Task structure unification system for e-commerce recommendation
By using a unified task structure system for e-commerce recommendations and leveraging the MiXun Card module to visualize user needs and product characteristics as structured information, the system solves the problem of disconnect between manual operations and algorithms, enabling more accurate personalized recommendations and intelligent adjustments.
Patent Information
- Application Number
- CN202411172847.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-26
- Publication Date
- 2025-01-10
AI Technical Summary
In existing e-commerce recommendation technologies, the disconnect between human operation and algorithms makes it difficult to respond in a timely manner to the differences, ambiguities, and temporariness of buyer needs. The gap between sellers' products and user needs, as well as the information asymmetry in media's "planting the seed" and "harvesting" needs, hinder the intelligent integration of the overall e-commerce ecosystem.
The system adopts a unified task structure for e-commerce recommendations. Through the MiXun Card module, user needs and product characteristics are visualized into structured information. Combined with manual operation and algorithm recommendation, a unified processing solution is formed. By utilizing the user need description end, product attribute description end, and intelligent recommendation end in the MiXun Card module, the optimal matching between user needs and product selling points is achieved.
It achieves a seamless integration of human operational strategies and algorithmic strategies, enabling timely responses to changes in consumer demand, narrowing the information gap between user needs and product feature descriptions, and improving the accuracy and intelligence of recommendations.
Smart Images
Figure CN119295169A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of commodity recommendation, and in particular relates to a unified task structure system for e-commerce recommendation. Background Art
[0002] Personalized recommendations in the e-commerce field are actually the process of matching users' actual needs with the use value of goods. The purpose of all algorithms is to find matching product attributes based on known user information, including behavioral information, by defining a series of methods, so as to recommend suitable products to users. The current popular method is to directly use machine learning to calculate the relationship between people and goods under the logical framework of collaborative filtering. Representative algorithms include DSSM, DFM, GemNN, etc. These solutions are based on full consideration of the shopping behavior of all users, and obtain the most suitable products for recommendation to users through product collaboration (Item-CF) and user collaboration (User-CF). In addition, there are also supplementary strategies such as manual operations and crowd packages to meet the promotion services of small and medium-sized e-commerce companies.
[0003] In the existing technology, recommendation algorithms have not systematically considered how to combine with manual operations and business needs, and have only been optimized around the algorithm itself. As deep learning models become more complex, models become increasingly difficult to explain, and it becomes more difficult to integrate recommendation algorithms and manual operations. As a result, current intelligent strategies mainly focus on data, models, evaluation, and feedback optimization, while manual operations are isolated and regarded as pre- or post-process auxiliary means, failing to play a key role. However, in actual operations, different service objects (buyers, sellers, and media) will put forward various differentiated demands, which will change with time and circumstances. Therefore, it is very important to integrate the wisdom and thinking of manual operations into the recommendation system. If the operation strategy cannot be effectively combined, it will be difficult to solve problems such as the following: 1. In response to the needs of buyers, it is necessary to use artificial knowledge to make recommendations and adjustments in a timely manner under different consumption times, purposes and objects, because pure models cannot reflect these differentiated needs in a timely manner.
[0004] 2. The ambiguity and temporary nature of buyer demand requires an effective combination of manual operations and algorithms to overcome the model underfitting problem caused by data sparsity.
[0005] 3. The gap between the seller’s product selling points and the actual needs of users requires necessary adjustments to the product characteristics through manual operations.
[0006] 4. The media’s “planting grass” and “pulling grass” needs require a comprehensive understanding of product characteristics and user needs through data analysis, and act as a bridge between users and merchants.
[0007] Due to the disconnect between algorithms and operations, there is a lack of effective bridge between current technology and actual business operations, which hinders the intelligent integration of the entire e-commerce ecosystem; at the same time, it also makes it difficult for retailers with weak technology to benefit from big data. Therefore, it is necessary to coordinate operational strategies with algorithm strategies to better understand user needs and make up for the information gap between user demands and product selling points. To this end, we propose a unified task structure system for e-commerce recommendations. Summary of the invention
[0008] The purpose of this invention is to provide a unified task structure system for e-commerce recommendation. The actual solution of this invention is how to effectively combine manual operation with recommendation technology. It concretizes user needs and product selling points into an expression that humans can understand, and integrates them into complex recommendation systems and intelligent algorithms; by solving the above problems, we can more effectively match product characteristics with actual user needs.
[0009] The technical solution adopted by the present invention is as follows: The task structure unified system of e-commerce recommendation includes a power card module, which includes a user demand description end, a product attribute description end and an intelligent recommendation end. The intelligent recommendation end first obtains user demand information from the user demand description end and product attribute information from the product attribute description end, and then the intelligent recommendation end recommends products to the user according to the intelligent recommendation paradigm.
[0010] Preferably, the intelligent recommendation paradigm is: ;in, ; It describes the needs of current users , and the actual selling points of the product Based on , that is, artificial operation strategy, business strategy or technical strategy to obtain new user needs And the actual selling point of the product , and make it fit Be further improved.
[0011] Preferably, the commodity attribute end constructs a function model: ; The product understanding part of the product attribute description end is to observe the characteristics of the product And all observable collections of items Based on the function , get the real demand that the product can meet .
[0012] Preferably, the user demand description end constructs a function model: ; The user understanding part of the user demand description is based on the user's current demands. As well as the actual attributes and characteristics of users Based on the function , get the user's current real needs .
[0013] Preferably, the intelligent recommendation paradigm Represents the actual action selected in the model recommendation algorithm or manual operation strategy or product strategy , so as to find the optimal match between user needs and product selling points.
[0014] Preferably, The inherent attributes of the product are obtained through product information analysis and are embodied in the product attribute description database ASP. The product attribute knowledge base SKG is obtained through the statistics of the characteristics of a large number of products and the relationship between their characteristics; The Agent strategy based on the RAG system is analyzed as well as , and finally get the actual selling points of the product and the characteristics that meet the current market , and structured information that can be understood by humans.
[0015] Preferably, The ICI is obtained by extracting the appeal interest card ICI through the user's current shopping behavior or the keyword set used for retrieval through a large language model. The AUP is a user profile extracted through observation of the overall behavior of past users. It is an agent based on a self-developed large model. as well as , and finally get the user's current actual needs , and structured information that can be understood by humans.
[0016] Preferably, the appeal interest card ICI and the product attribute description database ASP are both structured natural language description sets, and the names are normalized; the product attribute knowledge base SKG is a product attribute relationship network based on a graph structure, and each node is a related attribute of a product; the user profile AUP stores structured user descriptions, and the descriptions are normalized to a unified description standard; , After that, strategy a is used to complete the matching task, including: Manual operation: By defining the relationship between words, complete , Matching between; Statistical recommendation: obtained through statistics A statement closer to The demand; Model calculation: through batch and The shopping relationship is trained to develop a recommendation model, and the recommendation model is directly used for recommendation.
[0017] The technical effects achieved by an invention are: Compared with previous technologies, the methodology of the present invention actually elevates personalized recommendation from a pure algorithm problem to a system problem; by defining the power card module, the demand and product characteristics are visualized in a way that humans can understand, and in this way a complete set of processing solutions for manual operation, algorithm recommendation and statistical analysis is established; the manual operation strategy and the algorithm strategy can learn from each other and improve each other. On this basis, the needs of different levels in consumption can be met more timely, and merchants and media can make reasonable strategic adjustments to meet market needs, understand user needs more deeply, and make up for the information gap between user demands and product selling points. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] Figure 1 is a system block diagram of the unified task structure system recommended by e-commerce of the present invention; Figure 2 It is a system operation block diagram of the task structure unified system recommended by the e-commerce of the present invention; Figure 3 It is a system operation flow chart of the unified task structure system recommended by the e-commerce of the present invention. DETAILED DESCRIPTION
[0019] In order to make the purpose and advantages of the present invention more clearly understood, the present invention is specifically described below in conjunction with embodiments. It should be understood that the following text is only used to describe one or several specific embodiments of the present invention, and does not strictly limit the scope of protection of the specific claims of the present invention.
[0020] like Figure 1-3 As shown, the task structure of e-commerce recommendation is unified as a system, including a power card module. The power card module includes a user demand description end, a product attribute description end, and an intelligent recommendation end. The intelligent recommendation end first obtains the user demand information of the user demand description end and the product attribute information of the product attribute description end, and then the intelligent recommendation end recommends products to the user according to the intelligent recommendation paradigm.
[0021] The intelligent recommendation paradigm is: ;in, ; It describes the needs of current users , and the actual selling points of the product Based on , that is, artificial operation strategy, business strategy or technical strategy to obtain new user needs And the actual selling point of the product , and make it fit Be further improved.
[0022] Constructing function model on product attribute side: ; The product understanding part of the product attribute description side is to observe the characteristics of the product And all observable collections of items Based on the function , get the real demand that the product can meet .
[0023] The user demand description end builds a function model: ; The user understanding part of the user demand description is based on the user's current demands. As well as the actual attributes and characteristics of users Based on the function , get the user's current real needs .
[0024] Intelligent recommendation paradigm Represents the actual action selected in the model recommendation algorithm or manual operation strategy or product strategy , so as to find the optimal match between user needs and product selling points.
[0025] The inherent attributes of the product are obtained through product information analysis and are embodied in the product attribute description database ASP. The product attribute knowledge base SKG is obtained through the statistics of the characteristics of a large number of products and the relationship between their characteristics; The Agent strategy based on the RAG system is analyzed as well as , and finally get the actual selling points of the product and the characteristics that meet the current market , and structured information that can be understood by humans; it should be added here that the RAG system is the abbreviation of Retrieval Augmented Generation, which is retrieval enhanced generation; in fact, it does combine the two major functional components of "retrieval" and "generator" to handle complex information query and generation tasks; on the basis of the original data set of the large language model (LLM), the "retrieval" and "generation" capabilities of the AI model are enhanced by adding external data to improve the quality of information query and generation; the Agent strategy is a core model and methodology in the field of artificial intelligence. It is a software entity that simulates human intelligent behavior and can complete specific tasks autonomously and intelligently. It not only has basic environmental perception capabilities, can obtain and analyze the state information of the surrounding environment in real time, but also has a goal-driven mechanism, which can make behavioral decisions based on preset or dynamically adjusted goal orientations; in addition, the Agent also has the ability to select actions. Among multiple possible operation paths, it selects the optimal solution or approximate optimal solution to perform tasks through algorithm optimization and strategy planning.
[0026] The ICI is obtained by extracting the appeal interest card ICI through the user's current shopping behavior or the keyword set used for retrieval through a large language model. The AUP is a user profile extracted through observation of the overall behavior of past users. It is an agent based on a self-developed large model. as well as , and finally get the user's current actual needs , and structured information that can be understood by humans.
[0027] like Figure 1-3 As shown in the figure, the appeal interest card ICI and the product attribute description database ASP are both structured natural language description sets, and the names are normalized; the product attribute knowledge base SKG is a product attribute relationship network based on a graph structure, and each node is a related attribute of the product; the user portrait AUP stores structured user descriptions, and the descriptions are normalized to a unified description standard; , After that, strategy a is used to complete the matching task, including: Manual operation: By defining the relationship between words, complete , Matching between; for example, anti-hair loss shampoo can correspond to the demands of middle-aged men's shampoo; statistical recommendation: get through statistics A statement closer to For example, windproof tents are statistically close to desert travel demands; model calculation: through batch and The shopping relationship is trained to develop a recommendation model, and the recommendation model is directly used for recommendation.
[0028] like Figure 1-3 In the present invention shown, based on the above system, manual operation can participate in the establishment of specifications when constructing user portrait AUP, product attribute description database ASP, product attribute knowledge base SKG and appeal interest card ICI to form a relatively stable demand expression system; it can also dynamically add operation strategies when matching power information cards to meet timely needs.
[0029] In the present invention, the user demands of e-commerce are visualized into the demand interest card ICI and the user profile AUP, which are specifically described as follows: The ICI is completely defined as a composite description of demand, including fields such as category, purpose or scenario, object, brand, product characteristics, and price range; the user profile AUP is a weekly, monthly, and annual demand compliance analysis of users; including: product category preference, price preference, sales preference, scenario purpose preference, purchase object preference, user type, user social attributes, and user interest preference; The selling points and characteristics of the products are visualized into the attribute description database ASP and the product attribute relationship database SKG, which are described as follows: The attribute description database ASP indexes the detailed attributes of various commodities, including: category, applicable scenario, price range, sales range, service object, commodity characteristics, brand, service life, conversion rate and conversion cycle, the possibility and cycle of commodity from recommendation to purchase; the attribute relationship database SKG establishes the relationship between commodity attributes, including the relationship between category and each brand, usage scenario, design characteristics, etc., or the category or brand relationship that may be expanded by the scenario, as well as the purchase relationship between categories; this data structure is constructed using the Bayes network to facilitate probability-based recommendations.
[0030] The matching process of the above two is turned into a process that human intelligence can participate in, that is, a process of seamlessly combining manual operation with big models, statistical analysis, and recommendation algorithms. In particular, the big model is used to conduct a deeper analysis of the product and extract structured information, and the NL2Json strategy is used to transform the universal description of the product and the long-term behavior description of the user into a deeper understanding; for example: "Hanwang Clear 66-inch E-ink screen reader E-paper eye protection reading reader E-book ink screen tablet E-book reading novel comics" analysis: { "category":"E-ink screen reader, e-paper, tablet e-book", "brand":"Hanwang", "audience":"Book lovers", "feature":"Eye protection, electronic paper, ink screen", "screen_size":"6 inches", "scenario":"Read books, novels, comics" } "I have seen many laptops with high-end graphics cards, several PS games, some toys and clothes for little girls, men's perfume, and several orders of leather shoes" analysis: { "demographics":{ "gender":"male", "age_range":"30-45", "marital_status":"married", "children":"oneyoungdaughter" }, "interests":{ "technology":"high", "gaming":"high", "fashion":"moderate", }, "personality":{ "family_oriented":"high", "tech_savvy":"high", }, } like Figure 1-3 As shown, in the present invention, compared with the previous technologies, this methodology actually elevates personalized recommendation from a pure algorithm problem to a system problem; by defining the power card module, the needs and product characteristics are visualized in a way that humans can understand, and in this way a complete set of processing solutions for manual operation, algorithm recommendation and statistical analysis is established; manual operation strategies and algorithm strategies can learn from each other and improve each other; on this basis, different levels of consumption needs can be met more promptly, and merchants and media can be cooperated with to make reasonable strategy adjustments to meet market needs.
[0031] like Figure 1-3As shown, the working principle of the present invention is as follows: For example, user A wants to buy shampoo, and his previous behavior shows that he is a middle-aged male who often participates in social activities; however, the shampoo product does not describe the brand and characteristics suitable for this type of user; by combining system recommendation and manual operation, we can determine that the "anti-slip" function is more in line with the needs of user A; user B is an expectant mother, and she needs to buy shockproof flat shoes. However, shockproof is not mentioned in the product information, and maternity shoes cannot fully meet the needs of user B; through algorithms and operation strategies, we can infer that "rubber soles" and "flat heels" are related to "shockproof", so as to find a suitable product set; in short, this technology organically integrates statistics, large models and manual operations. It narrows the conceptual gap between user demand description and product feature description, helps users find more suitable products, and also assists companies in finding real target customers.
[0032] The above is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications should also be considered as the protection scope of the present invention. The structures, devices and operating methods not specifically described and explained in the present invention shall be implemented according to the conventional means in the art unless otherwise specified and limited.
Claims
1. A unified system of task structures recommended by e-commerce companies, characterized by: It includes a power card module, which includes a user demand description end, a product attribute description end and an intelligent recommendation end. The intelligent recommendation end first obtains user demand information from the user demand description end and product attribute information from the product attribute description end, and then recommends products to the user according to the intelligent recommendation paradigm.
2. The task structure unified system for e-commerce recommendation according to claim 1, characterized in that: The intelligent recommendation paradigm is: ;in, ; It describes the needs of current users , and the actual selling points of the product Based on Get new user needs And the actual selling point of the product .
3. The task structure unified system for e-commerce recommendation according to claim 2, characterized in that: The product attribute end constructs a function model: ; The product understanding part of the product attribute description end is to observe the characteristics of the product And all observable collections of items Based on the function , get the real demand that the product can meet .
4. The task structure unified system for e-commerce recommendation according to claim 3, characterized in that: The user demand description end constructs a function model: ; The user understanding part of the user demand description is based on the user's current demands. As well as the actual attributes and characteristics of users Based on the function , get the user's current real needs .
5. The task structure unified system for e-commerce recommendation according to claim 4, characterized in that: The Intelligent Recommendation Paradigm Represents the actual action selected in the model recommendation algorithm or manual operation strategy or product strategy , so as to find the optimal match between user needs and product selling points.
6. The task structure unified system for e-commerce recommendation according to claim 5, characterized in that: The inherent attributes of the product are obtained through product information analysis and are embodied in the product attribute description database ASP. The product attribute knowledge base SKG is obtained through the statistics of the characteristics of a large number of products and the relationship between their characteristics; The Agent strategy based on the RAG system is analyzed as well as , and finally get the actual selling points of the product and the characteristics that meet the current market .
7. The task structure unified system for e-commerce recommendation according to claim 6, characterized in that: The ICI is obtained by extracting the appeal interest card ICI through the user's current shopping behavior or the keyword set used for retrieval through a large language model. The AUP is a user profile extracted through observation of the overall behavior of past users. It is an agent based on a self-developed large model. as well as , and finally get the user's current actual needs .
8. The task structure unified system for e-commerce recommendation according to claim 7, characterized in that: The appeal interest card ICI and the product attribute description database ASP are both well-structured natural language description sets, and the names are normalized; the product attribute knowledge base SKG is a product attribute relationship network based on a graph structure, and each node is a related attribute of a product; the user portrait AUP stores structured user descriptions, and the descriptions are normalized to a unified description standard; , After that, strategy a is used to complete the matching task, including: Manual operation: By defining the relationship between words, complete , Matching between; Statistical recommendation: obtained through statistics A statement closer to The demand; Model calculation: through batch and The shopping relationship is trained to develop a recommendation model, and the recommendation model is directly used for recommendation.