Click rate prediction model training and commodity recommendation method, equipment and medium

Abstract

The embodiment of the invention provides a click rate prediction model training and commodity recommendation method and device and a medium, and the method comprises the steps: collecting burying point information which comprises a first ID of a first commodity and a second ID of a second commodity; fusing the first commodity and the second commodity based on the first ID and the second ID to obtain a feature vector; constructing a sample data set based on the feature vector, the user feature information and the shop feature information; and training a preset model by using the sample data set until the preset model meets a preset condition, and obtaining the click rate prediction model. By using the method, the click rate of the first commodity and the second commodity can be accurately predicted.

Description

Technical Field

[0001] This application belongs to the field of data processing, relates to information recommendation technology, and particularly relates to a method, device, and medium for training a click-through rate prediction model and recommending products. Background Art

[0002] In the e-commerce business scenario, it has become an important recommendation application scenario to mine the combination method of products through algorithms and recommend other products that match when a user browses a certain product. Traditional recommendation systems usually recommend based on a click-through rate prediction model for a single product. However, in terms of predicting the conversion rate of combined products, the existing technology is relatively lacking. Existing click-through rate prediction models usually train the model by taking two products as independent features, and cannot accurately predict the click-through rate of the recommended products in this application scenario. Summary of the Invention

[0003] The embodiments of this application provide a method, device, and medium for training a click-through rate prediction model and recommending products, which can solve the technical problem of being unable to accurately predict the click-through rate of combined products.

[0004] In a first aspect of the embodiments of this application, a method for training a click-through rate prediction model is provided. The method includes: collecting buried point information, where the buried point information includes the first ID of the first product and the second ID of the second product; fusing the first product and the second product based on the first ID and the second ID to obtain a feature vector; constructing a sample data set based on the feature vector, user feature information, and store feature information; and training a preset model using the sample data set until the preset model meets the preset conditions to obtain the click-through rate prediction model.

[0005] In some possible implementation manners, the fusing the first product and the second product based on the first ID and the second ID to obtain a feature vector includes: obtaining a first vector based on the first ID; obtaining a second vector based on the second ID; and fusing the first vector and the second vector through a fusion framework to obtain the feature vector.

[0006] In some possible implementation manners, the fusing the first vector and the second vector through a fusion framework to obtain the feature vector includes: performing dimensionality reduction processing on the first vector through a first encoder in the fusion framework to obtain a first dimensionality-reduced vector; performing dimensionality reduction processing on the second vector through a second encoder in the fusion framework to obtain a second dimensionality-reduced vector; and processing the first dimensionality-reduced vector and the second dimensionality-reduced vector through a cross-fusion encoder in the fusion framework to obtain the feature vector.

[0007] In some possible embodiments, the process of obtaining the feature vector by processing the first dimension-reduced vector and the second dimension-reduced vector through the cross-fusion encoder includes: projecting the first dimension-reduced vector into a first intermediate feature vector through the first bridging layer in the fusion framework; projecting the second dimension-reduced vector into a second intermediate feature vector through the second bridging layer in the fusion framework; dynamically fusing the first intermediate feature vector and the second dimension-reduced vector through the first cross-attention layer in the cross-fusion encoder to obtain the first feature vector corresponding to the first commodity; and dynamically fusing the second intermediate feature vector and the first dimension-reduced vector through the second cross-attention layer in the cross-fusion encoder to obtain the second feature vector corresponding to the second commodity.

[0008] In some possible embodiments, the process of obtaining the feature vector by processing the first dimension-reduced vector and the second dimension-reduced vector through the cross-fusion encoder includes: projecting the first dimension-reduced vector into a first intermediate feature vector through the first bridging layer in the fusion framework; projecting the second dimension-reduced vector into a second intermediate feature vector through the second bridging layer in the fusion framework; enhancing the first context representation of the first intermediate feature vector through the first self-attention layer in the cross-fusion encoder; enhancing the second context representation of the second intermediate feature vector through the second self-attention layer in the cross-fusion encoder; dynamically fusing the first context representation and the second dimension-reduced vector through the first cross-attention layer in the cross-fusion encoder to obtain the first feature vector corresponding to the first commodity; and dynamically fusing the second context representation and the first dimension-reduced vector through the second cross-attention layer in the cross-fusion encoder to obtain the second feature vector corresponding to the second commodity.

[0009] In some possible embodiments, the process of training a preset model using the sample data set until the preset model meets a preset condition to obtain a click-through rate prediction model includes: obtaining a first fused feature vector based on the first feature vector, the user feature information, and the store feature information; obtaining a second fused feature vector based on the second feature vector, the user feature information, and the store feature information; and training the preset model based on the first fused feature vector and the second fused feature vector.

[0010] In some possible embodiments, training the preset model based on the first fusion feature vector and the second fusion feature vector includes: predicting the click-through rates of the user clicking to view the first product and the second product based on the first fusion feature vector and the second fusion feature vector; determining a loss value according to the annotation label of the buried point information and the click-through rate; training the preset model according to the loss value until the preset model meets the preset conditions, so as to obtain the click-through rate prediction model.

[0011] A second aspect of the embodiments of the present application provides a product recommendation method, where the product recommendation method includes: obtaining target user feature information; obtaining a first ID of a first product to be recommended and a second ID of a second product; based on the target user feature information, the first ID and the second ID, using the click-through rate prediction model to determine the probabilities of the target user clicking on the first product and the second product, where the click-through rate prediction model is obtained by training using the click-through rate prediction model training method as described above; and determining the first product and the second product as combined recommended products according to the probabilities, and recommending the combined recommended products to the target user.

[0012] A third aspect of the embodiments of the present application provides an electronic device, including: a memory, and a processor, where the processor executes computer-readable instructions stored in the memory to implement the click-through rate prediction model training method as described above, or to implement the product recommendation method as described above.

[0013] A fourth aspect of the embodiments of the present application provides a computer-readable storage medium, where the computer-readable storage medium stores at least one instruction, and when the at least one instruction is executed by a processor, it implements the click-through rate prediction model training method as described above, or implements the product recommendation method as described above.

[0014] In the embodiments of the present application, by fusing the first ID of the first product with the second ID of the second product to obtain a feature vector; based on the feature vector, user feature information, and store feature information, a high-quality sample data set is constructed, and by training the preset model until the preset conditions are met, a click-through rate that can estimate the user's clicks on the first product and the second product is obtained. This makes the click-through rate of the user clicking on the first product and the second product more accurate and improves the accuracy of determining the click-through rate. Description of the Drawings

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained according to these drawings.

[0016] Figure 1 Schematic diagram of the application environment of a click-through rate prediction model method provided by an embodiment of the present application.

[0017] Figure 2 Schematic flowchart of the click-through rate prediction model training method provided by an embodiment of the present application.

[0018] Figure 3 Schematic diagram of the fusion framework provided by an embodiment of the present application.

[0019] Figure 4 Flowchart of the product recommendation method provided by an embodiment of the present application.

[0020] Figure 5 Schematic diagram of the graphical user interface provided by an embodiment of the present application.

[0021] Figure 6 Schematic diagram of the click-through rate prediction model training device provided by an embodiment of the present application.

[0022] Figure 7 Schematic diagram of the product recommendation device provided by an embodiment of the present application.

[0023] Figure 8 Schematic diagram of the structure of an electronic device provided by an embodiment of the present application. Detailed implementation manners

[0024] In order to make the objectives, technical solutions, and advantages of the present application clearer, the present application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] It should be noted that, in the present application, "at least one" means one or more, and "a plurality" means two or more than two. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, A and B exist simultaneously, and B exists alone, where A and B may be singular or plural. The terms "first", "second", "third", "fourth", etc. (if any) in the specification, claims, and drawings of the present application are used to distinguish similar objects, rather than to describe a specific order or sequence.

[0026] In the embodiments of the present application, words such as "exemplary" or "for example" are used to represent examples, illustrations, or explanations. Any embodiment or design solution described as "exemplary" or "for example" in the embodiments of the present application should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Specifically, using words such as "exemplary" or "for example" is intended to present related concepts in a specific manner. Without conflict, the following embodiments and the features in the embodiments may be combined with each other.

[0027] Cross - border e - commerce can be operated through independent websites. Users can purchase different products on different independent websites. The user interface (UI) design provided by the stores running on independent websites is flexible and diverse, which is a key factor in measuring the personalized experience of the store. Due to the fact that this kind of interface design relies too much on the personal experience and subjectivity of merchants or professional UI designers and there is no unified evaluation standard, there may be frequent design changes, resulting in high labor costs and low efficiency of interface design.

[0028] Merchants can convert the independent website into a mobile application (App). Users can then install the App on their mobile phones and access the store by clicking on the App icon. When the App is running in the foreground of the electronic device, the electronic device can display the application interface of the App on the display screen, that is, the store page of the independent - website store. Users can interact with the App through the controls in the application interface of the App. In addition, merchants can also maintain the website form of the independent website, and users can access the store by entering the website address or clicking on the bookmark through the mobile - phone browser.

[0029] Among them, the application interface refers to the media interface for interaction and information exchange between the application program or operating system and the user, which can realize the conversion between the internal form of information and the form that the user can receive. The application interface is the source code written in specific computer languages such as Java and Extensible Markup Language (XML). The interface source code is parsed and rendered on the electronic device and finally presented as content that the user can recognize. The common manifestation form of the application interface is the graphic user interface (GUI), which refers to the application interface related to computer operation displayed in a graphic way. It can be visual interface elements such as text, icons, buttons, menus, tabs, text boxes, dialog boxes, status bars, navigation bars, Widgets, etc. displayed on the display screen of the electronic device. To solve the technical problem of being unable to accurately predict the click - through rate of combined products, the embodiments of this application provide a method for training a click - through rate prediction model, a product recommendation method, device, and medium, which can obtain a more accurate click - through rate of users clicking on the first product and the second product, improving the accuracy of determining the click - through rate. First, the application scenario is introduced below.

[0030] Figure 1 It is a schematic diagram of the application environment of the click - through rate prediction model training method provided by the embodiments of this application. As Figure 1As shown, the click-through rate prediction model training method and product recommendation method described in this application are applied to an e-commerce platform 10, which can be used to provide merchants' products and services to consumers. Although this disclosure contemplates using devices, systems, and processes to purchase products and services, for simplicity, the description herein will refer to products. Throughout this disclosure, all references to products should also be understood as references to products and / or services, including physical products, digital content, tickets, subscriptions, services to be provided, and so on.

[0031] "Merchants" and "consumers" can be understood as users who support the e-commerce environment, and all references to merchants and consumers throughout this disclosure should also be understood as references to the following: merchant users, prospective users, service provider users, corporate or enterprise users, information technology users, computing entity users, and so on.

[0032] The e-commerce platform 10 can provide a centralized system for providing online resources and facilities to merchants for managing their businesses. The facilities described herein can be partially or fully deployed by a machine that executes computer software, modules, program code, and / or instructions on one or more processors that can be part of the e-commerce platform 10 or external to the e-commerce platform 10. Merchants can utilize the e-commerce platform 10 to manage commerce with consumers, such as by interacting with consumers via communication 120 of the e-commerce platform 10, or any combination thereof.

[0033] The online store 110 can represent a multi-tenant facility that includes multiple virtual storefronts. In an embodiment, a merchant can manage one or more storefronts in the online store 110, such as via a merchant device 20 (e.g., a computer, laptop, mobile computing device, etc.).

[0034] In some embodiments, the e-commerce platform 10 may be implemented by a processing facility including a processor and a memory, the processing facility storing a set of instructions which when executed cause the e-commerce platform 10 to perform the e-commerce and support functions as described herein. The processing facility may be part of a server, a consumer device, a network infrastructure, a mobile computing platform, a cloud computing platform, a fixed computing platform, or other computing platform, and provides an electronic connection and communication among and between the electronic components of the e-commerce platform 10, the merchant device 20, the payment gateway, the application developer, the channel, the shipping provider, the consumer device 30, the point-of-sale device, and so on. The e-commerce platform 10 may be implemented as a cloud computing service, software as a service (SaaS), infrastructure as a service (IaaS), platform as a service (PaaS), desktop as a service (DaaS), managed software as a service (MSaaS), mobile backend as a service (MBaaS), information technology management as a service (ITMaaS), and so on. In some embodiments, elements of the e-commerce platform 10 may be implemented to operate on various platforms and operating systems such as iOS, Android, on the Web, and so on (e.g., the administrator 130 is implemented in multiple instances of a given online store for iOS, Android, and for the Web, each instance having similar functionality).

[0035] In some embodiments, the online store 110 can serve consumer devices 30 via a web page provided by the server of the e-commerce platform 10. The server can receive a request for the web page from a browser or other application installed on the consumer device 30, where the browser (or other application) connects to the server via an IP address that is obtained by translating a domain name. In return, the server sends back the requested web page. The web page can be written in Hypertext Markup Language (HTML), a templating language, JavaScript, etc., or any combination thereof, or the web page includes the above languages. For example, HTML is a computer language that describes the static information of a web page, such as the layout, format, and content of the web page. Website designers and developers can use a templating language to build web pages that combine static content (which is the same across multiple pages) and dynamic content (which varies from one page to the next). The templating language can make it possible to reuse the static elements that define the web page layout while dynamically populating the page with data from the online store. The static elements can be written in HTML, and the dynamic elements can be written in the templating language. The templating language elements in the file can act as placeholders so that the code in the file is compiled and sent to the consumer device 30, and then the templating language is replaced with data from the online store 110, such as when installing a theme. Templates and themes can take into account tags, objects, and filters. The consumer-side device web browser (or other application) then displays the page accordingly.

[0036] In some embodiments, a merchant can customize its online store 110 using a merchant-configurable domain name, a customizable HTML theme, etc. The merchant can customize the look and feel of its website via a theme system, such as where the merchant can select and change the look and feel of its online store 110 by changing its theme while having the same underlying product and business data shown within the product hierarchy of the online store. The theme can be further customized via a theme editor, which is a design interface that enables a user to flexibly customize the design of its website. The theme can also be customized using theme-specific settings that can change aspects such as specific colors, fonts, and pre-built layout schemes. The online store can implement a content management system for website content. The merchant can write blog posts or static pages and publish them to its online store 110, such as by publishing via a blog, article, etc., and configure the navigation menu. The merchant can upload images (e.g., images of products), videos, content, data, etc. to the e-commerce platform 10, such as for the system to store (e.g., store as data 140).

[0037] As described herein, the e-commerce platform 10 can provide trading facilities for products to merchants through multiple different channels including by telephone and through the physical POS devices described herein, through an online store 110. The e-commerce platform 10 can include business support services 150, an administrator 130, etc. associated with running an online business. The business management engine 160 includes the basic or "core" functions of the e-commerce platform 10, and as such, not all functions that support the online store 110 may be suitable for inclusion, as described herein.

[0038] The e-commerce platform 10 includes a click-through rate prediction engine 170. The click-through rate prediction engine 170 is an example of a computer-implemented system that obtains and fuses a first ID of a first product and a second ID of a second product, and trains the click-through rate of a combined recommended product of the first product and the second product that a user clicks based on the fusion result, user characteristic information, and store characteristic information. In an example, during the process of a merchant recommending a first product and a second product to a user, the merchant device 20 sends the first ID of the first product and the second ID of the second product to the click-through rate prediction engine 170. The click-through rate prediction engine 170 analyzes the click-through rate of the combined recommended product of the first product and the second product that the user clicks, and recommends products to the user based on the click-through rate.

[0039] The click-through rate prediction engine 170 includes a processor 171 and a memory 172. The processor 171 can be implemented by one or more processors that execute instructions stored in the memory 172. Alternatively, some or all of the processor 171 can be implemented using dedicated circuits, such as application-specific integrated circuits (ASICs), graphics processing units (GPUs), or programmed field-programmable gate arrays (FPGAs). The memory 172 stores a click-through rate prediction model. The memory 172 also stores a click-through rate prediction model generator that facilitates the generation of the click-through rate prediction model. Once the click-through rate prediction model is generated, the click-through rate prediction model can be updated periodically, for example, when market trends change and / or when new data becomes available.

[0040] In the case where the click-through rate prediction model is a machine learning algorithm or includes a machine learning algorithm, the click-through rate prediction model generator includes a training algorithm to train the machine learning algorithm. The training algorithm can involve, for example, supervised learning, unsupervised learning, or reinforcement learning. The training algorithm can be stored as instructions executed by the processor 171.

[0041] The memory 172 stores buried point data used by the click-through rate prediction model generator to train the machine learning algorithm. The buried point data is obtained from the store pages of the existing online store 110 in the e-commerce platform 10 and from the consumer device 30.

[0042] In some cases, the buried point data is divided into a training data set and a test data set. The click-through rate prediction model generator uses the training data to train a machine learning algorithm that implements the click-through rate prediction model. The test data set is used to test the accuracy of the machine learning algorithm after training.

[0043] It can be understood that the application scenarios illustrated in the embodiments of the present application do not constitute specific limitations. In other embodiments of the present application, the e-commerce platform 10 may further include more or fewer components, such as a network interface, a domain, a payment gateway, etc.

[0044] Exemplarily, taking an e-commerce application as an example for illustration. When a user performs interaction operations such as clicking, adding to the cart, purchasing, favoriting, and liking the first product in the graphical user interface displaying the first product through a client (such as the consumer device 30) installed with the e-commerce application, and the e-commerce application recommends a second product for the user to select, the user can perform interaction operations such as clicking, adding to the cart, purchasing, favoriting, and liking the first product in the graphical user interface displaying the first product and the second product. The e-commerce platform 10 collects buried point information related to the first product and the second product. The buried point information includes user characteristic information, first information corresponding to the first product, second information corresponding to the second product, and store characteristic information. The first information includes at least a first ID, and the second information includes at least a second ID. Subsequently, the e-commerce platform 10 fuses the first product and the second product based on the first ID and the second ID to obtain a feature vector, and then constructs a sample data set based on the feature vector, user characteristic information, and store characteristic information; trains a preset model using the sample data set until the preset model meets the preset conditions to obtain a click-through rate prediction model; the e-commerce platform 10 uses the click-through rate prediction model to determine the probability that the target user clicks on the first product and the second product; the e-commerce platform 10 determines the first product and the second product as combined recommended products according to this probability, and recommends the combined recommended products to the target user. In this way, the e-commerce platform 10 can determine the probability that the user clicks on the first product and the second product according to the user characteristic information and the fusion feature information of the first product and the second product, and then recommend the combined recommended products after combining the first product and the second product to the user according to this probability, so as to recommend combined products that meet the expectations to the user.

[0045] The following will take the computer program product running on an e-commerce platform (such as Figure 1 the e-commerce platform 10) as an example for illustration. Please refer to Figure 2 As shown, it is a schematic flowchart of the click-through rate prediction model training method provided by the embodiments of the present application. In an embodiment of the present application, it includes the following steps: S201: Collect buried point information, where the buried point information includes the first ID of the first product and the second ID of the second product.

[0046] In the embodiments of the present application, clickstream data is collected, and the clickstream data includes the first information of the first product and the second information of the second product. For example, the exposure information and click information of the combination pair composed of the first product and the second product. Among them, the first product and the second product can be products. For example, when the first product is a mobile phone, the second product can be headphones; when the first product is a top, the second product can be pants or a skirt; when the first product is a tissue, the second product can be laundry detergent. The first information includes the first ID and the first attributes of the first product; the second information includes the second ID and the second attributes of the second product. The first attributes may include information such as the type of the first product, the link to the product details page, and parameters; the second attributes may include information such as the type of the second product, the link to the product details page, and parameters. For example, when the first product is a camera, the first attributes may include that the product type corresponding to the camera is an electronic product, and may also include the brand information of the camera and the parameter information of the camera. The user can perform a click operation on the attribute information of the first product to enable the server to display the product details page of the first product, and the product details page of the first product includes the product details information of the first product. The user clicks to view the second product, which means that the user performs a click operation on the attribute information of the second product to enable the server to display the product details information of the second product.

[0047] In the embodiments of the present application, the clickstream data may further include user characteristic information and store characteristic information. Among them, the user characteristic information includes the user's account information, user age, user gender, user purchasing power, user's historical purchase records, user preference for specific product categories, preference tags, click history, etc. Of course, the user characteristic information may further include other content, and the embodiments of the present invention do not make specific limitations thereto. The store characteristic information includes store level, operating country, operating category, etc. Of course, the store characteristic information may further include other content, and the embodiments of the present invention do not make specific limitations thereto.

[0048] S202: Fuse the first product and the second product based on the first ID and the second ID to obtain a feature vector.

[0049] In the embodiments of the present application, in order to estimate the click-through rate after combining the first product and the second product, the first product and the second product can be fused first to obtain a feature vector. Among them, the first product and the second product can be fused through a fusion framework. As Figure 3 shown, it is a schematic diagram of the fusion framework provided by the embodiments of the present application.

[0050] In the embodiment of the present application, the fusion framework includes a first encoder, a second encoder, and a cross-fusion encoder. Among them, the first encoder is used to encode the information corresponding to the first commodity; the second encoder is used to encode the information corresponding to the second commodity; the cross-fusion encoder is used to fuse the processing results of the first encoder and the second encoder to obtain the feature vectors corresponding to the first commodity and the second commodity.

[0051] In the embodiment of the present application, the first encoder is a twelve-layer network structure, including six self-attention layers and six feed-forward networks. Each self-attention layer is one layer, and each feed-forward network is one layer. Among them, a feed-forward network is connected after each self-attention layer. The structure of the second encoder is the same as that of the first encoder and will not be described in detail here.

[0052] In the embodiment of the present application, the cross-fusion encoder is a six-layer network structure, and each layer includes self-attention, cross-attention, and a feed-forward network. Each layer is connected to the corresponding layer of the first encoder and the second encoder through a bridging layer. For example, the cross-fusion encoder in the embodiment of the present application includes a first self-attention layer, a second self-attention layer, a first cross-attention layer, a second cross-attention layer, a first feed-forward network, and a second feed-forward network.

[0053] In the embodiment of the present application, the fusion framework further includes a first bridging layer and a second bridging layer. After the top-layer network of the first encoder is connected to the first bridging layer, it is then connected to the first self-attention layer of the cross-fusion encoder, and the first self-attention layer is connected to the first feed-forward network through the first cross-attention layer; after the top-layer network of the second encoder is connected to the second bridging layer, it is then connected to the second self-attention layer of the cross-fusion encoder, and the second self-attention layer is connected to the second feed-forward network through the second cross-attention layer.

[0054] In some embodiments of the present application, after the top-layer network of the first encoder is connected to the first bridging layer, it can also be connected to the first cross-attention layer, and the first cross-attention layer is connected to the first feed-forward network; after the top-layer network of the second encoder is connected to the second bridging layer, it can also be connected to the second cross-attention layer, and the second cross-attention layer is connected to the second feed-forward network.

[0055] In the embodiment of the present application, fusing the first commodity and the second commodity based on the first ID and the second ID to obtain a feature vector includes: obtaining a first vector based on the first ID; obtaining a second vector based on the second ID; and fusing the first vector and the second vector through the fusion framework to obtain a feature vector.

[0056] In an embodiment of the present application, the fusion framework further includes an embedding layer. The embedding layer of the fusion framework can be used to vectorize the first ID to obtain the features of the first commodity represented in vector form; or the embedding layer of the fusion framework can be used to vectorize the second ID to obtain the features of the second commodity represented in vector form. Thus, processing the first ID through the embedding layer can obtain a first vector; processing the second ID through the embedding layer can obtain a second vector; and the fusion framework is used to fuse the first vector and the second vector to obtain a feature vector.

[0057] In an embodiment of the present application, fusing the first vector and the second vector through the fusion framework to obtain a feature vector includes: reducing the dimension of the first vector through the first encoder in the fusion framework to obtain a first dimension-reduced vector; reducing the dimension of the second vector through the second encoder in the fusion framework to obtain a second dimension-reduced vector; and processing the first dimension-reduced vector and the second dimension-reduced vector through the cross-fusion encoder in the fusion framework to obtain a feature vector.

[0058] In an embodiment of the present application, processing the first dimension-reduced vector and the second dimension-reduced vector through the cross-fusion encoder in the fusion framework to obtain a feature vector includes: projecting the first dimension-reduced vector into a first intermediate feature vector through the first bridging layer in the fusion framework; projecting the second dimension-reduced vector into a second intermediate feature vector through the second bridging layer in the fusion framework; dynamically fusing the first intermediate feature vector and the second dimension-reduced vector through the first cross-attention layer in the cross-fusion encoder to obtain a first feature vector corresponding to the first commodity; and dynamically fusing the second intermediate feature vector and the first dimension-reduced vector through the second cross-attention layer in the cross-fusion encoder to obtain a second feature vector corresponding to the second commodity.

[0059] In an embodiment of the present application, the output of the i-th layer of the first encoder will be projected to the input of the i-th layer of the cross-fusion encoder through the first bridging layer, so that the local features of the first commodity are injected into the cross-fusion encoder, and the characteristics of the first commodity are determined to be retained during fusion. Similarly, the output of the j-th layer of the second encoder will be projected to the input of the j-th layer of the cross-fusion encoder through the second bridging layer, so that the local features of the second commodity are injected into the cross-fusion encoder, and the characteristics of the second commodity are determined to be retained during fusion.

[0060] In the embodiments of the present application, the first intermediate feature vector and the second dimensionality-reduced vector are dynamically fused through the first cross-attention layer in the cross-fusion encoder to obtain the first feature vector corresponding to the first product. The detailed features (such as color and texture) of the first product and the second product can be fused through the low-level network in the cross-fusion encoder, and then the semantic features (such as category and function) of the first product and the second product can be fused through the high-level network in the cross-fusion encoder. Then, the features obtained by processing the first product through the first encoder through the first bridging layer and the features obtained by processing the second product through the second encoder through the second bridging layer are respectively fused with the cross-fusion encoder, and the deep interaction features integrating the first product and the second product are output through the top-level network of the cross-fusion encoder, enabling effective bottom-up cross-product alignment and fusion between products, which is convenient for subsequent click-through rate prediction.

[0061] In the embodiments of the present application, in order to enhance the expression of the key features of the first product and the second product, suppress noise, and improve the feature robustness, the first ID and the second ID can also be fused through the self-attention layer to obtain a feature vector. Specifically, the process of obtaining a feature vector by processing the first dimensionality-reduced vector and the second dimensionality-reduced vector through the cross-fusion encoder in the fusion framework may further include: projecting the first dimensionality-reduced vector into a first intermediate feature vector through the first bridging layer in the fusion framework; projecting the second dimensionality-reduced vector into a second intermediate feature vector through the second bridging layer in the fusion framework; enhancing the first context representation of the first intermediate feature vector through the first self-attention layer in the cross-fusion encoder; enhancing the second context representation of the second intermediate feature vector through the second self-attention layer in the cross-fusion encoder; dynamically fusing the first context representation and the second dimensionality-reduced vector through the first cross-attention layer in the cross-fusion encoder to obtain the first feature vector corresponding to the first product; and dynamically fusing the second context representation and the first dimensionality-reduced vector through the second cross-attention layer in the cross-fusion encoder to obtain the second feature vector corresponding to the second product.

[0062] In the embodiments of the present application, a feature vector for fusing the first product and the second product can be obtained based on the first feature vector and the second feature vector. For example, by concatenating the first feature vector and the second feature vector, a feature vector for fusing the first product and the second product can be obtained.

[0063] S203: Construct a sample data set based on the feature vector, user feature information, and store feature information.

[0064] In an embodiment of the present application, a sample data set is constructed based on feature vectors, user feature information, and store feature information. Among them, the sample data set includes multiple groups of sample data, and each group of sample data includes a feature vector, user feature information, and store feature information. The feature vector is a fusion feature vector of a first product and a second product.

[0065] In an embodiment of the present application, click-through information can be collected based on the information viewed by multiple users in historical periods. For example, assume that when user 1 views the first product in a historical period, the electronic device displays five second products that can be paired with the first product, and the user performs a click operation on the second product to view the detailed information of the second product. Accordingly, 5 groups of sample data can be determined, and the 5 groups of sample data can be as shown in Table 1: Table 1 Among them, each row in the table can be used as a piece of labeled sample data. Table 1 only shows 5 groups of sample data by way of example and does not limit the sample data. In actual application, the number of groups of sample data obtained can be several thousand groups, tens of thousands of groups, etc.

[0066] S204: Train a preset model using the sample data set until the preset model meets the preset conditions to obtain a click-through rate prediction model.

[0067] In an embodiment of the present application, after the sample data set is constructed, the sample data set can be used to train a preset model until the preset model meets the preset conditions to obtain a click-through rate prediction model. Specifically, a first fusion feature vector is obtained based on the first feature vector of the first product, user feature information, and store feature information; a second fusion feature vector is obtained based on the second feature vector of the second product, user feature information, and store feature information; the preset model is trained based on the first fusion feature vector and the second fusion feature vector until the preset model meets the preset conditions to obtain a click-through rate prediction model.

[0068] In an embodiment of the present application, training the preset model based on the first fusion feature vector and the second fusion feature vector until the preset model meets the preset conditions to obtain a click-through rate prediction model includes: predicting the click-through rates of the user clicking on the first product and the second product based on the first fusion feature vector and the second fusion feature vector; determining the total loss value according to the annotation label and click-through rate of the click-through information; training the preset model according to the total loss value until the preset model meets the preset conditions to obtain a click-through rate prediction model.

[0069] In this embodiment, the annotation label can be represented by a preset value. For example, 0 or 1. For example, the annotation label "1" can be used to represent that the user clicks on the first product and the second product, and the annotation label "0" can be used to represent that the user will not click on the first product or the second product.

[0070] In this embodiment, since the total loss value is calculated from the true annotation labels of the buried point information and the predicted click-through rate, the total loss value can reflect the gap between the prediction of the model and the true annotation labels.

[0071] In some embodiments of the present application, the loss value can be calculated by various methods based on the annotation labels and the click-through rate, and the present application does not limit the type of the loss value. For example, the loss value can be the cross-entropy loss.

[0072] For example, the calculation method of the loss value can refer to the following formula: ; where L represents the loss value, represents the number of annotation labels, represents the th annotation label among the N annotation labels, represents the click-through rate corresponding to the th annotation label among the N click-through rates.

[0073] In this embodiment, since the loss value is calculated from the annotation labels and the corresponding click-through rates, the loss value can reflect the gap between the prediction of the model and the true annotation labels.

[0074] In some embodiments of the present application, the preset model can be a wide-deep model, and the process of training the preset model includes a forward propagation process and a backward propagation process. Among them, the wide-deep model includes a wide model and a deep model. In some embodiments, the framework for offline training of the preset model can adopt the parameter server architecture of TensorFlow. The model parameters are stored on the training server, and the training server can be an electronic device with a Central Processing Unit (CPU). The training process can be implemented by the worker side, and the worker side is an electronic device with a Graphics Processing Unit (GPU). During the training process, the worker side obtains the latest model parameters from the server, then performs forward and backward propagation in parallel on multiple GPU cards of the local machine, then integrates the gradients on multiple GPU cards, and finally calls the optimization algorithm to send the integrated gradients back to the server. An asynchronous update strategy is adopted among the workers.

[0075] In some embodiments of the present application, the training process of the model can use a supervised training method. Specifically, the training method may include: dividing the sample data set into a training data set and a test data set, and using the data of the current batch in the training data set to iteratively update and train a preset model; using the test data set to determine whether the preset model updated each time meets the preset conditions. If the preset conditions are not met, the next batch of data is used to perform the next update on the preset model according to an optimization algorithm (such as the gradient descent algorithm); repeating the above steps until the preset model meets the preset conditions. Among them, the preset conditions may be that the loss value corresponding to the loss function is less than a preset loss threshold, the model performance (such as accuracy, recall rate, etc.) reaches a preset performance threshold, the number of iterations of the model reaches a preset number threshold, etc. Among them, the above various thresholds can be set according to actual needs, and the present application does not make specific limitations on this.

[0076] The method provided by the embodiments of the present application fuses the first ID of the first commodity and the second ID of the second commodity to obtain a feature vector; based on the feature vector, user feature information, and store feature information, a high-quality sample data set is constructed, and by training the preset model until the preset conditions are met, a model capable of predicting the click-through rate of the user on the first commodity and the second commodity is obtained. According to this click-through rate, a commodity combination composed of the first commodity and the second commodity is recommended to the user, thereby improving the accuracy of commodity combination recommendation and providing a more personalized and demand-satisfying shopping recommendation service for the user.

[0077] After the click-through rate prediction model is trained, the click-through rate prediction model can be used for commodity recommendation. The commodity recommendation method provided by the embodiments of the present application can run on a user terminal (such as Figure 1 the consumer device 30 shown).

[0078] Figure 4 It is a flowchart of the commodity recommendation method provided by the embodiments of the present application. The commodity recommendation method specifically includes the following steps. According to different requirements, the order of the steps in this flowchart can be changed, and some can be omitted.

[0079] S401, obtain target user feature information.

[0080] In the embodiments of the present application, the user feature information includes the user's account information, user age, user gender, user purchasing power, user's historical purchase records, user's preferred sub-categories, preference tags, click history, etc.

[0081] S402, obtain the first ID of the first commodity to be recommended and the second ID of the second commodity.

[0082] In an embodiment of the present application, when a user inputs relevant search information to obtain a corresponding first commodity, first information of the first commodity is acquired. The first information includes information such as the first ID of the first commodity, the type to which the first commodity belongs, the link to the detail page, and parameters. When the user selects the first commodity and enters the commodity shopping cart page and / or the commodity settlement page, a second commodity is recommended to the user, and second information of the second commodity is acquired. The second information includes information such as the second ID of the second commodity, the type to which the second commodity belongs, the link to the detail page, and parameters.

[0083] S403. Based on the target user characteristic information, the first ID, and the second ID, use a click-through rate prediction model to determine the probabilities that the target user clicks to view the first commodity and the second commodity.

[0084] In an embodiment of the present application, based on the target user characteristic information, the first ID, and the second ID, use a click-through rate prediction model to determine the probabilities that the target user clicks to view the first commodity and the second commodity. The training method of the click-through rate prediction model refers to the process above Figure 2 and will not be elaborated here.

[0085] S404. Determine the first commodity and the second commodity as combined recommended commodities according to the probabilities, and recommend the combined recommended commodities to the target user.

[0086] In some embodiments of the present application, the first commodity and the second commodity can be determined as combined recommended commodities according to the probabilities. For example, if the probability is greater than or equal to a preset probability, it is determined that the probability that the target user purchases the combined recommended commodities is relatively high, and the combined recommended commodities are recommended to the target user; if the probability is less than the preset probability, it is determined that the probability that the target user purchases the combined recommended commodities is relatively low, and the combined recommended commodities are not recommended to the target user.

[0087] In some embodiments, the second commodity can also be correspondingly displayed at a preset position of the first commodity in the graphical user interface. The preset position can be set according to actual needs, and information such as the image and title of the recommended second commodity can be displayed at the preset position together. For example Figure 5 as shown, a combined recommended commodity composed of a first commodity being a dress and a second commodity being overalls is simultaneously displayed in the graphical user interface for the target user to select.

[0088] The method provided by the embodiment of the present application can be applicable to a commodity recommendation scenario where the user has a purchase intention, such as a commodity shopping cart page and a commodity settlement page, and can achieve fast and accurate recommendation of complementary commodities for the current commodity.

[0089] Figure 6It is a structural diagram of a click-through rate prediction model training device provided by an embodiment of the present application. The click-through rate prediction model training device 600 may include multiple functional modules composed of computer program segments. The computer programs of each program segment in the click-through rate prediction model training device 600 may be stored in the memory of a computer device and executed by at least one processor to perform the function of model training (see details in Figure 2 description).

[0090] In this embodiment, according to the functions it performs, the click-through rate prediction model training device 600 may be divided into multiple functional modules. The functional modules may include: an acquisition module 601 and a processing module 602. As used in this application, a module refers to a series of computer program segments that can be executed by at least one processor and can complete a fixed function, and are stored in the memory. In this embodiment, the click-through rate prediction model training device 600 may be used to implement the click-through rate prediction model training method as Figure 2 shown. As Figure 6 shown, the click-through rate prediction model training device 600 is applied to a computer device (such as the server as Figure 1 shown), and this click-through rate prediction model training device 600 includes: The acquisition module 601 is used to acquire buried point information, where the buried point information includes the first ID of the first commodity and the second ID of the second commodity; the processing module 602 is used to fuse the first commodity and the second commodity based on the first ID and the second ID to obtain a feature vector; the processing module 602 is used to construct a sample data set based on the feature vector, as well as the user feature information and store feature information in the buried point information; the processing module 602 is used to train a preset model using the sample data set until the preset model meets the preset conditions to obtain the click-through rate prediction model.

[0091] In some embodiments of the present application, the processing module 602 is further used to obtain a first vector based on the first ID; obtain a second vector based on the second ID; and fuse the first vector and the second vector through a fusion framework to obtain the feature vector.

[0092] In some embodiments of the present application, the processing module 602 is further used to perform dimensionality reduction processing on the first vector through the first encoder in the fusion framework to obtain a first dimensionality reduction vector; perform dimensionality reduction processing on the second vector through the second encoder in the fusion framework to obtain a second dimensionality reduction vector; and process the first dimensionality reduction vector and the second dimensionality reduction vector through the cross-fusion encoder in the fusion framework to obtain the feature vector.

[0093] In some embodiments of the present application, the processing module 602 is further configured to project the first dimensionality-reduced vector into a first intermediate feature vector through the first bridging layer in the fusion framework; project the second dimensionality-reduced vector into a second intermediate feature vector through the second bridging layer in the fusion framework; dynamically fuse the first intermediate feature vector and the second dimensionality-reduced vector through the first cross-attention layer in the cross-fusion encoder to obtain a first feature vector corresponding to the first commodity; and dynamically fuse the second intermediate feature vector and the first dimensionality-reduced vector through the second cross-attention layer in the cross-fusion encoder to obtain a second feature vector corresponding to the second commodity.

[0094] In some embodiments of the present application, the processing module 602 is further configured to project the first dimensionality-reduced vector into a first intermediate feature vector through the first bridging layer in the fusion framework; project the second dimensionality-reduced vector into a second intermediate feature vector through the second bridging layer in the fusion framework; enhance the first context representation of the first intermediate feature vector through the first self-attention layer in the cross-fusion encoder; enhance the second context representation of the second intermediate feature vector through the second self-attention layer in the cross-fusion encoder; dynamically fuse the first context representation and the second dimensionality-reduced vector through the first cross-attention layer in the cross-fusion encoder to obtain a first feature vector corresponding to the first commodity; and dynamically fuse the second context representation and the first dimensionality-reduced vector through the second cross-attention layer in the cross-fusion encoder to obtain a second feature vector corresponding to the second commodity.

[0095] In some embodiments of the present application, the processing module 602 is further configured to obtain a first fusion feature vector based on the first feature vector, the user feature information, and the store feature information; obtain a second fusion feature vector based on the second feature vector, the user feature information, and the store feature information; and train the preset model based on the first fusion feature vector and the second fusion feature vector until the preset model meets the preset conditions to obtain a click-through rate prediction model.

[0096] In some embodiments of the present application, the processing module 602 is further configured to predict the click-through rates of the user to view the first commodity and the second commodity based on the first fusion feature vector and the second fusion feature vector; determine a loss value according to the annotation label of the buried point information and the click-through rate; and train the preset model according to the loss value until the preset model meets the preset conditions to obtain the click-through rate prediction model.

[0097] Figure 7It is a structural diagram of a product recommendation device provided by an embodiment of the present application. The product recommendation device 700 may include multiple functional modules composed of computer program segments. The computer programs of each program segment in the product recommendation device 700 may be stored in the memory of the computer device and executed by at least one processor to execute (see details in Figure 4 description) the product recommendation function.

[0098] In this embodiment, according to the functions it executes, the product recommendation device 700 may be divided into multiple functional modules. The functional modules may include: an acquisition module 701 and a processing module 702. A module referred to in the present application means a series of computer program segments that can be executed by at least one processor and can complete fixed functions, and are stored in the memory.

[0099] In this embodiment, the product recommendation device 700 may be used to implement the product recommendation method as shown in Figure 4 . As shown in Figure 7 , the product recommendation device 700 is applied to a computer device. The product recommendation device 700 includes: an acquisition module 701 for acquiring target user feature information; the acquisition module 701 is further used to acquire the first ID of the first product to be recommended and the second ID of the second product; a processing module 702 is used to determine the probabilities that the target user clicks on the first product and the second product based on the target user feature information, the first ID, and the second ID by using a click-through rate prediction model; and determine the first product and the second product as combined recommended products according to the probabilities, and recommend the combined recommended products to the target user.

[0100] Another embodiment of the present application further provides an electronic device. Figure 1 The application environment shown is only an example. In some other exemplary embodiments, the computer program product for implementing the click-through rate prediction model training and product recommendation method of the embodiments of the present application may also run on any electronic device with sufficient computing power (such as the electronic device shown in Figure 8 ) to execute each step of the click-through rate prediction model training and product recommendation method, so as to provide the functions of click-through rate prediction model training and product recommendation.

[0101] Please refer to Figure 8 , which is a schematic structural diagram of an electronic device provided by an embodiment of the present application. As shown in Figure 8As shown, in an embodiment of the present application, the electronic device 1000 can be a mobile phone, a tablet computer, a smart wearable device, an Augmented Reality (AR) / Virtual Reality (VR) device, a laptop computer, a netbook, etc. The embodiments of the present application do not impose any restrictions on the specific type of the electronic device 1000.

[0102] As Figure 8 shown, the electronic device 1000 may include, but is not limited to, a communication module 101, a memory 102, a processor 103, an Input / Output (I / O) interface 104, and a bus 105. The processor 103 is respectively coupled to the communication module 101, the memory 102, and the I / O interface 104 through the bus 105.

[0103] Those skilled in the art can understand that the schematic diagram is only an example of the electronic device 1000, and does not constitute a limitation on the electronic device 1000. It may include more or fewer components than shown in the figure, or combine certain components, or different components. For example, the electronic device 1000 may further include a network access device, etc.

[0104] The communication module 101 may include a wired communication module and / or a wireless communication module. The wired communication module may provide one or more of the solutions for wired communication such as Universal Serial Bus (USB), Controller Area Network (CAN), etc. The wireless communication module may provide one or more of the solutions for wireless communication such as Wireless Fidelity (Wi-Fi), Bluetooth (BT), a mobile communication network, Frequency Modulation (FM), near field communication (NFC), Infrared (IR) technology, etc.

[0105] The memory 102 can be used to store computer-readable instructions and / or modules. By running or executing the computer-readable instructions and / or modules stored in the memory 102, and by invoking the data stored in the memory 102, the processor 103 realizes various functions of the electronic device 1000. The memory 102 mainly includes a program storage area and a data storage area. Among them, the program storage area can store an operating system, application programs required for at least one function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created according to the use of the electronic device 1000. The memory 102 can include non-volatile and volatile memories, such as: hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one magnetic disk storage device, flash memory device, or other storage devices.

[0106] The memory 102 can be an external memory and / or an internal memory of the electronic device 1000. Further, the memory 102 can be a memory in physical form, such as a memory stick, a TF card (Trans-flash Card), and so on.

[0107] The processor 103 can be a central processing unit (CPU), or can also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The processor 103 is the operation core and control center of the electronic device 1000, connecting various parts of the entire electronic device 1000 through various interfaces and lines, and executing the operating system of the electronic device 1000 and various installed application programs, program codes, etc.

[0108] Exemplarily, the computer-readable instructions can be divided into one or more modules / sub-modules / units. One or more modules / sub-modules / units are stored in the memory 102 and executed by the processor 103 to complete the present application. One or more modules / sub-modules / units can be a series of computer-readable instruction segments capable of performing specific functions, and these computer-readable instruction segments are used to describe the execution process of the computer-readable instructions in the electronic device 1000. For example, the computer-readable instructions can be divided into multiple modules of the above click-through rate prediction model training and application device.

[0109] If the modules / units integrated in the electronic device 1000 are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, to implement all or part of the processes in the above-mentioned embodiment methods of the present application, it can also be completed by computer-readable instructions instructing relevant hardware. The computer-readable instructions can be stored in a computer-readable storage medium. When the computer-readable instructions are executed by the processor, the steps of the above-mentioned various method embodiments can be implemented.

[0110] Among them, the computer-readable instructions include computer-readable instruction codes, and the computer-readable instruction codes can be in the form of source code, object code, executable files, or some intermediate forms, etc. The computer-readable medium can include: any entity or device capable of carrying the computer-readable instruction codes, recording media, USB flash drives, mobile hard disks, magnetic disks, optical disks, computer memories, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory).

[0111] Combined with Figures 2 to 8 , the memory 102 in the electronic device 1000 stores computer-readable instructions, and the processor 103 can execute the computer-readable instructions stored in the memory 102 to implement the click-through rate prediction model training and application method as Figures 2 to 8 shown. Specifically, for the specific implementation method of the above computer-readable instructions by the processor 103, reference can be made to the description of the relevant steps in Figures 2 to 8 the corresponding embodiment, which will not be elaborated here.

[0112] The I / O interface 104 is used to provide a channel for user input or output. For example, the I / O interface 104 can be used to connect various input and output devices, such as a mouse, a keyboard, a touch device, a display screen, etc., so that the user can input information or visualize the information.

[0113] The bus 105 is at least used to provide a communication channel for mutual communication among the communication module 101, the memory 102, the processor 103, and the I / O interface 104 in the electronic device 1000.

[0114] In several embodiments provided in the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of modules is only a logical function division, and there may be other division methods in actual implementation.

[0115] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0116] In addition, in each embodiment of the present application, the functional modules can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of a combination of hardware and software functional modules.

[0117] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present application is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present application. Any reference signs in the claims should not be regarded as limiting the claimed claims.

[0118] In addition, obviously, the term "including" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices can also be implemented by one unit or device through software or hardware. The terms such as "first" and "second" are used to represent names and do not represent any specific order.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit them. Although the present application has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. A click rate prediction model training method, characterized in that: The method comprises: Collecting tracking information, wherein the tracking information includes a first ID of the first product and a second ID of the second product; Based on the first ID and the second ID, the first product and the second product are merged to obtain a feature vector; Based on the feature vector, and the user feature information and the store feature information in the embedding information, construct a sample data set; The preset model is trained using the sample data set until the preset model meets preset conditions to obtain the click rate prediction model.

2. The click rate prediction model training method according to claim 1, characterized in that: The fusing the first product and the second product based on the first ID and the second ID to obtain a feature vector includes: Obtain a first vector based on the first ID; Obtain a second vector based on the second ID; The first vector and the second vector are fused through a fusion framework to obtain the feature vector.

3. The click rate prediction model training method according to claim 2, characterized in that: The fusing the first vector and the second vector through a fusion framework to obtain the feature vector includes: Performing dimensionality reduction processing on the first vector by a first encoder in the fusion framework to obtain a first reduced dimensionality vector; Performing dimensionality reduction processing on the second vector by a second encoder in the fusion framework to obtain a second reduced dimensionality vector; The first reduced dimensionality vector and the second reduced dimensionality vector are processed by a cross-fusion encoder in the fusion framework to obtain the feature vector.

4. The click rate prediction model training method according to claim 3, characterized in that: The processing of the first dimensionality reduction vector and the second dimensionality reduction vector by a cross-fusion encoder to obtain the feature vector comprises: Projecting the first dimension reduction vector into a first intermediate feature vector through a first bridge layer in the fusion framework; Projecting the second dimension reduction vector into a second intermediate feature vector through a second bridge layer in the fusion framework; Dynamically fusing the first intermediate feature vector and the second dimensionality reduction vector through a first cross attention layer in the cross fusion encoder to obtain a first feature vector corresponding to the first product; The second intermediate feature vector and the first dimensionality reduction vector are dynamically fused through a second cross-attention layer in the cross-fusion encoder to obtain a second feature vector corresponding to the second product.

5. The click rate prediction model training method according to claim 3, characterized in that: The processing of the first dimensionality reduction vector and the second dimensionality reduction vector by a cross-fusion encoder to obtain the feature vector comprises: Projecting the first dimension reduction vector into a first intermediate feature vector through a first bridge layer in the fusion framework; Projecting the second dimension reduction vector into a second intermediate feature vector through a second bridge layer in the fusion framework; enhancing a first contextual representation of the first intermediate feature vector by a first self-attention layer in the cross-fusion encoder; enhancing a second contextual representation of the second intermediate feature vector by a second self-attention layer in the cross-fusion encoder; Dynamically fusing the first context representation and the second dimensionality reduction vector through a first cross attention layer in the cross fusion encoder to obtain a first feature vector corresponding to the first product; The second context representation and the first dimensionality reduction vector are dynamically fused through a second cross-attention layer in the cross-fusion encoder to obtain a second feature vector corresponding to the second product.

6. The click rate prediction model training method according to claim 4 or 5, characterized in that: The using the sample data set to train a preset model until the preset model meets a preset condition to obtain a click rate prediction model includes: Obtaining a first fused feature vector based on the first feature vector, the user feature information, and the store feature information; Obtaining a second fused feature vector based on the second feature vector, the user feature information, and the store feature information; The preset model is trained based on the first fused feature vector and the second fused feature vector.

7. The click rate prediction model training method according to claim 6, characterized in that: The training of the preset model based on the first fused feature vector and the second fused feature vector comprises: Predicting the click rate of users clicking to view the first product and the second product based on the first fused feature vector and the second fused feature vector; Determine the loss value according to the annotation label of the embedding point information and the click rate; The preset model is trained according to the loss value until the preset model meets the preset conditions, thereby obtaining the click rate prediction model.

8. A product recommendation method, characterized in that: The product recommendation method comprises: Obtain target user feature information; Obtain the first ID of the first product to be recommended and the second ID of the second product; Based on the target user feature information, the first ID and the second ID, using a click-through rate prediction model to determine the probability that the target user clicks on the first product and the second product, wherein the click-through rate prediction model is obtained by training using the click-through rate prediction model training method according to any one of claims 1 to 7; and The first product and the second product are determined as a combination of recommended products according to the probability, and the combination of recommended products is recommended to the target user.

9. An electronic device, characterized in that: include: Memory, and A processor, wherein the processor executes computer-readable instructions stored in the memory to implement the click-through rate prediction model training method according to any one of claims 1 to 7, or implements the product recommendation method according to claim 8.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one instruction, and when the at least one instruction is executed by the processor, it implements the click-through rate prediction model training method as described in any one of claims 1 to 7, or implements the product recommendation method as described in claim 8.

Images