A product recommendation method and system based on double tower model

Abstract

The present invention provides a commodity recommendation method and system based on a dual-tower model, the method comprising: obtaining user input data and commodity input data, mapping the user input data to a user embedding vector based on a user tower in the dual-tower model, and mapping the commodity input data to a commodity embedding vector based on a commodity tower in the dual-tower model; preliminarily filtering all commodities according to the user input data, commodity input data, user embedding vectors and commodity embedding vectors to obtain candidate commodities; calculating the similarity between the user embedding vector and the commodity embedding vector of each candidate commodity; and recommending candidate commodities with a similarity greater than a preset similarity threshold to the user. The present invention analyzes and processes the user input data and commodity data based on the dual-tower model according to the user browsing and purchasing behavior data and commodity data, and calculates the similarity between each commodity and the user input data, recommends commodities to the user, and achieves the accuracy of commodity recommendation.

Description

Technical Field

[0001] The present invention relates to the technical field of commodity recommendation, and more specifically, to a commodity recommendation method and system based on a double-tower model. Background Art

[0002] With the rapid development of online education platforms, recommendation systems play an increasingly important role in improving user experience and promoting sales of products and courses. However, existing recommendation systems still have many deficiencies in personalized recommendations, user intent recognition, and semantic understanding. In particular, when dealing with dynamic changes in user needs, they often show problems such as inaccurate recommendations and untimely responses. How to combine user profiling, information retrieval, multi-round dialogue technology, and large models to improve the accuracy of recommendations and user satisfaction has become a current research hotspot.

[0003] Internationally, recommendation algorithms that combine deep learning and natural language processing (NLP) are gradually becoming the mainstream. For example, Google's YouTube recommendation system optimizes personalized recommendations by combining large-scale user data with deep learning models; Netflix improves the relevance of recommended content by combining user historical data with multi-layer models. However, these systems still face challenges in dealing with user semantic ambiguity or incomplete information. Summary of the invention

[0004] The present invention proposes a product recommendation method based on a twin-tower model, which processes user input data and product input data to achieve a higher recommendation efficiency. The present invention aims at the technical problems existing in the prior art and provides a product recommendation method and system based on a twin-tower model, which solves the problem of insufficient recommendation accuracy in the prior art.

[0005] According to a first aspect of the present invention, a product recommendation method based on a double tower model is provided, comprising:

[0006] Obtain user input data and product input data;

[0007] Mapping the user input data into a user embedding vector based on the user tower in the dual-tower model, and mapping the commodity input data into a commodity embedding vector based on the commodity tower in the dual-tower model;

[0008] Preliminarily filtering all commodities according to the user input data, commodity input data, user embedding vector and commodity embedding vector to obtain candidate commodities;

[0009] Calculating the similarity between the user embedding vector and the product embedding vector of each candidate product;

[0010] Recommend candidate products whose similarity is greater than a preset similarity threshold to users.

[0011] According to a second aspect of the present invention, a product recommendation method system based on a double tower model is provided, comprising:

[0012] The acquisition module is used to acquire user input data and product input data;

[0013] A mapping module, configured to map the user input data into a user embedding vector based on a user tower in a dual-tower model, and to map the commodity input data into a commodity embedding vector based on a commodity tower in a dual-tower model;

[0014] A filtering module, used to perform preliminary filtering on all commodities according to the user input data, commodity input data, user embedding vector and commodity embedding vector to obtain candidate commodities;

[0015] A calculation module, used to calculate the similarity between the user embedding vector and the product embedding vector of each candidate product;

[0016] The recommendation module is used to recommend candidate products whose similarity is greater than a preset similarity threshold to users.

[0017] The present invention provides a product recommendation method and system based on a dual-tower model, which obtains user input data and product input data, and maps the user input data to a user embedding vector and the product input data to a product embedding vector based on the user tower and the product tower in the dual-tower model respectively; performs preliminary filtering on all products according to the user input data, the product input data, the user embedding vector and the product embedding vector to obtain candidate products; calculates the similarity between the user embedding vector and the product embedding vector of each candidate product; and recommends candidate products with a similarity greater than a preset similarity threshold to the user. The present invention analyzes and processes the user input data and the product data based on the dual-tower model according to the user browsing and purchasing behavior data and the product data, and calculates the similarity between each product and the user input data, recommends products to the user, and achieves the accuracy of product recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A flow chart of a commodity recommendation method based on a double tower model provided by the present invention;

[0019] Figure 2 A schematic diagram of the structure of a product recommendation system based on a double-tower model provided by the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. In addition, the technical features in the various embodiments or single embodiments provided by the present invention can be arbitrarily combined with each other to form a feasible technical solution. This combination is not subject to the constraints of the sequence of steps and / or the structural composition mode, but must be based on the ability of ordinary technicians in this field to achieve. When the combination of technical solutions is contradictory or cannot be achieved, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0021] In view of the shortcomings of traditional recommendation methods, the development direction of recommendation systems in the future will gradually move towards a more intelligent and personalized direction, especially combining user portraits, contextual information and large model generation technology to build a recommendation system with multi-round interaction and real-time feedback capabilities has become a research hotspot.

[0022] The present invention can not only improve the recommendation effect of e-commerce and online education platforms, but also has a wide range of application value. In the field of online education, it can better match user learning needs and improve the accuracy of course recommendations and user participation.

[0023] Figure 1 A flowchart of a product recommendation method based on a double tower model provided by the present invention is shown in FIG. Figure 1 As shown, the method includes:

[0024] Step 1: Obtain user input data and product input data, wherein the user input data includes user basic information and user historical behavior data, and the product input data includes product structured attributes and product unstructured description.

[0025] It is understandable that based on the user's online browsing and purchasing behavior records, a user profile can be constructed. The so-called user profile is the user's interests and hobbies.

[0026] User portrait is the core of recommendation, which builds a personalized profile for each user by analyzing the user's basic information and behavior data. User basic information usually includes gender, age, location, and interest preferences. These structured data can be processed through scalar search. At the same time, the user's historical behavior data (such as search history, browsing history, and purchase history) is mostly unstructured or semi-structured data, which is more suitable for capturing user preferences and potential needs through vector search.

[0027] (1) User basic information and scalar search.

[0028] Basic user information is clear and structured data that can be used as the basis for filtering recommended content. For example, information such as gender, age, and location can help the system quickly filter out products or content that are not relevant to user needs. Scalar search can efficiently perform Boolean queries (such as "gender=Female") through inverted indexes, ensuring that the system takes into account the user's basic needs when making recommendations.

[0029] Basic user information (such as gender and age) can be used as preliminary filtering conditions. For example, the system can exclude products that do not meet gender requirements based on the user's gender information (such as automatically filtering out men's clothing when a female user searches). In addition, scalar search can also process users' explicit requirements, such as price range, product brand, and geographic location, to help further narrow the scope of recommendations.

[0030] (2) User historical behavior records and vector search

[0031] Most of the user's historical behavior records (such as browsing history, search history, and purchase history) are unstructured data and cannot be processed through simple Boolean queries. At this time, vector search can play a role. By converting these behavior records into vectors, the system can capture the user's implicit interests and potential needs. Vector search can use semantic similarity (such as cosine similarity) to infer the user's interests, and by comparing the similarity between the user's current query and historical behavior, it can generate personalized recommendations that are more in line with the user's preferences. Historical behavior records can be vectorized through an embedding model to convert the user's behavior information into a vector representation in a high-dimensional space. The formula is expressed as:

[0032] ;

[0033] in, Is a user The image vector, It is the user's behavior and characteristic data.

[0034] Vector search combines preferences captured in historical behavior records with current queries to help the system identify products or content that best match user needs. For example, when a user browses a certain category of products multiple times, the system can use vector search to identify content that is semantically related to these products and recommend them to the user. The user's historical purchase records can also provide important feedback signals to help predict products or services that the user may be interested in in the future.

[0035] Step 2: Mapping the user input data into a user embedding vector based on the user tower in the dual-tower model, and mapping the commodity input data into a commodity embedding vector based on the commodity tower in the dual-tower model.

[0036] It is understandable that the core of personalized recommendation lies in how to efficiently combine user basic information (structured data) and user historical behavior data (unstructured data) to provide users with accurate and dynamic recommendations. To this end, the present invention adopts a dual tower model to achieve a seamless combination of scalar search and vector search by generating embedding vectors for users and products respectively.

[0037] The dual-tower model includes two independent embedding towers, the user tower and the product tower, which process the vector representation of users and products respectively to ensure similarity matching between user basic information and product attributes.

[0038] (1) User Tower: Processes user input data, including basic user information (such as gender, age, interests, etc.) and user historical behavior data (such as browsing history, purchase history), and generates user embedding vectors.

[0039] The formula is:

[0040]

[0041] in, It is user input data, including basic user information (such as gender, age) and historical behavior data (such as browsing and purchase records), through the embedding function of the user tower Map user features to vectors .

[0042] In addition to historical behavior, the system also dynamically updates the embedding vector in the user tower based on the user's current needs (such as the most recently searched keywords or clicked products). ,This adjustment ensures that the recommended content not only conforms to the user's historical behavior, but also dynamically responds to current needs.

[0043]

[0044] in, Indicates the current needs of users. is the embedding function of the current requirement, by adjusting the weights and , dynamically balancing the impact of users’ historical behavior and current needs.

[0045] By combining historical behavior with immediate needs, the system can recommend products that are more suitable for the current situation. For example, if a user recently searched for "light jackets", the system will prioritize products that meet this need and further personalize the recommendation results by referring to the user's previous preferences (such as "summer clothing").

[0046] (2) Item Tower: Converts the structured attributes of items (such as price, category, and brand) and unstructured descriptions (such as product reviews and product introduction) into vectors.

[0047] The formula is:

[0048]

[0049] in, It is the product input data, including product structured attributes (such as price, brand) and product unstructured descriptions (such as product introduction, comments, etc.), through the embedding function of the product tower Map product features into product embedding vectors .

[0050] Step 3: Preliminarily filter all commodities according to the user input data, commodity input data, user embedding vector and commodity embedding vector to obtain candidate commodities.

[0051] It is understandable that the above steps obtain user input data and product input data, and map the user input data and product input data into user embedding vectors and product embedding vectors based on the user tower and product tower in the dual-tower model, respectively. This step performs a preliminary filtering of all products browsed or purchased by the user based on these data to obtain candidate products that are subsequently recommended to the user.

[0052] Scalar search performs preliminary filtering based on basic user information to ensure that the recommended products meet the user's basic needs (such as gender, age, preferences, etc.). Vector search dynamically matches the user's potential interests among these candidate products based on the user's historical behavior to ensure the relevance and accuracy of the recommended content.

[0053] Exemplarily, the preliminarily filtering all commodities according to the user input data, commodity input data, user embedding vectors and commodity embedding vectors to obtain candidate commodities includes:

[0054] S31, calculating a first score of each product relative to user demand based on the user input data and the product input data and based on a scalar search module in a product tower;

[0055] The first score of each product relative to the user's needs is calculated as follows: the product tower in the dual-tower model processes the basic attributes of the product (such as price, category, brand, etc.) through the scalar search module, and combines the embedding vector generated based on the user's basic information in the user tower to perform preliminary filtering of the products. In this way, it is ensured that the recommended products meet the user's basic needs (such as gender, age, price range, etc.). The formula for calculating the first score is as follows:

[0056] ;

[0057] in, is the scalar search weight matrix, and Enter data for users and goods respectively.

[0058] For example, if the user's basic information is "female, 30-40 years old", the system will first filter out products related to women in this age group, such as clothing, maternal and child products, etc.

[0059] S32, calculating a second score of each product relative to the user's needs based on the user embedding vector and the product embedding vector and using a vector search module in the product tower;

[0060] Among them, under the framework of the dual-tower model, different weights can be set for scalar search and vector search respectively, and the scalar search weight and vector search weight can be flexibly adjusted according to user needs and scenarios.

[0061] When the scalar search weight is high, the recommendation results are more likely to be based on the user's basic information (such as gender, age, brand preference, etc.) and the current clear needs, such as the price and brand specified by the user. This is suitable for scenarios where users have clear needs.

[0062] When the vector search weight is high, the recommendation results focus more on the user's historical behavior and potential interests. This is suitable for scenarios where the user's search semantics are vague or there is no clear demand.

[0063] By fusing and ranking scalar search and vector search results through the dual-tower model, algorithms such as Rank Reciprocal Fusion (RRF) can be used to weightedly merge the results of scalar search and vector search to ensure that both the user's current needs are met and the user's potential interests are captured.

[0064] The calculation formula for the second score is as follows:

[0065] ;

[0066] in, Search weight matrix for vector, and They are the dynamically updated user embedding vector and product embedding vector respectively.

[0067] S33, calculating a total score of each commodity relative to the user's needs based on the first score and the second score of each commodity relative to the user's needs.

[0068] The formula for calculating the total score of each product relative to user needs is:

[0069] ;

[0070] in, and They are the weight coefficients for scalar search and vector search, respectively, and are dynamically adjusted according to the actual scenario.

[0071] S34, screening all commodities according to the total score of each commodity relative to the user's needs to obtain candidate commodities.

[0072] Among them, S33 calculates the total score of each product for the user's needs, and takes the products with total scores greater than a preset score threshold as candidate products for subsequent recommendation.

[0073] Step 4: Calculate the similarity between the user embedding vector and the product embedding vector of each candidate product.

[0074] It is understandable that among the candidate products after the initial screening by scalar search, the user tower in the dual tower model combines the user's historical behavior (such as past search history, browsing history, purchase history) and the user's current needs to generate a user embedding vector , and the product embedding vector generated by the product tower Perform similarity calculations to further refine recommendations.

[0075] Through the dual-tower model, the similarity of user and product features is calculated in the same embedding space, thereby achieving an efficient combination of scalar search and vector search.

[0076] The similarity between the user embedding vector and the item embedding vector is usually calculated using cosine similarity:

[0077] ;

[0078] in, and are the dynamically updated user embedding vector and the item embedding vector of the candidate item, respectively. for and The cosine similarity of express and The inner product between Representation vector The norm of .

[0079] Step 5: Recommend candidate products whose similarity is greater than a preset similarity threshold to the user.

[0080] It can be understood that the similarity between the product embedding vector and the user embedding vector of each candidate product is calculated through step 4, and the products with similarity greater than the preset similarity threshold are recommended to the user as target products.

[0081] Exemplarily, the product recommendation method further includes:

[0082] Based on user basic information, user historical behavior data, product structured attributes and the relationship between products, a knowledge graph of user interests is established; according to the user's query request, a target product is searched in the knowledge graph; products related to the target product are searched in the knowledge graph, and the target product and products related to the target product are recommended to the user.

[0083] The present invention proposes a knowledge graph that combines user behavior data and product data to improve the accuracy of the recommendation system. Specifically, the knowledge graph establishes a complex relationship network between users and products by performing association analysis on the structured attributes of products (such as category, brand, price, etc.) and the behavior data of users (such as browsing history, search history, purchase behavior). The knowledge graph is dynamically updated in this process to capture changes in user interests. For example, after a user's purchase history is updated, the nodes and edges in the knowledge graph will be adjusted accordingly to optimize the recommendation results.

[0084] The knowledge graph not only supports preliminary recommendations based on user queries, but also continuously optimizes the recommended content through multiple rounds of interactions. Through multiple rounds of dialogue with users, the knowledge graph can gradually identify the user's real needs and further refine the recommendations based on contextual information. For example, when a user first queries "baby stroller", the system can infer through the knowledge graph that the user may also have potential needs for "baby seats" or other childcare products. This association analysis can significantly improve the accuracy and relevance of recommendation results.

[0085] For example, when a user first searches for "baby strollers", the system can not only directly recommend related products, but also associate other similar products (such as baby seats, baby clothes, etc.) through the knowledge graph. In the subsequent process of multiple rounds of dialogue, the system can further narrow the recommendation range based on the user's answer and provide more personalized suggestions.

[0086] After constructing the knowledge graph of user interests, the target product is found according to the query conditions entered by the user. Based on the target product, the products related to the target product are searched in the knowledge graph of user interests, and the target product and the related products are recommended to the user together.

[0087] The system can infer important information that the user did not mention in the initial query by analyzing the user's identity, context, and historical behavior. For example, when a female user searches for children's products, the system can infer that she is buying for her child based on her behavior and rewrite the question accordingly to make the recommended content more accurate.

[0088] By building a knowledge graph of product labels and descriptions, content related to the target product in the search results can be recommended to users. Knowledge graphs are an important tool currently used to enhance the intelligence of recommendation systems. By structuring products, user behaviors, and contextual information, knowledge graphs can provide rich semantic relationships and provide strong support for question rewriting and user intent clarification in multi-round conversations.

[0089] In addition, the knowledge graph plays a role of "completion" in the process of rewriting query conditions. When the user's query information is incomplete or vague, the system can use the knowledge graph to supplement important information that the user may have missed. For example, when a user searches for "children's school supplies", the knowledge graph can infer that the user is buying for preschool children by identifying the user's historical purchase records or behavior patterns, and actively ask whether a specific category needs to be recommended, such as "jigsaw puzzles" or "preschool education books."

[0090] In addition, the system can also automatically generate follow-up recommended questions based on the knowledge graph to further guide users to clarify their intentions. For example, if the user's initial query is "buy toys for children", the system can use the product attributes in the knowledge graph to ask questions such as "Is it for a boy or a girl?" or "How old is the child?" to make recommendations more accurately.

[0091] See also Figure 2 , a product recommendation system based on a double tower model provided by the present invention, the system comprising:

[0092] The acquisition module 201 is used to acquire user input data and product input data;

[0093] A mapping module 202, configured to map the user input data into a user embedding vector based on the user tower in the twin tower model, and to map the commodity input data into a commodity embedding vector based on the commodity tower in the twin tower model;

[0094] A filtering module 203 is used to perform preliminary filtering on all commodities according to the user input data, commodity input data, user embedding vector and commodity embedding vector to obtain candidate commodities;

[0095] A calculation module 204 is used to calculate the similarity between the user embedding vector and the product embedding vector of each candidate product;

[0096] The recommendation module 205 is used to recommend candidate commodities with a similarity greater than a preset similarity threshold to the user.

[0097] It can be understood that the product recommendation system based on the twin-tower model provided by the present invention corresponds to the product recommendation method based on the twin-tower model provided in the aforementioned embodiments. The relevant technical features of the product recommendation system based on the twin-tower model can refer to the relevant technical features of the product recommendation method based on the twin-tower model, which will not be repeated here.

[0098] The embodiment of the present invention provides a product recommendation method and system based on a dual-tower model, which obtains user input data and product input data, and maps the user input data to a user embedding vector and the product input data to a product embedding vector based on the user tower and the product tower in the dual-tower model respectively; performs preliminary filtering on all products according to the user input data, the product input data, the user embedding vector and the product embedding vector to obtain candidate products; calculates the similarity between the user embedding vector and the product embedding vector of each candidate product; and recommends candidate products with a similarity greater than a preset similarity threshold to the user. The present invention analyzes and processes the user input data and the product data based on the dual-tower model according to the user browsing and purchasing behavior data and the product data, and calculates the similarity between each product and the user input data, recommends products to the user, and achieves the accuracy of product recommendation.

[0099] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and for parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0100] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0101] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0102] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0103] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0104] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0105] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A product recommendation method based on the twin tower model, characterized in that: include: Obtain user input data and product input data; Mapping the user input data into a user embedding vector based on the user tower in the dual-tower model, and mapping the commodity input data into a commodity embedding vector based on the commodity tower in the dual-tower model; Preliminarily filtering all commodities according to the user input data, commodity input data, user embedding vector and commodity embedding vector to obtain candidate commodities; Calculating the similarity between the user embedding vector and the product embedding vector of each candidate product; Recommend candidate products whose similarity is greater than a preset similarity threshold to users; The user input data includes user basic information and user historical behavior data, the commodity input data includes commodity structured attributes and commodity unstructured descriptions, the user basic information includes gender, age and interests, the user historical behavior data includes browsing records and purchase records, and the user tower based on the double tower model maps the user input data into a user embedding vector, including: ； in, is the user input data, is the embedding function of the user tower, Embedding vectors for users; The method further includes dynamically updating the user embedding vector according to the current needs of the user: ； in, Indicates the user's current needs. is the embedding function of the user's current needs, and is the weight coefficient, dynamically balancing the weight of the user's historical behavior and current needs. Embedding vector for the updated user; The preliminarily filtering all commodities according to the user input data, commodity input data, user embedding vector and commodity embedding vector to obtain candidate commodities includes: Calculate a first score of each product relative to user demand based on the user input data and the product input data and based on a scalar search module in the product tower; According to the user embedding vector and the product embedding vector, the second score of each product relative to the user demand is calculated based on the vector search module in the product tower; Calculate the total score of each product relative to the user's needs based on the first score and the second score of each product relative to the user's needs; All products are screened according to the total score of each product relative to user needs to obtain candidate products.

2. The commodity recommendation method according to claim 1, characterized in that: The structured attributes of the product include product price, product category and product brand, the unstructured description of the product includes product introduction and product review, and the product tower in the double tower model maps the product input data into a product embedding vector, including: ； in, is the product input data, is the embedding function of the commodity tower, Embedding vector for products.

3. The commodity recommendation method according to claim 1, characterized in that: The expression of the first score is: ； in, is the scalar search weight matrix, and Enter data for users and goods respectively; The expression of the second score is: ； in, Search weight matrix for vector, and They are the dynamically updated user embedding vector and product embedding vector respectively; Calculate the total score of each product relative to user needs: ； in, and They are the weight coefficients for scalar search and vector search, respectively, and are dynamically adjusted according to the actual scenario.

4. The commodity recommendation method according to claim 1, characterized in that: The calculating the similarity between the user embedding vector and the product embedding vector of each candidate product includes: ； in, and are the dynamically updated user embedding vector and the item embedding vector of the candidate item, respectively. for and The cosine similarity of express and The inner product between Representation vector The norm of Represents the norm of vector V.

5. The commodity recommendation method according to claim 1, characterized in that: The method further comprises: Build a knowledge graph of user interests based on basic user information, user historical behavior data, product structured attributes, and the relationship between products; According to the user's query request, search for the target product in the knowledge graph; Search the knowledge graph for products associated with the target product, and recommend the target product and products associated with the target product to the user.

6. A product recommendation system based on the double tower model, characterized in that: include: The acquisition module is used to acquire user input data and product input data; A mapping module, configured to map the user input data into a user embedding vector based on a user tower in a dual-tower model, and to map the commodity input data into a commodity embedding vector based on a commodity tower in a dual-tower model; A filtering module, used to perform preliminary filtering on all commodities according to the user input data, commodity input data, user embedding vector and commodity embedding vector to obtain candidate commodities; A calculation module, used to calculate the similarity between the user embedding vector and the product embedding vector of each candidate product; A recommendation module, used to recommend candidate products whose similarity is greater than a preset similarity threshold to users; The user input data includes user basic information and user historical behavior data, the commodity input data includes commodity structured attributes and commodity unstructured descriptions, the user basic information includes gender, age and interests, the user historical behavior data includes browsing records and purchase records, and the user tower based on the double tower model maps the user input data into a user embedding vector, including: ； in, is the user input data, is the embedding function of the user tower, Embedding vectors for users; It also includes dynamically updating the user embedding vector according to the user's current needs: ； in, Indicates the user's current needs. is the embedding function of the user's current needs, and is the weight coefficient, dynamically balancing the weight of the user's historical behavior and current needs. Embedding vector for the updated user; The preliminarily filtering all commodities according to the user input data, commodity input data, user embedding vector and commodity embedding vector to obtain candidate commodities includes: Calculate a first score of each product relative to user demand based on the user input data and the product input data and based on a scalar search module in the product tower; According to the user embedding vector and the product embedding vector, the second score of each product relative to the user demand is calculated based on the vector search module in the product tower; Calculate the total score of each product relative to the user's needs based on the first score and the second score of each product relative to the user's needs; All products are screened according to the total score of each product relative to user needs to obtain candidate products.

Images