Article attribute recommendation model training, generation method and device, and medium

Abstract

The present disclosure provides a kind of training of article attribute recommendation model, article recommendation attribute generation method and device and storage medium, wherein the training method comprises: determining first associated article information based on first article picture and article labeled attribute;First image association feature information is generated based on first associated article information;Obtain the first article category information and the first article total attribute information corresponding to the article;Using first article picture, first image association feature information, first article category information, first article total attribute information and article labeled attribute and article labeled attribute value, the article attribute recommendation model is trained.The present disclosure obtains the individualized recommendation of attribute based on article original drawing and picture overall and local association information etc., improves the accuracy of article attribute recommendation, generalization ability, stability and article exposure rate.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and apparatus for training an item attribute recommendation model, generating item recommendation attributes, and a storage medium. Background Technology

[0002] In existing technologies, item attribute identification generally employs two technical solutions: 1. Training a classification algorithm or generative algorithm model using labeled image samples, and then using the trained model to identify the item attributes corresponding to the main item image; 2. Searching for identical or highly similar items in the existing inventory of a sales platform based on the main item image, and then predicting the item attributes based on the attributes of identical or highly similar items. The image recognition algorithm-based solution is ineffective for item attributes that are not visually distinguishable from images or for attributes that are occluded. For example, it is difficult to determine the sleeve length of folded clothing from an image; or for material attributes, it is difficult for non-professionals to determine them solely from an image. The image-based approach of finding identical or similar items to predict attributes heavily relies on the recall quality of identical items and the accuracy of the attributes themselves. If the platform's inventory of items is not rich enough, and there are no identical or highly similar items, recalling only some items with similar attributes will cause inaccurate predictions of other attributes. Furthermore, neither of the above two solutions can effectively handle some item attributes that are seasonal or highly subjective. For example, for an item that is suitable for spring or autumn, merchants may select spring when listing it in spring and autumn when listing it in autumn in order to increase the item's exposure. Summary of the Invention

[0003] In view of this, one technical problem to be solved by the present invention is to provide a method and apparatus for training an item attribute recommendation model, generating item recommendation attributes, and a storage medium.

[0004] According to a first aspect of this disclosure, a training method for an item attribute recommendation model is provided, comprising: generating training samples based on a first item image corresponding to the item, item annotation attributes, and item annotation attribute values; obtaining the training samples and determining first associated item information based on the first item image and the item annotation attributes; generating first image association feature information based on the first associated item information; obtaining first item category information and all attribute information of the first item corresponding to the item; and training the item attribute recommendation model using the first item image, the first image association feature information, the first item category information, the first item all attribute information, the item annotation attributes, and the item annotation attribute values ​​to obtain the trained item attribute recommendation model.

[0005] Optionally, the first associated item information includes: first overall associated item information and first partial associated item information; determining the first associated item information based on the first item image and the item annotation attributes includes: using an overall image association model and based on the first item image to generate a first vector; obtaining the first overall associated item information corresponding to the item in a first item vector library based on the first vector; using an partial image association model and based on the first item image and the item annotation attributes to generate a second vector; and obtaining the first partial associated item information corresponding to the item in a second item vector library based on the second vector.

[0006] Optionally, obtaining the first overall associated item information corresponding to the item in the first item vector library based on the first vector includes: obtaining one or more first item vectors that are closest to the first vector in the first item vector library; obtaining the first item SKU information corresponding to the first item vector as the first overall associated item information; obtaining the first local associated item information corresponding to the item in the second item vector library based on the second vector includes: obtaining one or more second item vectors that are closest to the second vector in the second item vector library; obtaining the second item SKU information corresponding to the second item vector as the local associated item information.

[0007] Optionally, the first image association feature information includes: a first overall image association item vector and a first local image association item vector; generating the first image association feature information based on the first associated item information includes: obtaining first item information and a first attribute value corresponding to the first overall image association item information; generating the first overall image association item vector based on the first item information and the first attribute value; obtaining second attribute information and second attribute value corresponding to the local image association item information; and generating the first local image association item vector based on the first attribute information and the second attribute value.

[0008] Optionally, obtaining the first item category information and all attribute information of the first item corresponding to the item includes: obtaining the first item category information based on user settings information or using a category model; obtaining all attribute information of the first item corresponding to the first item category information; wherein, the first item attribute information includes: all item attributes and all corresponding attribute values.

[0009] Optionally, training the item attribute recommendation model using the first item image, the first image association feature information, the first item category information, all attribute information of the first item, and the item annotation attributes and item annotation attribute values ​​includes: using the item attribute recommendation model and based on the first item image, the first image association feature information, the first item category information, and all attribute information of the first item to obtain recommended attributes and corresponding recommended attribute values; and training the item attribute recommendation model based on the comparison results between the recommended attributes, the recommended attribute values, and the item annotation attributes and item annotation attribute values, in order to adjust the item attribute recommendation model.

[0010] Optionally, the item attribute recommendation model includes: an item image model branch corresponding to the first item image, an image association model branch corresponding to the first image association feature information, an item category model branch corresponding to the first item category information, an item attribute model branch corresponding to all attribute information of the first item, an attribute analysis model block, and an attribute output branch; wherein, the vector data output by the item image model branch, the image association model branch, the item category model branch, and the item attribute model branch are obtained respectively; the vector data is input into the attribute analysis model block; the data output by the attribute analysis model block is input into the attribute output branch, and the attribute output branch outputs the attribute value with the highest score and the corresponding attribute.

[0011] Optionally, the item image model branch includes: a ResNet-S1 network in the first layer and a ResNet network in the second layer; wherein, a ResNet network is constructed for each attribute; the item category model branch includes: a bottom-level embedding network and an upper-level ConvNet network; the item attribute model branch includes: a bottom-level embedding network and an upper-level ConvNet network; the image association model branch includes: a bottom-level ConvNet network and an upper-level FCNet network. The attribute analysis model block includes: an FCNet sub-network; wherein, an FCNet sub-network is constructed for each attribute; the attribute output branch includes: a softmax network.

[0012] According to a second aspect of this disclosure, a method for generating item recommendation attributes is provided, comprising: acquiring a second item image, item attributes, and item attribute values ​​corresponding to an item to be processed; determining second associated item information based on the second item image and the item attributes; generating second image association feature information based on the second associated item information; acquiring second item category information and all attribute information of the second item corresponding to the item to be processed; and generating recommendation attributes and corresponding recommendation attribute values ​​using a trained item attribute recommendation model based on the second item image, the second image association feature information, the second item category information, and the all attribute information of the second item; wherein the item attribute recommendation model is trained using the above-described training method.

[0013] Optionally, the second associated item information includes: second overall associated item information and second partial associated item information; determining the second associated item information based on the second item image and the attributes of the item to be processed includes: using an overall image association model and based on the second item image to generate a third vector; obtaining the second overall associated item information corresponding to the item to be processed from a third item vector library based on the third vector; using an partial image association model and based on the second item image and the attributes of the item to be processed to generate a fourth vector; obtaining the second partial associated item information corresponding to the item to be processed from a fourth item vector library based on the fourth vector.

[0014] Optionally, obtaining the second overall associated item information corresponding to the item to be processed from the third item vector library based on the third vector includes: obtaining one or more third item vectors that are closest to the third vector in the third item vector library; obtaining the third item SKU information corresponding to the third item vector as the second overall associated item information; obtaining the second local associated item information corresponding to the item to be processed from the fourth item vector library based on the fourth vector includes: obtaining one or more fourth item vectors that are closest to the fourth vector in the fourth item vector library; obtaining the fourth item SKU information corresponding to the fourth item vector as the second local associated item information.

[0015] Optionally, the second image association feature information includes: a second overall image association item vector and a second local image association item vector; generating the second image association feature information based on the second associated item information includes: obtaining second item information and a third attribute value corresponding to the overall image association item information; generating the second overall image association item vector based on the second item information and the third attribute value; obtaining fourth attribute information and a fourth attribute value corresponding to the second image local image association item information; and generating the second local image association item vector based on the fourth attribute information and the fourth attribute value.

[0016] Optionally, obtaining the second item category information and all attribute information of the second item corresponding to the item to be processed includes: obtaining the second item category information based on user settings or using a category model; obtaining all attribute information of the second item corresponding to the second item category information; wherein, the all attribute information of the second item includes: all item attributes and all corresponding attribute values.

[0017] Optionally, the recommended attribute and the recommended attribute value are sent to the item terminal; the user's adoption rate information for the recommended attribute and the recommended attribute value is determined, and the user's preference information is obtained based on the adoption rate information, so as to label the training samples based on the preference information.

[0018] According to a third aspect of this disclosure, a training apparatus for an item attribute recommendation model is provided, comprising: a training sample generation module, configured to generate training samples based on a first item image corresponding to the item, item annotation attributes, and item annotation attribute values; an associated item determination module, configured to acquire the training samples and determine first associated item information based on the first item image and the item annotation attributes; an associative feature generation module, configured to generate first image associative feature information based on the first associated item information; a category attribute acquisition module, configured to acquire first item category information and all attribute information of the first item corresponding to the item; and a recommendation model training module, configured to train the item attribute recommendation model using the first item image, the first image associative feature information, the first item category information, the first item all attribute information, the item annotation attributes, and the item annotation attribute values, to obtain the trained item attribute recommendation model.

[0019] Optionally, the first associated item information includes: first overall associated item information and first partial associated item information; the associated item determination module is used to generate a first vector based on the first item image using an overall image association model; obtain the first overall associated item information corresponding to the item in a first item vector library based on the first vector; generate a second vector based on the first item image and the item's labeled attributes using a partial image association model; and obtain the first partial associated item information corresponding to the item in a second item vector library based on the second vector.

[0020] Optionally, the associated item determination module is specifically used to obtain one or more first item vectors that are closest to the first vector in the first item vector library; obtain the first item SKU information corresponding to the first item vector as the first overall associated item information; obtain one or more second item vectors that are closest to the second vector in the second item vector library; and obtain the second item SKU information corresponding to the second item vector as the local associated item information.

[0021] Optionally, the first image association feature information includes: an overall image association item vector and a local image association item vector; the association feature generation module is used to obtain first item information and a first attribute value corresponding to the first overall image association item information; generate the overall image association item vector based on the first item information and the first attribute value; obtain second attribute information and second attribute value corresponding to the local image association item information; and generate the local image association item vector based on the first attribute information and the second attribute value.

[0022] Optionally, the category attribute acquisition module is used to acquire the first item category information based on user settings or using a category model; acquire all attribute information of the first item corresponding to the first item category information; wherein, the first item all attribute information includes: all item attributes and all corresponding attribute values.

[0023] Optionally, the recommendation model training module is used to obtain recommended attributes and corresponding recommended attribute values ​​using the item attribute recommendation model and based on the first item image, the first image association feature information, the first item category information, and all attribute information of the first item; and to train the item attribute recommendation model based on the comparison results between the recommended attributes, the recommended attribute values, the item annotation attributes, and the item annotation attribute values, so as to adjust the item attribute recommendation model.

[0024] According to a fourth aspect of this disclosure, an item recommendation attribute generation apparatus is provided, comprising: a feature acquisition module for acquiring a second item image, an item attribute, and an item attribute value corresponding to an item to be processed; an association determination module for determining second associated item information based on the second item image and the item attribute; an association generation module for generating second image association feature information based on the second associated item information; a category determination module for acquiring second item category information and all attribute information of the second item corresponding to the item to be processed; and an attribute generation module for generating recommended attributes and corresponding attribute values ​​using a trained item attribute recommendation model based on the second item image, the second image association feature information, the second item category information, and the second item attribute information; wherein the item attribute recommendation model is trained using the above-described training method.

[0025] Optionally, the second associated item information includes: second overall associated item information and second partial associated item information; the association determination module is used to generate a third vector based on the second item image using an overall image association model; obtain the second overall associated item information corresponding to the item to be processed from a third item vector library based on the third vector; generate a fourth vector based on the second item image and the attributes of the item to be processed using an partial image association model; and obtain the second partial associated item information corresponding to the item to be processed from a fourth item vector library based on the fourth vector.

[0026] Optionally, the association determination module is specifically used to obtain one or more third item vectors that are closest to the third item vector in the third item vector library; obtain the third item SKU information corresponding to the third item vector as the second overall associated item information; obtain one or more fourth item vectors that are closest to the fourth item vector in the fourth item vector library; and obtain the fourth item SKU information corresponding to the fourth item vector as the second local associated item information.

[0027] Optionally, the second image association feature information includes: a second overall image association item vector and a second local image association item vector; the association generation module is used to obtain second item information and a third attribute value corresponding to the overall image association item information; generate the second overall image association item vector based on the second item information and the third attribute value; obtain fourth attribute information and a fourth attribute value corresponding to the second image local image association item information; and generate the second local image association item vector based on the fourth attribute information and the fourth attribute value.

[0028] Optionally, the category determination module is used to obtain the second item category information based on user settings or by using a category model; and to obtain all attribute information of the second item corresponding to the second item category information; wherein, the all attribute information of the second item includes: all item attributes and all corresponding attribute values.

[0029] Optionally, the attribute sending module is used to send the recommended attribute and the recommended attribute value to the item terminal; the sample update module is used to determine the user's adoption rate information for the recommended attribute and the recommended attribute value, obtain the user's preference information based on the adoption rate information, and use the preference information to label the training samples.

[0030] According to a fifth aspect of this disclosure, a training apparatus for an item attribute recommendation model is provided, comprising: a memory; and a processor coupled to the memory, the processor being configured to perform the method described above based on instructions stored in the memory.

[0031] According to a sixth aspect of this disclosure, an applicator for generating item recommendation attributes is provided, comprising: a memory; and a processor coupled to the memory, the processor being configured to perform the method described above based on instructions stored in the memory.

[0032] According to a seventh aspect of this disclosure, a computer-readable storage medium is provided that stores computer instructions which, when executed by a processor, perform the method described above.

[0033] This disclosure discloses a training method and apparatus for an item attribute recommendation model, an item recommendation attribute generation method, and a storage medium. It obtains personalized attribute recommendations based on the original image of the item and overall and local image association information, improving the accuracy, generalization ability, stability, and item exposure rate of item attribute recommendations, thus enhancing the user experience. It addresses the shortcomings of existing image recognition algorithms in recognizing occluded or difficult-to-identify attributes, as well as their reliance on the sufficiency and accuracy of the platform's existing item database. Furthermore, it reduces the resource consumption for model training and supports the stable expansion of new attributes under a unified model. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1This is a flowchart illustrating an embodiment of a training method for an item attribute recommendation model based on the present disclosure;

[0036] Figure 2 This is a schematic diagram of the process for determining associated item information in one embodiment of the training method for the item attribute recommendation model according to this disclosure;

[0037] Figure 3 This is a schematic diagram of the process for generating associated item vectors in one embodiment of the training method for the item attribute recommendation model according to the present disclosure;

[0038] Figure 4 This is a schematic diagram of the process for generating an overall image-associated item vector in one embodiment of the training method for an item attribute recommendation model according to this disclosure;

[0039] Figure 5 This is a schematic diagram of the process for generating locally associated item vectors in one embodiment of the training method for the item attribute recommendation model according to the present disclosure;

[0040] Figure 6 This is a schematic diagram of the sample labeling process in one embodiment of the training method for the item attribute recommendation model according to this disclosure;

[0041] Figure 7 This is a schematic diagram of the structure of an item attribute recommendation model in one embodiment of the training method for the item attribute recommendation model according to the present disclosure;

[0042] Figure 8 This is a schematic diagram of the process of training an item attribute recommendation model according to one embodiment of the training method of the item attribute recommendation model of this disclosure;

[0043] Figure 9 This is a flowchart illustrating an embodiment of the item recommendation attribute generation method according to the present disclosure;

[0044] Figure 10 This is a schematic diagram illustrating the process of labeling training samples in one embodiment of the item recommendation attribute generation method according to this disclosure;

[0045] Figure 11 A schematic diagram of a module of an embodiment of a training apparatus for an item attribute recommendation model based on the present disclosure;

[0046] Figure 12 This is a schematic diagram of a module of an embodiment of an item recommendation attribute generation apparatus according to the present disclosure;

[0047] Figure 13 A schematic diagram of a module for another embodiment of a training apparatus for an item attribute recommendation model based on the present disclosure;

[0048] Figure 14 This is a schematic diagram of a module of another embodiment of the item recommendation attribute generation apparatus according to the present disclosure. Detailed Implementation

[0049] The present disclosure will now be described more fully with reference to the accompanying drawings, which illustrate exemplary embodiments of the present disclosure. The technical solutions of the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present disclosure, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present disclosure without creative effort are within the scope of protection of the present disclosure.

[0050] The terms "first" and "second" used in the following text are only used to describe the differences between them and have no other special meaning.

[0051] Image recognition is a core branch of artificial intelligence, and item image attribute recognition is a crucial technological foundation in the e-commerce field. Item image attribute recognition technology plays a vital role in merchants' new product launches and inventory item attribute inspections, saving significant manual costs associated with attribute entry and verification.

[0052] The steps for object image attribute recognition in the existing technology are as follows: input an object image through the input terminal, determine the category to which the object image belongs, obtain the attribute to be predicted under the category, and select the attribute value corresponding to the object image from all candidate values ​​under the attribute.

[0053] For example, by inputting a main image of a garment, the system can identify the item as belonging to the Apparel - Men's Wear - Jacket category and obtain the attributes to be predicted, including: collar type, sleeve type, sleeve length, material, applicable age, placket, target audience, style, pattern, color, and popular elements. Taking sleeve length as an example, the system selects the attribute value that best matches the item from candidate attribute values ​​under sleeve length attributes such as long sleeve, short sleeve, sleeveless, three-quarter sleeve, seven-eighths sleeve, and nine-eighths sleeve.

[0054] Figure 1 This is a flowchart illustrating an embodiment of the training method for an item attribute recommendation model according to the present disclosure, as shown below. Figure 1 As shown:

[0055] Step 101: Generate training samples based on the first item image corresponding to the item, the item annotation attributes, and the item annotation attribute values.

[0056] In one embodiment, the user annotates the first item image corresponding to the item to generate training samples. The item can be of various types, such as clothing, appliances, and mobile phones. For clothing, the corresponding item annotation attributes include collar type, sleeve type, sleeve length, and material. For the sleeve length attribute, the corresponding item annotation attribute values ​​include short sleeve, sleeveless, three-quarter sleeve, and seven-eighths sleeve.

[0057] Step 102: Obtain training samples and determine the information of the first associated item based on the first item image and the item annotation attributes.

[0058] Step 103: Generate first image association feature information based on the first associated item information.

[0059] Step 104: Obtain the first item category information and all attribute information of the first item corresponding to the item.

[0060] In one embodiment, the first item's total attribute information includes all item attributes of the first item and all corresponding attribute values. For example, item attributes include sleeve length, collar type, pattern, target audience, etc.; for the sleeve length attribute, attribute values ​​include long sleeve, short sleeve, three-quarter sleeve, etc.; for the collar type attribute, attribute values ​​include: round neck, square neck, stand collar, etc.

[0061] Step 105: Use the first item image, the first image association feature information, the first item category information, all attribute information of the first item, as well as the item annotation attributes and item annotation attribute values ​​to train the item attribute recommendation model to obtain the trained item attribute recommendation model.

[0062] The first associated item information includes the first overall associated item information and the first partial associated item information. Figure 2 This is a schematic diagram illustrating the process of determining associated item information in one embodiment of the training method for the item attribute recommendation model according to this disclosure, as shown below. Figure 2 As shown:

[0063] Step 201: Using the overall image association model and based on the first item image, generate the first vector. The overall image association model can be any model, such as a convolutional neural network (CNN) model.

[0064] Step 202: Based on the first vector, obtain the first overall associated item information corresponding to the item from the first item vector library.

[0065] For example, in the first item vector library, determine one or more first item vectors that are closest to the first vector, and obtain the first item SKU (Stock Keeping Unit) information corresponding to the first item vector, as the first overall associated item information. SKU is the stock keeping unit used to distinguish items; the first item SKU information is the SKU information set for the first item in the inventory, used to distinguish items.

[0066] Step 203: Using an image local association model and based on the first item image and its labeled attributes, generate a second vector. The image local association model can be of various types, such as a convolutional neural network (CNN) model.

[0067] Step 204: Based on the second vector, obtain the first local associated item information corresponding to the item in the second item vector library.

[0068] For example, in the second item vector library, one or more second item vectors that are closest to the second item vector are identified, and the SKU information of the second item corresponding to the second item vector is obtained as the first local associated item information. The second item SKU information is the SKU information set for the second items in the inventory, used to distinguish items.

[0069] In one embodiment, the first image association feature information includes a first overall image association item vector and a first local image association item vector. Figure 3 This is a schematic diagram illustrating the process of generating associated item vectors in one embodiment of the training method for the item attribute recommendation model according to this disclosure, as shown below. Figure 3 As shown:

[0070] Step 301: Obtain the first item information and the first attribute value corresponding to the first overall associated item information.

[0071] In one embodiment, an item information database is constructed, whose data includes item names, attributes, and information such as recent exposure, clicks, follows, add-to-cart, purchases, and reviews. A merchant information database is also constructed, which includes merchant traffic, sales volume, advertising investment, store rating, and the merchant's preference for various attribute values ​​when populating attributes across different categories.

[0072] The first item information can be one or more of the following: SKU information, item name, item number, etc.; the first attribute value includes data such as clicks, follows, sales, and add-to-cart for the first item, as well as the merchant's preference for the above attribute data.

[0073] Step 302: Based on the first item information and the first attribute value, generate the overall associated item vector of the first image. Various existing methods can be used to generate the overall associated item vector of the first image.

[0074] Step 303: Obtain the second attribute information and second attribute value corresponding to the partial associated item information of the image.

[0075] In one embodiment, taking a T-shirt as an example, the second attribute includes sleeve length, collar type, pattern, target audience, etc.; for the sleeve length attribute, the second attribute value includes long sleeve, short sleeve, three-quarter sleeve, etc.; for the collar type attribute, the second attribute value includes: crew neck, square neck, stand collar, etc.

[0076] Step 304: Based on the first attribute information and the second attribute value, generate a first local associated item vector. Various existing methods can be used to generate the first local associated item vector.

[0077] Figure 4 This is a schematic diagram illustrating the process of generating an overall image-associated item vector in one embodiment of the training method for the item attribute recommendation model according to this disclosure, as shown below. Figure 4 As shown:

[0078] Step 401: Enter the main image of the item. The main image is the image of the first item.

[0079] Step 402: Construct the ability to embed the main image of an item into a finite-dimensional vector space.

[0080] Step 403: Retrieve the K vectors with the highest similarity to the embedded main image vectors and their corresponding main images from the existing item main image vector library.

[0081] Step 404: Return the SKU (stock keeping unit) ID associated with these K main charts.

[0082] Figure 5 This is a schematic diagram illustrating the process of generating locally associated item vectors in one embodiment of the training method for the item attribute recommendation model according to this disclosure, as shown below. Figure 5 As shown:

[0083] Step 501: Enter the main image of the item and attribute A.

[0084] Step 502: Construct the ability to embed local information of an object into a finite-dimensional vector space.

[0085] Step 503: Construct the ability to embed the item master image and attributes into a finite-dimensional vector space.

[0086] Step 504: Retrieve the K main images with the highest similarity to attribute A of the input main image from the existing item main image vector library.

[0087] Step 505: Return the sku_id associated with these K main graphs.

[0088] In one embodiment, an item information database is constructed, whose data includes item names, attributes, and information such as recent exposure, clicks, follows, add-to-cart, purchases, and reviews. A merchant information database is also constructed, which includes merchant traffic, sales volume, advertising investment, store rating, and the merchant's preference for each attribute value when populating attributes across different categories. Finally, a category information database is constructed, whose data includes the attribute set for each category and candidate attribute values ​​for each attribute.

[0089] An attribute value information database is constructed. This database contains data including the number of items containing a given attribute value across the entire site, the number of times users searched for that attribute value, the number of times the item with that attribute value was recommended by the system, the number of times it was adopted, the adoption rate (the percentage of adoptions for that attribute value across all attributes), and a sliding window processing method based on time windows. Relevant information can be obtained by searching the item information database, merchant information database, category information database, and attribute value information database.

[0090] The ResNet50 model T is used to embed the main image p of an item into a vector, denoted as T(p). All existing main images of items are also embedded into the vector space and stored as an image vector library (the first item vector library) V1. Given an item main image p, model T maps p to vector T(p). Vector retrieval is then used to search the vector library V1 for the K vectors with the largest cosine value of T(p), and the set of sku_ids associated with the images corresponding to these K vectors is returned, denoted as S1.

[0091] Using deep learning technology, a local image association model L is constructed that embeds the main image of an item (first item image) and its attributes into a finite-dimensional vector space. Through model L, the main image of an item (first item image) p and its attributes A can be embedded into a vector, denoted as L(p, A). Based on this model, existing main images of items and their attributes are embedded into the vector space and stored as a vector library (second item vector library) V2.

[0092] Given an item image p, map the item image p and its attribute A to a vector L(p, A). Using vector retrieval, search the vector library V2 for the K vectors with the largest cosine value of L(p, A), and return the set of sku_ids associated with the images corresponding to the K second-ranked vectors, denoted as . .

[0093] The system retrieves item information and attribute values ​​corresponding to the SKUs in set S1 from both the item information database and the merchant information database. Item information includes the SKU, item name, and item number; attribute values ​​include data such as clicks, follows, sales, and add-to-cart activity, as well as the merchant's preference for each attribute. Using existing feature engineering techniques, a feature vector xi (the first overall associated item information) is generated for each SKU i and attribute value a, and K vectors are concatenated into x(a) = (x1, x2, ..., xk). This process is repeated for all attribute values ​​{a1, a2, ..., am, b1, b2, ...}.

[0094] For attribute A, retrieve the set from the item information database. The SKU contains item information, attribute values, and other information. Using feature engineering techniques, a feature vector yi (locally associated item information) is generated corresponding to (SKU i, attribute value a), and K vectors are concatenated into y = (x1, x2, ..., xK). This process is repeated for all attributes {A, B, C, D, ...} and all attribute values ​​{a1, a2, ..., am, b1, b2, ...} under each attribute.

[0095] In one embodiment, obtaining the first item category information and all attribute information of the first item corresponding to the item can be achieved using multiple methods. The first item category information can be obtained based on user settings or using a category model. The first item's all attribute information corresponding to the first item category information can then be obtained; this information includes all item attributes and their corresponding attribute values.

[0096] For example, all item attributes (first item category information) can be obtained through category IDs filled in by merchants or predicted by category models, denoted as {A, B, C, D, ...}. The category model can be a CNN model, etc. The attribute values ​​under each attribute are obtained as {a1, a2, ..., am, b1, b2, ...} (all attribute information of the first item), and the feature z(a) of each attribute value. Taking a T-shirt as an example, {A, B, C, D, ...} can be {A: sleeve length, B: collar type, C: pattern, D: target audience, ...}. {a1, a2, ..., am, b1, b2, ...} can be {a1: long sleeve, a2: short sleeve, a3: three-quarter sleeve, a4: seven-eighth sleeve, ..., b1: crew neck, b2: square neck, b3: stand collar, ...,}.

[0097] In one embodiment, such as Figure 6As shown, the training samples come from two parts: the first part consists of samples annotated by industry experts, which are of high quality, high cost, and few in number; the second part consists of attributes filled in after final adoption or modification by merchants, which have relatively lower annotation quality than the first part, are cheaper to obtain, more numerous, and have a personalized effect. To improve sample quality, an active learning approach is adopted. During model training, samples whose predicted values ​​do not match the actual values ​​are manually annotated again to generate a high-quality secondary sample library, thereby improving accuracy.

[0098] In one embodiment, the item attribute recommendation model includes an item image model branch corresponding to the first item image, an image association model branch corresponding to the first image association feature information, an item category model branch corresponding to the first item category information, an item attribute model branch corresponding to all attribute information of the first item, an attribute analysis model block, and an attribute output branch.

[0099] The vectors output from the item image model branch, image association model branch, item category model branch, and item attribute model branch are input into the attribute analysis model block. The data output from the attribute analysis model block is input into the attribute output branch. The attribute output branch outputs the attribute value with the highest score and the corresponding attribute.

[0100] The item image model branch includes a ResNet-S1 network in the first layer and a ResNet (residual neural network) network in the second layer; a ResNet network is built for each attribute. The item category model branch includes an embedding network at the bottom layer and a ConvNet network at the top layer. The item attribute model branch includes an embedding network at the bottom layer and a ConvNet (conventional neural network) network at the top layer. The image association model branch includes a ConvNet network at the bottom layer and an FCNet network at the top layer. The attribute analysis model block includes an FCNet (fully connected neural network) sub-network; an FCNet sub-network is built for each attribute. The attribute output branch includes a softmax network.

[0101] like Figure 7 As shown, the input to the item attribute recommendation model includes the item's main image with uniform pixel values ​​(e.g., 256*256), category ID, attribute value ID, attribute value features, and features generated by image association, each corresponding to one of four scores. Each branch outputs a (128-dimensional) vector.

[0102] In the image input branch, i.e. the item image model branch, the underlying ResNet-S1 network is shared. To increase model compatibility and ensure the stability of old attributes during training and updates after adding new attributes, the model adopts a Tree-Net network structure in the L1 network block, and a private ResNet network is built for each attribute.

[0103] In the category branch, i.e., the item category model branch, an embedding network is first constructed to embed the category ID into a (128-dimensional) vector space, and a ConvNet network is built on top of it. In the attribute value branch, i.e., the item attribute model branch, an embedding network is first constructed to embed the attribute value ID into a (128-dimensional) vector space, and a ConvNet network is built on top of it, taking the attribute value embedding vector and the basic features of the attribute value as input.

[0104] In the image association feature branch, i.e., the image association model branch, the bottom layer uses ConvNet, and the upper layer uses FCNet. The hidden layers output from the four branches are fed into the upper-layer model. Similar to the L1 network block, the L2 network block also builds an FCNet sub-network for each attribute, corresponding one-to-one with the ResNet sub-network of L1. When a new attribute is added to the model, a new sub-network is added to both L1 and L2. Training for this attribute ensures that the sub-networks corresponding to other attributes do not participate in the training, thus guaranteeing the stability of the predictions for other attributes.

[0105] Figure 8 This is a schematic diagram illustrating the process of training an item attribute recommendation model according to one embodiment of the training method for the item attribute recommendation model of this disclosure, as shown below. Figure 8 As shown:

[0106] Step 801: Using the item attribute recommendation model and based on the first item image, the first image association feature information, the first item category information, and all attribute information of the first item, obtain the recommended attributes and the corresponding recommended attribute values.

[0107] Step 802: Based on the comparison results between the recommended attributes and their values, and the item label attributes and their values, the item attribute recommendation model is trained for further adjustments. Various existing training and parameter adjustment methods can be employed.

[0108] Figure 9 This is a flowchart illustrating an embodiment of the item recommendation attribute generation method according to the present disclosure, as follows: Figure 9 As shown:

[0109] Step 901: Obtain the second item image, the item attributes, and the item attribute values ​​corresponding to the item to be processed.

[0110] Step 902: Determine the information of the second associated item based on the second item image and the attributes of the item to be processed.

[0111] Step 903: Generate second image association feature information based on the second associated item information.

[0112] Step 904: Obtain the second item category information and all attribute information of the second item corresponding to the item to be processed.

[0113] In one embodiment, the complete attribute information of the second item includes all the item attributes of the second item and all the corresponding attribute values. For example, the item attributes include sleeve length, collar type, pattern, target audience, etc.; for the sleeve length attribute, the attribute values ​​include long sleeve, short sleeve, three-quarter sleeve, etc.; for the collar type attribute, the attribute values ​​include: round neck, square neck, stand collar, etc.

[0114] Step 905: Using the trained item attribute recommendation model and based on the second item image, the second image association feature information, the second item category information, and all attribute information of the second item, generate recommended attributes and corresponding recommended attribute values; wherein, the item attribute recommendation model is trained by the training method in any of the above embodiments.

[0115] For example, input the image p, category id, attribute value id, attribute value feature z(a), and image association feature (x, y) into a trained program such as... Figure 7 The item attribute recommendation model shown above processes each attribute using a softmax network. It then outputs the score for each attribute (e.g., sleeve length) for all attribute values ​​(long sleeve, short sleeve, sleeveless, three-quarter sleeve, seven-eighths sleeve, nine-eighths sleeve), and finally outputs the attribute value with the highest score.

[0116] The second associated item information includes second overall associated item information and second partial associated item information. Using an image-based overall association model and based on the second item image, a third vector is generated. Based on this third vector, the second overall associated item information corresponding to the item to be processed is retrieved from a third item vector library. For example, one or more third item vectors closest to the third vector are retrieved from the third item vector library, and the corresponding third item SKU information is obtained as the second overall associated item information.

[0117] Using an image-based local association model and based on the second item image and the attributes of the item to be processed, a fourth vector is generated. Based on this fourth vector, the corresponding second locally associated item information is retrieved from a fourth item vector library. For example, one or more fourth item vectors closest to the fourth vector are retrieved from the fourth item vector library, and the corresponding fourth item SKU information is obtained as the second locally associated item information.

[0118] The second image associative feature information includes the overall associative item vector and the second local associative item vector. Third item information and third attribute values ​​corresponding to the overall associative item information are obtained. Based on the second item information and third attribute values, the overall associative item vector for the second image is generated. The second item information includes the item's SKU information, item name, item number, etc., and the second attribute value information includes data such as item clicks, follows, sales, and add-to-cart activity, as well as information such as the merchant's preference for each attribute data. Fourth attribute information and fourth attribute values ​​corresponding to the local associative item information of the second image are obtained. Based on the fourth attribute information and fourth attribute values, the second local associative item vector is generated.

[0119] Based on user settings or by using a category model, obtain the category information of the second item and obtain all attribute information of the second item corresponding to the category information of the second item; wherein, the all attribute information of the second item includes: all item attributes and all corresponding attribute values.

[0120] Figure 10 This is a schematic diagram illustrating the process of labeling training samples in one embodiment of the item recommendation attribute generation method according to this disclosure, as shown below. Figure 10 As shown:

[0121] Step 1001: Send the recommended attributes and recommended attribute values ​​to the item terminal.

[0122] Step 1002: Determine the user's adoption rate information for the recommended attributes and their values, obtain the user's preference information based on the adoption rate information, and use this preference information to label the training samples.

[0123] For example, the predicted results output by the item attribute recommendation model {such as sleeve length - long sleeve, collar type - crew neck, pattern - anime / cartoon, etc.} are recommended to the item terminal where merchants publish items. The item terminal can be a PC, mobile phone, or other similar devices. After seeing the displayed results, merchants can choose to adopt the recommended attribute values ​​or modify them. Merchant behavior data is recorded to calculate the accuracy and adoption rate of the recommended attributes. Based on the adoption rate, merchant attribute value preferences are obtained, user preference information is calculated, and training samples are labeled based on the preference information.

[0124] In one embodiment, such as Figure 11 As shown, this disclosure provides a training device 110 for an item attribute recommendation model, including a training sample generation module 1101, an associated item determination module 1102, an associative feature generation module 1103, a category attribute acquisition module 1104, and a recommendation model training module 1105. The training sample generation module 1101 generates training samples based on a first item image corresponding to the item, item annotation attributes, and item annotation attribute values. The associated item determination module 1102 acquires the training samples and determines first associated item information based on the first item image and item annotation attributes.

[0125] The associative feature generation module 1103 generates first image associative feature information based on the first associated item information. The category attribute acquisition module 1104 acquires the first item category information and all attribute information of the first item corresponding to the item. The recommendation model training module 1105 uses the first item image, the first image associative feature information, the first item category information, the first item all attribute information, as well as the item annotation attributes and item annotation attribute values ​​to train the item attribute recommendation model to obtain the trained item attribute recommendation model.

[0126] In one embodiment, the first associated item information includes first overall associated item information and first partial associated item information. The associated item determination module 1102 uses an overall image association model and, based on the first item image, generates a first vector, and then retrieves the first overall associated item information corresponding to the item from the first item vector library based on the first vector. The associated item determination module 1102 uses an partial image association model and, based on the first item image and item annotation attributes, generates a second vector, and then retrieves the partial associated item information corresponding to the item from the second item vector library based on the second vector.

[0127] The associated item determination module 1102 retrieves one or more first item vectors that are closest to the first vector from the first item vector library, and retrieves the first item SKU information corresponding to the first item vector as the first overall associated item information. The associated item determination module 1102 retrieves one or more second item vectors that are closest to the second vector from the second item vector library, and retrieves the second item SKU information corresponding to the second item vector as the local associated item information.

[0128] The first image association feature information includes an overall image association item vector and a local image association item vector. The association feature generation module 1103 acquires first item information and a first attribute value corresponding to the first overall image association item information, and generates an overall image association item vector based on the first item information and the first attribute value. The association feature generation module 1103 acquires second attribute information and second attribute values ​​corresponding to the local image association item information, and generates a local image association item vector based on the first attribute information and the second attribute value.

[0129] The category attribute acquisition module 1104 acquires the first item category information based on user settings or by using a category model, and acquires all attribute information of the first item corresponding to the first item category information. The first item's all attribute information includes all item attributes and their corresponding all attribute values.

[0130] The recommendation model training module 1105 uses an item attribute recommendation model and, based on the first item image, first image association feature information, first item category information, and all attribute information of the first item, obtains recommended attributes and corresponding recommended attribute values. The recommendation model training module 1105 trains the item attribute recommendation model based on the comparison results between the recommended attributes and recommended attribute values ​​and the item labeled attributes and item labeled attribute values, in order to adjust and process the item attribute recommendation model.

[0131] In one embodiment, such as Figure 12 As shown, this disclosure provides an item recommendation attribute generation device 120, including a feature acquisition module 1201, an association determination module 1202, an association generation module 1203, a category determination module 1204, an attribute generation module 1205, an attribute sending module 1206, and a sample update module 1207.

[0132] The feature acquisition module 1201 acquires the second item image, the item attributes, and the item attribute values ​​corresponding to the item to be processed. The association determination module 1202 determines the second associated item information based on the second item image and the item attributes. The association generation module 1203 generates second image association feature information based on the second associated item information.

[0133] The category determination module 1204 acquires the second item category information and all attribute information of the second item corresponding to the item to be processed. The attribute generation module 1205 uses a trained item attribute recommendation model and generates recommended attributes and corresponding attribute values ​​based on the second item image, second image association feature information, second item category information, and all attribute information of the second item. The item attribute recommendation model is trained using the training method in any of the above embodiments.

[0134] In one embodiment, the second associated item information includes second overall associated item information and second partial associated item information. The association determination module 1202 uses an overall image association model and, based on the second item image, generates a third vector, and uses this third vector to retrieve the second overall associated item information corresponding to the item to be processed from a third item vector library. The association determination module 1202 uses an partial image association model and, based on the second item image and the attributes of the item to be processed, generates a fourth vector, and uses this fourth vector to retrieve the second partial associated item information corresponding to the item to be processed from a fourth item vector library.

[0135] The association determination module 1202 retrieves one or more third item vectors that are closest to the third item vector from the third item vector library, and retrieves the third item SKU information corresponding to the third item vector as the second overall associated item information. The association determination module 1202 also retrieves one or more fourth item vectors that are closest to the fourth item vector from the fourth item vector library, and retrieves the fourth item SKU information corresponding to the fourth item vector as the second local associated item information.

[0136] The second image association feature information includes the overall associated item vector and the second local associated item vector. The association generation module 1203 acquires the second item information and the third attribute value corresponding to the overall associated item information, and generates the overall associated item vector of the second image based on the second item information and the third attribute value. The association generation module 1203 acquires the fourth attribute information and the fourth attribute value corresponding to the local associated item information of the second image, and generates the second local associated item vector based on the fourth attribute information and the fourth attribute value.

[0137] The category determination module 1204 obtains the second item category information based on user settings or by using a category model, and obtains all attribute information of the second item corresponding to the second item category information. The all attribute information of the second item includes all item attributes and all corresponding attribute values.

[0138] The attribute sending module 1206 sends the recommended attributes and recommended attribute values ​​to the item terminal. The sample update module 1207 determines the user's adoption rate information for the recommended attributes and recommended attribute values, obtains the user's preference information based on the adoption rate information, and uses the preference information to label the training samples.

[0139] Figure 13 This is a schematic diagram of a module for another embodiment of a training apparatus for an item attribute recommendation model based on this disclosure. (See diagram below.) Figure 13 As shown, the device may include a memory 131, a processor 132, a communication interface 133, and a bus 134. The memory 131 is used to store instructions, and the processor 132 is coupled to the memory 131. The processor 132 is configured to execute the training method of the item attribute recommendation model described above based on the instructions stored in the memory 131.

[0140] The memory 131 can be high-speed RAM, non-volatile memory, or a memory array. The memory 131 may also be divided into blocks, and these blocks can be combined into virtual volumes according to certain rules. The processor 132 can be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the training method of the item attribute recommendation model disclosed herein.

[0141] Figure 14 This is a schematic diagram of a module of another embodiment of the item recommendation attribute generation apparatus according to this disclosure. Figure 14 As shown, the device may include a memory 141, a processor 142, a communication interface 143, and a bus 144. The memory 141 is used to store instructions, and the processor 142 is coupled to the memory 141. The processor 142 is configured to execute the above-described item recommendation attribute generation method based on the instructions stored in the memory 141.

[0142] The memory 141 can be a high-speed RAM, non-volatile memory, or a memory array. The memory 141 may also be divided into blocks, and these blocks can be combined into virtual volumes according to certain rules. The processor 142 can be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the article recommendation attribute generation method of this disclosure.

[0143] In one embodiment, this disclosure provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the training method for the item attribute recommendation model as described in any of the preceding embodiments.

[0144] The training of the item attribute recommendation model, the method and apparatus for generating item recommendation attributes, and the storage medium in the above embodiments obtain personalized recommendations of attributes based on the original image of the item and overall and local image association information, which improves the accuracy, generalization ability, stability, and exposure rate of item attribute recommendations, thereby improving the user experience. It solves the problems of poor recognition effect of existing image recognition algorithms on occluded attributes or difficult-to-recognize attributes, as well as the dependence on the sufficiency of the platform's existing item database and the accuracy of the item information in the existing item database. It reduces the resource consumption of model training and supports the stable expansion of new attributes under a unified model.

[0145] The methods and systems of this disclosure may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the methods is for illustrative purposes only, and the steps of the methods of this disclosure are not limited to the order specifically described above unless otherwise specifically stated. Furthermore, in some embodiments, this disclosure may also be implemented as a program recorded on a recording medium, the program including machine-readable instructions for implementing the methods according to this disclosure. Thus, this disclosure also covers recording media storing programs for performing the methods according to this disclosure.

[0146] The description in this disclosure is provided for illustrative and descriptive purposes only and is not intended to be exhaustive or to limit the disclosure to its forms. Many modifications and variations will be apparent to those skilled in the art. The embodiments were chosen and described in order to better illustrate the principles and practical application of this disclosure and to enable those skilled in the art to understand this disclosure and to design various embodiments with various modifications suitable for a particular purpose.

Claims

1. A training method for an item attribute recommendation model, comprising: Training samples are generated based on the first item image corresponding to the item, the item's labeled attributes, and the item's labeled attribute values. The training samples are obtained, and the first associated item information is determined based on the first item image and the item annotation attributes. The first associated item information includes: first overall associated item information and first partial associated item information. First image association feature information is generated based on the first associated item information; Obtain the first item category information and all attribute information of the first item corresponding to the item; The first item image, the first image association feature information, the first item category information, all attribute information of the first item, the item annotation attributes and the item annotation attribute values ​​are used to train the item attribute recommendation model to obtain the trained item attribute recommendation model.

2. The method as described in claim 1, wherein determining the first associated item information based on the first item image and the item annotation attributes includes: The first vector is generated using the overall image association model and based on the first item image; Based on the first vector, obtain the first overall associated item information corresponding to the item from the first item vector library; A second vector is generated using an image local association model and based on the first item image and the item's labeled attributes; Based on the second vector, obtain the first local associated item information corresponding to the item from the second item vector library.

3. The method as described in claim 2, wherein obtaining the first overall associated item information corresponding to the item in the first item vector library based on the first vector includes: Retrieve one or more first item vectors that are closest to the first vector from the first item vector library; Obtain the first item SKU information corresponding to the first item vector, and use it as the first overall associated item information; The step of obtaining the first local associated item information corresponding to the item in the second item vector library based on the second vector includes: Retrieve one or more second item vectors that are closest to the second vector from the second item vector library; Obtain the second item SKU information corresponding to the second item vector, and use it as the first local associated item information.

4. The method as described in claim 2, wherein the first image association feature information includes: The overall associative object vector and the first local associative object vector of the first image; The generation of the first image association feature information based on the first associated item information includes: Obtain the first item information and the first attribute value corresponding to the first overall associated item information; Based on the first item information and the first attribute value, generate the overall associated item vector of the first image; Obtain the second attribute information and second attribute value corresponding to the first local associated item information; Based on the second attribute information and the second attribute value, the first local associated item vector is generated.

5. The method as described in claim 1, wherein obtaining the first item category information and all attribute information of the first item corresponding to the item includes: The first item category information is obtained based on user settings or by using a category model; Obtain all attribute information of the first item corresponding to the first item category information; wherein, all attribute information of the first item includes: all item attributes and all corresponding attribute values.

6. The method as described in claim 1, wherein training the item attribute recommendation model using the first item image, the first image association feature information, the first item category information, all attribute information of the first item, and the item annotation attributes and the item annotation attribute values ​​includes: Using the item attribute recommendation model and based on the first item image, the first image association feature information, the first item category information, and all attribute information of the first item, we obtain recommended attributes and corresponding recommended attribute values. Based on the comparison results of the recommended attributes, the recommended attribute values, the item label attributes, and the item label attribute values, the item attribute recommendation model is trained and then adjusted.

7. The method according to any one of claims 1 to 6, wherein, The item attribute recommendation model includes: an item image model branch corresponding to the first item image, an image association model branch corresponding to the first image association feature information, an item category model branch corresponding to the first item category information, an item attribute model branch corresponding to all attribute information of the first item, an attribute analysis model block, and an attribute output branch. Specifically, the vector data output by the item image model branch, the image association model branch, the item category model branch, and the item attribute model branch are obtained respectively; Input the vector data into the attribute analysis model block; The data output by the attribute analysis model block is input into the attribute output branch, and the attribute output branch outputs the attribute value with the highest score and the corresponding attribute.

8. The method of claim 7, wherein, The item image model branch includes: a ResNet-S1 network in the first layer and a ResNet network in the second layer; wherein, a ResNet network is constructed for each attribute. The item category model branch includes: a bottom-level embedding network and an upper-level ConvNet network; The item attribute model branch includes: a bottom-level embedding network and an upper-level ConvNet network; The image association model branches include: the ConvNet network at the bottom layer and the FCNet network at the top layer; The attribute analysis model block includes: an FCNet sub-network; wherein, an FCNet sub-network is constructed for each attribute; The attribute output branches include: a softmax network.

9. A method for generating recommended attributes for items, comprising: Get the image of the second item corresponding to the item to be processed, the attributes of the item to be processed, and the attribute values ​​of the item to be processed; The second associated item information is determined based on the second item image and the attributes of the item to be processed; Generate second image association feature information based on the second associated item information; Obtain the second item category information and all attribute information of the second item corresponding to the item to be processed; The trained item attribute recommendation model is used to generate recommended attributes and corresponding recommended attribute values ​​based on the second item image, the second image association feature information, the second item category information, and all attribute information of the second item. The item attribute recommendation model is trained using the training method described in any one of claims 1 to 8.

10. The method of claim 9, wherein the second associated item information includes: Second overall associated item information and second partial associated item information; The step of determining the second associated item information based on the second item image and the attributes of the item to be processed includes: The third vector is generated using the overall image association model and based on the second item image; Based on the third vector, obtain the second overall associated item information corresponding to the item to be processed from the third item vector library; A fourth vector is generated using a local image association model and based on the second item image and the attributes of the item to be processed. Based on the fourth vector, obtain the second local associated item information corresponding to the item to be processed from the fourth item vector library.

11. The method of claim 10, wherein obtaining the second overall associated item information corresponding to the item to be processed from the third item vector library based on the third vector includes: Obtain one or more third item vectors that are closest to the third vector from the third item vector library; Obtain the third item SKU information corresponding to the third item vector, and use it as the second overall associated item information; The step of obtaining the second local associated item information corresponding to the item to be processed from the fourth item vector library based on the fourth vector includes: Obtain one or more fourth item vectors that are closest to the fourth vector from the fourth item vector library; Obtain the fourth item SKU information corresponding to the fourth item vector, and use it as the second local associated item information.

12. The method of claim 10, wherein the second image association feature information includes: The overall associative object vector and the second local associative object vector of the second image; The generation of second image association feature information based on the second associated item information includes: Obtain the second item information and the third attribute value corresponding to the second overall associated item information; Based on the second item information and the third attribute value, generate the overall associated item vector of the second image; Obtain the fourth attribute information and fourth attribute value corresponding to the partial associated item information of the second image; Based on the fourth attribute information and the fourth attribute value, the second local associated item vector is generated.

13. The method as described in claim 9, wherein obtaining the second item category information and all attribute information of the second item corresponding to the item to be processed includes: The second item category information is obtained based on user settings or by using a category model; Obtain all attribute information of the second item corresponding to the second item category information; wherein, all attribute information of the second item includes: all item attributes and all corresponding attribute values.

14. The method of claim 9, further comprising: Send the recommended attributes and their values ​​to the product terminal; Determine the user's adoption rate information for the recommended attribute and the value of the recommended attribute, obtain the user's preference information based on the adoption rate information, and use the preference information to label the training samples.

15. A training device for an item attribute recommendation model, comprising: The training sample generation module is used to generate training samples based on the first item image corresponding to the item, the item annotation attributes, and the item annotation attribute values; The associated item determination module is used to acquire the training samples and determine the first associated item information based on the first item image and the item annotation attributes; wherein, the first associated item information includes: first overall associated item information and first partial associated item information; The associative feature generation module is used to generate first image associative feature information based on the first associated item information; The category attribute acquisition module is used to acquire the first item category information and all attribute information of the first item corresponding to the item. The recommendation model training module is used to train the item attribute recommendation model using the first item image, the first image association feature information, the first item category information, all attribute information of the first item, the item annotation attributes and the item annotation attribute values, so as to obtain the trained item attribute recommendation model.

16. The apparatus of claim 15, wherein, The associated item determination module is used to generate a first vector based on the first item image using an overall image association model; obtain first overall associated item information corresponding to the item from a first item vector library based on the first vector; generate a second vector based on the first item image and the item's labeled attributes using a partial image association model; and obtain first partial associated item information corresponding to the item from a second item vector library based on the second vector.

17. The apparatus of claim 16, wherein, The associated item determination module is specifically used to: obtain one or more first item vectors that are closest to the first vector in the first item vector library; obtain the first item SKU information corresponding to the first item vector as the first overall associated item information; obtain one or more second item vectors that are closest to the second vector in the second item vector library; and obtain the second item SKU information corresponding to the second item vector as the first local associated item information.

18. The apparatus of claim 16, wherein, The first image association feature information includes: overall image association item vector and local image association item vector; The associative feature generation module is used to obtain first item information and first attribute value corresponding to the first overall associative item information; generate the overall image associative item vector based on the first item information and the first attribute value; obtain second attribute information and second attribute value corresponding to the first local associative item information; and generate the local associative item vector based on the second attribute information and the second attribute value.

19. The apparatus of claim 15, wherein, The category attribute acquisition module is used to acquire the first item category information based on user settings or by using a category model. Obtain all attribute information of the first item corresponding to the first item category information; wherein, all attribute information of the first item includes: all item attributes and all corresponding attribute values.

20. The apparatus of claim 15, wherein, The recommendation model training module is used to obtain the recommended attributes and corresponding recommended attribute values ​​using the item attribute recommendation model and based on the first item image, the first image association feature information, the first item category information, and all attribute information of the first item. Based on the comparison results of the recommended attributes, the recommended attribute values, the item label attributes, and the item label attribute values, the item attribute recommendation model is trained and then adjusted.

21. An item recommendation attribute generation device, comprising: The feature acquisition module is used to acquire the second item image, the item attributes, and the item attribute values ​​corresponding to the item to be processed. The association determination module is used to determine the second associated item information based on the second item image and the attributes of the item to be processed; The association generation module is used to generate second image association feature information based on the second associated item information; The category determination module is used to obtain the second item category information and all attribute information of the second item corresponding to the item to be processed; The attribute generation module is used to generate recommended attributes and corresponding attribute values ​​using a trained item attribute recommendation model and based on the second item image, the second image association feature information, the second item category information, and all attribute information of the second item. The item attribute recommendation model is trained using the training method described in any one of claims 1 to 8.

22. The apparatus of claim 21, wherein the second associated item information includes: Second overall associated item information and second partial associated item information; The association determination module is used to generate a third vector based on the second item image using an overall image association model; obtain second overall associated item information corresponding to the item to be processed from a third item vector library based on the third vector; generate a fourth vector based on the second item image and the attributes of the item to be processed using an image local association model; and obtain second local associated item information corresponding to the item to be processed from a fourth item vector library based on the fourth vector.

23. The apparatus of claim 22, wherein, The association determination module is specifically used to obtain one or more third item vectors that are closest to the third item vector in the third item vector library; obtain the third item SKU information corresponding to the third item vector as the second overall associated item information; obtain one or more fourth item vectors that are closest to the fourth item vector in the fourth item vector library; and obtain the fourth item SKU information corresponding to the fourth item vector as the second local associated item information.

24. The apparatus of claim 22, wherein the second image association feature information includes: The overall associative object vector and the second local associative object vector of the second image; The association generation module is used to obtain second item information and third attribute values ​​corresponding to the overall associated item information; generate an overall associated item vector of the second image based on the second item information and the third attribute values; obtain fourth attribute information and fourth attribute values ​​corresponding to the local associated item information of the second image; and generate a second local associated item vector based on the fourth attribute information and the fourth attribute values.

25. The apparatus of claim 21, wherein, The category determination module is used to obtain the second item category information based on user settings or by using a category model; Obtain all attribute information of the second item corresponding to the second item category information; wherein, all attribute information of the second item includes: all item attributes and all corresponding attribute values.

26. The apparatus of claim 21, further comprising: An attribute sending module is used to send the recommended attribute and the recommended attribute value to the product terminal; The sample update module is used to determine the user's adoption rate information for the recommended attribute and the recommended attribute value, obtain the user's preference information based on the adoption rate information, and use the preference information to label the training samples.

27. A training device for an item attribute recommendation model, comprising: Memory; And a processor coupled to the memory, the processor being configured to perform the method as described in any one of claims 1 to 8 based on instructions stored in the memory.

28. An item recommendation attribute generation device, comprising: Memory; and a processor coupled to the memory, the processor being configured to perform the method as described in any one of claims 9 to 14 based on instructions stored in the memory.

29. A computer-readable storage medium storing computer instructions that, when executed by a processor, implement the method as described in any one of claims 1 to 14.

Images