, rearranged, decomposed, combined, or deleted. Furthermore, the steps, measures, and schemes in the various operations, methods, and processes in the prior art that are open source and in this application may also be alternated, changed, rearranged, decomposed, combined, or deleted.

[0135] The above description is only a partial implementation method of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A store product recommendation method, characterized in that: include: Respond to user search requests and obtain search text entered by the user; Extracting keywords from the search text, and determining corresponding categories according to the search text; In the first stage, a store-level recall operation is performed on the store feature library to recall some stores whose store features match the keywords and the categories to which they belong, thereby forming a store recommendation list; In the second stage, a product level recall operation is performed on the product feature data of the products owned by the stores in the store recommendation list in the product feature library, and some products whose product features match the keywords and the categories to which they belong are recalled to form a product recommendation list.

2. The store product recommendation method according to claim 1, characterized in that: Before responding to user search requests, including: Extract product information of each product from the product database; Calling the big model to extract keywords and predict the categories of the commodity information of each commodity as corresponding commodity features in the commodity feature data of the commodity; The product feature data is translated into multiple languages ​​and stored in a product feature library, so that when the recall operation is performed, feature matching corresponding to the keyword and the category to which it belongs can be achieved based on the same language.

3. The store product recommendation method according to claim 1, characterized in that: Before responding to user search requests, including: Get the categories and keywords of all products in each store in the product feature database; According to the preset filtering rules, a preset number of categories and keywords are filtered out as the categories and keywords corresponding to the store information of the store, and the associated store unique identifier is stored in the store feature library as the corresponding store feature in the store feature data.

4. The store product recommendation method according to claim 3, characterized in that: In the first stage, a store-level recall operation is performed on the store feature library to recall some stores whose store features match the keywords and the categories to which they belong, forming a store recommendation list, including: Store corresponding store features in the store feature data as nodes of the first graph database in the first graph database, and establish corresponding edge connections between the nodes according to predefined business logic; According to the keywords and categories of the search text input by the user, a specific graph matching algorithm is used to match the first graph database, and some stores whose store characteristics match the keywords and categories are screened out to form a store recommendation list.

5. The store product recommendation method according to claim 4, characterized in that: After filtering out some stores whose characteristics match the keywords and categories, including: Calculating a text relevance score between a search text input by a user and store feature data of each store in the store by using an information retrieval algorithm; Determine whether the category of the search text input by the user matches the category in the store feature data of each store. If so, the category relevance score is recorded as 1, otherwise, it is recorded as 0; Determine whether the keyword of the search text input by the user matches the keyword in the store feature data of each store. If so, the keyword relevance score is recorded as 1, otherwise it is recorded as 0; According to the corresponding text relevance scores, category relevance scores and keyword relevance scores of the various stores, the total relevance scores of the various stores are calculated, and according to the total relevance scores, the various stores are sorted to form a store recommendation list.

6. The store product recommendation method according to claim 4, characterized in that: After filtering out some stores whose characteristics match the keywords and categories, including: Calculating the similarity between the store features of each store in the store and the keywords and categories of the search text input by the user; According to the calculated similarity, the stores are sorted to form a store recommendation list.

7. The store product recommendation method according to claim 6, characterized in that: Calculating the similarity between the store features of each store in the store and the keywords and categories of the search text input by the user, including: Get the keywords and categories of the search text entered by the user; Corresponding to the second graph database constructed based on the store feature data of each store in the store feature library, the keywords and the categories to which the search text belongs are constructed as new nodes of the second graph database to update the second graph database; The updated second graph database is input into the graph neural network model pre-trained using the store feature data to predict the similarity between the keywords and categories of the search text and the feature data of each store.

8. A store product recommendation device, characterized in that: include: A search text acquisition module, configured to respond to a user search request and acquire a search text input by the user; A search text processing module, configured to extract keywords from the search text and determine corresponding categories according to the search text; The first-stage recall module is configured to perform a store-level recall operation on the store feature library in the first stage, recalling some stores whose store features match the keywords and the categories to which they belong, and forming a store recommendation list; The second-stage recall module is configured to perform a product-level recall operation on the product feature data of the products of the stores in the store recommendation list in the product feature library in the second stage, and recall some products whose product features match the keywords and the categories to which they belong, to form a product recommendation list.

9. A store product recommendation device, comprising a central processing unit and a memory, characterized in that: The central processing unit is used to call and run the computer program stored in the memory to execute the steps of the method according to any one of claims 1 to 7.

10. A non-volatile readable storage medium, characterized in that: It stores a computer program implemented according to the method described in any one of claims 1 to 7 in the form of computer-readable instructions, and when the computer program is called and executed by a computer, the steps included in the corresponding method are executed.

Images