Commodity related word recommendation method and device

Abstract

The invention relates to a commodity related word recommendation method and device. The method comprises the steps that multiple scene domains inside and outside a commodity platform and targets corresponding to the scene domains are determined; according to the multiple scene domains and the corresponding targets, determining a corresponding recommendation implementation mode; performing word recommendation operation according to the corresponding recommendation implementation mode to obtain a word recommendation result; and according to the word recommendation result, a shopping guide center and a pallet center are adopted to execute personalized recommendation corresponding to the current user for the user portrait and the user growth source. Starting from a search growth scene of shopping guide query recommendation, on the basis of global information multi-dimensional modeling and different-domain information scene transfer modeling, a large model is utilized to optimize word recommendation capability, general basic capability construction is carried out, full-link efficiency improvement of users in different life cycles is continuously optimized, and then search penetration and search growth are improved.

Description

Technical Field

[0001] The present application relates to the field of e-commerce technology, and in particular to a method and device for recommending product-related terms. Background Art

[0002] In the process of recommending words related to e-commerce products, search guides are an important part of search. They are the most direct and effective means to improve search penetration and achieve search growth within the product platform. They are responsible for stimulating users' search interests and play an important role in helping users efficiently describe their needs and stimulate new user needs.

[0003] Currently, major search platforms usually have their own product shopping guide word recommendation solutions. However, the goal of existing solutions is mainly to increase the search UV (Unique Visitor) of the shopping guide domain. They do not start from the overall search growth of the application (including search penetration, search acceptance, and search UV). They are usually unable to utilize global information to balance the information gap inside and outside the product platform to accurately capture user interests and preferences. The problems faced by existing solutions include: (1) they cannot provide differentiated word recommendation services for users in different life cycles (newcomers, low activity, and high activity); (2) they cannot understand the differentiated intentions of users in different scenario domains; (3) they inefficiently accept user behavior on the search results page, which brings a poor user experience. Summary of the Invention

[0004] In response to the technical problems faced by the existing technology, this application provides a recommendation solution for product-related words. Starting from the search growth scenario of shopping guide query recommendation, based on multi-dimensional modeling of global information and transfer modeling of heterogeneous information scenarios, it uses a large model to optimize the word recommendation capabilities, build general basic capabilities, and continuously optimize the full-link efficiency of users in different life cycles, thereby improving search penetration and search growth.

[0005] According to a first aspect of the present application, a method for recommending product-related words is provided, characterized by comprising:

[0006] Determine multiple scenario domains within and outside the product platform and the goals corresponding to the multiple scenario domains;

[0007] Determine corresponding recommended implementation methods based on the various scenario domains and corresponding goals;

[0008] Perform a word recommendation operation according to the corresponding recommendation implementation method to obtain a word recommendation result; and

[0009] According to the word recommendation results, the shopping guide center and the product tray center are used to perform personalized recommendations corresponding to the current user based on the user portrait and user growth source.

[0010] According to a second aspect of the present application, there is provided a device for recommending product-related words, characterized by comprising:

[0011] A first determination module is configured to determine multiple scene domains within and outside the commodity platform and targets corresponding to the multiple scene domains;

[0012] A second determination module is configured to determine corresponding recommended implementation methods based on the multiple scenario domains and corresponding objectives;

[0013] an obtaining module, configured to perform a word recommendation operation according to the corresponding recommendation implementation method to obtain a word recommendation result; and

[0014] The execution module is used to execute personalized recommendations corresponding to the current user based on the word recommendation results, using the shopping guide center and the product tray center for user portraits and user growth sources.

[0015] According to a third aspect of the present application, an electronic device is provided, comprising a memory storing one or more programs and a processor electrically coupled to the memory and configured to execute the one or more programs to perform any method or step or combination thereof in the present application.

[0016] According to a fourth aspect of the present application, a computer program product is provided, comprising a computer program, which enables the computer to execute any one of the methods provided in the first aspect when the computer program is run on the computer.

[0017] According to a fifth aspect of the present application, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it causes the execution of any method or step or a combination thereof in the present application.

[0018] The above and other aspects and implementations thereof are described in more detail in the drawings, the description, and the claims.

[0019] According to the solution provided by this application, on the one hand, not only various post-search solutions are proposed based on the user's behavior in each scenario domain, but also scenario transfer targets are constructed in each scenario domain, new shopping guide search touchpoints are built in different scenario domains, and information in each domain is integrated and divided. Combined with the user growth scenario characteristics of the product platform outside the different domains, users are attracted to search; on the other hand, interest transfer in each scenario domain is modeled. Compared with the query recommendation model of other platforms, this application not only models the word click efficiency, but also uses the natural reasoning advantages of the large model and the optimization goals of the full-link multi-scenario to stabilize the balance and prosperity of the market (the global stability of the platform's core indicators) and the ecology, bringing about an increase in the conversion efficiency of users in the market from access to key behavior links. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without exceeding the scope of protection required by this application.

[0021] Figure 1 This is the overall architecture diagram of the product-related word recommendation system of this application.

[0022] Figure 2 This is a diagram of the overall architecture of the infrastructure within the commodity platform according to an embodiment of the present application.

[0023] Figure 3 It is a flowchart of the differentiated construction of internal and external scene domains of the commodity platform according to the embodiment of the present application.

[0024] Figure 4 It is a framework diagram of user intent modeling according to an embodiment of the present application.

[0025] Figure 5 This is a flowchart of generating query statements using a large model according to an embodiment of the present application.

[0026] Figure 6 This is a flowchart of a method for recommending product-related words according to an embodiment of the present application.

[0027] Figure 7 This is a flowchart of a method for recommending product-related words according to another embodiment of the present application.

[0028] Figure 8 This is a flowchart of a method for recommending product-related words according to another embodiment of the present application.

[0029] Figure 9 This is a flowchart of a method for recommending product-related words according to another embodiment of the present application.

[0030] Figure 10 2 is a schematic diagram of a device for recommending product-related words according to an embodiment of the present application.

[0031] Figure 11 2 is a schematic diagram of a device for recommending product-related words according to another embodiment of the present application.

[0032] Figure 12 2 is a schematic diagram of a device for recommending product-related words according to another embodiment of the present application.

[0033] Figure 13 2 is a schematic diagram of a device for recommending product-related words according to another embodiment of the present application.

[0034] Figure 14 This is a structural diagram of an electronic device provided by this application. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0036] Overall, the solution of this application achieves the goal of search growth through differentiated construction of scene domains inside and outside the commodity platform and infrastructure construction within the commodity platform.

[0037] In a specific embodiment, "differentiated construction of scene domains inside and outside the product platform" can refer to designing different recommendation strategies based on the different locations of the user's environment (inside / outside the product platform). For example, for the construction of the external scene domain: before the user enters the application (APP), a search engine will display "winter sweatshirt clearance sale" to stimulate the user with price and attract clicks with content; for the construction of the internal scene domain: after the user enters the application (APP), the background side of the homepage will recommend "polar fleece sweatshirt" based on the user's previous search page behavior to maintain consumption intention.

[0038] In a specific embodiment, "infrastructure construction within the commodity platform" can refer to optimizing the quality of database queries based on user search and click intent, combining user portraits, increasing link recall diversity, upgrading the sorting model, and optimizing the reordering mechanism based on user movement. Figure 2 , including infrastructure construction, recall construction, ranking construction and strategy.

[0039] In a specific embodiment, through the infrastructure construction within the product platform, the optimization of the search conversion path after the user enters the application can be more effective, which can quickly increase the UV value. At the same time, combined with the differentiated construction of internal and external scene domains, it can attract traffic from the external scene domain to increase new search users, the internal scene takes over the conversion, and the data is fed back to the external scene domain for precise delivery.

[0040] Figure 1 This is the overall architecture diagram of the product-related word recommendation system of this application. Figure 1In order to build the ability to recommend shopping guide terms across multiple dimensions, we leverage user demand graphs and multi-link target optimization to achieve the goal of attracting new users and promoting active users. We dig deeper into multiple areas. In the scenario dimension, we leverage a multi-touchpoint model based on the characteristics of different user growth scenarios to achieve a user-ready-to-search effect. In the user dimension, we drive the integration of general algorithm capabilities upward. This requires not only constructing feature samples within the product platform's external domain and training meta-learning architectures for new, low-activity users, a key target group for search recommendation user growth, but also continuously optimizing the full-link efficiency of medium- and high-activity users. Leveraging the inherent reasoning advantages of large models and full-link, multi-scenario optimization goals, we aim to maintain a balanced prosperity within the broader market and the ecosystem. Search guides, in particular, involve attracting users to click on words and enter searches at various entry points. Their primary function is to stimulate user interest before searching, provide convenient search guidance during searching, and help users clarify their needs and discover potential needs after searching, thereby promoting search user growth and increasing user search stickiness. Figure 1 The "dual-circulation-in-one" shopping guide center and the pallet center are designed to provide services in various scenario domains (scenario dimension) and various types of users (user dimension). It is necessary to build these two centers to play a connecting role. Among them, the shopping guide center is a modular center with the core goals of user traffic operation and conversion efficiency improvement. It focuses on demand-side data mining and scenario-based reach, and emphasizes user traffic and conversion, involving recommendation algorithms, promotional strategies, content recommendations, etc. The pallet center is a modular center with commodity supply optimization and supply chain efficiency as its core, focusing on structural adaptation and agile response on the supply side. It usually refers to the commodity management module, covering product selection, inventory, supply chain and other links.

[0041] In some embodiments, product-related words may include product-related search terms, keywords, hot search terms, etc., and this application does not impose any restrictions on this.

[0042] In some embodiments, the overall solution of this application includes three aspects: differentiated construction of scene domains inside and outside the commodity platform, precise user intent modeling, and optimization of large-model recommendation word expressions. Among them, in the differentiated construction of scene domains inside and outside the commodity platform, through the infrastructure construction within the commodity platform and the differentiated construction of scene domains inside and outside the commodity platform, based on the user's search and click intention, combined with user portraits, the link recall diversity is increased, the sorting model is upgraded, and the user movement optimization and re-arrangement mechanism is combined to build accurate recommendations of user interests. From a technical point of view, in terms of query quality infrastructure construction, user behavior strategy recall, and sorting model upgrades, the search side's background, search discovery and recommendation side's weather vane, first guess hot words and other query recommendation products are optimized. On this basis, combined with the characteristics of user growth scenarios outside the commodity platform in different domains, the user demand map and multi-touch mode are used to achieve the effect of user search and use. Among them, different model architectures are constructed for different users on the recall and sorting sides. In the precise modeling of user intentions and the optimization of the expression of recommendation words of large models, we build the full-link shopping guide word recommendation capability of multi-dimensional entities, and the two-way circulation of the shopping guide center and the pallet center. We use the user demand map and multi-link target optimization to achieve the goal of attracting new customers and promoting activation. We use the natural reasoning advantages of large models and the optimization goals of full-link multiple scenarios to stabilize the balance and prosperity of the market and the ecology.

[0043] The three aspects of differentiated construction of scene domains inside and outside the commodity platform, precise user intent modeling, and optimization of the expression of large-model recommendation words form a parallel collaborative relationship, which has both functional independence and system complementarity. Specifically, the differentiated construction of scene domains inside and outside the commodity platform focuses on multi-dimensional product layout, and is responsible for building differentiated distribution strategies and infrastructure construction for products inside and outside the commodity platform; precise user intent modeling focuses on user intent modeling, and realizes accurate demand capture through strategic-level scene domain interest migration modeling; optimization of large-model recommendation word expression focuses on semantic optimization, and builds a multi-dimensional semantic expression optimization system. The three form a technical closed loop of "spatial layout-demand analysis-semantic enhancement", and ultimately work together to build an intelligent shopping guide system ecosystem through the ecological deployment of the product matrix, in-depth modeling of user intent, and continuous optimization of the expression system.

[0044] Differentiated development within and outside the product platform typically involves selecting different solutions to meet user needs for different scenarios. This includes developing the homepage's background, developing search and discovery capabilities, and developing external traffic. The goal is to increase search unique visitors (UVs) and attract users to click on recommended keywords. This differentiated development includes analyzing scenario characteristics, building a user needs map, implementing a dual-hub circulation mechanism, and facilitating cross-scenario collaboration. Among them, scenario characteristic analysis can refer to dividing different scenario domains inside and outside the terminal (such as homepage background, search discovery, product platform off-site traffic, etc.), and clarifying the user behavior characteristics and goals of each scenario domain; building a user demand map refers to generating user demand labels through data collection (search records, click behavior, etc.), and associating multiple scenario touchpoints; the dual-hub circulation mechanism includes a shopping guide hub and a merchandise tray hub. The shopping guide hub selects people and matches personalized recommendation pools (such as coupons, products) based on user portraits and sources (such as new users / old customers). The merchandise tray hub dynamically adjusts the product exposure priority according to traffic distribution (such as hot-selling traffic, long-tail product supplementation); cross-scenario collaboration refers to designing jump paths between scenario domains (such as product platform off-site traffic → search → transaction) to ensure seamless connection of user needs.

[0045] Differentiated development within and outside the product platform's scenarios combines the specific characteristics of user growth scenarios within and outside the product platform, leveraging user demand maps and a multi-touchpoint model to achieve instant search. A two-way loop between the shopping guide hub and the assortment hub allows for personalized recommendations and exclusive engagement based on user profiles and user growth sources, while also regulating traffic within the assortment hub. Figure 3 This is a flow chart of the differentiated construction of internal and external scene domains of the commodity platform according to the embodiment of this application. Figure 3 As shown, the process includes clarifying the scenario domain types inside and outside the commodity platform and the corresponding goals of each scenario domain; determining the corresponding recommendation implementation methods based on the scenario domains and the corresponding goals; performing word recommendation operations according to the corresponding recommendation implementation methods to obtain word recommendation results; based on the word recommendation results, using the shopping guide center and the shelf center to perform personalized recommendations corresponding to the current user based on the user portrait and user growth source.

[0046] In one embodiment, the homepage shading side mainly uses the user's search terms, clicks, favorites, purchases, and other behaviors in various scenario domains to build a differentiated query recommendation interest model. With "scenario prediction-behavior perception-dynamic response" as the core logic, a seamless "use and search" experience can be created. The corresponding implementation method of the homepage shading side is to generate behavior-scenario domain transfer data, and by integrating multi-dimensional user behavior data with real-time scenario domain prediction technology, a closed-loop response from behavior analysis to precise search guidance is completed, including:

[0047] 1. Behavioral data tracking: Collect user behavior data such as browsing time on the homepage, click hot spots, and sliding tracks;

[0048] 2. Scenario domain transfer analysis: Identify the user's jump path from the homepage to the search / product details page, generate behavior-scenario domain transfer data, and extract high-frequency triggering scenarios (such as "entering keywords after clicking the search icon").

[0049] After generating the behavior-scenario domain transfer data, real-time interest matching can be achieved, such as predicting demand based on current behavior (for example, a user repeatedly browses shoe products → the background shows "sports shoes"); it can also achieve scene domain association recommendations. For example, when a user returns to the background side from the image search scene, the background should recommend the query corresponding to the product the user just searched.

[0050] When the context domain is the background texture of the homepage, the corresponding goal is to achieve a user-instant-search effect based on user behavior and the transition between context domains. The corresponding implementation method is to construct behavior-context domain transition data. Generating behavior-context domain transition data includes: collecting user behavior data; and generating behavior-context domain transition data based on the behavior data and the user's jump path from the homepage to the search / product details page.

[0051] In one embodiment, the purpose of the search discovery domain is to use information inside and outside the commodity platform to determine user demand preferences, model user interests, and introduce multi-dimensional interests. The search discovery domain is a page belonging to the search middle page. The basic principle of implementation is similar to that of the shading, but this function emphasizes the user's search intent. Compared with the shading, the difference is that the search discovery domain focuses on the user's search behavior and search intent, performs "query to query" (query2query) recall based on the search terms, and recommends words for different scenarios to the user. The corresponding implementation method of the search discovery domain is to generate multi-dimensional interest graph data.

[0052] In a specific embodiment, generating multi-dimensional interest graph data includes cross-end data integration, multi-dimensional interest modeling and / or search discovery word generation. In some embodiments, cross-end data integration may include aggregating search behavior within the commodity platform and user behavior data outside the commodity platform to build a user interest pool; multi-dimensional interest modeling may include short-term interest modeling (recent search terms, such as "Dragon Boat Festival rice dumpling gift box") and long-term interest modeling (historical category preferences, such as maternal and child products). Search discovery word generation may include integrating interest tags and sorting by weight, such as "Users often buy pet products → Recommend 'cat food'", and may also combine popularity-related data (hot search list) with personalized tags for mixed display, such as based on 30% hot search list and 70% personalized tags.

[0053] Generating multi-dimensional interest graph data may include: building a user interest pool by aggregating user behavior data within and outside the product platform; and / or performing user interest modeling based on the user's short-term interest data and long-term interest data; and / or generating search discovery words based on user tags (including interest tags and personalized tags) and / or popularity-related data.

[0054] In one specific embodiment, the integrated search and recommendation domain reconstructs shopping guide products (including trending indicators and first-guess hot words) in the first-guess scenario, aiming to stimulate demand and guide search transactions. The implementation of the integrated search and recommendation domain includes generating a personalized deep model based on data related to users' potential purchase intentions. This data can be based on users' purchasing behavior or browsing data to predict the products that users may be most interested in.

[0055] In a specific embodiment, the off-platform traffic flow domain of the commodity platform adds personalized search term touchpoints for pushes with different user growth intentions, and can combine AI (artificial intelligence) technology to understand user intentions and recommend rich queries to achieve cross-commodity platform references. The implementation method of the off-platform traffic flow domain of the commodity platform includes user intent mapping. In some embodiments, user intent mapping is adopted, including: determining push strategies based on different intent types; and / or parsing user historical behavior data to extract potential demand words; and / or executing a jump to the current search page to pre-fill recommended words in response to the user clicking on the push.

[0056] In a specific embodiment, intent types may include explicit intent and ambiguous intent. For example, an explicit intent may be that a user has clicked on the same product multiple times but has not completed the transaction, and the push strategy is to push "exclusive discount + precise search terms" (such as "click to search [xx mobile phone direct discount of 500]"); an ambiguous intent may be that a user has browsed the product category but has not refined it, and the push strategy is to push "scenario-based guide words" (such as "summer outfit guide").

[0057] In another specific embodiment, NLP (natural language processing) can be used to analyze user historical behaviors (such as "Which projector is the best"), extract potential demand words (such as "home projector review"), and generate high-conversion queries.

[0058] In another specific embodiment, after the user clicks on a push, the page is redirected to pre-fill recommended words and corresponding products are displayed (for example, after searching for "affordable projector", high-cost-effective products are displayed first).

[0059] In addition, this solution has built a shopping guide acceptance platform within the product platform, and personalized product acceptance will use the shopping guide capabilities within the product platform to guide searches by calling the interests of customers outside the product platform.

[0060] In some embodiments, user precision intention modeling builds the full-link shopping guide word recommendation capability of multi-dimensional entities, introduces more high-quality behaviors and enhances the accuracy of capturing the intentions behind user behaviors, and uses user demand maps and multi-link target optimization to achieve the goal of attracting new users and promoting activation, stabilizing the balance and prosperity of the market and the ecology. Executing user precision intention modeling includes basic vocabulary capacity building, demand capacity upgrading, and differentiation of user scenario interest transfer. User precision intention modeling can refer to Figure 4 .

[0061] In a specific embodiment, for the construction of basic vocabulary capabilities, the vocabulary is simplified by constructing a multi-dimensional query semantic analysis, the text quality of the candidate pool query is improved, and the user's basic shopping guide experience is improved. The construction of basic vocabulary capabilities can include two stages: the first stage (multi-dimensional semantic understanding construction), through the basic layer (part-of-speech tagging + entity recognition), the intention layer (search type classification), and the emotional layer (tendency judgment) to achieve deep semantic analysis, the training data uses historical user search word data, and integrates post-click behavior data (such as purchase rate). At the same time, a label system covering product dimensions (category / price range), behavior dimensions (decision-making stage), and time and space dimensions (season / geographic location) is constructed. For example, "winter thick down jacket" is parsed into a label chain of [apparel-winter clothing] [high customer order] [mid-term decision-making]. The second stage (lexicon simplification): Deduplication is achieved through the edit distance algorithm (for example, ED (edit distance) < 3 and semantic similarity > 0.85 are merged) and synonym library mapping (such as "laptop" → "notebook computer"), intercepting invalid queries (such as garbled characters) and XGBoost model prediction of low-quality words (features include exposure conversion ratio).

[0062] In a specific embodiment, for demand capability upgrades, for zero-few behaviors, the user's basic identity information is used to model group personalization to attract users to click and enter the search. Among them, "zero-few behaviors" can refer to the low activity and weak demand expression characteristics shown by users in product scenarios. For example, "effective operations ≤ 5 times in the past 30 days" is defined as "zero-few behaviors". Although users with zero-few behaviors have fewer behaviors on the product platform, their basic user identity information exists, such as gender, age, and identity (young women, middle-aged women, cross-border e-commerce, etc.). By collecting the basic identity information of users with zero-few behaviors (such as old customers), GMM (Gaussian Mixture Model Clustering) clustering is performed according to age, gender, and identity, and the recommended words obtained by aggregating the identity information of users with zero-few behaviors are pushed to new users with the same identity.

[0063] In a specific embodiment, for the differentiation of user scenario interest transfer, if there is expression behavior on the platform, the user intention is refined based on different scenarios, and multi-scenario modeling is used using a ranking model to stimulate users to initiate searches, establish interest transfer relationships between scenario domains, and determine the degree of refinement of word recommendations in conjunction with behavioral data.

[0064] Among them, using the ranking model for multi-scenario modeling may include:

[0065] 1. Cross-scenario data alignment

[0066] a. Build a unified ID system: user ID, product ID, scenario ID (e.g. scenario code: search = 001, recommend = 002)

[0067] b. Data association storage: Associating multi-scenario exposure and click data through behavioral logs (such as the user's complete link from search → recommendation → purchase).

[0068] 2. Feature Stratification

[0069] a. Globally shared features: user profile (e.g., age / gender), product attributes (e.g., category / price), and real-time behavior (e.g., clicks in the last hour)

[0070] b. Scenario-specific features:

[0071] Search scenario: query word intent classification

[0072] Recommendation scenario: user long-term and short-term interest vectors

[0073] 3. Model building: multi-scenario joint modeling framework

[0074] a. Bottom-layer shared layer: Use Transformer to perform cross-scenario feature cross-pollination (such as calculating the attention weight of user historical behavior and product attributes)

[0075] b. Scene adaptation layer:

[0076] MMoE (Multi-gate Mixture-of-Experts) structure: Each scene corresponds to an independent gating network, and the expert model weight is dynamically assigned.

[0077] c. Output layer: output sorting by scene

[0078] For user behavior data, such as click-after-search and search-after-search data, we can generate data based on user click and search behaviors. For example, if a user searches for "men's sports shoes," clicks on the product "Nike shoes" on the search results page, and then initiates a new search for "Nike shoes," we can build a search-after-search recall link for "men's sports shoes" -> "Nike shoes" based on the user's behavior.

[0079] In some embodiments, performing precise user intent modeling includes:

[0080] Improve the text quality of candidate pot queries by building a multi-dimensional query semantic parser to simplify the vocabulary; and / or

[0081] Pushing the recommended words determined by the identity information of the zero-less behavior user to a new user with the same identity information; and / or

[0082] The degree of refinement of word recommendations is determined based on the interest transfer relationship between scene domains and user behavior data.

[0083] In some embodiments, as described in the user precise intent modeling section, in order to stimulate user clicks, the solution adopted by this application is to "introduce more high-quality behaviors and enhance the accuracy of capturing the intent behind user behaviors", and propose recall solutions such as search after click and search after purchase that are in line with user behavior patterns. In addition, the expression of recommended words is also one of the key factors in stimulating user clicks. Therefore, based on the guessed intent, this application stimulates the user's search intent by optimizing the expression of recommended words, thereby increasing click-through rate and increasing search penetration. The large model recommended word expression optimization solution proposed in this application is as follows: Figure 5 shown.

[0084] Based on the user's existing intention (long-term and short-term), for long-term and short-term recall and weak personalized recall queries, the query trend rewriting and rising hot spot generation e-commerce word method of the large model is used to replace the recalled query. (1) Data preparation: First, collect and organize high-quality search terms, including recommended terms pushed by the product platform and high-frequency search terms entered by users inside and outside the product platform, as the basis for retrieval. (2) Integrated large model: Select a pre-trained large language model that meets the task requirements. Preferably, it can be fine-tuned to better adapt to the language style and professional terminology in a specific field. Among them, the pre-trained large language model can include an LLM (Large Language Model) model. By inputting user preference information or context, the model can generate more personalized and natural recommendations. (3) Combined with the RAG (Retrieval-Augmented Generation) framework, using existing basic real-time data, the retrieval model and the generation model are combined to incubate differentiated hot spots inside and outside the product platform. Through the above steps, the large model is effectively used in combination with RAG technology to provide users with more accurate and attractive recommendations, thereby improving the user experience.

[0085] In one specific implementation, the method of optimizing the expression of recommendation words includes: collecting search words input by users in the shopping guide behavior of the product platform and recommendation words pushed by the platform; inputting the recommendation words and the search words into a pre-trained large language model, and outputting the recommendation words optimized by the large language model.

[0086] Based on the above-mentioned solution, according to one aspect of the present application, a method for recommending product-related words is provided. Figure 6 This is a flow chart of a method for recommending product-related words according to an embodiment of the present application. Figure 6 As shown, the method includes the following steps:

[0087] Step S601, determining multiple scene domains inside and outside the commodity platform and the targets corresponding to the multiple scene domains;

[0088] Step S602: determining corresponding recommended implementation methods based on the multiple scenario domains and corresponding goals;

[0089] Step S603, performing a word recommendation operation according to the corresponding recommendation implementation method to obtain a word recommendation result; and

[0090] Step S604: Based on the word recommendation results, the shopping guide center and the product tray center are used to perform personalized recommendations corresponding to the current user according to the user portrait and user growth source.

[0091] In an optional embodiment, the scenario domain includes the homepage background side, the search discovery domain, the search recommendation integrated scenario domain and / or the product platform external traffic domain, and step S602 includes:

[0092] For the homepage shading side, generating behavior-scenario domain transfer data;

[0093] For the search discovery domain, generate multi-dimensional interest graph data to determine user demand preferences;

[0094] For the search and recommendation integrated scenario domain, generating a personalized deep model by combining relevant data on the user's potential purchase intention; and

[0095] For the external import domain of the commodity platform, user intention mapping is adopted to achieve cross-commodity platform reference.

[0096] In an optional embodiment, generating behavior-scenario domain transfer data includes:

[0097] Collect user behavior data; and

[0098] The behavior-scenario domain transfer data is generated based on the user behavior data and the user's jump path from the home page to the search / product details page.

[0099] In an optional embodiment, generating multi-dimensional interest graph data includes:

[0100] Build a user interest pool by aggregating user behavior data on and off the product platform; and / or

[0101] Performing user interest modeling based on the user's short-term interest data and long-term interest data; and / or

[0102] Generate search discovery terms based on user tags and / or popularity-related data.

[0103] In an optional embodiment, the adopting user intent mapping includes:

[0104] Determine push strategies based on different intent types; and / or

[0105] Analyze user historical behavior data and extract potential demand words; and

[0106] In response to the user clicking on the push, the current search page is redirected to pre-fill the recommended words.

[0107] Figure 7 This is a flow chart of a method for recommending product-related words according to another embodiment of the present application. Figure 6 compared to, Figure 7 Steps S701 to S704 are the same as Figure 6Steps S601 to S604 are the same except that: Figure 7 The method also includes:

[0108] Step S705: Execute precise user intent modeling.

[0109] In an optional embodiment, step S705 may include:

[0110] Improve the text quality of candidate pot queries by building a multi-dimensional query semantic parser to simplify the vocabulary; and / or

[0111] Pushing the recommended words determined by the identity information of the zero-less behavior user to a new user with the same identity information; and / or

[0112] The degree of refinement of word recommendations is determined based on the interest transfer relationship between scene domains and user behavior data.

[0113] Figure 8 This is a flowchart of a product recommendation method according to another embodiment of the present application. Figure 6 compared to, Figure 8 Steps S801 to S804 are the same as Figure 6 Steps S601 to S604 are the same except that: Figure 8 The method also includes:

[0114] Step S805: collecting search terms entered by users and recommended terms pushed by the platform during shopping guide activities on the product platform; and

[0115] Step S806: input the recommended word and the search word into a pre-trained large language model, and output the recommended word optimized by the large language model.

[0116] Figure 9 This is a flow chart of a method for recommending product-related words according to one embodiment of the present application. Figure 6 compared to, Figure 9 Steps S901 to S904 are the same as Figure 6 Steps S601 to S604 are the same except that: Figure 9 The method also includes:

[0117] Step S905 , based on the RAG technology, according to the optimized recommendation words and using existing basic real-time data outside the commodity platform, determine the optimized recommendation words inside and outside the commodity platform.

[0118] According to another aspect of the present application, a product recommendation device is provided. Figure 10 FIG. 1 is a schematic diagram of a device for recommending product-related words according to an embodiment of the present application. Figure 10As shown, the device includes:

[0119] A first determination module 1001 is configured to determine multiple scene domains within and outside the commodity platform and targets corresponding to the multiple scene domains;

[0120] A second determining module 1002 is configured to determine corresponding recommended implementation methods based on the multiple scenario domains and corresponding objectives;

[0121] An obtaining module 1003 is configured to perform a word recommendation operation according to the corresponding recommendation implementation method to obtain a word recommendation result; and

[0122] The execution module 1004 is used to perform personalized recommendations corresponding to the current user based on the word recommendation results, using the shopping guide center and the product tray center for user portraits and user growth sources.

[0123] In an optional embodiment, the scenario domain includes the homepage background side, the search discovery domain, the search recommendation integrated scenario domain and / or the product platform external traffic domain, and the second determination module 1002 includes:

[0124] A first generating unit is configured to generate behavior-scenario domain transfer data for the background texture side of the homepage;

[0125] A second generating unit is configured to generate multi-dimensional interest graph data for the search discovery domain to determine user demand preferences;

[0126] A third generating unit is configured to generate a personalized deep model for the search recommendation integrated scenario domain in combination with data related to the user's potential purchase intention; and

[0127] The fourth generating unit is configured to use user intention mapping for the external import domain of the commodity platform to achieve cross-commodity platform reference.

[0128] In an optional embodiment, the first generating unit is configured to:

[0129] Collect user behavior data; and

[0130] The behavior-scenario domain transfer data is generated based on the user behavior data and the user's jump path from the home page to the search / product details page.

[0131] In an optional embodiment, the second generating unit is configured to:

[0132] Build a user interest pool by aggregating user behavior data on and off the product platform; and / or

[0133] Performing user interest modeling based on the user's short-term interest data and long-term interest data; and / or

[0134] Generate search discovery terms based on user tags and / or popularity-related data.

[0135] In an optional embodiment, the fourth generating unit is configured to:

[0136] Determine push strategies based on different intent types; and / or

[0137] Analyze user historical behavior data and extract potential demand words; and

[0138] In response to the user clicking on the push, the current search page is redirected to pre-fill the recommended words.

[0139] Figure 11 Schematic diagram of a device for recommending product-related words according to another embodiment of the present application. Figure 10 compared to, Figure 11 Modules 1101 to 1104 and Figure 10 Modules 1001 to 1004 are the same except that Figure 11 The device also includes:

[0140] The intention modeling module 1105 is used to perform accurate user intention modeling.

[0141] In an optional embodiment, the intent modeling module 1105 can be used to:

[0142] Improve the text quality of candidate pot queries by building a multi-dimensional query semantic parser to simplify the vocabulary; and / or

[0143] Pushing the recommended words determined by the identity information of the zero-less behavior user to a new user with the same identity information; and / or

[0144] The degree of refinement of word recommendations is determined based on the interest transfer relationship between scene domains and user behavior data.

[0145] Figure 12 Schematic diagram of a device for recommending product-related words according to another embodiment of the present application. Figure 10 compared to, Figure 12 Modules 1201 to 1204 and Figure 10 Modules 1001 to 1004 are the same except that Figure 12 The device also includes:

[0146] The collection module 1205 is used to collect the search terms entered by users and the recommended terms pushed by the platform during the shopping guide activities on the commodity platform; and

[0147] The output module 1206 is configured to input the recommended words and the search words into a pre-trained large language model, and output the recommended words optimized by the large language model.

[0148] Figure 13 Schematic diagram of a device for recommending product-related words according to another embodiment of the present application. Figure 10 compared to, Figure 13 Modules 1301 to 1304 and Figure 10 Modules 1001 to 1004 are the same except that Figure 13 The device also includes:

[0149] The third determining module 1305 is configured to determine the optimized recommendation words within and outside the commodity platform based on the RAG technology, according to the optimized recommendation words and using existing basic real-time data outside the commodity platform.

[0150] The product-related word recommendation method and device provided in this application, targeting the multi-scenario shopping guide model, proposes a solution for global information modeling and scenario-domain basic link optimization through global scenario strategy modeling and basic capability building within the product platform. This solution has at least the following advantages:

[0151] (1) Differentiated construction of different internal and external scene domains: Optimize the layout of each scene domain, combine the characteristics of the scene domain itself, identify the correlation between different scenes and the user interest transfer pattern, and use the user demand map and multi-touch mode to achieve the effect of user search and use.

[0152] (2) Construction of accurate user intentions: Introduce more high-quality behaviors and enhance the accuracy of capturing the intentions behind user behaviors. At the same time, optimize the existing personalized recommendation system, develop more accurate and personalized recommendation algorithms, and consider other important factors such as residence time and conversion rate to predict users' potential intentions.

[0153] (3) Optimization of the expression of assisting words in the large model: Using the large model LLM and the retrieval system RAG, differentiated hot spots are incubated inside and outside the product platform, triggering highly timely content in multiple shopping guide scenarios to stimulate user clicks.

[0154] Through the above methods, more accurate and comprehensive information services are provided, effectively shortening the time users spend looking for the content they need and increasing satisfaction. As user stickiness increases and conversion rates improve, search penetration and search growth are driven, thereby bringing a positive impact on the overall conversion of the market.

[0155] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0156] It should be noted that for the aforementioned method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in this specification are all optional embodiments, and the actions and modules involved are not necessarily required by this application.

[0157] In the several embodiments provided in this application, it should be understood that the disclosed devices can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical connection or other forms.

[0158] See Figure 14 , Figure 14 An electronic device is provided, comprising a processor and a memory. The memory stores computer instructions or one or more programs. When the computer instructions or one or more programs are executed by the processor, the processor executes the computer instructions to achieve the following Figures 6 to 9 The method and refinement scheme shown.

[0159] It should be understood that the above-described device embodiments are merely illustrative, and the devices disclosed herein may also be implemented in other ways. For example, the division of units / modules described in the above-described embodiments is merely a logical functional division, and actual implementations may employ alternative divisions. For example, multiple units, modules, or components may be combined or integrated into another system, or some features may be omitted or not implemented.

[0160] In addition, unless otherwise specified, the functional units / modules in the various embodiments of the present invention may be integrated into a single unit / module, each unit / module may exist physically separately, or two or more units / modules may be integrated together. The aforementioned integrated units / modules may be implemented in the form of hardware or software program modules.

[0161] If the integrated unit / module is implemented in hardware, the hardware may be a digital circuit, an analog circuit, or the like. The physical implementation of the hardware structure includes, but is not limited to, transistors, memristors, and the like. Unless otherwise specified, the processor or chip may be any appropriate hardware processor, such as a CPU, GPU, FPGA, DSP, and ASIC. Unless otherwise specified, the on-chip cache, off-chip memory, and storage may be any appropriate magnetic storage medium or magneto-optical storage medium, such as resistive random access memory (RRAM), dynamic random access memory (DRAM), static random access memory (SRAM), enhanced dynamic random access memory (EDRAM), high-bandwidth memory (HBM), hybrid memory cube (HMC), and the like.

[0162] If the integrated unit / module is implemented in the form of a software program module and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a memory, including a number of instructions for enabling a computer electronic device (which can be a personal computer, a server or a network electronic device, etc.) to perform all or part of the steps of the method described in each embodiment of the present disclosure. The aforementioned memory includes: various media that can store program codes, such as a U disk, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk.

[0163] The embodiment of the present application further provides a computer-readable storage medium storing one or more computer programs, which, when executed by multiple processors, causes the processors to execute the following Figures 6 to 9 The method and refinement scheme shown.

[0164] An embodiment of the present application further provides a computer program product, comprising a computer program, which enables the computer to execute the method of any of the above embodiments when the computer program is run on a computer.

[0165] References to features, advantages, or similar language throughout this specification do not imply that all features and advantages achievable with this solution are included or embodied in any single implementation thereof. Rather, language referring to features and advantages is understood to mean that a specific feature, advantage, or characteristic described in connection with an embodiment is included in at least one embodiment of this solution. Therefore, discussions of features, advantages, and similar language throughout this specification may, but do not necessarily, refer to the same embodiment.

[0166] Furthermore, the features, advantages, and characteristics of the present invention may be combined in any suitable manner in one or more embodiments. Based on the description herein, one of ordinary skill in the relevant art will recognize that the present invention may be practiced without one or more of the specific features or advantages of a particular embodiment. In other cases, additional features and advantages may be realized in a particular embodiment that is not presented in all embodiments of the present invention.

[0167] The embodiments of the present application are described in detail above. Specific examples are used herein to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only intended to help understand the method and core ideas of the present application. At the same time, changes or modifications made by those skilled in the art based on the ideas of the present application, the specific implementation methods, and the scope of application of the present application, all fall within the scope of protection of the present application. In summary, the contents of this specification should not be construed as limiting the present application.

Claims

1. A method for recommending product-related words, characterized in that: include: Determine multiple scenario domains within and outside the product platform and the goals corresponding to the multiple scenario domains; Determine corresponding recommended implementation methods based on the various scenario domains and corresponding goals; Perform a word recommendation operation according to the corresponding recommendation implementation method to obtain a word recommendation result; as well as According to the word recommendation results, the shopping guide center and the product tray center are used to perform personalized recommendations corresponding to the current user based on the user portrait and user growth source.

2. The method according to claim 1, wherein The scenario domains include the homepage background side, the search discovery domain, the search recommendation integrated scenario domain and / or the product platform external traffic domain. The determination of corresponding recommendation implementation methods based on the multiple scenario domains and corresponding goals includes: For the homepage shading side, generating behavior-scenario domain transfer data; For the search discovery domain, generate multi-dimensional interest graph data to determine user demand preferences; For the search and recommendation integrated scenario domain, generating a personalized deep model by combining data related to the user's potential purchase intention; and / or For the external import domain of the commodity platform, user intention mapping is adopted to achieve cross-commodity platform reference.

3. The method according to claim 2, wherein Generating behavior-scenario domain transfer data includes: Collect user behavior data; and The behavior-scenario domain transfer data is generated based on the user behavior data and the user's jump path from the home page to the search / product details page.

4. The method according to claim 2, wherein Generating multi-dimensional interest graph data includes: Build user interest pools by aggregating user behavior data on and off the platform; and / or Performing user interest modeling based on the user's short-term interest data and long-term interest data; and / or Generate search discovery terms based on user tags and / or popularity-related data.

5. The method according to claim 2, wherein The user intent mapping includes: Determine push strategies based on different intent types, and / or analyze user historical behavior data to extract potential demand words; and In response to the user clicking on the push, the current search page is redirected to pre-fill the recommended words.

6. The method according to claim 1, wherein Also includes: Perform precise user intent modeling, including: Improve the text quality of candidate pot queries by building a multi-dimensional query semantic parser to simplify the vocabulary; and / or Pushing recommended words determined by the identity information of low-activity users to new users with the same identity information; and / or The degree of refinement of word recommendations is determined based on the interest transfer relationship between scene domains and user behavior data.

7. The method according to claim 6, wherein The lexicon simplification by constructing multi-dimensional query semantic analysis includes: Parsing the query semantics in multiple dimensions to obtain parsed data; and The parsed data is simplified.

8. The method according to claim 1, wherein Also includes: Collect the search terms entered by users during shopping guides on product platforms and the recommended terms pushed by the platforms; The recommended words and the search words are input into a pre-trained large language model, and the recommended words optimized by the large language model are output.

9. The method according to claim 8, wherein Also includes: Based on the RAG technology, the optimized recommendation words within and outside the platform are determined according to the optimized recommendation words and by utilizing the existing basic real-time data outside the platform.

10. A device for recommending product-related words, characterized in that: include: A first determination module is configured to determine multiple scene domains within and outside the commodity platform and targets corresponding to the multiple scene domains; A second determination module is configured to determine corresponding recommended implementation methods based on the multiple scenario domains and corresponding objectives; An acquisition module, configured to perform a word recommendation operation according to the corresponding recommendation implementation method to obtain a word recommendation result; as well as The execution module is used to execute personalized recommendations corresponding to the current user based on the word recommendation results, using the shopping guide center and the product tray center for user portraits and user growth sources.

11. An electronic device, characterized in that: The invention comprises a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the method according to any one of claims 1 to 9 when executing the computer program in the memory.

12. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 9 is implemented.

Images