Multi-product selling combination generation method, commodity recommendation method, device, medium and equipment

Abstract

The invention discloses a multi-product selling combination generation method, a commodity recommendation method and device, a medium and equipment. The multi-product selling combination generation method comprises the steps of obtaining commodity information of a to-be-sold commodity; generating on-sale commodities corresponding to the to-be-on-sale commodities according to the commodity information through a pre-trained on-sale commodity generation model, and constructing a multi-commodity on-sale combination comprising the to-be-on-sale commodities and the on-sale commodities; according to the commodity combination information corresponding to the multi-product on-sale combination, establishing on-sale combination quality inspection cue words, and through a pre-trained on-sale combination quality inspection model, according to the on-sale combination quality inspection cue words, performing quality inspection on the multi-product on-sale combination; and determining the multi-product selling combination meeting the quality inspection condition as a target multi-product selling combination. Through a double-stage architecture of model generation and quality inspection verification, the benign balance of efficiency-quality-experience is realized, and a better technical solution is provided for a multi-product selling strategy in the fields of e-commerce and retail.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a multi-product bundled combination generation method, a commodity recommendation method, a device, a medium and equipment. BACKGROUND

[0002] In the field of e-commerce and retail, the generation efficiency and rationality of multi-product bundled combination as a strategy to improve user stickiness have a direct impact on user experience.

[0003] However, the current technical practice has significant shortcomings: on the one hand, the traditional bundled combination construction is highly dependent on manual experience, and the operation personnel need to complete the commodity screening and binding based on industry cognition and subjective judgment. This method is not only inefficient, but also difficult to adapt to the dynamic update needs of a large number of commodity categories due to the limited coverage of manpower, especially in scenarios such as seasonal commodity adjustment or new product listing, the cost of manual optimization is high and the response is lagging; on the other hand, although some automatic systems try to generate bundled combinations through preset rule engines, such as binding commodities based on "high-frequency co-purchase" and other simple logic, such methods are essentially driven by static rules and lack the ability to dynamically mine complex commodity relationships, making it difficult to capture implicit associations such as complementary products and substitutes, and also difficult to meet individual user needs, resulting in insufficient diversity of combinations; more importantly, existing technologies generally lack a rationality verification mechanism for bundled combinations, and combinations generated by manual experience or rule engines may contain logical contradictions such as functional conflicts, unreasonable price gradients, or user experience problems such as forced bundling of low-demand commodities, but due to the lack of quantitative quality inspection means, such unreasonable combinations are easily directly flowed into the sales link, ultimately causing user loss and resource waste.

[0004] The above defects collectively result in a long-term contradiction between efficiency and quality in the generation of multi-product bundled combinations. SUMMARY

[0005] Therefore, the embodiments of the present application provide a multi-product bundled combination generation method, a commodity recommendation method, a device, a medium and equipment.

[0006] According to one aspect of the present application, a multi-product bundled combination generation method is provided, comprising: obtaining commodity information of a to-be-bundled commodity; generating a bundled commodity corresponding to the to-be-bundled commodity according to the commodity information through a pre-trained bundled commodity generation model, and constructing a multi-product bundled combination containing the to-be-bundled commodity and the bundled commodity; constructing a bundled combination quality inspection prompt word according to the commodity combination information corresponding to the multi-product bundled combination, and performing quality inspection on the multi-product bundled combination through a pre-trained bundled combination quality inspection model according to the bundled combination quality inspection prompt word; The multi-product combination meeting the quality inspection condition is determined as a target multi-product combination.

[0007] According to another aspect of the present application, a commodity recommendation method is provided, comprising: obtaining a target multi-product combination generated in advance; performing commodity recommendation according to the target multi-product combination.

[0008] According to still another aspect of the present application, a multi-product combination generation device is provided, comprising: a commodity information acquisition module configured to acquire commodity information of a commodity to be sold in combination; a combination construction module configured to generate a combination commodity corresponding to the commodity to be sold in combination according to the commodity information through a pre-trained combination commodity generation model, and construct a multi-product combination containing the commodity to be sold in combination and the combination commodity; a combination inspection module configured to construct a combination inspection prompt word according to commodity combination information corresponding to the multi-product combination, and perform quality inspection on the multi-product combination according to the combination inspection prompt word through a pre-trained combination inspection model; a combination generation module configured to determine the multi-product combination meeting the quality inspection condition as a target multi-product combination.

[0009] According to still another aspect of the present application, a commodity recommendation device is provided, comprising: a combination acquisition module configured to acquire a target multi-product combination generated in advance; a commodity recommendation module configured to perform commodity recommendation according to the target multi-product combination.

[0010] According to still another aspect of the present application, a storage medium having a computer program stored thereon is provided, the program being executed by a processor to implement the multi-product combination generation method.

[0011] According to still another aspect of the present application, a computer device is provided, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, the processor executing the program to implement the multi-product combination generation method.

[0012] By the technical scheme, the multi-product cross-selling combination generation method, the commodity recommendation method, the device, the medium and the equipment provided by the embodiment of the application first determine cross-selling commodities according to commodity information of to-be-cross-sold commodities by a cross-selling commodity generation model to construct a multi-product cross-selling combination, replace subjective judgment of artificial experience and static logic of a rule engine, improve efficiency and coverage of combination generation, and reduce artificial optimization cost. Secondly, the cross-selling combination is verified by a cross-selling combination quality inspection model to avoid negative influence of unreasonable combinations on user experience and reduce user loss and resource waste. Finally, while improving operation efficiency, the rationality of the cross-selling combination and user satisfaction are ensured, the dual-stage architecture of'model generation + quality inspection verification' is implemented, the benign balance of 'efficiency-quality-experience' is achieved, and a more optimal technical solution is provided for a multi-product cross-selling strategy in the e-commerce and retail fields.

[0013] The above description is only a summary of the technical scheme of the application, in order to more clearly understand the technical means of the application, the specific embodiments of the application can be implemented according to the content of the specification, and in order to make the above and other purposes, characteristics and advantages of the application more obvious and easy to understand, the following specifically describes the embodiments of the application. BRIEF DESCRIPTION OF DRAWINGS

[0014] The drawings described herein are used to provide further understanding of the application, constitute a part of the application, the schematic embodiments of the application and the description thereof are used to explain the application, and do not constitute an improper limitation on the application. In the drawings: Figure 1 A flowchart of a multi-product cross-selling combination generation method provided by the embodiment of the application is shown; Figure 2 A flowchart of a training method of a picture reconstruction model provided by the embodiment of the application is shown; Figure 3 A flowchart of a training method of a cross-selling commodity text generation model provided by the embodiment of the application is shown; Figure 4 A flowchart of a training method of a cross-selling combination quality inspection model provided by the embodiment of the application is shown; Figure 5 A flowchart of a commodity recommendation method provided by the embodiment of the application is shown; Figure 6 A structural diagram of a multi-product cross-selling combination generation device provided by the embodiment of the application is shown; Figure 7 A structural diagram of a commodity recommendation device provided by the embodiment of the application is shown. DETAILED DESCRIPTION

[0015] Hereinafter, the present application will be described in detail with reference to the accompanying drawings and in conjunction with embodiments. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0016] In view of the problems of insufficient user behavior or insufficient graph knowledge under the medium and long tail in the prior art, and the limitation of end-to-end generation technology to the early technical framework, the present application introduces a cross-selling commodity generation model to generate a cross-selling commodity combination, so as to alleviate the problem of sparse user behavior and knowledge under the medium and long tail. And in the cross-selling combination quality inspection stage, a multi-modal large model is introduced to further improve the quality inspection accuracy. Specifically: 1. Insufficient user behavior or insufficient graph knowledge under the medium and long tail: Insufficient user behavior: For popular commodities (head commodities), since there are a large amount of user behavior data available for analysis, it is easy to form relatively stable association rules. However, for medium and long tail commodities (i.e. commodities with low sales and low attention), it is difficult to extract effective association patterns from the relatively small amount of user behavior data. This means that the accuracy and confidence of the recommendation system for such commodities will be greatly reduced.

[0017] Insufficient graph knowledge: The construction of a knowledge graph depends on the connection between existing commodity information and attributes. For medium and long tail commodities, due to the lack of sufficient attention and description, their connection in the knowledge graph will also be relatively sparse, which makes it difficult to obtain reliable associated commodity recommendations through graph reasoning.

[0018] 2. End-to-end generation technology is limited by early technical framework: Technical framework limitation: The technical framework ten years ago is relatively backward compared to now, especially in the field of deep learning. The early Encoder-Decoder architecture may not fully utilize the latest neural network structures and technical improvements, such as attention mechanisms, Transformer models, etc. These new developments can significantly improve the expressiveness and generalization ability of the model.

[0019] Underfitting & Overfitting: Even if an end-to-end method is used, for medium and long tail commodities, due to the small amount of available data, the model may still face the problem of overfitting, resulting in low accuracy of the generated recommendation results.

[0020] In the present embodiment, a multi-product cross-selling combination generation method is provided, as shown in Figure 1 The method comprises: Step 101, obtaining commodity information of commodities to be cross-sold.

[0021] In the embodiments of the present application, for the to-be-cross-selling goods that need to be combined for cross-selling combination construction, the goods information of the to-be-cross-selling goods is collected, so as to construct the cross-selling combination according to the goods information. The goods information can specifically include basic attributes (such as category, function, price, etc.), and can also include sales data (such as historical sales volume, user evaluation), user behavior data (such as browsing record, co-purchase frequency) and other multi-dimensional information, providing data support for subsequent model analysis of goods correlation.

[0022] In step 102, the pre-trained cross-selling goods generation model generates the corresponding cross-selling goods of the to-be-cross-selling goods according to the goods information, and constructs a multi-product cross-selling combination containing the to-be-cross-selling goods and the cross-selling goods.

[0023] In the embodiments of the present application, the pre-trained cross-selling goods generation model can automatically mine the implicit correlation (such as complementary goods, substitute goods, scene matching, etc.) between goods from a large amount of historical data, breaking through the subjective limitations of traditional manual experience and the static logic of rule engine. For example, the model can identify the scene correlation between “swimwear” and “sunscreen”, or the functional complementary relationship between “coffee machine” and “coffee beans”, thereby dynamically generating cross-selling combinations that meet the potential needs of users. Thus, the generation efficiency of the cross-selling combination is improved, a wider range of product categories is covered, and it is helpful to quickly adapt to dynamic scenarios such as new product listing and seasonal changes, solving the problem of limited coverage of manual experience and the lack of dynamic mining ability of rule engine.

[0024] In step 103, cross-selling combination quality inspection prompt words are constructed according to the goods combination information corresponding to the multi-product cross-selling combination, and a pre-trained cross-selling combination quality inspection model is used to perform quality inspection on the multi-product cross-selling combination according to the cross-selling combination quality inspection prompt words.

[0025] In the embodiments of the present application, in order to ensure the quality of the multi-product cross-selling combination, the pre-trained cross-selling combination quality inspection model can be used to perform quality inspection on the multi-product cross-selling combination. For example, for the generated multi-product cross-selling combination, quality inspection prompt words are constructed in dimensions including “functional compatibility” (such as functional conflict between “electric kettle” and “gas stove”), “price gradient rationality” (such as user perception imbalance of high-price main goods combined with ultra-low-price cross-selling goods), “user demand matching degree” (such as bundling low-frequency demand goods), and the like, and a pre-trained quality inspection model is used for quantitative verification. The quality inspection model automatically identifies logical contradictions (such as functional conflicts) or user experience problems (such as bundling low-frequency demand goods) in the combination through natural language processing or structured data analysis, making up for the defects of the prior art that lack quantitative quality inspection means. For example, the model can detect the seasonal matching contradiction between “down jacket” and “short-sleeved T-shirt”, or the unreasonable price gradient problem between “high-end camera” and “cheap tripod”, thereby avoiding unreasonable combinations from flowing into the sales link.

[0026] Step 104: determining the multi-product bundle combination meeting the quality inspection condition as a target multi-product bundle combination.

[0027] In the embodiment of the application, after quality inspection, the combination meeting the quality inspection condition is selected as the final target multi-product bundle combination to ensure the quality of the final multi-product bundle combination. Thus, through the two-stage architecture of "model generation + quality inspection verification", the efficiency and quality of multi-product bundle combination generation are improved, and the problem of "efficiency and quality contradiction" in the traditional method is solved.

[0028] By applying the technical solution of the embodiment, first, the bundle product generation model determines the bundle product according to the product information of the product to be bundled, constructs a multi-product bundle combination, replaces the subjective judgment of artificial experience and the static logic of rule engine, improves the efficiency and coverage of combination generation, and reduces the cost of artificial optimization. Secondly, the bundle combination quality inspection model verifies the combination rationality, avoids the negative impact of unreasonable combination on user experience, reduces user loss and resource waste. Finally, while improving the operation efficiency, the rationality and user satisfaction of the bundle combination are guaranteed, through the two-stage architecture of "model generation + quality inspection verification", the "efficiency-quality-experience" is balanced, and a better technical solution is provided for the multi-product bundle strategy in the e-commerce and retail field.

[0029] In the embodiment of the application, the product information includes product picture information, and the bundle product generation model includes a picture reconstruction model. In step 102, the pre-trained bundle product generation model generates the bundle product corresponding to the product to be bundled according to the product information, including: inputting the product picture information into the picture reconstruction model to perform product picture segmentation on the product picture through the image reconstruction model, and reconstructing the product picture information based on the segmented product picture block to predict the missing product picture corresponding to the product picture information; and determining the product corresponding to the missing product picture as the bundle product corresponding to the product to be bundled.

[0030] In this embodiment, the commodity information can be commodity picture information, and the upsell commodity generation model realizes intelligent generation of upsell commodities by integrating a picture reconstruction model. The specific process is as follows: first, input the commodity picture information of the to-be-upsold commodity into the pre-trained picture reconstruction model, the model decomposes the commodity picture into multiple semantically meaningful regions (such as commodity main body, accessories, background elements, etc.) through image segmentation technology (such as semantic segmentation or instance segmentation based on convolutional neural network), and obtains the segmented commodity picture block; then, the model performs picture reconstruction based on these segmentation blocks, and predicts the "missing" part in the original picture (for example, if the original picture only shows the commodity main body, the model can infer the usually matched accessories or complementary products) through a generative adversarial network (GAN) or an autoencoder (Autoencoder) algorithm; finally, the physical commodity corresponding to the predicted missing commodity picture is determined as the upsell commodity of the to-be-upsold commodity, and a multi-product upsell combination containing both is constructed. This scheme uses visual information to mine implicit correlations by introducing a picture reconstruction model. For example, for clothing commodities, the model can identify the matching relationship between "top" and "bottom" through picture segmentation, or predict the style matching between "handbag" and "shoes" through reconstruction; for home commodities, the model can capture the scene co-occurrence rules of "sofa" and "tea table". This upsell generation method based on visual correlation not only improves the diversity of the combination (such as discovering color or style matching that is difficult for humans to detect), but also enhances the rationality of the combination (such as avoiding functionally conflicting visual mismatched combinations). In addition, the automatic processing of the picture reconstruction model reduces the cost of manual screening of upsell commodities, improves the user appeal and commercial conversion potential of the upsell combination through deep mining of visual information, while maintaining the efficiency and adaptability of the generation process.

[0031] In the embodiments of the present application, optionally, as shown in Figure 2 The training step of the picture reconstruction model includes: Step 201, obtaining a combination picture sample corresponding to an upsell commodity combination sample, wherein the combination picture sample contains commodity combination pictures of each commodity in the upsell commodity combination sample; Step 202, performing picture segmentation on the commodity combination picture to determine the independent commodity picture of each commodity in the commodity combination picture, and performing masking processing on at least one independent commodity picture in the commodity combination picture to obtain a commodity missing picture corresponding to the commodity combination picture; Step 203, training the picture reconstruction model according to the commodity missing picture and the commodity combination picture.

[0032] In this embodiment, a combination picture sample library containing multi-product co-occurrence relationships is constructed by collecting successful multi-product bundle combinations in real sales scenarios (such as user high-frequency co-purchase combinations, operation preferred combinations) or pre-constructed potential multi-product bundle combinations corresponding picture data. For example, a clothing combination may include a "T-shirt + jeans + sneakers" combination picture, a home combination may include a "sofa + coffee table + carpet" scene picture, and a dish combination may include a "fish-flavored shredded pork + rice + cola" set picture. Using instance segmentation technology (such as Segment Anything Model), each product in the combination picture is independently separated to generate an accurate contour mask (Mask) for each product, resulting in independent pictures of each product. For example, a combination picture containing "phone + earphone + power bank" is segmented into three independent product pictures. Randomly cover at least one product in the combination picture (such as covering the "earphone"), generate a product missing picture. Simulate a scenario where some products are visible and the missing products need to be predicted, so that the model learns the implicit association between products. The product missing picture is used as input, and the original combination picture is used as target output to train the picture reconstruction model (such as MAE, Masked Autoencoder, a self-supervised learning method based on the Transformer architecture, mainly used in the field of vision, especially image processing tasks). In traditional autoencoders, the entire input image is encoded into a low-dimensional latent representation, and then the original image is reconstructed by the decoder. However, MAE selectively masks a portion of the input image, only encoding the unmasked part, and then uses the decoder to restore the masked image part. The workflow of MAE: Masking process: First, divide the input image into a series of image blocks (for example, use fixed-size patches), then randomly select a part of the image blocks to mask (i.e. hide the information of this part of the image blocks). Encoding process: Use an encoder (usually a pre-trained Transformer encoder) to encode the unmasked image blocks to generate a latent representation. Decoding process: The encoded representation is fed into a decoder (also a Transformer structure), and the task of the decoder is to predict the content of the masked image blocks based on the encoded image block information. Loss calculation: By comparing the decoder's prediction of the image blocks with the actual masked image blocks, a loss function (such as mean square error MSE) is calculated to optimize the model parameters. MAE is an efficient self-supervised learning method that learns image features by masking part of the image and predicting those parts. This method not only significantly reduces the demand for computing resources, but also improves the performance of the model through large-scale unlabeled data set learning. Thus, the model learns to reconstruct the complete combination picture from the missing picture, automatically capturing the visual association rules between products (such as color matching, functional complementarity, scene co-occurrence, etc.).For example, the model can learn that "fish-flavored shredded pork" is often paired with "rice", or that "sports shoes" are consistent in style with "gym wear". This training method, through a closed-loop learning of "segmentation-masking-reconstruction", enables the model to dynamically mine product associations from the visual dimension, solving the problems of low efficiency and poor adaptability of traditional methods, and improving the rationality and diversity of combinations through data-driven methods, ultimately achieving a dual improvement in the efficiency and quality of multi-product combination generation.

[0033] In the embodiment of the present application, optionally, step 203 comprises: performing missing product prediction on the missing product image sample by an initial image reconstruction model; performing image difference calculation and category similarity calculation on the predicted missing product image and the masked product image in the product combination image respectively, calculating the loss of the initial image reconstruction model according to the image difference calculation result and the category similarity calculation result, and performing parameter training on the initial image reconstruction model according to the loss, and determining the image reconstruction model based on the initial image reconstruction model after parameter training.

[0034] In this embodiment, the specific process of training the picture reconstruction model includes: inputting the generated missing commodity picture into the initial picture reconstruction model, and the model predicts the complete picture of the masked commodity through an encoder-decoder structure. For example, if the masking process covers the "fish-flavored shredded pork" in the combined picture, the model needs to generate a prediction result that is highly similar to the original "fish-flavored shredded pork" picture. Further, the visual difference between the predicted picture and the real masked commodity picture is quantified by a pixel-level indicator (such as mean square error MSE, structural similarity index SSIM). For example, if the predicted picture color or shape deviates greatly from the real picture, the MSE value will increase significantly. At the same time, the pre-trained classification model (such as ResNet-50) is used to extract the feature vectors of the predicted picture and the real picture, and the category consistency is evaluated by indicators such as cosine similarity. For example, if the missing stir-fried dish "fish-flavored shredded pork" is incorrectly generated as a western dish "hamburger", the similarity of its feature vector to the real "fish-flavored shredded pork" dish category will be lower. The picture difference loss and the category similarity loss are weighted and fused into a total loss function, and the model parameters are adjusted through back propagation. For example, the total loss can be defined as: Ltotal = a • Lrecon + (1-a) • Lcategory, where a is the weight coefficient, Lrecon is the reconstruction loss of the picture difference, and Lcategory is the category similarity loss. The model improves the visual fidelity and semantic rationality of the predicted picture by minimizing the total loss. This training method introduces the category similarity loss to force the model to generate visually realistic pictures while maintaining semantic consistency with the real commodity. This design avoids the generation of unrelated commodities in the bundle sale combination (such as "rice + coke" combination incorrectly generating "hamburger"), and improves the logical rationality of the combination. Through the training of real and potential bundle sale combination samples, the model can learn the complex rules of multiple commodity co-occurrence. For example, in the "sofa + tea table + carpet" combination, the model not only needs to predict the masked "carpet" picture, but also needs to ensure that its style (such as modern minimalist) is consistent with "sofa" and "tea table" through the category similarity loss. This training method enables the model to capture implicit associations that are difficult for humans to define, break through the simple logic of rule engines, and generate more diverse combination options. The model needs to meet the requirements of visual reconstruction and semantic consistency during training, which indirectly provides a data foundation for subsequent bundle sale combination quality inspection. For example, if the model cannot accurately predict the missing "short-sleeved T-shirt" picture in the "down jacket + short-sleeved T-shirt" combination (due to seasonal conflict leading to low category similarity), the combination may be automatically filtered during the training phase, reducing the generation of unreasonable combinations and reducing the cost of the quality inspection phase.

[0035] Specifically, the initial picture reconstruction model trained based on the parameters determines the picture reconstruction model, including: obtaining a bundled goods combination enhanced sample, wherein the bundled goods combination enhanced sample includes a combination score of an enhanced bundled goods combination and an enhanced goods combination picture containing each goods in the enhanced bundled goods combination; setting a learning rate lower than the parameter training for the initial picture reconstruction model trained based on the parameters; performing picture segmentation on the enhanced goods combination picture to determine an independent goods picture of each goods in the enhanced goods combination picture, and performing masking processing on at least one independent goods picture in the enhanced goods combination picture to obtain an enhanced goods missing picture corresponding to the enhanced goods combination picture; performing missing goods prediction on the enhanced goods missing picture by the initial picture reconstruction model trained based on the parameters, and according to the picture difference and the category similarity between the predicted missing goods picture and the masked goods picture in the enhanced goods combination picture, combining the combination score corresponding to the enhanced goods combination picture, and performing parameter tuning on the initial picture reconstruction model trained based on the parameters to determine the picture reconstruction model.

[0036] In this embodiment, the combination quality can also be quantified based on service indicators (such as user click rate, purchase conversion rate, and bundled rationality score marked by operation), the enhanced bundled goods combination (used for enhancing training of the picture reconstruction model) is scored, and the picture reconstruction model is enhanced trained according to the combination score and the enhanced goods combination picture. Fine-tune on the basis of the initial model parameters to avoid destroying the learned features while improving the adaptability of the model to the enhanced sample. Specifically, the learning rate can be adjusted to 1 / 10 or lower (such as from 0.001 to 0.0001) of the initial training stage. Instance segmentation technology is used to separate each goods in the combination picture to generate independent goods pictures. At least one goods (such as "jeans") is randomly covered to generate an enhanced goods missing picture to simulate the "partially visible" scene in the user's perspective. The enhanced goods missing picture is input into the initial picture reconstruction model trained based on the parameters to predict the complete picture of the masked goods. The picture difference loss, the category similarity loss, and the combination score are weighted and fused, and the model parameters are updated through back propagation to make the model generate high-score combinations while reconstructing the picture. By introducing the combination score (such as conversion rate) as the optimization target, the model not only focuses on visual reconstruction accuracy, but also directly relates to service indicators. For example, the model may learn that the "sports shoes + fitness clothes" combination needs to be generated preferentially due to high conversion rate, while the "dress shoes + sports socks" combination is suppressed due to low score.

[0037] In the embodiment of the present application, optionally, the commodity information includes commodity text information, and the cross-selling commodity generation model includes a cross-selling commodity text generation model; and the step of generating the cross-selling commodity corresponding to the to-be-cross-sold commodity through the pre-trained cross-selling commodity generation model according to the commodity information in step 102 includes: in the case that the to-be-cross-sold commodity is pre-provided with a candidate commodity set, obtaining user browsing history corresponding to the to-be-cross-sold commodity and the candidate commodity set, and screening the cross-selling commodity from the candidate commodity set through the cross-selling commodity text generation model according to the commodity text information and the user browsing history information; in the case that the to-be-cross-sold commodity is not pre-provided with a candidate commodity set, generating the cross-selling commodity corresponding to the to-be-cross-sold commodity through the cross-selling commodity text generation model according to the commodity information.

[0038] In this embodiment, the commodity information can be text information (such as commodity title, description, category label, etc.), and the cross-selling commodity generation model can be a cross-selling commodity text generation model, which can be a LLM model customized based on a general LLM model. When there is a pre-determined candidate commodity set, the main role of LLM is to screen and optimize. In this stage, the customized LLM will receive information from user browsing history, main commodity A and general LLM as input. The customized LLM will evaluate the candidate commodity set based on these information and select those B commodities that can best match commodity A. The customized LLM here can be a specially trained model that can understand user behavior patterns, commodity characteristics and market trends, etc. to make appropriate recommendation decisions. Finally, after screening, one or more groups of cross-selling commodities B related to commodity A will be generated. If there is no candidate commodity set, LLM needs to generate commodity B from scratch. At this time, the customized LLM will completely rely on the knowledge provided by the general LLM for generation. The general LLM is a pre-trained large-scale language model that contains rich semantic understanding and generation capabilities. In this process, the customized LLM will fine-tune the general LLM to adapt to specific recommendation tasks. The fine-tuning process means that the customized LLM will adjust the parameters of the general LLM according to the relevant information of commodity A, so as to generate commodity B related to commodity A and having commercial value.

[0039] In the embodiment of the present application, optionally, as shown in Figure 3 the training step of the cross-selling commodity text generation model includes: Step 301, obtaining a cross-selling commodity combination sample, wherein the cross-selling commodity combination sample includes commodity information of a cross-sold commodity sample, expected description information of the cross-sold commodity sample, and a sample acquisition label, the sample acquisition label is used to identify whether the cross-selling commodity combination sample is a real cross-selling combination or a potential cross-selling combination; Step 302, add a trainable low-rank matrix between preset levels in a general large model, freeze model parameters in the general large model except the low-rank matrix, train the general large model by using the bundled commodity combination sample to optimize the low-rank matrix, and obtain a bundled commodity text generation model containing the trained low-rank matrix.

[0040] Specifically, the bundled commodity combination sample is obtained by a preset bundled commodity combination obtaining method, wherein the bundled commodity combination obtaining method includes at least one of user behavior association, knowledge graph reasoning, pattern rule mining, and machine learning model generation.

[0041] In this embodiment, the training scheme of the bundled commodity text generation model realizes efficient mining and personalized adaptation of commodity text association through multi-source sample collection and low-rank matrix fine-tuning technology. Specifically, the training process first constructs a training data set through diversified bundled commodity combination sample obtaining methods, including user behavior association analysis, knowledge graph reasoning, pattern rule mining, and machine learning model generation, etc., covering high-frequency combined commodities in real scenarios (such as user historical purchases of “mobile phone + mobile phone shell”) and potential combinations inferred by algorithms (such as “coffee machine + coffee beans” based on functional complementarity). The sample not only contains the text information of the bundled commodities (such as commodity title, description, and attribute label), but also annotates the expected description information of the bundled commodities (such as “anti-drop mobile phone shell” needs to contain “adaptable model, anti-drop material” and other keywords) and sample acquisition labels (identifying real or potential bundled combinations). This data construction strategy ensures that the model can not only learn the explicit association in user actual purchase behavior, but also capture the implicit association based on commodity attributes, providing a rich semantic basis for subsequent training. In the model training stage, the low-rank matrix (LoRA) technology is used to adapt and optimize the general large model. Specifically, a trainable low-rank matrix is inserted between the preset levels (such as attention layers) of the general large model, and other parameters are frozen, and only the low-rank matrix is fine-tuned by the bundled commodity combination sample. This design reduces the training cost, and only a small number of original model parameters need to be adjusted to complete the adaptation, while retaining the powerful natural language understanding ability of the general large model. During the training process, the model takes the text information of the bundled commodities and the sample labels as input, generates the description of the bundled commodities, and calculates the loss (such as cross-entropy loss) with the expected description information, and only updates the low-rank matrix parameters through back propagation.

[0042] In the embodiment of the present application, optionally, the step 103 of constructing the bundled combination quality inspection prompt word according to the commodity combination information corresponding to the multi-commodity bundled combination comprises: determining the commodity combination information according to the commodity information and the commodity picture of each commodity in the multi-commodity bundled combination, and constructing the bundled combination quality inspection prompt word according to a preset bundled combination quality inspection prompt template and the commodity combination information, wherein the preset bundled combination quality inspection prompt template comprises scoring point information.

[0043] In this embodiment, first, the commodity combination information containing multi-dimensional features is constructed according to the commodity information (such as text description, category label, price interval) and the commodity picture of the to-be-bundled commodity and the bundled commodity. For example, for the combination of "smartphone + wireless earphone", the commodity information may include text labels such as "5G phone" and "active noise reduction earphone". Then, the combination information is input into a preset bundled combination quality inspection prompt template. The template integrates the core scoring points of quality inspection through structured design, such as functional compatibility (such as "whether the earphone supports the Bluetooth protocol of the phone"), price gradient rationality (such as "whether the price of the earphone matches the price of the phone"), user demand matching degree (such as "whether it contains accessories that are frequently purchased by users"), scene consistency (such as "whether a business phone is matched with a simple style earphone"), and the like. The template converts these scoring points into natural language instructions, such as "please evaluate the functional compatibility, price gradient rationality, and user demand matching degree of the following combination, and give a score of the combination". The finally generated quality inspection prompt word combines multi-modal combination information and structured evaluation standards, and is input into the pre-trained quality inspection model, so that the quality inspection model can quantitatively score and generate optimization suggestions from the aspects of semantics, vision, and service logic. Thus, by integrating commodity text and picture information, the quality inspection model can capture both explicit associations (such as text label matching) and implicit associations (such as visual style coordination), thereby improving the accuracy of quality inspection.

[0044] In the embodiment of the present application, optionally, as shown in Figure 4 The training step of the bundled combination quality inspection model comprises at least one of the following: Step 401, training the multi-modal model according to a first image-text pair sample to train the multi-modal model to associate the image-text pair, wherein the first image-text pair sample comprises an image and a text having an association relationship; Step 402, constructing a visual language understanding task according to a second image-text pair sample, and training the multi-modal model through the visual language understanding task, wherein the visual language understanding task comprises at least one of a visual question answering task, an image description generation task, an image positioning task, and a document understanding task; In step 403, the multi-product bundle combination sample is marked with quality inspection information, and the multi-modal model is trained through the multi-product bundle combination sample with quality inspection information marking. The quality inspection information marking includes the scoring reasons and dimension scores corresponding to multiple scoring dimensions, and a comprehensive quality inspection score.

[0045] In this embodiment, the training and quality inspection process of the multi-modal model is implemented through three core steps. The overall scheme is based on multi-modal data, combined with visual language understanding tasks and quality inspection annotation information, to build a complete training framework from basic capabilities to service adaptation. First, the multi-modal model is trained through the first image-text pair sample to establish the basic capability of image-text association. The model uses a contrastive learning framework to map the output features of the image encoder and the text encoder to the same semantic space. In this stage, the model is first pre-trained through a large number of image-text pairs to learn the basic visual and language representation. This usually involves a large amount of image-text pair data scraped from the Internet, such as image title pairs, image-text parallel web content, etc. The model can use Vision Transformer (ViT) as the visual encoder to extract image features, while the language model part is based on the Transformer architecture. The training goals may include predicting the context of the image, fill-in-the-blank, masked language modeling, contrastive learning, etc., to enable the model to understand and associate the relationship between images and text. Second, the visual language understanding task is constructed through the second image-text pair sample to further improve the cross-modal reasoning capability of the model. The training data in this stage is usually annotated, which is used to let the model learn how to perform specific visual language understanding tasks. The fine-tuning stage may also include freezing certain layers and only training specific parts, such as optimizing only the language model and the adapter module, to reduce the computational cost and maintain the generalization ability of the model. The visual question answering task requires the model to answer natural language questions based on the image content, such as "What color is the product in the image?" or "Which scenarios is the product suitable for?" The model needs to combine the visual features of the image (such as color, shape) and the semantic features of the text (such as keywords in the question) to generate accurate answers. The image description generation task requires the model to automatically generate text descriptions based on the image, such as generating a description of "western main course, better with red wine" for a "grilled steak" image. These tasks are implemented through sequence-to-sequence models (such as BLIP) or cross-attention mechanisms (such as VisualBERT), enabling the model to capture the implicit associations between images and text, such as the complementary functionality of "steak" and "red wine" or the scene consistency of "sports shoes" and "fitness wear". Finally, the model is service-adapted through multi-product bundled combination samples annotated with quality inspection information. Quality inspection annotations include multiple dimensions, such as functional compatibility (whether the products can be used together), price gradient rationality (whether the bundled product prices match the main product), user demand matching degree (whether it meets user's common purchase habits), etc. During training, the model inputs the image-text information of the bundled combination into a multi-task learning framework, with each quality inspection dimension corresponding to a sub-task, and through a shared bottom-level feature extractor (such as ResNet or BERT) and a task-specific prediction head (such as a fully connected layer), the joint optimization is realized. For example, the functional compatibility task judges whether the products can be bundled through a classification loss, the price gradient task predicts the price matching degree through a regression loss, and the comprehensive score integrates the output of each dimension.This design enables the model to optimize multiple service objectives simultaneously, avoiding bias caused by single tasks. Through the above training method, the model's capabilities are gradually improved through multi-stage training, both possessing general image-text understanding capabilities and serving the upsell quality inspection scenario. Second, the multi-dimensional design of quality inspection annotation can cover key factors such as function, price, and user behavior, solving the problem of strong subjectivity and single dimension coverage in traditional quality inspection. In addition, the multi-task learning framework reduces computational cost by sharing feature extractors, while achieving dimension decoupling through task-specific heads, improving training efficiency and model generalization ability.

[0046] In one specific application scenario, assume that a certain catering platform needs to generate upsell goods (such as side dishes, drinks) for the "black pepper steak" main course and verify the rationality of the set meal. During the model training phase, the training data of the picture reconstruction model may include set meal pictures of real user high-frequency orders (such as "steak set meal" containing pictures of steak, wine, and salad combinations), and through segmentation and masking processing (such as covering the wine glass), the model learns the visual association between "steak" and "wine" (such as tableware, liquid color). The text generation model is trained through multiple sources of samples: for example, extracting the co-occurrence behavior of "steak + wine" from user order records (real combination), or reasoning "western main course often matches wine" through knowledge graph (potential combination), and then fine-tuning the general large model (such as BERT) using low-rank matrix, so that it can generate "wine" and other upsell suggestions according to the text description of "black pepper steak". In the quality inspection link, construct quality inspection prompt words according to the generated set meal (such as "black pepper steak + wine + garlic bread"): combine the dish picture and text information (dish name, taste description), input the preset template (such as "Please evaluate the taste coordination, price gradient rationality, and user demand matching degree, and give a score for this set meal"). The quality inspection model can analyze whether the black pepper flavor of the steak and the tannin of the wine are coordinated (taste dimension), whether the total price of the set meal is lower than the single-point discount price (price dimension), and whether the user has historically often ordered "garlic bread" (demand dimension), and finally output the score and optimization suggestions (such as "the wine can be replaced by a more suitable 'Merlot' variety"). Through multi-modal analysis of vision and text, it avoids unreasonable combinations caused by relying on a single data source (such as traditional methods may ignore the visual coordination of "steak sauce color and side dishes"); through the combination of user behavior and knowledge reasoning, it covers real order habits (such as the high-frequency combination of "steak + wine") and excavates potential reasonable combinations (such as the functional complementarity of "steak + salad to relieve greasiness"); through structured quality inspection prompt words, it ensures that the set meal evaluation covers key dimensions such as taste, price, and user preference, solving the problem of strong subjectivity in traditional quality inspection. Ultimately, this solution can help the catering platform generate set meals that better meet user needs (such as improving average order value and satisfaction), while filtering unreasonable combinations (such as "steak + sweet soy milk" taste conflict) to improve set meal generation effectiveness.

[0047] In the embodiments of the present application, optionally, the multi-product bundled combination includes a dish set meal combination; the method further includes: according to the target dish set meal combination, recommending a dish set meal to a merchant terminal, so that the merchant sets an offline dish set meal and / or an online dish set meal based on the recommended dish set meal.

[0048] In this embodiment, after generating the target dish set meal combination, the dish set meal can be recommended to the merchant terminal, so that the merchant can adjust the menu of the restaurant and the set meal of the takeout platform according to the recommended dish set meal, and help the merchant to find dish set meals that meet the needs of users.

[0049] In the embodiments of the present application, optionally, the method further includes: performing dish set meal calorie identification on the target dish set meal combination.

[0050] In this embodiment, the target dish set meal combination can also be subjected to calorie identification, so that users can select set meals that meet their own food calorie needs, further improving the user experience.

[0051] Further, as a refinement and expansion of the above embodiment, in order to fully describe the specific implementation process of the present embodiment, a commodity recommendation method is provided, as shown in Figure 5 The method includes: Step 501, obtaining a pre-generated target multi-product bundled combination; Step 502, according to the target multi-product bundled combination, performing commodity recommendation.

[0052] Specifically, the target multi-product bundled combination is determined by the multi-product bundled combination generation method as described above. According to the target multi-product bundled combination, performing commodity recommendation includes: in response to a triggering behavior of a target commodity, obtaining a recommended multi-product bundled combination containing the target commodity in the target multi-product bundled combination; performing commodity combination recommendation based on the recommended multi-product bundled combination, or performing commodity recommendation based on the commodities in the recommended multi-product bundled combination; wherein the triggering behavior of the target commodity includes at least one of a commodity browsing behavior, an adding behavior, and an order browsing behavior of the target commodity.

[0053] In the embodiments of the present application, the target multi-product bundle combination is constructed in advance by the multi-product bundle combination generation method. For example, in the catering scenario, combinations such as "black pepper steak + red wine + garlic bread" and "roasted chicken with vanilla + cream mushroom soup + tiramisu" are generated. These combinations include not only real combinations that are frequently purchased together by users (such as those mined from user order records), but also potential reasonable combinations generated by knowledge graph reasoning or models (such as algorithm recommendations based on taste complementation and scene matching). When a user has a triggering behavior (such as browsing the product detail page, adding to cart, or viewing historical orders) on a target product (such as "black pepper steak"), the recommended combination containing the product can be retrieved. The recommendation logic is divided into two categories: combination recommendation: directly show the complete bundle set meal (such as "black pepper steak + red wine + garlic bread"), and also label combination discount information (such as "save 20 yuan on set meal price"). Single product recommendation: split out single products that are strongly associated with the target product (such as "red wine" or "garlic bread") from the combination, and display them in the form of "often bought together" or "combination recommendation". Specifically, the product combination recommendation based on the recommended multi-product bundle combination, or the product recommendation based on the products in the recommended multi-product bundle combination, includes: obtaining user feature data, wherein the user feature data includes user historical behavior and / or user portrait; based on the user feature data and the recommended multi-product bundle combination, performing product combination recommendation or product recommendation of the products in the recommended multi-product bundle combination. During the recommendation process, user historical behavior (such as the user's past purchase of "steak + salad" rather than "steak + red wine") can also be combined for personalized adjustment, and combinations that match the user's preferences are preferentially displayed. For example, if the user is price-sensitive, recommend a set meal that includes cost-effective side dishes; if the user is quality-oriented, preferentially display combinations that include high-end red wine.

[0054] In the embodiments of the present application, the target multi-product bundle combination includes a target dish set meal; and the product recommendation according to the target multi-product bundle combination further includes: obtaining the calorie information of the target dish set meal, and performing dish set meal recommendation based on the calorie information of the target dish set meal.

[0055] In this embodiment, for pre-generated dish packages (such as "black pepper steak + red wine + garlic bread"), a nutrition database (such as the Chinese Food Composition Table) or dish ingredient analysis model can also be called to analyze the raw materials of each dish in the package (such as the weight of the steak, the volume of the red wine), the cooking method (such as frying, baking), and based on the standard heat calculation formula (such as the heat of 100 grams of food material x weight), the total heat of the package and the heat distribution of each dish are generated. For example, "herb roasted chicken package" may contain "roasted chicken (1200kcal / each) + butter mushroom soup (300kcal / bowl) + tiramisu (400kcal / portion)", and the total heat is 1900kcal and the proportion of each dish is labeled. In the recommendation stage, the user's health preferences and package heat information can be combined for dynamic adaptation. If the user's historical behavior shows concern for low-carb diet (such as frequent search for "light food" "fat loss meal"), or through user portrait recognition as a health management group (such as fitness enthusiasts, diabetes patients), low-calorie packages (such as "roasted chicken breast package" total heat 800kcal) are preferred, and a "low-carb preferred" label is marked on the recommendation interface. Through the calculation and display of heat data, the problem of ignoring user health needs in traditional recommendations is solved.

[0056] Further, as Figure 1 The specific implementation of the method, the embodiment of the application provides a multi-product combination generation device, as shown in the figure, the device comprises: Figure 6 A commodity information acquisition module for acquiring commodity information of a commodity to be sold together; A combination construction module for generating a combination commodity corresponding to the commodity to be sold together according to the commodity information through a pre-trained combination commodity generation model, and constructing a multi-product combination comprising the commodity to be sold together and the combination commodity; A combination quality inspection module for constructing a combination quality inspection prompt word according to the commodity combination information corresponding to the multi-product combination, and performing quality inspection on the multi-product combination through a pre-trained combination quality inspection model according to the combination quality inspection prompt word; A combination generation module for determining the multi-product combination that meets the quality inspection condition as a target multi-product combination.

[0057] In the embodiment of the application, optionally, the commodity information comprises commodity picture information, and the combination commodity generation model comprises a picture reconstruction model; the combination construction module is configured to: input the commodity picture information into the picture reconstruction model to perform commodity picture segmentation on the commodity picture through the image reconstruction model, and perform picture reconstruction on the commodity picture information based on the segmented commodity picture blocks to predict the missing commodity picture corresponding to the commodity picture information; ​Determine the commodity corresponding to the missing commodity picture as the cross-selling commodity corresponding to the to-be-cross-sold commodity.

[0058] In the embodiments of the present application, the cross-selling combination construction module is configured to: Obtain a combination picture sample corresponding to a cross-selling commodity combination sample, wherein the combination picture sample comprises commodity combination pictures of each commodity in the cross-selling commodity combination sample. Perform picture segmentation on the commodity combination pictures to determine independent commodity pictures of each commodity in the commodity combination pictures, and perform masking processing on at least one independent commodity picture in the commodity combination pictures to obtain a commodity missing picture corresponding to the commodity combination picture. Train the picture reconstruction model according to the commodity missing picture and the commodity combination picture.

[0059] In the embodiments of the present application, the cross-selling combination construction module is configured to: Perform missing commodity prediction on the commodity missing picture sample through an initial picture reconstruction model. Perform picture difference calculation and category similarity calculation on the predicted missing commodity picture and the masked commodity picture in the commodity combination picture, respectively, calculate the loss of the initial picture reconstruction model according to the picture difference calculation result and the category similarity calculation result, and perform parameter training on the initial picture reconstruction model according to the loss, to determine the picture reconstruction model based on the parameter trained initial picture reconstruction model.

[0060] In the embodiments of the present application, the cross-selling combination construction module is configured to: Obtain a cross-selling commodity combination enhancement sample, wherein the cross-selling commodity combination enhancement sample comprises combination scores of enhanced cross-selling commodity combinations and enhanced commodity combination pictures comprising each commodity in the enhanced cross-selling commodity combinations. Set a learning rate lower than the parameter training for the initial picture reconstruction model after the parameter training. Perform picture segmentation on the enhanced commodity combination pictures to determine independent commodity pictures of each commodity in the enhanced commodity combination pictures, and perform masking processing on at least one independent commodity picture in the enhanced commodity combination pictures to obtain enhanced commodity missing pictures corresponding to the enhanced commodity combination pictures. Perform missing commodity prediction on the enhanced commodity missing pictures through the initial picture reconstruction model after the parameter training, and perform parameter optimization on the initial picture reconstruction model after the parameter training according to the picture difference and the category similarity between the predicted missing commodity picture and the masked commodity picture in the enhanced commodity combination picture, in combination with the combination scores corresponding to the enhanced commodity combination pictures, to determine the picture reconstruction model.

[0061] In the embodiment of the present application, optionally, the commodity information includes commodity text information, and the bundled commodity generation model includes a bundled commodity text generation model; the bundled combination construction module is configured to: In a case where the to-be-bundled commodity is pre-provided with a candidate commodity set, obtaining user browsing history corresponding to the to-be-bundled commodity and the candidate commodity set, and filtering the bundled commodity from the candidate commodity set according to the commodity text information and the user browsing history information by using the bundled commodity text generation model; In a case where the to-be-bundled commodity is not pre-provided with a candidate commodity set, generating the bundled commodity corresponding to the to-be-bundled commodity according to the commodity information by using the bundled commodity text generation model.

[0062] In the embodiment of the present application, optionally, the bundled combination construction module is configured to: obtain bundled commodity combination samples, wherein the bundled commodity combination samples include commodity information of a to-be-bundled commodity sample, expected description information of a bundled commodity sample, and a sample acquisition label, and the sample acquisition label is used to identify whether the bundled commodity combination sample is a real bundled combination or a potential bundled combination; add a trainable low-rank matrix between preset levels in a general large model, freeze model parameters in the general large model except the low-rank matrix, train the general large model by using the bundled commodity combination samples to optimize the low-rank matrix, and obtain a bundled commodity text generation model including the trained low-rank matrix.

[0063] In the embodiment of the present application, optionally, the bundled combination construction module is configured to: obtain bundled commodity combination samples in multiple application scenarios by using a preset bundled commodity combination acquisition method, wherein the bundled commodity combination acquisition method includes at least one of user behavior association, knowledge graph reasoning, pattern rule mining, and machine learning model generation.

[0064] In the embodiment of the present application, optionally, the bundled combination quality inspection module is configured to: determine the commodity combination information according to commodity information and commodity pictures of each commodity in the multi-commodity bundled combination, and construct the bundled combination quality inspection prompt words according to a preset bundled combination quality inspection prompt template and the commodity combination information, wherein the preset bundled combination quality inspection prompt template includes scoring point information.

[0065] In the embodiment of the present application, optionally, the bundled combination quality inspection module is configured to: The multimodal model is trained based on the first image-text pair sample to train the association relationship between the multimodal model and the image-text pair, wherein the first image-text pair sample includes images and text with an association relationship; A visual language understanding task is constructed based on the second image and text sample, and a multimodal model is trained through the visual language understanding task. The visual language understanding task includes at least one of a visual question answering task, an image description generation task, an image localization task, and a document understanding task. Quality inspection information is labeled on multi-product bundled combination samples, and a multimodal model is trained using multi-product bundled combination samples with quality inspection information labels. The quality inspection information labels include the scoring reasons and dimension scores corresponding to multiple scoring dimensions, and the comprehensive quality inspection score.

[0066] Optionally, in this embodiment of the application, the multi-item bundled sale combination includes a set meal combination; the bundled sale combination generation module is used for: Based on the target menu combination, menu recommendations are sent to the merchant's terminal so that the merchant can set up offline and / or online menus based on the recommended menu combinations.

[0067] Optionally, in this embodiment of the application, the bundling combination generation module is used to: Identify the calorie content of the target dish set meal.

[0068] Furthermore, as Figure 5 In terms of specific implementation of the method, this application provides a multi-product bundled sales combination generation device, such as... Figure 7 As shown, the device includes: The bundled product acquisition module is used to acquire pre-generated target multi-product bundled product combinations; The product recommendation module is used to recommend products based on the target multi-product bundled combination.

[0069] Optionally, in this embodiment of the application, the product recommendation module is further configured to: In response to a triggering action on the target product, a recommended multi-product bundled package containing the target product is obtained from the target multi-product bundled package; product bundle recommendations are made based on the recommended multi-product bundled package, or product recommendations are made based on the products in the recommended multi-product bundled package; wherein, the triggering action on the target product includes at least one of the following: product browsing action, adding to cart action, and order browsing action.

[0070] Optionally, in this embodiment of the application, the product recommendation module is further configured to: Acquire user characteristic data, wherein the user characteristic data includes user historical behavior and / or user profile; Based on the user feature data and the recommended multi-product combination, a commodity combination recommendation is performed or a commodity recommendation of commodities in the recommended multi-product combination is performed.

[0071] In the embodiments of the present application, optionally, the target multi-product combination comprises a target dish set meal; and the commodity recommendation module is further configured to: obtain calorie information of the target dish set meal, and perform dish set meal recommendation based on the calorie information of the target dish set meal.

[0072] In the embodiments of the present application, optionally, the target multi-product combination is determined by the multi-product combination generation method.

[0073] It should be noted that other corresponding descriptions of each functional unit involved in the multi-product combination generation device and the commodity recommendation device provided in the embodiments of the present application can be referred to the corresponding descriptions in the methods, which will not be described here. Figures 1 to 5

[0074] The embodiments of the present application also provide a computer device, which can be a personal computer, a server, a network device, etc. The computer device includes a bus, a processor, a memory, and a communication interface, and can also include an input / output interface and a display device. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium. The database of the computer device is configured to store location information. The network interface of the computer device is configured to communicate with an external terminal through a network connection. The computer program is executed by the processor to implement the steps in the method embodiments.

[0075] Those skilled in the art can understand that the structure of the computer device described above is only part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components, or combine certain components, or have a different component arrangement.

[0076] In one embodiment, a computer readable storage medium is provided, which can be non-volatile or volatile, and has a computer program stored thereon. The computer program is executed by the processor to implement the steps in the above method embodiments.

[0077] In one embodiment, a computer program product is provided, which includes a computer program. The computer program is executed by the processor to implement the steps in the above method embodiments. ​

[0078] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties.

[0079] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. Any reference to memory, database or other medium used in the embodiments provided by the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The database involved in the embodiments provided by the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, etc., without being limited thereto. The processor involved in the embodiments provided by the present application can be a general-purpose processor, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto.

[0080] The technical features of the above embodiments can be combined arbitrarily. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, but as long as the combinations of the technical features do not exist, they should be considered as the scope of the present application.

[0081] The above-described embodiments are merely illustrative of several embodiments of the present application, and the description is relatively specific and detailed, but should not be understood as a limitation on the scope of the patent. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.

Claims

1. A method for generating multi-product bundled sales combinations, characterized in that, The method includes: Obtain product information for items to be bundled; The pre-trained bundling product generation model generates the bundling products corresponding to the product to be bundled based on the product information, and constructs a multi-product bundling combination that includes the product to be bundled and the bundling products. Based on the product combination information corresponding to the multi-product bundled sale combination, a quality inspection prompt word for the bundled sale combination is constructed, and the multi-product bundled sale combination is inspected based on the quality inspection prompt word for the bundled sale combination using a pre-trained quality inspection model for the bundled sale combination. The multi-product bundled sales combination that meets the quality inspection conditions is identified as the target multi-product bundled sales combination.

2. The method according to claim 1, characterized in that, The product information includes product image information, and the bundled product generation model includes an image reconstruction model; the pre-trained bundled product generation model generates the bundled products corresponding to the product to be bundled based on the product information, including: The product image information is input into the image reconstruction model to segment the product image using the image reconstruction model, and the product image information is reconstructed based on the segmented product image blocks to predict the missing product image corresponding to the product image information. The product corresponding to the missing product image is identified as the bundled product corresponding to the product to be bundled.

3. The method according to claim 2, characterized in that, The training steps of the image reconstruction model include: Obtain a combination image sample corresponding to the bundled product combination sample, wherein the combination image sample includes product combination images of each product in the bundled product combination sample; The product combination image is segmented to determine the independent product image of each product in the product combination image, and at least one independent product image in the product combination image is masked to obtain the missing product image corresponding to the product combination image; The image reconstruction model is trained based on the missing product image and the combined product image.

4. The method according to claim 3, characterized in that, The step of training the image reconstruction model based on the missing product image samples and the combined product images includes: The missing product images are reconstructed using an initial image reconstruction model to predict the missing products. Image difference calculation and category similarity calculation are performed on the predicted missing product images and the masked product images in the product combination images, respectively. The loss of the initial image reconstruction model is calculated based on the image difference calculation results and the category similarity calculation results. The parameters of the initial image reconstruction model are trained based on the loss. The image reconstruction model is determined based on the initial image reconstruction model after parameter training.

5. The method according to claim 4, characterized in that, The initial image reconstruction model, trained based on parameters, determines the image reconstruction model, including: Obtain enhanced samples of bundled product combinations, wherein the enhanced samples of bundled product combinations include a combination score of the enhanced bundled product combination and an enhanced product combination image containing each product in the enhanced bundled product combination; Set the learning rate for the initial image reconstruction model after parameter training to be lower than that used during parameter training. The enhanced product combination image is segmented to determine the independent product image of each product in the enhanced product combination image, and at least one independent product image in the enhanced product combination image is masked to obtain the enhanced product missing image corresponding to the enhanced product combination image; The initial image reconstruction model, trained with parameters, is used to predict missing products in the enhanced product images. Based on the image differences and category similarities between the predicted missing product images and the masked product images in the enhanced product combination images, combined with the combination score corresponding to the enhanced product combination images, the parameters of the initial image reconstruction model are fine-tuned to determine the image reconstruction model.

6. A product recommendation method, characterized in that, The method includes: Obtain pre-generated target multi-item bundled sales combinations; Based on the target multi-product bundled sales combination, product recommendations are made.

7. A multi-product bundled sales combination generation device, characterized in that, The device includes: The product information acquisition module is used to acquire product information of the products to be bundled. The bundled product construction module is used to generate bundled products corresponding to the product to be bundled based on the product information using a pre-trained bundled product generation model, and to construct a multi-product bundled product that includes the product to be bundled and the bundled products. The bundled sales quality inspection module is used to construct bundled sales quality inspection prompts based on the product combination information corresponding to the multi-product bundled sales combination, and to perform quality inspection on the multi-product bundled sales combination based on the bundled sales quality inspection prompts using a pre-trained bundled sales quality inspection model. The bundled product generation module is used to determine the multi-product bundled product combinations that meet the quality inspection conditions as the target multi-product bundled product combinations.

8. A product recommendation device, characterized in that, The device includes: The bundled product acquisition module is used to acquire pre-generated target multi-product bundled product combinations; The product recommendation module is used to recommend products based on the target multi-product bundled combination.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 5.

10. A computer device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 5.

Images