Method and apparatus for recommending matching goods

Abstract

The application discloses a method and device for recommending matching commodities, and relates to the technical field of computers. A specific embodiment of the method comprises the following steps: collecting demand information of a user, and generating a demand label according to a collection result; inquiring whether a historical user portrait exists; in response to the absence of the historical user portrait, generating a user portrait according to user demand; in response to the existence of the historical user portrait, generating a user portrait according to user demand and the historical user portrait; generating a demand vector according to the demand label and the user portrait; inputting the demand vector into a large model that has been pre-learned in context, and outputting a commodity matching scheme that comprises at least two categories. The embodiment reduces the randomness of demand information collection, and improves the matching degree and the recommendation success rate of commodity recommendation.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method and apparatus for recommending product combinations. Background Technology

[0002] In the current offline home furnishing sales scenario, the following problems exist: when sales staff collect demand information, the information is incomplete or non-standard, resulting in low matching degree of subsequent recommended products; demand information and user information are difficult to structure and reuse; sales staff cannot quickly recommend cross-category matching products, for example: after a user purchases tiles, the sales staff has difficulty quickly recommending matching tile adhesive, grout, skirting boards, lighting fixtures, etc. Summary of the Invention

[0003] In view of this, embodiments of the present invention provide a method and apparatus for recommending complementary products, which can reduce the randomness of demand information collection and improve the matching degree and recommendation success rate of product recommendations.

[0004] To achieve the above objectives, according to one aspect of the present invention, a method for recommending complementary products is provided, comprising: Collect user demand information and generate demand tags based on the collection results; Check if a historical user profile exists; if no historical user profile exists, generate a user profile based on user needs; if a historical user profile exists, generate a user profile based on user needs and the historical user profile. Generate a demand vector based on demand tags and user profiles; Input the demand vector into a large model that has undergone context learning beforehand, and output product combination schemes that include at least two product categories.

[0005] Optionally, user demand information may be collected, including: Key requirements can be determined through at least one of the following methods: obtaining dialogue content from users through a question-and-answer format, and determining key requirements based on the user's dialogue content; obtaining intentional images provided by users, and obtaining key requirements based on the shape, size, color, and pattern of the items in the intentional images; obtaining drawn images provided by users, and determining key requirements based on the size, position, and area of ​​the items in the drawn images; or showing users a questionnaire containing question options, and determining key requirements based on the answers returned by users. Summarize all key requirements, and if there are duplicates among the key requirements, retain any one of the duplicates to obtain the collection results.

[0006] Optionally, after outputting product combination schemes that include at least two product categories, the following may also be included: Check whether the product combination scheme conforms to the preset output structure; In response to non-compliance with the output structure, at least two product entities of different categories are matched in the pre-built knowledge graph based on user needs and user history data; and a product combination scheme is determined based on the product entities.

[0007] Optionally, after outputting product combination schemes that include at least two product categories, the following may also be included: Check the store rules for each item in the product combination scheme; If any item in a product combination scheme does not meet the store's rules, the corresponding product combination scheme will be filtered out from all product combination schemes.

[0008] Optionally, after outputting product combination schemes that include at least two product categories, the following may also be included: Determine real-time user data and historical transaction data; Input product combination schemes, real-time user data, and historical transaction data into a deep learning neural network with at least two objectives, and output the predicted values ​​of product combination schemes for at least two objectives; Calculate the Pareto frontier for each product combination based on the predicted values, and then rank the product combinations according to the Pareto frontier.

[0009] Optionally, after outputting product combination schemes that include at least two product categories, the following may also be included: Push product combination suggestions to users; The system obtains at least one of the following actions taken by the user after receiving the push notification: clicking a product link, purchasing a product, or rating a product. Determine the context information based on the operation; The large model is invoked to learn based on context information, and the learned large model is used to update the large model that has been pre-learned based on context.

[0010] Secondly, embodiments of the present invention provide an apparatus for recommending complementary products, comprising: The data acquisition module is used to collect user demand information and generate demand tags based on the collection results. The user profile generation module is used to query whether a historical user profile exists; if no historical user profile exists, it generates a user profile based on the user's needs; if a historical user profile exists, it generates a user profile based on the user's needs and the historical user profile. The demand vector generation module is used to generate demand vectors based on demand tags and user profiles. The product pairing output module is used to input the demand vector into a large model that has undergone context learning in advance, and output product pairing schemes that include at least two product categories.

[0011] Thirdly, embodiments of the present invention provide an electronic device, including: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the methods of any of the above embodiments.

[0012] Fourthly, embodiments of the present invention provide a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements the method of any of the above embodiments.

[0013] Fifthly, embodiments of the present invention provide a computer program product, including a computer program, wherein the computer program, when executed by a processor, implements the method of any of the above embodiments.

[0014] One embodiment of the above invention has the following advantages or beneficial effects: by collecting user demand information and generating demand tags based on the collection results, standardized user demand information is obtained; user profiles are generated based on historical user profiles and / or current user demands, improving the reusability of demand information and user information; by inputting the demand vector generated based on demand tags and user profiles into a large model that has undergone pre-contextual learning, product matching schemes can be obtained, resulting in product combinations with a high degree of matching. Therefore, the method and apparatus for recommending product combinations through the above embodiments can reduce the randomness of demand information collection and improve the matching degree and recommendation success rate of product recommendations.

[0015] The further effects of the aforementioned unconventional alternative methods will be explained below in conjunction with specific implementation methods. Attached Figure Description

[0016] The accompanying drawings are provided to better understand the invention and are not intended to unduly limit the scope of the invention. Wherein: Figure 1 This is a schematic diagram of the main flow of a method for recommending product combinations according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating a method for recommending complementary products according to another embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the process of a sales associate recommending complementary products according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the demand information collection process provided in one embodiment of the present invention; Figure 5 This is a schematic diagram of store rules provided in one embodiment of the present invention; Figure 6 This is a flowchart illustrating the multi-objective sorting of product pairing schemes provided in one embodiment of the present invention; Figure 7This is a schematic diagram of the main modules of a demand information collection device provided in one embodiment of the present invention; Figure 8 This is an exemplary system architecture diagram in which embodiments of the present invention can be applied; Figure 9 This is a schematic diagram of the structure of a computer system suitable for implementing terminal devices or servers of the present invention. Detailed Implementation

[0017] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of the present invention, including various details to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0018] It should be noted that the collection, gathering, updating, analysis, processing, use, transmission, and storage of user personal information involved in this disclosed technical solution all comply with relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. Necessary measures are taken to prevent unauthorized access to user personal information data and to safeguard user personal information security, network security, and national security. Users are informed and their consent or authorization is obtained. Where applicable, user personal information undergoes de-identification and / or anonymization and / or encryption technical processing. For example, after collecting user request information, we will use technical means to de-identify the data.

[0019] In order to accurately describe the technical content of this invention, the terms used in this specification are explained or defined as follows: 1) Apriori Algorithm: This is an association rule learning algorithm used to discover frequent itemsets in transaction data. By setting a minimum support threshold, it filters out product combinations that appear frequently enough in the dataset, and then generates association rules that meet the minimum confidence and lift requirements to explore the pairing relationships between products.

[0020] 2) Multi-gated hybrid expert model (MMoE): This is a multi-objective learning model architecture that learns expert knowledge for different objectives by sharing underlying feature representations and setting up multiple expert networks. Each objective is equipped with an independent gating network, and the outputs of multiple expert networks are dynamically weighted and combined to achieve multi-objective collaborative learning.

[0021] 3) Non-dominated sorting genetic algorithm (NSGA-II): This is an evolutionary algorithm for multi-objective optimization problems. It iteratively generates a set of approximate optimal solutions through genetic operations such as selection, crossover, and mutation. Non-dominated solutions are those that cannot be further improved without compromising other objectives. It is used to search for the Pareto front under multiple operational constraints.

[0022] 4) Top-p kernel sampling: This is a decoding strategy used during text generation to sample from the batch of candidate words with the highest probability when the model outputs the next word.

[0023] 5) Uplift Model: This is a causal inference model used to predict the incremental effect of intervention measures (such as push notifications) on user behavior. By predicting the incremental conversion rate (ΔCVR) generated after a push notification, it determines whether the push can significantly increase the probability of a transaction. If the incremental conversion rate is lower than a preset threshold, the push is suppressed.

[0024] 6) Hot Reloading (TF-Serving): This is a model deployment and update mechanism that allows new versions of models to be dynamically loaded without restarting the service. Through hot reloading, sorting models can be updated online, and can be gradually rolled out using an A / B testing framework to ensure the new model is stable before a full deployment.

[0025] Figure 1 This is a method for recommending product combinations provided in one embodiment of the present invention, such as... Figure 1 As shown, it includes the following steps: Step S101: Collect user demand information and generate demand tags based on the collection results.

[0026] Demand information collection refers to the collection of demand information related to the products a user intends to purchase, obtained from users through demand information collection tools such as mobile terminal software / web pages, voice / image acquisition devices, etc. Demand information can be multimodal, such as voice (dialogue), images (e.g., renderings, hand-drawn illustrations, photographs), natural language text (e.g., text dialogue content), and structured data / text (e.g., standardized text collected through questionnaires).

[0027] In some embodiments, collecting user demand information includes: determining key demands through at least one of the following methods: obtaining dialogue content from the user through a question-and-answer format, and determining key demands based on the user's dialogue content; obtaining an intention image provided by the user, and obtaining key demands based on the shape, size, color, and pattern of the items in the intention image; obtaining a drawing image provided by the user, and determining key demands based on the size, position, and area of ​​the items in the drawing image; showing the user a questionnaire containing question options, and determining key demands based on the user's returned answers; summarizing all key demands, and retaining any item from the duplicate content among the key demands, to obtain the collection result.

[0028] Key requirements refer to specific content within the demand information, including but not limited to: style preferences (such as "modern minimalist style," "new Chinese style," "Nordic style," etc.), color preferences, budget range, apartment area and dimensions, space list (purpose, type), product brand preferences, promotional sensitivity, construction period requirements, product environmental protection level, and past purchase records. Key requirements can be pre-categorized, and demand information can be directly mapped to relevant categories when acquiring key requirements; for example, if a user's demand information is "prefers a simpler style," then after feature extraction and synonym extraction, the key requirement of style preference can be mapped to "modern minimalist style."

[0029] In some embodiments, generating requirement tags based on the collection results includes: preprocessing the results of requirement information collection, and then mapping the collection results to standardized requirement tags through feature extraction, conflict resolution, and fusion decision-making, or mapping key requirements to requirement tags. Preprocessing may include: speech-to-text conversion, feature extraction, intent recognition, dimension annotation, data cleaning, normalization, and standardization.

[0030] Requirement tags are a set of structured vectors, with each dimension corresponding to a key requirement. For example, requirement tag T = {style, budget range, apartment size, space list, brand preference, promotional sensitivity, construction period requirements, environmental protection level}. The generated requirement tags are associated with the user for easy modification and retrospective analysis.

[0031] Step S102: Query whether a historical user profile exists; if no historical user profile exists, generate a user profile based on user needs; if a historical user profile exists, generate a user profile based on user needs and the historical user profile.

[0032] In some embodiments, historical user profiles include long-term user profiles and short-term user profiles. The long-term user profile is generated based on user behavior records such as purchasing and browsing products within a first time period; the short-term user profile is generated based on user behavior records such as purchasing and browsing products within a second time period. A user profile is a set of feature vectors.

[0033] In response to the absence of historical user profiles, user profiles are generated based on user needs, including: obtaining contextual features based on the demand vector and contextual background information other than key demands, inputting the contextual features into a multilayer perceptron, and generating user profiles.

[0034] In response to the existence of historical user profiles, a user profile is generated based on user needs and historical user profiles. This includes weighting short-term user profiles, long-term user profiles, and user profiles generated based on user needs to obtain a new user profile. Preferably, dynamic weights for weighting are learned through an attention gating network. The newly generated user profile can be used to replace or modify historical user profiles in the database.

[0035] Step S103: Generate a demand vector based on demand tags and user profiles.

[0036] In some embodiments, demand tags and user profiles are input into a pre-trained deep learning neural network (such as a Transformer network with self-attention mechanism, a Deep Interest Network, etc.) to generate a demand vector. Preferably, the deep learning neural network used to generate the demand vector can encode the input using a dual-tower structure and be trained through contrastive learning (InfoNCE). The training data includes positive and negative samples, and the loss function is a binary classification cross-entropy. For example, positive samples are product combinations purchased within 7 days of a user's visit to the store (such as tiles + tile adhesive + grout); negative samples are randomly sampled product combinations from the same store that were not purchased.

[0037] Step S104: Input the demand vector into a large model that has undergone context learning in advance, and output a product combination scheme that includes at least two product categories.

[0038] The large model used in this embodiment of the invention refers to a large language model; before using the large model to output product matching schemes, the large model can be processed by pre-training, context learning, fine-tuning and other methods.

[0039] Preferably, prompt word templates are pre-injected into the large model for contextual learning. The prompt word templates may include: role settings, user profiles, requirement tags, generation requirements, example demonstrations, constraints, and guidance on reasoning steps.

[0040] In some embodiments, after outputting a product combination scheme that includes at least two product categories, the method further includes: checking whether the product combination scheme conforms to a preset output structure; in response to not conforming to the output structure, matching product entities of at least two product categories in a pre-built knowledge graph based on user needs and user historical data; and determining the product combination scheme based on the product entities.

[0041] One method to check whether the product pairing scheme conforms to the preset output structure is to use JSON Schema regular expression validation to verify whether the output conforms to the preset JSON structure, such as containing specific fields.

[0042] The entity types in a pre-built knowledge graph can include, but are not limited to: product entities, space entities, style entities, user entities, and rule entities. Among them, product entities refer to specific stock keeping units (SKUs); space entities refer to home spaces (such as living rooms and bedrooms); style entities refer to home decoration styles; user entities refer to pre-set user profiles and preference tags, which can be set according to needs after statistical analysis of historical data in the database; and rule entities refer to matching rules and taboo rules. The entities in the knowledge graph have topological relationships. For example, there is a matching relationship between product entities and rule entities, a constraint relationship between rule entities and style entities, a suitability relationship between product entities and space entities, and an inclusion relationship between space entities and user entities (used to represent a type of user's space needs and preferences), etc.

[0043] Knowledge graphs can be pre-built through the following steps: mapping product, space, style, user, and rule data to corresponding entities; where the data can be structured or unstructured. For example, structured product data can be used to build a category tree based on product categories, and then product entities and their contained product attribute information (such as brand, material, size, price, etc.) can be constructed based on the number of categories; unstructured data such as product detail pages, user reviews, design cases, and live streaming scripts can be extracted and mapped to corresponding entities using technologies such as Bidirectional Encoder Representations from Transformers Name Entity Recognition (BERT-NER).

[0044] Optionally, the entities and topological relationships of the knowledge graph can be constructed using an expert rule base. For example, a corresponding style can be set based on expert knowledge (e.g., a modern minimalist style avoiding overly ornate carvings), and a confidence level can be configured for each rule. Frequent itemset mining can be performed on users' historical purchase records using algorithms (e.g., the Apriori algorithm), and the quality of each rule can be evaluated based on the mining results. Rules that meet the standards are then incorporated into the knowledge graph. Rule entities can be incrementally updated based on new database changes (e.g., user actions, inquiries, browsing history, transaction records, etc.). For rules that contradict each other during updates, methods such as timestamp comparison, confidence weighting, multi-source verification, voting mechanisms, and rule inference engine monitoring can be used to handle these issues.

[0045] Based on user needs and historical user data, at least two categories of product entities are matched in a pre-built knowledge graph, including: determining initial product entities based on user needs and historical user data; starting from the initial product entities, expanding through the topological relationships in the knowledge graph to obtain associated product entities that have a matching relationship with the initial entities; and filtering the obtained associated product entities to obtain product entities of at least two categories.

[0046] In some embodiments, after outputting product combination schemes that include at least two product categories, the method further includes: querying the store rules for each product in the product combination scheme; and filtering out the corresponding product combination scheme from all product combination schemes in response to any product in the product combination scheme not meeting the store rules.

[0047] These store rules may include, but are not limited to, inventory constraint rules, price calculation rules, design matching rules, environmental protection level filtering rules, and other customizable rules. For example, an inventory constraint rule may require that the items in a product combination scheme have a real-time inventory greater than zero in the store or be available for next-day delivery from a regional warehouse; if any item in a product combination scheme does not meet the above inventory constraint rule, then that product combination scheme will be filtered out from all product combination schemes.

[0048] In some embodiments, after outputting product combination schemes that include at least two product categories, the method further includes: determining real-time user data and historical transaction data; inputting the product combination schemes, real-time user data, and historical transaction data into a deep learning neural network with at least two objectives, and outputting predicted values ​​of the product combination schemes for at least two objectives; calculating the Pareto front of each product combination scheme based on the predicted values, and ranking the product combination schemes based on the Pareto fronts.

[0049] Preferably, the deep learning neural network can employ a multi-gate mixture of experts (MMoE) model. This model shares underlying feature representations and sets up multiple expert networks to learn expert knowledge for different objectives. These networks are then dynamically weighted and combined through gating networks to achieve multi-objective collaborative learning. The objectives may include click-through rate (CTR) prediction, conversion rate (CVR) prediction, and gross merchandise volume (GMV) prediction. The result for each prediction objective is a predicted value; the GMV prediction is calculated by multiplying the CVR by the merchandise price.

[0050] Real-time user data includes, but is not limited to, the user's current location, real-time weather information, and holiday status; historical transaction data refers to historical purchase records.

[0051] In some embodiments, the Pareto front of each product mix is ​​calculated based on the predicted values, including: constructing a multi-objective optimization problem with GMV maximization, CTR not lower than a first threshold, CVR not lower than a second threshold, and category diversity not lower than a third threshold as multi-objective constraints; solving the Pareto front using a Non-dominated Sorting Genetic Algorithm II (NSGA-II) to generate a set of approximately optimal solutions; and ranking the product mixes according to each non-dominated and dominated solution in the optimal solution set. Here, the non-dominated solutions in the Pareto front are solutions that cannot be further improved without compromising other objectives, and are ranked before the dominated solutions.

[0052] Category diversity can be calculated using the following three indicators: category coverage, which is the proportion of categories covered by the product mix to the total number of categories, with a value range of [0,1]; category distribution entropy, which is the uniformity of the category distribution of the product mix; and category Gini coefficient, which is the balance of the category distribution of the product mix. Category diversity is obtained by weighted summation of category coverage, category distribution entropy, and category Gini coefficient.

[0053] In some embodiments, after outputting a product combination scheme that includes at least two product categories, the method further includes: pushing the product combination scheme to the user; obtaining at least one of the following operations performed by the user after receiving the push: clicking a product link, purchasing a product, or rating a product; determining context information based on the operation; calling a large model to learn based on the context information, and using the learned large model to update the large model that has been pre-learned for context.

[0054] The process of pushing product combination schemes to users includes: pushing to users through at least one of the following channels: sales associate's mobile terminal device, store large screen, and user's WeChat. Optionally, the push channel can be selected according to the preset channel priority. The sales associate's mobile terminal device is used to display the 3D combination scheme in real time and supports the preview of the augmented reality tiling effect; the store large screen is used to display the scheme effect diagram; and the user's WeChat is used to push the mini-program link so that users and their families can browse, modify, and place orders together. Optionally, fatigue control can also be implemented before pushing: the number of times the same user receives pushes from the same channel within a preset time window shall not exceed a preset threshold, and / or an incremental conversion rate prediction model is used to predict the incremental conversion rate generated after the push. If the incremental conversion rate is lower than the preset threshold, the push is suppressed.

[0055] Based on the operation, determine the context information, including: writing the data generated by the user operation to the streaming processing (such as Flink) job in real time via a message queue; filtering high-quality interaction samples obtained from the operation data and writing positive sample prompts, and writing low-conversion samples (such as CVR<5%) to negative sample prompts. Call the large model to learn based on the context information, including: calling the large model to learn the context based on the positive and negative sample prompts, and updating the large model that has been pre-learned for context.

[0056] The written data can be used for periodic incremental updates to user profiles; for incremental training of a deep learning neural network for generating demand vectors using a parameter server architecture; and for hot-loading (TF-Serving) of the algorithm for ranking product combination schemes, which can be updated through canary releases via an A / B testing framework.

[0057] In some embodiments, after obtaining at least one of the following operations performed by the user after receiving the push notification: clicking a product link, purchasing a product, or rating a product, the method further includes: determining the sales guide performance indicators based on the results of demand collection and the operations performed by the user after receiving the push notification, and providing business guidance based on the performance indicators.

[0058] The performance indicators may include, but are not limited to, at least one of the following: completeness of demand collection, which measures the fill rate of key demands collected, and is considered qualified when the fill rate reaches a preset threshold; adoption rate of recommended solutions, which represents the degree of user approval of product combination solutions, and is considered as adoption when a user clicks, favorites, or shares any product combination solution; conversion rate, which measures the proportion of users who make a purchase after returning to the store to the total number of customers visiting the store; cross-selling rate, which represents the ability to sell across categories, and is determined by counting the number of cross-category product SKUs in each order; and average transaction time, which measures the length of time from demand collection to payment completion.

[0059] Business guidance is provided based on performance indicators, including: ranking sales associates according to performance indicators, and providing incentive resources to sales associates whose ranking is within a preset proportion. These incentive resources include, but are not limited to, cash rewards, coupons, additional commissions, and training opportunities; and triggering business guidance reminders in response to sales associates' performance indicators falling below a preset threshold.

[0060] Figure 2 This is another embodiment of the method for recommending complementary products, such as... Figure 2 As shown, it includes the following steps: Step S201: Collect user demand information and generate demand tags based on the collection results.

[0061] like Figure 3 As shown, sales associates can recommend complementary products to users using the method described in this embodiment.

[0062] First, sales associates can collect user needs information through mobile devices. For example... Figure 4 As shown, user needs information can be obtained in the form of voice streams, image streams, drawing tablets, dynamic questionnaires, etc.

[0063] For voice streaming scenarios, Automatic Speech Recognition (ASR) technology can be used to transcribe the conversation between the sales guide and the user in real time. Based on the transcribed text data, key summary text is generated through feature extraction. This summary text includes understanding the user's needs, such as the area to be discussed, preferred style, and brands needed for decoration.

[0064] For image streaming scenarios, the system performs image recognition (Computer Vision, CV) on user-provided floor plans, inspiration images, or other photos or pictures to automatically identify styles and color schemes, thereby determining the user's style or color requirements.

[0065] For hand-drawn input scenarios, sales guides or users can draw floor plans on a tablet, automatically obtaining dimensional information such as the area, perimeter, and location of doors and windows, saving and uploading it to the system; the system then determines the user's key requirements such as floor plan, area, and location of doors and windows based on the hand-drawn content.

[0066] In dynamic questionnaire scenarios, users can actively select and check key needs information through a pre-built dynamic questionnaire engine, or sales staff can fill in the questionnaire based on other needs information collection methods and then confirm the selection results with the user. For example, after a user selects "modern minimalist," the system will automatically ask follow-up questions such as "Do you prefer warm gray or cool gray?" and "Do you prefer a design without a main light?" The system retains the user's content data for each tag of needs.

[0067] After collecting the required information, the key requirements are cleaned and processed into a set of rule tags using a tag standardization engine. The processing steps are: data input, feature extraction, conflict resolution, fusion decision, and output. For example, if a user mentions "I like a simpler style" during communication, after feature extraction, the style can be categorized as modern minimalist; or the approximate area of ​​a room can be calculated based on the floor plan provided by the user to arrive at the area conclusion.

[0068] The final requirement tags are T = {style, budget range, apartment size, space list, brand preference, promotion sensitivity, construction period requirements, environmental protection level}. These requirement tags are then saved to the database based on the user identifier and the corresponding selected style benchmark for easy modification and retrospective analysis.

[0069] Step S202: Query whether a historical user profile exists; if no historical user profile exists, generate a user profile based on user needs; if a historical user profile exists, generate a user profile based on user needs and the historical user profile.

[0070] When a sales associate makes a purchase based on demand tags, the backend synchronously retrieves historical user profiles or responds to a query request to retrieve historical user profiles, including long-term user profiles (PL) and short-term user profiles (PS).

[0071] The PL (Personal Style) data encompasses user behavior over the past three months, including browsing, adding to cart, placing orders, and reviews, as well as information such as style preference vectors and price sensitivity. The style preference vector can be a set of weights corresponding to various styles, such as (0.82, 0.65), corresponding to the weights of "modern minimalist" and "Nordic style," respectively. Style preference vectors can be obtained by statistically mapping a user's historical style preferences; for example, if a user has browsed solid wood furniture multiple times and saved several Chinese-style items in the past three months, the style preference vector in the PL will correspondingly strengthen the weight of the Japanese style.

[0072] PS includes recent (e.g., within a week) online and offline behavioral data, such as product search terms, viewed product detail pages, and watched live stream records. For example, if a user repeatedly watched an introductory video for a gray sofa online three days before visiting the store, the PS will record this behavioral sequence and assign it a high time-series weight.

[0073] In response to the lack of historical user profiles, user profiles are generated directly based on user needs. That is, based on the demand label T collected during this demand collection process provided by the user during this visit, a vector corresponding to PL is generated. .

[0074] In response to the existence of historical user profiles, a user profile is generated based on user needs and historical user profiles, which involves fusing PS, PL, and T. Before profile fusion, the need labels T collected in this needs assessment are input as contextual features into the Multi-Layer Perception (MLP). In addition, the contextual features also include key intent information from the conversation history between the salesperson and the user, as well as the collection timestamp. PS, PL, and MLP(T) are then dynamically fused into an attention-gated network (Gate), as follows: ; Attention-gated networks adaptively calculate dynamic weight coefficients using the formula described above. , , .

[0075] Current user profile after merging The user identifier is stored in the database as the primary key for subsequent generation, sorting, and recommendation of product combination schemes. Meanwhile, It will also be written back to the user's long-term profile database and short-term profile database as the basis for the next profile update.

[0076] Step S203: Generate a demand vector based on demand tags and user profiles.

[0077] The demand labels and updated user profiles are input into a pre-trained deep learning neural network to generate demand vectors.

[0078] The deep learning neural network used to generate demand vectors (hereinafter referred to as the demand vector generator) can encode the input using a dual-tower structure and train through comparative learning. The loss function adopts the joint loss of binary cross-entropy and InfoNCE. The structure and training method of the demand vector generator are as follows.

[0079] A dual-tower encoding structure is used to encode demand tags and user profiles separately. The tag tower uses a Transformer network to encode demand tags. The Transformer network captures the dependencies between tag fields through a self-attention mechanism, outputting tag feature vectors. The user profile tower uses a Deep Interest Network (DIN) combined with a target attention mechanism to encode user profiles.

[0080] After the two towers complete their encoding, the label feature vector and the image feature vector are input into the fusion layer for processing. The fusion layer first concatenates the two feature vectors to form a joint feature vector; then, the joint feature vector is input into two MLP layers for nonlinear transformation to learn the high-order interaction features between the label and the image; finally, the feature vector output by the MLP is L2 normalized to obtain the final required vector.

[0081] The training process of the demand vector generator is as follows: Training data includes positive and negative samples. Positive samples are cross-category product combinations (such as tiles + tile adhesive + grout) that were actually purchased within 7 days of a user's visit to the store. These positive samples reflect the actual satisfaction of the user's needs. Negative samples are randomly sampled product combinations that were not purchased. The loss function uses a joint loss of binary cross-entropy (BCE) and contrastive learning loss (InfoNCE), where the temperature coefficient τ of InfoNCE can be set to 0.07. Binary cross-entropy is used to constrain the model's ability to distinguish between positive and negative samples; InfoNCE contrastive learning loss is used to narrow the distance between semantically similar positive sample pairs in the vector space, while widening the distance between negative sample pairs, thereby enhancing the discriminative and representational capabilities of the demand vector in the vector space.

[0082] Step S204: Input the demand vector into a large model that has undergone context learning in advance, and output a product combination scheme that includes at least two product categories.

[0083] The demand vector generated in step S203 is input into a large language model that has undergone pre-contextual learning. The large language model generates product combination schemes that include at least two product categories. Prompt word templates are pre-injected into the large language model to guide it in generating structured outputs that align with business objectives.

[0084] The large language foundation model used in this invention embodiment can be a large language model that has been fine-tuned or pre-trained based on corpus data from the home furnishing domain. The corpus includes, but is not limited to, product titles, product detail pages, decoration guides, live streaming scripts, and design specifications, enabling the model to possess product understanding, matching logic, and professional knowledge in the home furnishing domain. The model can support 128K context windows and can simultaneously process multi-source inputs such as user profiles, demand tags, demand vectors, store inventory, and promotional information.

[0085] Here is an example of a prompt word template injected into a large language model: Role: You are a senior interior designer.

[0086] Given: - User profile: {P_new} - This requirement is tagged with: {T} - Demand vector: {v_d} - Store inventory and promotions: {store_inventory} Target: 1) Generate 3 cross-category matching schemes, each containing 5-8 SKUs, covering basic decoration + main materials + soft decoration; 2) Provide a selling point title of no more than 40 characters and a design description of no more than 100 characters for each design scheme; 3) Output JSON format: {"schemes": [{"title": "...", "description": "...", "skus": [{"sku": "12345", "reason": "...", "confidence": 0.93}]}]} constraint: - All SKUs must be in stock at the store / region; - The total price is within the range of [0.8 × budget, 1.2 × budget]; - Maintain a consistent style; - Prioritize products with 3D renderings and support one-click order splitting.

[0087] Where {P_new} represents the user profile generated or updated in step S202. {T} represents the demand label generated in step S201, {v_d} represents the demand vector generated in step S203, and {store_inventory} represents the store inventory and promotion information obtained by the system in real time. The above placeholders are automatically filled in according to the data stored in the database during model inference.

[0088] A controllable generation strategy is employed during the generation process to balance the stability and diversity of the output. The temperature coefficient can be set to 0.2 to reduce the randomness of generation, ensuring stable and predictable recommendation results, and adapting to the deterministic requirements of the shopping guide scenario. The Top-p kernel sampling is set to 0.9 to retain high-probability words while also ensuring a certain level of diversity. The duplication penalty is set to 1.15 to avoid generating duplicate content and improve the information density of the output scheme.

[0089] Using the aforementioned prompt word templates and generation strategies, the large language model outputs product combination schemes that include at least two product categories, in a structured JSON format.

[0090] Step S205: Check whether the product combination scheme conforms to the preset output structure; in response to not conforming to the output structure, match at least two product entities of different categories in the pre-built knowledge graph based on user needs and user historical data; determine the product combination scheme based on the product entities.

[0091] The output of step S204 is checked using JSON Schema regular expression validation to verify whether the output conforms to the preset JSON structure (such as containing fields such as scheme_id, products, rationale).

[0092] If the product pairing scheme output by the large language model does not conform to the preset output structure (such as JSON parsing failure, missing required fields, or incorrect field types), a candidate completion mechanism based on the knowledge graph is triggered.

[0093] The entity types of the pre-built knowledge graph are shown in Table 1, including but not limited to: product entities, spatial entities, style entities, user entities, and rule entities.

[0094] Table 1 Knowledge Graph Entity Types

[0095] Among them, product entities refer to specific product inventory identifiers (SKUs); space entities refer to home spaces (such as living rooms, bedrooms, kitchens, bathrooms, etc.); style entities refer to home decoration styles (such as modern minimalist, Nordic style, Chinese style, industrial style, etc.); user entities refer to pre-set user profiles and preference tags, which can be set according to needs after statistical analysis of historical data in the database; and rule entities refer to matching rules and taboo rules.

[0096] Entities in a knowledge graph have topological relationships. For example, there is a matching relationship between product entities and rule entities, a constraint relationship between rule entities and style entities, a suitability relationship between product entities and space entities, and an inclusion relationship between space entities and user entities (used to represent the space needs and preferences of a type of user).

[0097] Knowledge graphs can be pre-built through the following steps: mapping data such as products, spaces, styles, users, and rules to corresponding entities. This data can be structured or unstructured. For structured product data, a category tree can be constructed based on product categories, and then product entities and their associated product attribute information (such as brand, material, size, price, etc.) can be built based on the category tree. For unstructured data such as product detail pages, user reviews, design examples, and live streaming scripts (as shown in Table 2), content can be extracted and mapped to corresponding entities using technologies such as Named Entity Recognition (BERT-NER).

[0098] Table 2 Unstructured Data

[0099] Matching rules can be constructed using an expert rule base (as shown in Table 3). For example, based on expert knowledge, rules can be set for style consistency (such as "avoid using elaborate carvings in modern minimalist style"), color coordination, functional complementarity, and budget constraints, and a confidence level can be configured for each rule.

[0100] Table 3 Expert Rule Base

[0101] The Apriori algorithm is used to mine frequent itemsets from users' historical purchase records (or other behavioral data such as browsing and adding to cart). Based on the mining results, the quality of each rule is evaluated (such as confidence and lift), and rules that meet the criteria are incorporated into the knowledge graph. For example, the rule "Users who buy tiles also buy tile adhesive" with a confidence greater than or equal to 0.7 and a minimum lift greater than or equal to 1.5 is incorporated into the knowledge graph.

[0102] Rule entities can be incrementally updated based on new database changes (such as user operations, demand inquiries, browsing history, transaction records, etc.). For rules that conflict during the update, conflict resolution can be achieved through methods such as timestamp comparison, confidence weighting, multi-source verification, voting mechanisms, and rule reasoning engine detection (as shown in Table 4).

[0103] Table 4 Conflict Resolution Strategies

[0104] After obtaining the knowledge graph, candidate completion is performed based on the knowledge graph, including the following steps: Determine the initial product entity based on user needs and user historical data. Starting from the initial product entity, expand it through the topological relationships in the knowledge graph to obtain related product entities that have matching relationships with the initial entity (such as matching products, alternative products of different brands in the same category, necessary accessories, etc.). Then, filter the related product entities obtained from the expansion, for example, limit the category to no more than 3 items and the style to no more than 2 items to ensure diversity. Finally, at least two categories of product entities are obtained as a candidate set, and the final product matching scheme is determined accordingly.

[0105] Step S206: Query the store rules for each product in the product combination scheme; in response to any product in the product combination scheme not meeting the store rules, filter out the corresponding product combination scheme from all product combination schemes.

[0106] like Figure 5 As shown, product combination schemes can be filtered according to store rules.

[0107] For example, store rules require that any item in a product combination scheme has a real-time inventory greater than zero in the store. If the item is not in stock in the store, it must meet the condition of being available from the regional warehouse the next day. Product combination schemes that do not meet the above store rules will be filtered out.

[0108] You can also request that the final price be determined based on the user's identity. For example, the current store promotional price can be used as the base price, and if the user is a member, a member discount can be added on top of the promotional price. All product prices in the plan are calculated and displayed according to this rule.

[0109] It can verify the compatibility of specifications between products in the same product combination scheme. For example, when the size of the floor tile is not less than 600mm×600mm, the matching tile adhesive grade must reach C2TE level. If the floor tile size in the product combination scheme meets this condition but the tile adhesive grade does not meet the standard, the scheme will be filtered out.

[0110] The store rules can also be filtered based on users' environmental protection requirements. For example, if a user requires all board products to meet the ENF environmental protection standard, then product combinations that include products below that standard will be filtered out.

[0111] Store rules can be flexibly configured and expanded according to business needs. Multiple rules can be selected for combination or logical checks as needed. For example, only products that meet inventory requirements can continue to be validated by subsequent store rules.

[0112] Step S207: Determine real-time user data and historical transaction data; input product combination schemes, real-time user data, and historical transaction data into a deep learning neural network with at least two objectives, and output the predicted values ​​of product combination schemes for at least two objectives; calculate the Pareto front of each product combination scheme based on the predicted values, and sort the product combination schemes according to the Pareto front.

[0113] like Figure 6 The filtered product combination schemes can be sorted using multi-objective sorting.

[0114] Specifically, the filtered product combination scheme, the user's real-time context (location, weather, holidays) and the sales guide's historical transaction data are input into a deep learning neural network model with at least two objectives (hereinafter referred to as the multi-objective model).

[0115] The multi-objective model employs a multi-gated hybrid expert model (MMoE), which shares underlying feature representations and sets up multiple expert networks to learn expert knowledge for different objectives. These networks are then dynamically weighted and combined through gating networks to achieve collaborative learning across multiple objectives. Objectives may include click-through rate (CTR) prediction, conversion rate (CVR) prediction, and gross merchandise volume (GMV) prediction. The result for each predicted objective is a predicted value; the GMV prediction is calculated by multiplying the CVR by the product price.

[0116] The estimated CTR, CVR, and GMV are input into the constraint optimization engine. A multi-objective optimization problem is constructed with GMV maximization, CTR not lower than a first threshold, CVR not lower than a second threshold, and category diversity not lower than a third threshold as multi-objective constraints. For example, the constraint optimization can be expressed as max GMV, st CTR≥0.10, CVR≥0.15, and category diversity≥0.5.

[0117] An evolutionary algorithm based on the Non-dominated Sorting Genetic Algorithm (NSGA-II) is used to solve the Pareto front online. Through iterative genetic operations such as selection, crossover, and mutation, a set of near-optimal solutions is generated, and the Pareto front is searched under constraints. Non-dominated solutions in the Pareto front are those that cannot be further improved without compromising other objectives. During sorting, non-dominated solutions are ranked before dominated solutions, and the top three solutions are returned.

[0118] In some embodiments, category diversity is calculated by combining the following three indicators: Category Coverage (Cov), which refers to the proportion of categories covered by the product combination scheme to the total number of categories, with a value range of [0,1]; Category distribution entropy (Ent) refers to the uniformity of the category distribution of product mix options; its value ranges from [0, log_n]. The Gini coefficient refers to the balance of product mix distribution across product categories; its value ranges from [0,1].

[0119] Product diversity can be calculated using the following formula: ; in, , , These are weighting coefficients, and the sum of the three is 1. Preferably, they can be set to... , , .

[0120] Below is a practical calculation example: Suppose a recommendation scheme contains 10 products, involving 5 categories, with the category distribution as follows: 4 products in category A, 3 products in category B, 1 product in category C, 1 product in category D, and 1 product in category E.

[0121] Calculate category coverage: ; Calculate the category distribution entropy: After standardization, the result is ; Calculate the Gini coefficient for each category: Gini = 1 - (0.4² + 0.3² + 3 × 0.2²) = 0.66, and the balance coefficient is 1 - Gini = 0.34. Calculate category diversity: .

[0122] Step S208: Push product combination schemes to users; obtain at least one of the following actions taken by users after receiving the push: clicking on product links, purchasing products, or rating products.

[0123] The top-ranked product combination schemes obtained in step S207 are pushed to users through a preset push channel. For example: the sales staff's mobile terminal device is used to display the 3D combination scheme in real time and supports augmented reality tiling effect preview; the store's large screen is used to display the scheme effect diagram; and the user's WeChat is used to push the mini-program link so that the user and their family can browse, modify, and place an order together.

[0124] Optionally, fatigue control can also be implemented before the push: the number of times the same user receives pushes from the same channel within a preset time window does not exceed a preset threshold, and / or the incremental conversion rate prediction model is used to predict the incremental conversion rate generated after the push. If the incremental conversion rate is lower than the preset threshold, the push is suppressed.

[0125] Step S209: Determine the context information based on the operation; call the large model to learn based on the context information, and use the learned large model to update the large model that has been pre-learned based on the context.

[0126] User-generated data is written to a streaming processing job (such as Flink) in real time via a message queue. High-quality interaction samples are selected from the obtained data and written as positive sample prompts, while low-conversion samples (such as CVR < 5%) are written as negative sample prompts. The large model is invoked to learn based on contextual information, including: invoking the large model to learn context based on positive and negative sample prompts, and updating the pre-learned large model.

[0127] The written data can be used for periodic incremental updates to user profiles; for incremental training of a deep learning neural network for generating demand vectors using a parameter server architecture; and for hot-loading (TF-Serving) of the algorithm for ranking product combination schemes, which can be updated through canary releases via an A / B testing framework.

[0128] Step S210: Based on the results of demand collection and the actions taken by users after receiving the push notification, determine the sales guide performance indicators and provide business guidance based on the performance indicators.

[0129] Based on the results of demand collection and the actions taken by users after receiving push notifications, sales guide performance indicators can be determined, which may include: Determine the completeness of the requirements collection, which is used to measure the fill rate of key requirements. A fill rate that reaches a preset threshold is considered qualified. The adoption rate of recommended solutions is determined to represent the degree of user approval of product combination schemes. A user's action of clicking, saving, or sharing any product combination scheme is considered as adoption. Determine the conversion rate to measure the proportion of customers who make a purchase after returning to the store out of the total number of customers who visit the store; The cross-selling rate is determined to characterize cross-category sales capability, and is determined by counting the number of cross-category product SKUs in each order; the average transaction time is used to measure the length of time from demand collection to payment completion.

[0130] Business guidance based on performance indicators may include: ranking sales associates according to performance indicators, and providing incentive resources to sales associates whose ranking is within a preset proportion. These incentive resources include, but are not limited to, cash rewards, coupons, additional commissions, and training opportunities; and triggering business guidance reminders in response to sales associates' performance indicators falling below a preset threshold.

[0131] Figure 7 This is a schematic diagram of the main modules of a demand information collection device provided in one embodiment of the present invention, as shown below. Figure 7 As shown, the demand information collection device 700 includes: The data acquisition module 701 is used to collect user demand information and generate demand tags based on the collection results. User profile generation module 702 is used to query whether a historical user profile exists; in response to the absence of a historical user profile, to generate a user profile based on user needs; in response to the existence of a historical user profile, to generate a user profile based on user needs and the historical user profile. The requirement vector generation module 703 is used to generate requirement vectors based on requirement tags and user profiles; The product pairing output module 704 is used to input the demand vector into a large model that has undergone context learning in advance, and output product pairing schemes that include at least two product categories.

[0132] Figure 8 An exemplary system architecture 800 is shown, in which the method or apparatus for demand information collection can be applied according to embodiments of the present invention.

[0133] like Figure 8 As shown, system architecture 800 may include terminal devices 801, 802, and 803, a network 804, and a server 805. Network 804 serves as the medium for providing communication links between terminal devices 801, 802, and 803 and server 805. Network 804 may include various connection types, such as wired or wireless communication links or fiber optic cables, etc.

[0134] Users can use terminal devices 801, 802, and 803 to interact with server 805 via network 804 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 801, 802, and 803, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social media platform software, etc.

[0135] Terminal devices 801, 802, and 803 can be various electronic devices with displays and web browsing capabilities, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0136] Server 805 can be a server that provides various services, such as a backend management server that supports shopping websites browsed by users using terminal devices 801, 802, and 803. The backend management server can analyze and process data such as the received demand information collection results, and feed back the processing results (such as product combination schemes) to the terminal devices.

[0137] It should be noted that the demand information collection method provided in the embodiments of the present invention is generally executed by server 805, and correspondingly, the demand information collection device is generally set in server 805.

[0138] It should be understood that Figure 8 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0139] The following is for reference. Figure 9 It shows a schematic diagram of the structure of a computer system 900 suitable for implementing a terminal device of the present invention. Figure 9The terminal device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0140] like Figure 9 As shown, the computer system 900 includes a central processing unit (CPU) 901, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 902 or programs loaded from storage section 908 into random access memory (RAM) 903. The RAM 903 also stores various programs and data required for the operation of the system 900. The CPU 901, ROM 902, and RAM 903 are interconnected via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.

[0141] The following components are connected to I / O interface 905: an input section 906 including a keyboard, mouse, etc.; an output section 907 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 908 including a hard disk, etc.; and a communication section 909 including a network interface card such as a LAN card, modem, etc. The communication section 909 performs communication processing via a network such as the Internet. A drive 910 is also connected to I / O interface 905 as needed. A removable medium 911, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 910 as needed so that computer programs read from it can be installed into storage section 908 as needed.

[0142] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 909, and / or installed from removable medium 911. When the computer program is executed by central processing unit (CPU) 901, it performs the functions defined above in the system of this invention.

[0143] It should be noted that the computer-readable medium shown in this invention can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this invention, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0144] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0145] The modules described in the embodiments of the present invention can be implemented in software or hardware. The described modules can also be housed in a processor; for example, a processor may be described as including a data acquisition module, a user profile generation module, a demand vector generation module, and a product pairing output module. The names of these modules do not necessarily limit the module itself; for example, the data acquisition module may also be described as "a module for collecting user demand information."

[0146] In another aspect, the present invention also provides a computer-readable medium, which may be included in the device described in the above embodiments; or it may exist independently and not assembled into the device. The computer-readable medium carries one or more programs that, when executed by the device, cause the device to perform the method of recommending complementary goods in the first aspect, including: Collect user demand information and generate demand tags based on the collection results; Check if a historical user profile exists; if no historical user profile exists, generate a user profile based on user needs; if a historical user profile exists, generate a user profile based on user needs and the historical user profile. Generate a demand vector based on demand tags and user profiles; Input the demand vector into a large model that has undergone context learning beforehand, and output product combination schemes that include at least two product categories.

[0147] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can occur depending on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for recommending product combinations, characterized in that, include: Collect user demand information and generate demand tags based on the collection results; Check if a historical user profile exists; In response to the absence of the historical user profile, a user profile is generated based on the user's needs; In response to the existence of the historical user profile, a user profile is generated based on the user needs and the historical user profile; Based on the aforementioned demand tags and user profiles, a demand vector is generated; The demand vector is input into a large model that has undergone context learning beforehand, and the output includes product combination schemes that include at least two product categories.

2. The method according to claim 1, characterized in that, Collect user needs information, including: Key requirements can be determined through at least one of the following methods: obtaining dialogue content from users through a question-and-answer format, and determining key requirements based on the user dialogue content; obtaining intentional images provided by users, and obtaining key requirements based on the shape, size, color, and pattern of the items in the intentional images; obtaining drawn images provided by users, and determining key requirements based on the size, position, and area of ​​the items in the drawn images; or showing users a questionnaire containing question options, and determining key requirements based on the answers returned by users. Summarize all key requirements, and if there are duplicates among the key requirements, retain any one of the duplicates to obtain the collection results.

3. The method according to claim 1, characterized in that, After outputting product combination schemes that include at least two product categories, the method further includes: Check whether the product combination scheme conforms to the preset output structure; In response to a situation that does not conform to the output structure, at least two product entities of different categories are matched in a pre-built knowledge graph based on the user's needs and historical data; and the product pairing scheme is determined based on the product entities.

4. The method according to claim 1, characterized in that, After outputting product combination schemes that include at least two product categories, the method further includes: Query the store rules for each product in the product combination scheme; If any product in the product combination scheme does not meet the store rules, the corresponding product combination scheme will be filtered out from all product combination schemes.

5. The method according to claim 1, characterized in that, After outputting product combination schemes that include at least two product categories, the method further includes: Determine real-time user data and historical transaction data; The product combination scheme, real-time user data, and historical transaction data are input into a deep learning neural network targeting at least two objectives, and the product combination scheme is output as a predicted value for at least two objectives. Calculate the Pareto front for each product combination based on the predicted values, and then rank the product combinations according to the Pareto front.

6. The method according to claim 1, characterized in that, After outputting product combination schemes that include at least two product categories, the method further includes: The product combination scheme will be pushed to the user; The system acquires at least one of the following actions performed by the user after receiving the push notification: clicking a product link, purchasing a product, or rating a product. Based on the operation, determine the context information; The large model is invoked to learn based on the context information, and the learned large model is used to update the large model that has undergone context learning in advance.

7. A device for recommending complementary products, characterized in that, include: The data acquisition module is used to collect user demand information and generate demand tags based on the collection results. The user profile generation module is used to query whether historical user profiles exist. In response to the absence of the historical user profile, a user profile is generated based on the user's needs; In response to the existence of the historical user profile, a user profile is generated based on the user needs and the historical user profile; The demand vector generation module is used to generate demand vectors based on the demand tags and user profiles. The product pairing output module is used to input the demand vector into a large model that has undergone context learning in advance, and output product pairing schemes that include at least two product categories.

8. An electronic device, characterized in that, include: One or more processors; Storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-6.

9. A computer-readable medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-6.

Images