Data processing method and related product

Abstract

The invention discloses a data processing method and a related product. The method comprises the following steps: acquiring a sample object and article pair containing sample object data and sample article data corresponding to a sample object; performing embedding representation processing on the sample object data based on the object tower model to obtain a sample object embedding vector, performing embedding representation processing on the sample object data based on the object tower model to obtain a sample object embedding vector, and determining object object similarity loss according to the sample object embedding vector and the sample object embedding vector; performing object type discrimination processing on the sample object embedding vector based on an object discriminator to obtain a prediction probability that the sample object belongs to a new business object type, and generating an object tower generation loss according to the prediction probability; performing parameter adjustment processing on the object tower model according to the object item similarity loss and the object tower generation loss to obtain a target object tower model; the target object tower model is used for generating a business object embedding vector corresponding to the business object.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of recommendation systems, and in particular to a data processing method and related products. BACKGROUND

[0002] In the field of digital services such as e-commerce and content distribution, recommendation systems are the core technology for improving user experience and business conversion efficiency. The core logic of the current mainstream recommendation system is: by collecting user behavior data (such as clicks, collections, purchases, etc.) to extract user preference features, combining item attributes (such as category, label, price, etc.) to build an association model, and using collaborative filtering, deep learning, and other algorithms to achieve accurate recommendations. The performance of the recommendation system is highly dependent on sufficient accumulation of user data, because the richer the data dimension and the longer the time span, the more accurate the model's portrayal of user preferences, and the stronger the relevance and effectiveness of the recommendation results. However, new users have extremely scarce effective behavior data due to first-time access or extremely low interaction frequency, resulting in incomplete user profile core dimensions (such as interest inclination, consumption preference, etc.). This data deficiency makes it difficult for the recommendation model to capture the potential preferences of new users and establish effective associations between users and items, ultimately resulting in low recommendation accuracy. SUMMARY

[0003] The present application provides a data processing method and related products, wherein the related products include a data processing device, an electronic device, a computer readable storage medium, and a computer program product.

[0004] In a first aspect, a data processing method is provided, comprising:

[0005] Obtaining a sample object-item pair; the sample object-item pair contains sample object data corresponding to a sample object and sample item data;

[0006] Performing embedding representation processing on the sample object data based on an object tower model to obtain a sample object embedding vector, performing embedding representation processing on the sample item data based on an item tower model to obtain a sample item embedding vector, and determining an object-item similarity loss according to the sample object embedding vector and the sample item embedding vector;

[0007] Performing object type discrimination processing on the sample object embedding vector based on an object discriminator to obtain a prediction probability that the sample object belongs to a new business object type, and generating an object tower generation loss according to the prediction probability;

[0008] The object tower model is parameter-tuned based on the object item similarity loss and the object tower generation loss to obtain the target object tower model; the target object tower model is used to generate the business object embedding vector corresponding to the business object; the business object includes new business objects and existing business objects; the feature distribution pattern of the business object embedding vector corresponding to the new business object is similar to the feature distribution pattern of the business object embedding vector corresponding to the existing business object.

[0009] In any embodiment of this application, the sample object data includes N sample object attribute data; where N is a positive integer;

[0010] The embedding representation processing of the sample object data based on the object pyramid model to obtain the sample object embedding vector includes:

[0011] Iterate through the attribute data of the N sample objects to obtain the attribute data of the i-th sample object; i is a positive integer less than or equal to N;

[0012] If the attribute data of the i-th sample object is numerical data, then the attribute data of the i-th sample object is subjected to numerical standardization to obtain a standardized value, and the standardized value is determined as the preprocessed feature sub-data corresponding to the attribute data of the i-th sample object.

[0013] If the attribute data of the i-th sample object is discrete data, then the attribute data of the i-th sample object is encoded to obtain a low-dimensional representation vector, and the low-dimensional representation vector is determined as the preprocessed feature sub-data corresponding to the attribute data of the i-th sample object.

[0014] When all N sample object attribute data have been traversed, the preprocessed feature sub-data corresponding to the N sample object attribute data are concatenated to obtain the preprocessed feature data corresponding to the sample object data.

[0015] The sample object embedding vector is obtained by performing feature mapping on the preprocessed feature data corresponding to the sample object data through the object tower model.

[0016] In any embodiment of this application, the number of sample object item pairs is M; where M is an integer greater than 1.

[0017] The step of determining the object-item similarity loss based on the sample object embedding vector and the sample item embedding vector includes:

[0018] Determine the positive and negative sample item data corresponding to the sample object data in each of the M sample object item pairs; the positive sample item data corresponding to a sample object data refers to the sample item data in the M sample object item pairs that belong to the same sample object item pair as the sample object data; the negative sample item data corresponding to a sample object data refers to the sample item data in the M sample object item pairs that do not belong to the same sample object item pair as the sample object data.

[0019] Obtain the inner product similarity between the embedding vector of the sample object associated with the same sample object data and the embedding vector of the sample item corresponding to the positive sample item data, and obtain the inner product similarity of the positive samples corresponding to M sample object data respectively;

[0020] Obtain the inner product similarity between the embedding vector of the sample object associated with the same sample object data and the embedding vector of the sample item corresponding to the negative sample item data, and obtain the inner product similarity of the negative samples corresponding to M sample object data respectively;

[0021] The object-item similarity loss is determined based on the positive and negative inner product similarities corresponding to the M sample object data.

[0022] In conjunction with any embodiment of this application, determining the object-item similarity loss based on the positive sample inner product similarity corresponding to M sample object data and the negative sample inner product similarity corresponding to M sample object data includes:

[0023] Logarithmic processing is performed on the positive sample inner product similarity corresponding to each sample object data to obtain the positive sample log value corresponding to each sample object data. The mean of the positive sample log value corresponding to each sample object data is then processed to obtain the first expected negative value.

[0024] Logarithmic operations are performed on the differences between the preset parameters of the negative samples and the similarity of the inner product of the negative samples corresponding to each sample object data to obtain the logarithmic value of the negative samples corresponding to each sample object data. The mean of the logarithmic values ​​of the negative samples corresponding to each sample object data is then applied to obtain the second expected negative value.

[0025] Add the first expected negative value and the second expected negative value to obtain the core recall loss value;

[0026] Logarithmic values ​​are applied to the differences between the positive and negative inner product similarities for each sample object data to obtain the logarithmic difference values ​​for each sample object data. The mean of these logarithmic difference values ​​is then applied to obtain the personalized ranking loss value for the object.

[0027] The core recall loss value and the personalized ranking loss value of the object are weighted and summed to obtain the object-item similarity loss.

[0028] In conjunction with any embodiment of this application, the step of performing parameter tuning on the object tower model based on the object item similarity loss and the object tower generation loss to obtain the target object tower model includes:

[0029] The object item similarity loss and the object tower generation loss are weighted and summed to obtain the model weighted total loss corresponding to the object tower model.

[0030] Based on the model weighted total loss and loss minimization update direction corresponding to the object tower model, the object tower model is subjected to parameter tuning to obtain the parameter-tuned object tower model.

[0031] If it is determined that the model weighted total loss corresponding to the object tower model satisfies the loss convergence condition, then the parameter-tuned object tower model is determined as the target object tower model.

[0032] In conjunction with any embodiment of this application, the method further includes:

[0033] The item tower model is parametrically tuned based on the object item similarity loss and the loss minimization update direction to obtain the parametrically tuned item tower model.

[0034] The object discriminator is tuned based on the object-item similarity loss and the loss maximization update direction to obtain the tuned discriminator model.

[0035] If it is determined that the model weighted total loss corresponding to the object tower model does not meet the loss convergence condition, then the parameter-tuned object tower model, the parameter-tuned item tower model, and the parameter-tuned discriminator model are further iteratively trained based on the sample object item pairs to obtain the model weighted total loss corresponding to the parameter-tuned object tower model.

[0036] Based on the model weighted total loss and loss minimization update direction corresponding to the parameter-tuned object tower model, the parameter-tuned object tower model is subjected to parameter tuning to obtain a new parameter-tuned object tower model.

[0037] If it is determined that the model weighted total loss corresponding to the parameter-tuned object tower model satisfies the loss convergence condition, then the new parameter-tuned object tower model is determined as the target object tower model.

[0038] In conjunction with any embodiment of this application, the method further includes:

[0039] When a request for item recommendation is received from the first business object, the first object data of the first business object is obtained;

[0040] Based on the target object tower model, the first object data is embedded to obtain the first object embedding vector;

[0041] Item retrieval processing is performed based on the first object embedding vector to obtain an item recommendation set, which is then sent to the first business object. The higher the position of an item in the item recommendation set, the higher the similarity between the inner product of its item embedding vector and the first object embedding vector.

[0042] Secondly, a data processing apparatus is provided, comprising:

[0043] The acquisition module is used to acquire sample object-item pairs; the sample object-item pair includes sample object data and sample item data corresponding to the sample object;

[0044] The embedding representation module is used to perform embedding representation processing on the sample object data based on the object pyramid model to obtain the sample object embedding vector, and to perform embedding representation processing on the sample item data based on the item pyramid model to obtain the sample item embedding vector. The object-item similarity loss is determined based on the sample object embedding vector and the sample item embedding vector.

[0045] The discrimination module is used to perform object type discrimination processing on the embedding vector of the sample object based on the object discriminator, obtain the predicted probability that the sample object belongs to the new business object type, and generate the object tower generation loss based on the predicted probability;

[0046] The parameter tuning module is used to perform parameter tuning on the object tower model based on the object item similarity loss and the object tower generation loss to obtain the target object tower model; the target object tower model is used to generate the business object embedding vector corresponding to the business object; the business object includes new business objects and existing business objects; the feature distribution pattern of the business object embedding vector corresponding to the new business object is similar to the feature distribution pattern of the business object embedding vector corresponding to the existing business object.

[0047] In any embodiment of this application, the sample object data includes N sample object attribute data; where N is a positive integer;

[0048] The embedding representation processing of the sample object data based on the object pyramid model yields the sample object embedding vector. Specifically, the embedding representation module is used for:

[0049] Iterate through the attribute data of the N sample objects to obtain the attribute data of the i-th sample object; i is a positive integer less than or equal to N;

[0050] If the attribute data of the i-th sample object is numerical data, then the attribute data of the i-th sample object is subjected to numerical standardization to obtain a standardized value, and the standardized value is determined as the preprocessed feature sub-data corresponding to the attribute data of the i-th sample object.

[0051] If the attribute data of the i-th sample object is discrete data, then the attribute data of the i-th sample object is encoded to obtain a low-dimensional representation vector, and the low-dimensional representation vector is determined as the preprocessed feature sub-data corresponding to the attribute data of the i-th sample object.

[0052] When all N sample object attribute data have been traversed, the preprocessed feature sub-data corresponding to the N sample object attribute data are concatenated to obtain the preprocessed feature data corresponding to the sample object data.

[0053] The sample object embedding vector is obtained by performing feature mapping on the preprocessed feature data corresponding to the sample object data through the object tower model.

[0054] In any embodiment of this application, the number of sample object item pairs is M; where M is an integer greater than 1.

[0055] The step of determining the object-item similarity loss based on the sample object embedding vector and the sample item embedding vector, wherein the embedding representation module is specifically used for:

[0056] Determine the positive and negative sample item data corresponding to the sample object data in each of the M sample object item pairs; the positive sample item data corresponding to a sample object data refers to the sample item data in the M sample object item pairs that belong to the same sample object item pair as the sample object data; the negative sample item data corresponding to a sample object data refers to the sample item data in the M sample object item pairs that do not belong to the same sample object item pair as the sample object data.

[0057] Obtain the inner product similarity between the embedding vector of the sample object associated with the same sample object data and the embedding vector of the sample item corresponding to the positive sample item data, and obtain the inner product similarity of the positive samples corresponding to M sample object data respectively;

[0058] Obtain the inner product similarity between the embedding vector of the sample object associated with the same sample object data and the embedding vector of the sample item corresponding to the negative sample item data, and obtain the inner product similarity of the negative samples corresponding to M sample object data respectively;

[0059] The object-item similarity loss is determined based on the positive and negative inner product similarities corresponding to the M sample object data.

[0060] In conjunction with any embodiment of this application, the step of determining the object-item similarity loss based on the positive sample inner product similarity and the negative sample inner product similarity corresponding to the M sample object data, wherein the embedding representation module is specifically used for:

[0061] Logarithmic processing is performed on the positive sample inner product similarity corresponding to each sample object data to obtain the positive sample log value corresponding to each sample object data. The mean of the positive sample log value corresponding to each sample object data is then processed to obtain the first expected negative value.

[0062] Logarithmic operations are performed on the differences between the preset parameters of the negative samples and the similarity of the inner product of the negative samples corresponding to each sample object data to obtain the logarithmic value of the negative samples corresponding to each sample object data. The mean of the logarithmic values ​​of the negative samples corresponding to each sample object data is then applied to obtain the second expected negative value.

[0063] Add the first expected negative value and the second expected negative value to obtain the core recall loss value;

[0064] Logarithmic values ​​are applied to the differences between the positive and negative inner product similarities for each sample object data to obtain the logarithmic difference values ​​for each sample object data. The mean of these logarithmic difference values ​​is then applied to obtain the personalized ranking loss value for the object.

[0065] The core recall loss value and the personalized ranking loss value of the object are weighted and summed to obtain the object-item similarity loss.

[0066] In any embodiment of this application, the parameter tuning module for obtaining the target object tower model by performing parameter tuning on the object tower model based on the object item similarity loss and the object tower generation loss, wherein the parameter tuning module is specifically used for:

[0067] The object item similarity loss and the object tower generation loss are weighted and summed to obtain the model weighted total loss corresponding to the object tower model.

[0068] Based on the model weighted total loss and loss minimization update direction corresponding to the object tower model, the object tower model is subjected to parameter tuning to obtain the parameter-tuned object tower model.

[0069] If it is determined that the model weighted total loss corresponding to the object tower model satisfies the loss convergence condition, then the parameter-tuned object tower model is determined as the target object tower model.

[0070] In any embodiment of this application, the apparatus further includes an iterative training module.

[0071] The iterative training module is used to tune the item tower model based on the similarity loss of the object items and the loss minimization update direction, so as to obtain the tuned item tower model.

[0072] The object discriminator is tuned based on the object-item similarity loss and the loss maximization update direction to obtain the tuned discriminator model.

[0073] If it is determined that the model weighted total loss corresponding to the object tower model does not meet the loss convergence condition, then the parameter-tuned object tower model, the parameter-tuned item tower model, and the parameter-tuned discriminator model are further iteratively trained based on the sample object item pairs to obtain the model weighted total loss corresponding to the parameter-tuned object tower model.

[0074] Based on the model weighted total loss and loss minimization update direction corresponding to the parameter-tuned object tower model, the parameter-tuned object tower model is subjected to parameter tuning to obtain a new parameter-tuned object tower model.

[0075] If it is determined that the model weighted total loss corresponding to the parameter-tuned object tower model satisfies the loss convergence condition, then the new parameter-tuned object tower model is determined as the target object tower model.

[0076] In any embodiment of this application, the apparatus further includes a recommendation module.

[0077] The recommendation module is used to obtain the first object data of the first business object when it receives an item recommendation request from the first business object;

[0078] Based on the target object tower model, the first object data is embedded to obtain the first object embedding vector;

[0079] Item retrieval processing is performed based on the first object embedding vector to obtain an item recommendation set, which is then sent to the first business object. The higher the position of an item in the item recommendation set, the higher the similarity between the inner product of its item embedding vector and the first object embedding vector.

[0080] Thirdly, an electronic device is provided, comprising: a processor and a memory, the memory being used to store computer program code, the computer program code including computer instructions, wherein, when the processor executes the computer instructions, the electronic device performs as described in the first aspect and any of its embodiments.

[0081] Fourthly, another electronic device is provided, comprising: a processor, a transmitting device, an input device, an output device, and a memory, the memory being used to store computer program code, the computer program code including computer instructions, wherein, when the processor executes the computer instructions, the electronic device performs as described in the first aspect and any of its embodiments.

[0082] Fifthly, a computer-readable storage medium is provided, wherein a computer program is stored therein, the computer program including program instructions that, when executed by a processor, cause the processor to perform the first aspect and any of its embodiments described above.

[0083] In a sixth aspect, a computer program product is provided, the computer program product comprising a computer program or instructions that, when the computer program or instructions are executed on a computer, cause the computer to perform the first aspect described above and any of its embodiments.

[0084] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this application.

[0085] In this application, sample object-item pairs are obtained; each sample object-item pair includes sample object data and sample item data corresponding to the sample object; the sample object data is embedded using an object pyramid model to obtain sample object embedding vectors, and the sample item data is embedded using an item pyramid model to obtain sample item embedding vectors; the object-item similarity loss is determined based on the sample object embedding vectors and sample item embedding vectors; the sample object embedding vectors are then processed by an object discriminator to determine the object type, obtaining the predicted probability that the sample object belongs to a new business object type, and the object pyramid generation loss is generated based on the predicted probability; the object pyramid model is then tuned based on the object-item similarity loss and the object pyramid generation loss to obtain the target object pyramid model. The target object pyramid model is used to generate the business object embedding vectors corresponding to the business objects; the business objects include new business objects and existing business objects; the feature distribution patterns of the business object embedding vectors corresponding to new business objects are similar to those of the business object embedding vectors corresponding to existing business objects. The method provided in this application embodiment jointly updates the object tower model by using object-item similarity loss and object tower generation loss. This ensures that the final target object tower model can guarantee the accuracy of object-item correlation prediction while using the mature embedding of existing business objects to guide the embedding generation of new business objects. This achieves the supplementation and optimization of new business object embedding information and solves the problem of low recommendation accuracy caused by the scarcity of new business object behavioral data. Attached Figure Description

[0086] To more clearly illustrate the technical solutions in the embodiments of this application or the background art, the accompanying drawings used in the embodiments of this application or the background art will be described below.

[0087] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the specification, serve to explain the technical solutions of this application.

[0088] Figure 1 This application provides a schematic diagram of a network architecture.

[0089] Figure 2 A flowchart illustrating a data processing method provided in an embodiment of this application;

[0090] Figure 3 A schematic diagram of the model structure of an adversarial-dual-tower recall model provided in an embodiment of this application;

[0091] Figure 4 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application;

[0092] Figure 5 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0093] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0094] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0095] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0096] The embodiments of this application are described below with reference to the accompanying drawings.

[0097] Please see Figure 1 , Figure 1 This is a schematic diagram of a network architecture provided in an embodiment of this application. Figure 1 As shown, this network architecture may include a server 2000 and a terminal cluster. The terminal cluster may specifically include one or more terminals; the number of terminals in the terminal cluster is not limited here. Figure 1 As shown, the multiple terminals may specifically include terminal 3000a, terminal 3000b, terminal 3000c, ..., terminal 3000n; terminal 3000a, terminal 3000b, terminal 3000c, ..., terminal 3000n can be directly or indirectly connected to server 2000 via wired or wireless communication, so that each terminal can interact with server 2000 through the network connection. It is understood that the above network connection is not limited to a specific connection method; it can be directly or indirectly connected via wired communication, wireless communication, or other methods. This application does not impose any restrictions on this.

[0098] Each terminal in the terminal cluster can include: smartphones, tablets, laptops, desktop computers, smart voice interaction devices, smart home appliances (e.g., smart TVs), wearable devices, in-vehicle terminals, aircraft, and other smart terminals with data processing capabilities. It should be understood that, as... Figure 1 Each terminal in the terminal cluster shown can have an application client installed. When the application client runs on each terminal, it can interact with the aforementioned... Figure 1 Data exchange occurs between the 2000 servers shown.

[0099] Specifically, the application client can include in-vehicle clients, smart home clients, entertainment clients (e.g., game clients), multimedia clients (e.g., video clients), social clients, and information clients (e.g., news clients) with data processing capabilities. In this embodiment, the application client can be integrated into a client (e.g., a social client), or it can be a standalone client (e.g., a news client). This embodiment does not limit the type of application client.

[0100] Among them, such as Figure 1 The server 2000 shown can be the server corresponding to the application client. The server 2000 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0101] For ease of understanding, the embodiments of this application may be described in detail below. Figure 1 One of the multiple terminals shown is selected as the first terminal. For example, in the embodiments of this application, a terminal can be selected as the first terminal. Figure 1 The terminal 3000a shown serves as the first terminal, which can be equipped with an application client capable of data processing. In this case, the first terminal can interact with the server 2000 via the application client. Furthermore, the user corresponding to the first terminal can be referred to as the first object.

[0102] like Figure 1As shown, the first terminal can generate an item recommendation request through a running application client and then send the request to the server 2000. Upon receiving the request, the server 2000 obtains the first object data corresponding to the first object, performs embedding representation processing on the first object data based on the target object pyramid model to obtain the first object embedding vector, and then performs item recall processing based on the first object embedding vector to obtain an item recommendation set. This set is then sent to the first terminal so that the terminal can display the optimized content recommendation set in order. Specifically, in the item recommendation set, the higher the position of an item, the higher the similarity between its embedding vector and the first object's embedding vector should be. This ensures that items with high relevance to the first object are prioritized for display. Among them, the target object tower model is an object preference prediction model for extracting object interest features provided in the embodiments of this application. In order to overcome the problem of significant missing profile features caused by the scarcity of new business object behavior data, this application can introduce an object discriminator to perform adversarial learning when training the target object tower model. This allows the new business object embedding output by the target object tower model to absorb the feature distribution of the existing business object embedding, thereby alleviating the problem of missing new business object embedding information and ensuring that the recommendation accuracy of the final recommendation result is high, whether it is a new business object or an existing business object.

[0103] The data processing method for training the target object tower model provided in this application can be specifically as follows: Obtaining sample object-item pairs; each sample object-item pair includes sample object data and sample item data corresponding to the sample object; performing embedding representation processing on the sample object data based on the object tower model to obtain sample object embedding vectors, and performing embedding representation processing on the sample item data based on the item tower model to obtain sample item embedding vectors; determining the object-item similarity loss based on the sample object embedding vectors and sample item embedding vectors; performing object type discrimination processing on the sample object embedding vectors based on an object discriminator to obtain the predicted probability that the sample object belongs to a new business object type; generating the object tower generation loss based on the predicted probability; and tuning the object tower model based on the object-item similarity loss and the object tower generation loss to obtain the target object tower model. The target object tower model is used to generate the business object embedding vectors corresponding to the business objects; the business objects include new business objects and existing business objects; the feature distribution patterns of the business object embedding vectors corresponding to new business objects are similar to the feature distribution patterns of the business object embedding vectors corresponding to existing business objects.

[0104] Furthermore, embodiments of this application provide a data processing method, wherein the execution subject of the data processing method is a data processing device, which can be any electronic device capable of executing the technical solutions disclosed in the embodiments of this application. Optionally, the data processing device can be one of the following: a computer, a server. It should be understood that embodiments of this application can also be implemented by a processor executing computer program code. The data processing method provided in embodiments of this application is used to train a target object pyramid model, which is used to generate business object embedding vectors for business objects. These business object embedding vectors are used to characterize the interests and preferences of business objects, thereby recommending highly relevant items to business objects based on these embedding vectors. In this application, items include: text, video, images, audio, graphic notes, and goods.

[0105] Please see Figure 2 , Figure 2 This is a flowchart illustrating a data processing method provided in an embodiment of this application.

[0106] S101. Obtain sample object item pairs; the sample object item pairs include sample object data and sample item data corresponding to the sample object.

[0107] Specifically, a sample object-item pair contains sample object data corresponding to a sample object and sample item data corresponding to a sample item. It can be understood that the sample object and the sample item are strongly correlated, such as the sample object may have purchased the sample item or the sample object may have collected the sample item.

[0108] Specifically, a sample object can refer to a user (most often a single user, but in some special cases it may be a user group). Sample object data refers to the characteristic information related to the sample object (user), used to characterize the attributes or behaviors of the sample object. Sample object data can include basic attribute data: such as static information such as age, gender, region, registration time, etc.; behavioral data: such as the sample object's historical interaction records with sample items belonging to the same sample object item pair (clicks, favorites, purchases, browsing time, etc.), or interaction data with other similar items; preference characteristic data: such as the sample object's already labeled interest tags ("likes science fiction movies", "prefers low-sugar foods"), and potential preferences mined through historical data (such as "frequent purchase of maternity and baby products").

[0109] Specifically, sample items can refer to tradable goods or browseable text, videos, images, and other content; this application does not impose any restrictions. Taking goods as an example, sample item data can refer to characteristic information related to the sample item (goods) used to characterize the item's attributes. Sample item data can include basic attribute data such as item category ("mobile phone", "novel", "snack"), brand, price, and specifications (such as mobile phone memory and screen size); content characteristic data such as keywords for text-based items ("suspense" and "urban" tags for novels) and visual characteristics for image-based items ("casual style" and "red" for clothing); and related characteristic data such as the item's popularity (sales volume and clicks) and similarity to other items (such as "item B that is often purchased together with item A"). For example, a feasible sample object pair may contain sample object data for sample object A: age 28, gender female, 3 clicks on dresses in history, preference for "commuter style" and sample item data for sample item B: category is dress, style is "commuter", price is 399 yuan, and sales volume in the past 30 days is 1200 pieces.

[0110] Specifically, in the actual model training process, the number of sample object pairs can be set based on the actual training accuracy, such as 10,000, 20,000, etc., and this application does not impose any restrictions here.

[0111] S102. The sample object data is embedded and represented based on the object pyramid model to obtain the sample object embedding vector. The sample item data is embedded and represented based on the item pyramid model to obtain the sample item embedding vector. The object-item similarity loss is determined based on the sample object embedding vector and the sample item embedding vector.

[0112] Specifically, the object tower model and the item tower model can be collectively referred to as the dual-tower recall model. Their main function is to map objects and items to the same vector space, and then measure the matching degree between them using vector similarity. That is, when determining whether an object and an item are similar, they do not directly calculate the interaction relationship between them. Instead, they first transform the object features and item features into dense vectors of the same dimension through two independent neural networks (i.e., the object tower model and the item tower model). Then, they calculate the similarity between the two vectors using methods such as cosine similarity and dot product. The higher the similarity, the stronger the object's potential interest in that item. It can be understood that the basic structure of the object tower model and the item tower model can adopt the dual-tower architecture of DSSM (Deep Structured Semantic Models), where the two towers are symmetrical but their parameters are independent.

[0113] Specifically, the sample object data contains N sample object attribute data. A feasible implementation process for embedding the sample object data into an embedding vector based on the object pyramid model can be as follows: traverse the N sample object attribute data to obtain the i-th sample object attribute data, where i is a positive integer less than or equal to N; if the i-th sample object attribute data is numerical, perform numerical standardization on it to obtain a standardized value, and determine the standardized value as the preprocessed feature sub-data corresponding to the i-th sample object attribute data; if the i-th sample object attribute data is discrete, encode it to obtain a low-dimensional representation vector, and determine the low-dimensional representation vector as the preprocessed feature sub-data corresponding to the i-th sample object attribute data; after traversing all N sample object attribute data, concatenate the preprocessed feature sub-data corresponding to each of the N sample object attribute data to obtain the preprocessed feature data corresponding to the sample object data; perform feature mapping on the preprocessed feature data corresponding to the sample object data using the object pyramid model to obtain the sample object embedding vector. Numerical data refers to data that can be quantified, represented numerically, and supports mathematical operations, such as object age and item price. Numerical standardization involves converting numerical data to a uniform magnitude range to avoid affecting model training stability due to numerical differences. For example, an object's age of 18, compressed to the range [0,1], yields a standardized value of 0.2. Discrete data refers to data types with finite values, enumerability, and clear intervals between adjacent values, which cannot be infinitely subdivided, such as object gender and item category. A feasible implementation of the encoding process involves first performing one-hot encoding on the discrete data, and then converting the encoded result into a fixed-dimensional vector embedding through an embedding layer, thereby achieving a low-dimensional dense representation of discrete features. A feasible implementation of using an object pyramid model to perform feature mapping on the preprocessed feature data corresponding to the sample object data to obtain sample object embedding vectors involves inputting the preprocessed feature data into the DNN network (generator) of the object pyramid model, and then outputting a sample object embedding vector of dimension d through a nonlinear transformation of multiple fully connected layers. d is a positive integer.

[0114] It is understandable that the specific implementation process of embedding the sample item data into a sample item embedding vector based on the item pyramid model can refer to the above-mentioned feasible implementation process of generating sample object embedding vectors, and will not be elaborated here.

[0115] Specifically, a feasible implementation process for determining the object-item similarity loss based on the sample object embedding vector and the sample item embedding vector can be as follows: Identify the positive and negative sample item data corresponding to the sample object data contained in M ​​sample object-item pairs; a positive sample item data corresponding to a sample object data refers to the sample item data belonging to the same sample object-item pair as the sample object data in the M sample object-item pairs; a negative sample item data corresponding to a sample object data refers to the sample item data not belonging to the same sample object-item pair as the sample object data in the M sample object-item pairs; obtain the inner product similarity between the sample object embedding vector associated with the same sample object data and the sample item embedding vector corresponding to the positive sample item data, thus obtaining the positive inner product similarity for each of the M sample object data; obtain the inner product similarity between the sample object embedding vector associated with the same sample object data and the sample item embedding vector corresponding to the negative sample item data, thus obtaining the negative inner product similarity for each of the M sample object data; determine the object-item similarity loss based on the positive and negative inner product similarities corresponding to the M sample object data. For ease of understanding, assume there are M sample object pairs, including sample object pair D1 [sample object data A1, sample item data B1], sample object pair D2 [sample object data A2, sample item data B2], and sample object pair D3 [sample object data A3, sample item data B3]. Taking sample object data A1 as an example, its corresponding positive sample item data is sample item data B1, and its corresponding negative sample item data are sample item data B2 and sample item data B3. The inner product similarity is a core indicator for measuring the consistency of direction and the correlation of numerical scale between two vectors. Essentially, it quantifies the "strength of positive correlation" between the two vectors through the inner product operation. Taking sample object data A1 as an example again, its corresponding positive sample inner product similarity is the inner product similarity between the sample object embedding vector C1 corresponding to sample object data A1 and the sample item embedding vector E1 corresponding to sample item data B1. Its corresponding negative sample inner product similarity includes the inner product similarity between sample object embedding vector C1 and the sample item embedding vector E2 corresponding to sample item data B2, and the inner product similarity between sample object embedding vector C1 and the sample item embedding vector E3 corresponding to sample item data B3.

[0116] Specifically, after obtaining the positive and negative inner product similarities corresponding to M sample object data, the DNN (Deep Neural Network) parameters of the object tower model and item tower model can be updated using the real interaction labels between objects and items (such as clicks, views, etc.) as supervision, through backpropagation of cross-entropy loss and Bayesian Personalized Ranking (BPR) loss. This ensures that the model retains the predictive ability of DSSM for object-item correlation. Therefore, a feasible implementation process for determining the object-item similarity loss based on the positive and negative inner product similarities corresponding to M sample object data can be as follows: Perform logarithmic processing on the positive inner product similarity corresponding to each sample object data to obtain the logarithmic value of the positive samples for each sample object data; perform mean processing on the logarithmic value of the positive samples for each sample object data to obtain the first expected negative value; perform logarithmic processing on the difference between the negative sample preset parameter and the negative inner product similarity corresponding to each sample object data to obtain the negative sample value for each sample object data. For each sample object data point, the negative sample logarithmic values ​​are averaged to obtain a second expected negative value. The first and second expected negative values ​​are then added together to obtain the core recall loss value. For each sample object data point, the difference between the positive and negative inner product similarities is logarithmically calculated to obtain the logarithmic difference between positive and negative sample differences. This logarithmic difference is then averaged to obtain the personalized ranking loss value. Finally, the core recall loss value and the personalized ranking loss value are weighted and summed to obtain the object-item similarity loss.

[0117] For ease of understanding, the process of determining the core loss value of the recall described above can be represented by the following formula (1):

[0118]

[0119] Where u refers to the embedding vector of the sample object data, i refers to the embedding vector of the positive sample item data corresponding to the sample object data, and D(u,i) represents the inner product similarity of the positive sample data. This indicates that the logarithmic similarity of the inner product of each positive sample is averaged. That is, the first expected negative value; j refers to the sample item embedding vector corresponding to the negative sample item data corresponding to the sample object data, and D(u,j) represents the inner product similarity of the negative samples corresponding to the sample object data. This represents the mean of the logarithmic differences in the inner product similarity of each negative sample. That is, the second expected value; This refers to the core loss value of recall.

[0120] The process of determining the personalized ranking loss value of the above objects can be expressed by the following formula (2):

[0121]

[0122] Where u refers to the sample object embedding vector corresponding to the sample object data, i refers to the sample item embedding vector corresponding to the positive sample item data corresponding to the sample object data, and j refers to the sample item embedding vector corresponding to the negative sample item data corresponding to the sample object data. This represents the positive inner product similarity between sample object data. σ(·) represents the inner product similarity of negative samples corresponding to sample object data, and is the sigmoid function (σ(x)=1 / (1+e)). -x ))), This refers to the loss value for personalized sorting of objects.

[0123] The process of obtaining the object similarity loss described above can be represented by the following formula (3):

[0124]

[0125] in, This refers to the object-item similarity loss, where α, β ∈ [0, 1] are balancing coefficients used to adjust the weights of the two objectives. This refers to the core loss value of recall. This refers to the loss value for personalized sorting of objects.

[0126] S103. Based on the object discriminator, the object type of the sample object embedding vector is processed to obtain the predicted probability that the sample object belongs to the new business object type, and the object tower generation loss is generated according to the predicted probability.

[0127] Specifically, the object discriminator is used to distinguish object types, that is, to determine whether the sample object belongs to a new business object (new user) or an existing business object (old user). This application utilizes the adversarial training mechanism of the Generative Adversarial Network (GAN) to use the object pyramid model as the generator of the GAN, and then connects it to the object discriminator to judge the object type of the sample object, thereby optimizing the object embedding vector. A feasible implementation process for determining the object pyramid generation loss can be represented by the following formula (4):

[0128]

[0129] Where D(u) is the output probability of the object discriminator, and its value belongs to [0,1]. It can be stipulated that the closer the value is to 0, the more likely the sample object belongs to the existing business object, and the closer the value is to 1, the more likely the sample object belongs to the new business object. This refers to the object pyramid generation loss. The object pyramid generation loss can be used to update the parameters of both the object pyramid model and the object discriminator. However, due to the adversarial training mechanism of adversarial networks, the corresponding loss update methods for the two should be completely opposite. Because the object pyramid model, as a generator, needs to deceive the object discriminator, making it unable to distinguish whether a sample object is a new business object or an existing business object, therefore, based on... When updating the parameters of the object tower model, the minimum should be minimized. Conversely, for the object discriminator, it should strive to correctly determine whether the current sample object is a new business object or an existing business object. Therefore, based on... When updating the parameters of the object discriminator, the maximum value should be maximized.

[0130] Optionally, sample object behavior data can be obtained from the sample object data corresponding to the sample object. If the sample object behavior data exceeds the object behavior data threshold, the object type label corresponding to the sample object is determined to be the existing business object type label; if the sample object behavior data does not exceed the object behavior data threshold, the object type label corresponding to the sample object is determined to be the new business object type label. For example, sample object behavior data can be the object registration time or the object usage duration. When the registration time of the object is less than seven days or the usage duration of the object is less than 2 hours, the sample object can be considered a new business object.

[0131] S104. The object tower model is subjected to parameter tuning based on the object item similarity loss and the object tower generation loss to obtain the target object tower model; the target object tower model is used to generate the business object embedding vector corresponding to the business object; the business object includes new business objects and existing business objects; the feature distribution pattern of the business object embedding vector corresponding to the new business object is similar to the feature distribution pattern of the business object embedding vector corresponding to the existing business object.

[0132] Specifically, a feasible implementation process for tuning the object tower model based on the object-item similarity loss and the object tower generation loss to obtain the target object tower model can be as follows: The object-item similarity loss and the object tower generation loss are weighted and summed to obtain the weighted total loss of the object tower model; the object tower model is then tuned based on the weighted total loss and the loss minimization update direction to obtain the tuned object tower model; if the weighted total loss of the object tower model satisfies the loss convergence condition, the tuned object tower model is determined as the target object tower model. Here, the loss minimization update direction refers to minimizing the weighted total loss of the adjusted model.

[0133] It is understandable that the process of determining the model's weighted total loss can be represented by the following formula (5):

[0134]

[0135] in, That is, the model weighted total loss. That is, the loss of similarity between the aforementioned objects and items. That is, the loss generated by the above object tower, where δ is the weighting parameter.

[0136] Optionally, the item pyramid model is hyperparameterized based on the object-item similarity loss and the loss minimization update direction to obtain a hyperparameterized item pyramid model; the object discriminator is hyperparameterized based on the object-item similarity loss and the loss maximization update direction to obtain a hyperparameterized discriminator model; if it is determined that the weighted total loss of the object pyramid model does not meet the loss convergence condition, iterative training is continued on the hyperparameterized object pyramid model, the hyperparameterized item pyramid model, and the hyperparameterized discriminator model based on sample object-item pairs to obtain the weighted total loss of the hyperparameterized object pyramid model; the hyperparameterized object pyramid model is hyperparameterized based on the weighted total loss of the hyperparameterized object pyramid model and the loss minimization update direction to obtain a new hyperparameterized object pyramid model; if it is determined that the weighted total loss of the hyperparameterized object pyramid model meets the loss convergence condition, the new hyperparameterized object pyramid model is determined as the target object pyramid model. Here, the loss maximization update direction refers to the adjusted object discriminator causing the object-item similarity loss to increase, i.e., maximize.

[0137] Optionally, when a request for item recommendation is received from the first business object, the first object data of the first business object is obtained; the first object data is processed by embedding representation based on the target object pyramid model to obtain the first object embedding vector; the item retrieval process is performed based on the first object embedding vector to obtain the item recommendation set, and the item recommendation set is sent to the first business object; the higher the position of the item in the item recommendation set, the higher the similarity between the inner product of the item embedding vector and the first object embedding vector.

[0138] In this embodiment, sample object-item pairs are obtained; each sample object-item pair includes sample object data and sample item data corresponding to the sample object; the sample object data is embedded using an object pyramid model to obtain sample object embedding vectors, and the sample item data is embedded using an item pyramid model to obtain sample item embedding vectors; the object-item similarity loss is determined based on the sample object embedding vectors and sample item embedding vectors; the sample object embedding vectors are then processed by an object discriminator to determine the object type, obtaining the predicted probability that the sample object belongs to a new business object type; the object pyramid generation loss is generated based on the predicted probability; and the object pyramid model is tuned based on the object-item similarity loss and the object pyramid generation loss to obtain the target object pyramid model. The target object pyramid model is used to generate the business object embedding vectors corresponding to the business objects; the business objects include new business objects and existing business objects; the feature distribution patterns of the business object embedding vectors corresponding to new business objects are similar to those of the business object embedding vectors corresponding to existing business objects. The method provided in this application embodiment jointly updates the object tower model by using object-item similarity loss and object tower generation loss. This ensures that the final target object tower model can guarantee the accuracy of object-item correlation prediction while using the mature embedding of existing business objects to guide the embedding generation of new business objects. This achieves the supplementation and optimization of new business object embedding information and solves the problem of low recommendation accuracy caused by the scarcity of new business object behavioral data.

[0139] Furthermore, for ease of understanding, please refer to Figure 3 , Figure 3 This is a schematic diagram of the model structure of an adversarial-dual-tower recall model provided in an embodiment of this application. Figure 3 As shown, the adversarial-dual-tower recall model provided in this application is an improvement on the traditional DSSM model architecture. It optimizes the new user embedding vector by incorporating the adversarial training mechanism of a Generative Adversarial Network (GAN). This model retains the basic structure of item towers in DSSM, uses the item towers as generators in the GAN, adds an object discriminator to distinguish object types, and finally optimizes the model through bidirectional loss updates. The specific implementation of the model update mainly includes the following steps:

[0140] Step S31: Feature preprocessing.

[0141] Specifically, feature preprocessing refers to preprocessing user features (i.e., sample object data) and candidate item features (i.e., sample item data) separately to obtain preprocessed feature data to adapt to the model input requirements.

[0142] It's understandable that for numerical features (such as user age, item price, etc.), standardization (e.g., Z-score standardization) is used to convert feature values ​​to a uniform range, avoiding the impact of numerical differences on model training stability. For example, an age of 18 can be compressed to 0.2 (within the range [0, 1]) and directly used as input without additional conversion to a vector. For discrete features (such as user gender, item category, etc.), one-hot encoding is performed first, and then the encoding result is converted into a fixed-dimensional vector embedding through an embedding layer, achieving a low-dimensional dense representation of discrete features.

[0143] Step S32: Embedded vector generation.

[0144] Specifically, embedding vector generation refers to using a deep neural network (DNN) structure to map preprocessed object feature data and item feature data respectively, thereby generating corresponding embedding vectors: On the object side, the preprocessed object feature data is input into the DNN network (generator) of the object tower (also known as the user tower). Through nonlinear transformation of multiple fully connected layers, a sample object embedding vector (also known as the user embedding vector) of dimension d can be output; On the item side, the preprocessed item feature data is input into the DNN network of the item tower. Through transformation of multiple fully connected layers with the same structure as the object tower, a sample item embedding vector of dimension d is output. The sample object embedding vector and the sample item embedding vector have the same dimension to meet the needs of subsequent similarity calculation.

[0145] Step S33: Adversarial training and discriminator interaction.

[0146] Specifically, adversarial training and discriminator interaction refer to the adversarial training mechanism based on GAN, which uses an object discriminator to determine the object type and optimize the object embedding vector: the above sample object embedding vector is input into the object discriminator (a binary classifier composed of multiple DNNs), and the object discriminator outputs the category prediction probability. The category prediction probability can be used to determine the probability value of the object corresponding to the sample object embedding vector being a new business object; by calculating the cross-entropy loss (object discriminator loss) between the object discriminator output and the true type of the object (new business object / existing business object), supervised training for object type discrimination is achieved.

[0147] Step S34: Update model parameters.

[0148] Specifically, model parameter updating refers to jointly updating model parameters through a bidirectional loss function (adversarial generation loss and similarity loss), which includes three parts: discriminator loss, generator loss, and similarity loss, weighted and summed to obtain the final loss. Both the discriminator loss and generator loss can be the object pyramid generation loss mentioned above; however, the discriminator loss is updated to maximize the object pyramid generation loss, while the generator loss is updated to minimize it. The similarity loss is the object-item similarity loss mentioned above. The calculation process for the object pyramid generation loss and the object-item similarity loss can be found above. Figure 2 The specific description of step S103 in the corresponding embodiment will not be repeated here.

[0149] Specifically, after obtaining the loss, the model parameters can be updated. The object tower parameters can be updated based on the generator loss, the object-item similarity loss, and the direction of loss minimization. The specific update process can be found above. Figure 2 The specific description of step S104 in the corresponding embodiment will not be repeated here; at the same time, the parameters of the item tower can be updated according to the object item similarity loss and the loss minimization update direction. The specific update process can be found above. Figure 2 The specific description of step S104 in the corresponding embodiment will not be repeated here; at the same time, the parameters of the object discriminator can be updated according to the discriminator loss and the loss maximization update direction. The specific update process can be found above. Figure 2 The specific description of step S104 in the corresponding embodiment will not be repeated here.

[0150] It is understandable that the above steps achieve model convergence through iterative training. In each iteration, the adversarial loss and similarity loss are calculated and updated simultaneously, ultimately enabling the model to adapt the new business object embedding vector to the feature distribution of the existing business objects while ensuring the accuracy of object-item correlation prediction.

[0151] The method provided in this application involves using the object pyramid in the DSSM as a GAN generator to generate object embeddings. A discriminator is then concatenated with the generator for binary classification to determine whether an object is a new business object. Through GAN adversarial training, the mature embeddings of existing business objects are used to guide the generation of new business object embeddings, thereby supplementing and optimizing the embedding information of new business objects. The optimized object embeddings enable the DSSM to more accurately capture the interests of new business objects, reduce recall bias, and thus improve recommendation relevance and coverage.

[0152] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0153] The methods of the embodiments of this application have been described in detail above, and the apparatus of the embodiments of this application is provided below.

[0154] Please see Figure 4 , Figure 4 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application. The data processing device 1 includes: an acquisition module 11, an embedded representation module 12, a discrimination module 13, and a parameter tuning module 14. Specifically:

[0155] The acquisition module 11 is used to acquire sample object-item pairs; the sample object-item pairs include sample object data and sample item data corresponding to the sample object;

[0156] The embedding representation module 12 is used to perform embedding representation processing on the sample object data based on the object tower model to obtain the sample object embedding vector, and to perform embedding representation processing on the sample item data based on the item tower model to obtain the sample item embedding vector, and to determine the object-item similarity loss based on the sample object embedding vector and the sample item embedding vector.

[0157] The discrimination module 13 is used to perform object type discrimination processing on the embedding vector of the sample object based on the object discriminator, obtain the predicted probability that the sample object belongs to the new business object type, and generate the object tower generation loss based on the predicted probability.

[0158] The parameter tuning module 14 is used to perform parameter tuning on the object tower model based on the object item similarity loss and the object tower generation loss to obtain the target object tower model; the target object tower model is used to generate the business object embedding vector corresponding to the business object; the business object includes new business objects and existing business objects; the feature distribution pattern of the business object embedding vector corresponding to the new business object is similar to the feature distribution pattern of the business object embedding vector corresponding to the existing business object.

[0159] In any embodiment of this application, the sample object data includes N sample object attribute data; where N is a positive integer;

[0160] The embedding representation processing of the sample object data based on the object pyramid model yields the sample object embedding vector. The embedding representation module 12 is specifically used for:

[0161] Iterate through the attribute data of the N sample objects to obtain the attribute data of the i-th sample object; i is a positive integer less than or equal to N;

[0162] If the attribute data of the i-th sample object is numerical data, then the attribute data of the i-th sample object is subjected to numerical standardization to obtain a standardized value, and the standardized value is determined as the preprocessed feature sub-data corresponding to the attribute data of the i-th sample object.

[0163] If the attribute data of the i-th sample object is discrete data, then the attribute data of the i-th sample object is encoded to obtain a low-dimensional representation vector, and the low-dimensional representation vector is determined as the preprocessed feature sub-data corresponding to the attribute data of the i-th sample object.

[0164] When all N sample object attribute data have been traversed, the preprocessed feature sub-data corresponding to the N sample object attribute data are concatenated to obtain the preprocessed feature data corresponding to the sample object data.

[0165] The sample object embedding vector is obtained by performing feature mapping on the preprocessed feature data corresponding to the sample object data through the object tower model.

[0166] In any embodiment of this application, the number of sample object item pairs is M; where M is an integer greater than 1.

[0167] The embedding representation module 12, which determines the object-item similarity loss based on the sample object embedding vector and the sample item embedding vector, is specifically used for:

[0168] Determine the positive and negative sample item data corresponding to the sample object data in each of the M sample object item pairs; the positive sample item data corresponding to a sample object data refers to the sample item data in the M sample object item pairs that belong to the same sample object item pair as the sample object data; the negative sample item data corresponding to a sample object data refers to the sample item data in the M sample object item pairs that do not belong to the same sample object item pair as the sample object data.

[0169] Obtain the inner product similarity between the embedding vector of the sample object associated with the same sample object data and the embedding vector of the sample item corresponding to the positive sample item data, and obtain the inner product similarity of the positive samples corresponding to M sample object data respectively;

[0170] Obtain the inner product similarity between the embedding vector of the sample object associated with the same sample object data and the embedding vector of the sample item corresponding to the negative sample item data, and obtain the inner product similarity of the negative samples corresponding to M sample object data respectively;

[0171] The object-item similarity loss is determined based on the positive and negative inner product similarities corresponding to the M sample object data.

[0172] In conjunction with any embodiment of this application, the step of determining the object-item similarity loss based on the positive sample inner product similarity and the negative sample inner product similarity corresponding to the M sample object data, wherein the embedding representation module 12 is specifically used for:

[0173] Logarithmic processing is performed on the positive sample inner product similarity corresponding to each sample object data to obtain the positive sample log value corresponding to each sample object data. The mean of the positive sample log value corresponding to each sample object data is then processed to obtain the first expected negative value.

[0174] Logarithmic operations are performed on the differences between the preset parameters of the negative samples and the similarity of the inner product of the negative samples corresponding to each sample object data to obtain the logarithmic value of the negative samples corresponding to each sample object data. The mean of the logarithmic values ​​of the negative samples corresponding to each sample object data is then applied to obtain the second expected negative value.

[0175] Add the first expected negative value and the second expected negative value to obtain the core recall loss value;

[0176] Logarithmic values ​​are applied to the differences between the positive and negative inner product similarities for each sample object data to obtain the logarithmic difference values ​​for each sample object data. The mean of these logarithmic difference values ​​is then applied to obtain the personalized ranking loss value for the object.

[0177] The core recall loss value and the personalized ranking loss value of the object are weighted and summed to obtain the object-item similarity loss.

[0178] In any embodiment of this application, the parameter tuning module 14, which performs parameter tuning on the object tower model based on the object item similarity loss and the object tower generation loss to obtain the target object tower model, is specifically used for:

[0179] The object item similarity loss and the object tower generation loss are weighted and summed to obtain the model weighted total loss corresponding to the object tower model.

[0180] Based on the model weighted total loss and loss minimization update direction corresponding to the object tower model, the object tower model is subjected to parameter tuning to obtain the parameter-tuned object tower model.

[0181] If it is determined that the model weighted total loss corresponding to the object tower model satisfies the loss convergence condition, then the parameter-tuned object tower model is determined as the target object tower model.

[0182] In any embodiment of this application, the apparatus further includes an iterative training module 15.

[0183] The iterative training module is used to tune the item tower model based on the similarity loss of the object items and the loss minimization update direction, so as to obtain the tuned item tower model.

[0184] The object discriminator is tuned based on the object-item similarity loss and the loss maximization update direction to obtain the tuned discriminator model.

[0185] If it is determined that the model weighted total loss corresponding to the object tower model does not meet the loss convergence condition, then the parameter-tuned object tower model, the parameter-tuned item tower model, and the parameter-tuned discriminator model are further iteratively trained based on the sample object item pairs to obtain the model weighted total loss corresponding to the parameter-tuned object tower model.

[0186] Based on the model weighted total loss and loss minimization update direction corresponding to the parameter-tuned object tower model, the parameter-tuned object tower model is subjected to parameter tuning to obtain a new parameter-tuned object tower model.

[0187] If it is determined that the model weighted total loss corresponding to the parameter-tuned object tower model satisfies the loss convergence condition, then the new parameter-tuned object tower model is determined as the target object tower model.

[0188] In any embodiment of this application, the device further includes: a recommendation module 16.

[0189] The recommendation module is used to obtain the first object data of the first business object when it receives an item recommendation request from the first business object;

[0190] Based on the target object tower model, the first object data is embedded to obtain the first object embedding vector;

[0191] Item retrieval processing is performed based on the first object embedding vector to obtain an item recommendation set, which is then sent to the first business object. The higher the position of an item in the item recommendation set, the higher the similarity between the inner product of its item embedding vector and the first object embedding vector.

[0192] Figure 5 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. The electronic device 4 includes a processor 41 and a memory 42. Optionally, the electronic device 4 also includes an input device 43 and an output device 44. The processor 41, memory 42, input device 43, and output device 44 are coupled together via connectors, which include various interfaces, transmission lines, or buses, etc., and are not limited in this embodiment. It should be understood that in the various embodiments of this application, coupling refers to mutual connection in a specific way, including direct connection or indirect connection through other devices, such as through various interfaces, transmission lines, buses, etc.

[0193] Processor 41 may include one or more processors, such as one or more central processing units (CPUs). If the processor is a CPU, it may be a single-core CPU or a multi-core CPU. Optionally, processor 41 may be a processor group consisting of multiple CPUs, with the multiple processors coupled to each other via one or more buses. Optionally, the processor may also be other types of processors, etc., which are not limited in this embodiment.

[0194] The memory 42 can be used to store computer program instructions, as well as various types of computer program code, including program code for executing the present application. Optionally, the memory includes, but is not limited to, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), or compact disc read-only memory (CD-ROM), which is used for related instructions and data.

[0195] Input device 43 is used to input data and / or signals, and output device 44 is used to output data and / or signals. Input device 43 and output device 44 can be independent devices or an integrated device.

[0196] It is understood that in this embodiment of the application, the memory 42 can be used not only to store related instructions, but also to store related data. This embodiment of the application does not limit the specific data stored in the memory.

[0197] Understandable Figure 5 This is merely a simplified design of an electronic device. In practical applications, the electronic device may also include other necessary components, including, but not limited to, any number of input / output devices, processors, memories, etc., and all electronic devices that can implement the embodiments of this application are within the protection scope of this application.

[0198] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0199] Those skilled in the art will readily understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. Those skilled in the art will also readily understand that the various embodiments of this application have different focuses, and for the sake of convenience and brevity, the same or similar parts may not be repeated in different embodiments. Therefore, parts not described or not described in detail in one embodiment can be referred to the descriptions in other embodiments.

[0200] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0201] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0202] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0203] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., digital versatile discs (DVDs)), or semiconductor media (e.g., solid-state disks (SSDs)).

[0204] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A data processing method, characterized in that, include: Obtain sample object-item pairs; each sample object-item pair contains sample object data and sample item data corresponding to the sample object. The sample object data is embedded using the object pyramid model to obtain the sample object embedding vector, and the sample item data is embedded using the item pyramid model to obtain the sample item embedding vector. The object-item similarity loss is determined based on the sample object embedding vector and the sample item embedding vector. Based on the object discriminator, the object type is determined by the embedding vector of the sample object to obtain the predicted probability that the sample object belongs to the new business object type, and the object tower generation loss is generated based on the predicted probability. The object tower model is parameter-tuned based on the object item similarity loss and the object tower generation loss to obtain the target object tower model; the target object tower model is used to generate the business object embedding vector corresponding to the business object; the business object includes new business objects and existing business objects; the feature distribution pattern of the business object embedding vector corresponding to the new business object is similar to the feature distribution pattern of the business object embedding vector corresponding to the existing business object.

2. The method according to claim 1, characterized in that, The sample object data contains N sample object attribute data; where N is a positive integer. The embedding representation processing of the sample object data based on the object pyramid model to obtain the sample object embedding vector includes: Iterate through the attribute data of the N sample objects to obtain the attribute data of the i-th sample object; i is a positive integer less than or equal to N; If the attribute data of the i-th sample object is numerical data, then the attribute data of the i-th sample object is subjected to numerical standardization to obtain a standardized value, and the standardized value is determined as the preprocessed feature sub-data corresponding to the attribute data of the i-th sample object. If the attribute data of the i-th sample object is discrete data, then the attribute data of the i-th sample object is encoded to obtain a low-dimensional representation vector, and the low-dimensional representation vector is determined as the preprocessed feature sub-data corresponding to the attribute data of the i-th sample object. When all N sample object attribute data have been traversed, the preprocessed feature sub-data corresponding to the N sample object attribute data are concatenated to obtain the preprocessed feature data corresponding to the sample object data. The sample object embedding vector is obtained by performing feature mapping on the preprocessed feature data corresponding to the sample object data through the object tower model.

3. The method according to claim 1, characterized in that, The number of sample object pairs is M; M is an integer greater than 1. The step of determining the object-item similarity loss based on the sample object embedding vector and the sample item embedding vector includes: Determine the positive and negative sample item data corresponding to the sample object data in each of the M sample object item pairs; the positive sample item data corresponding to a sample object data refers to the sample item data in the M sample object item pairs that belong to the same sample object item pair as the sample object data; the negative sample item data corresponding to a sample object data refers to the sample item data in the M sample object item pairs that do not belong to the same sample object item pair as the sample object data. Obtain the inner product similarity between the embedding vector of the sample object associated with the same sample object data and the embedding vector of the sample item corresponding to the positive sample item data, and obtain the inner product similarity of the positive samples corresponding to M sample object data respectively; Obtain the inner product similarity between the embedding vector of the sample object associated with the same sample object data and the embedding vector of the sample item corresponding to the negative sample item data, and obtain the inner product similarity of the negative samples corresponding to M sample object data respectively; The object-item similarity loss is determined based on the positive and negative inner product similarities corresponding to the M sample object data.

4. The method according to claim 3, characterized in that, The step of determining the object-item similarity loss based on the positive sample inner product similarity and the negative sample inner product similarity corresponding to the M sample object data includes: Logarithmic processing is performed on the positive sample inner product similarity corresponding to each sample object data to obtain the positive sample log value corresponding to each sample object data. The mean of the positive sample log value corresponding to each sample object data is then processed to obtain the first expected negative value. Logarithmic operations are performed on the differences between the preset parameters of the negative samples and the similarity of the inner product of the negative samples corresponding to each sample object data to obtain the logarithmic value of the negative samples corresponding to each sample object data. The mean of the logarithmic values ​​of the negative samples corresponding to each sample object data is then applied to obtain the second expected negative value. Add the first expected negative value and the second expected negative value to obtain the core recall loss value; Logarithmic values ​​are applied to the differences between the positive and negative inner product similarities for each sample object data to obtain the logarithmic difference values ​​for each sample object data. The mean of these logarithmic difference values ​​is then applied to obtain the personalized ranking loss value for the object. The core recall loss value and the personalized ranking loss value of the object are weighted and summed to obtain the object-item similarity loss.

5. The method according to claim 1, characterized in that, The step of tuning the object tower model based on the object item similarity loss and the object tower generation loss to obtain the target object tower model includes: The object item similarity loss and the object tower generation loss are weighted and summed to obtain the model weighted total loss corresponding to the object tower model. Based on the model weighted total loss and loss minimization update direction corresponding to the object tower model, the object tower model is subjected to parameter tuning to obtain the parameter-tuned object tower model. If it is determined that the model weighted total loss corresponding to the object tower model satisfies the loss convergence condition, then the parameter-tuned object tower model is determined as the target object tower model.

6. The method according to claim 5, characterized in that, The method further includes: The item tower model is parametrically tuned based on the object item similarity loss and the loss minimization update direction to obtain the parametrically tuned item tower model. The object discriminator is tuned based on the object-item similarity loss and the loss maximization update direction to obtain the tuned discriminator model. If it is determined that the model weighted total loss corresponding to the object tower model does not meet the loss convergence condition, then the parameter-tuned object tower model, the parameter-tuned item tower model, and the parameter-tuned discriminator model are further iteratively trained based on the sample object item pairs to obtain the model weighted total loss corresponding to the parameter-tuned object tower model. Based on the model weighted total loss and loss minimization update direction corresponding to the parameter-tuned object tower model, the parameter-tuned object tower model is subjected to parameter tuning to obtain a new parameter-tuned object tower model. If it is determined that the model weighted total loss corresponding to the parameter-tuned object tower model satisfies the loss convergence condition, then the new parameter-tuned object tower model is determined as the target object tower model.

7. The method according to claim 1, characterized in that, The method further includes: When a request for item recommendation is received from the first business object, the first object data of the first business object is obtained; Based on the target object tower model, the first object data is embedded to obtain the first object embedding vector; Item retrieval processing is performed based on the first object embedding vector to obtain an item recommendation set, which is then sent to the first business object. The higher the position of an item in the item recommendation set, the higher the similarity between the inner product of its item embedding vector and the first object embedding vector.

8. A data processing apparatus, characterized in that, The device includes: The acquisition module is used to acquire sample object-item pairs; the sample object-item pair includes sample object data and sample item data corresponding to the sample object; The embedding representation module is used to perform embedding representation processing on the sample object data based on the object pyramid model to obtain the sample object embedding vector, and to perform embedding representation processing on the sample item data based on the item pyramid model to obtain the sample item embedding vector. The object-item similarity loss is determined based on the sample object embedding vector and the sample item embedding vector. The discrimination module is used to perform object type discrimination processing on the embedding vector of the sample object based on the object discriminator, obtain the predicted probability that the sample object belongs to the new business object type, and generate the object tower generation loss based on the predicted probability; The parameter tuning module is used to perform parameter tuning on the object tower model based on the object item similarity loss and the object tower generation loss to obtain the target object tower model; the target object tower model is used to generate the business object embedding vector corresponding to the business object; the business object includes new business objects and existing business objects; the feature distribution pattern of the business object embedding vector corresponding to the new business object is similar to the feature distribution pattern of the business object embedding vector corresponding to the existing business object.

9. An electronic device, characterized in that, include: A processor and a memory, the memory being used to store computer program code, the computer program code including computer instructions, wherein, when the processor executes the computer instructions, the electronic device performs the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method according to any one of claims 1 to 7.

11. A computer program product, characterized in that, The computer program product includes a computer program or instructions; when the computer program or instructions are executed on a computer, the computer causes the computer to perform the method according to any one of claims 1 to 7.

Images