Click rate prediction model processing method, content recommendation method, device and apparatus

Abstract

This application relates to a click-through rate (CTR) prediction model processing method, content recommendation method, apparatus, and device, belonging to the field of computer technology. The method includes: acquiring training data corresponding to sample content, including location features, non-location features, exposure location tags, and actual click-through rate (CTR); inputting the location features and non-location features into a CTR-location prediction model to obtain predicted CTR and predicted exposure location; calculating a loss function value using a cross-entropy loss function based on the exposure location tags, actual CTR, predicted CTR, and predicted location; training the CTR-location prediction model based on the loss function value; and constructing a CTR prediction model based on the trained CTR-location prediction model. This method optimizes the calculation of the loss function value, enabling the obtained CTR prediction model to improve the accuracy of CTR prediction while eliminating the influence of location bias on CTR prediction.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a click-through rate prediction model processing method, content recommendation method, apparatus and device. Background Technology

[0002] With the continuous development of network and computer technology, more and more users are obtaining information through the Internet, such as text, images, public accounts, videos, news, etc.

[0003] In related technologies, recommendation systems typically train a machine learning model based on the historical display and click data of each piece of content to predict the click rate of a particular user for a particular piece of content. When recommending content to a specific user, the machine learning model is used to predict the probability of each piece of content being clicked by the user, and the content to be pushed to that user is selected based on the predicted probability.

[0004] However, in recommendation systems, content that appears earlier in the list is usually more likely to be clicked than content that appears later in the list. This causes the model, which is trained based on historical information, to have a bias in its perception of user preferences, resulting in inaccurate predicted click-through rates. Summary of the Invention

[0005] This application provides a click-through rate (CTR) prediction model processing method, content recommendation method, apparatus, and device, which can improve the prediction accuracy of the generated CTR prediction model and thus improve the content recommendation effect. The technical solution is as follows:

[0006] On the one hand, a click-through rate prediction model processing method is provided, the method comprising:

[0007] Acquire training data corresponding to the sample content. The training data includes location features, non-location features, exposure location tags, and actual click-through rate. The non-location features include content features and user profile features. The location features are the features of the actual exposure location of the sample content, and the exposure location tags are tags constructed based on the actual exposure location of the sample content.

[0008] The location features and non-location features of the sample content are input into the click-rate-location prediction model to obtain the predicted click-rate and predicted exposure location of the sample content.

[0009] Based on the exposure location label of the sample content, the actual click-through rate, the predicted click-through rate of the sample content, and the predicted location of the sample content, the loss function value is calculated using the cross-entropy loss function.

[0010] The model parameters of the click-rate-location prediction model are updated based on the loss function value in order to train the click-rate-location prediction model.

[0011] Based on the trained click-rate-location prediction model, a click-rate prediction model is constructed; the click-rate prediction model is used to process the non-location features of the content to be recommended to obtain the predicted click-rate of the content to be recommended.

[0012] On the other hand, a content recommendation method is provided, the method comprising:

[0013] In response to receiving a content recommendation request from a target user, the non-location features of the content to be recommended are obtained, including content features and user profile features.

[0014] The non-location features are input into the click-through rate (CTR) prediction model to obtain the predicted CTR of the content to be recommended, output by the CTR prediction model. The CTR prediction model is a model built based on a trained CTR-location prediction model. The training method of the CTR-location prediction model includes: acquiring training data corresponding to the sample content, the training data including location features, non-location features, exposure location tags, and the actual CTR; the non-location features include content features and user profile features; the location features are the features of the actual exposure location of the sample content, and the exposure location tags are tags constructed based on the actual exposure location of the sample content; inputting the location features and the non-location features of the sample content into the CTR-location prediction model to obtain the predicted CTR and the predicted exposure location of the sample content; calculating the loss function value using the cross-entropy loss function based on the exposure location tags, the actual CTR, the predicted CTR, and the predicted location of the sample content; and updating the model parameters of the CTR-location prediction model based on the loss function value.

[0015] Content recommendations are made to the target user based on the predicted click-through rate of the content to be recommended.

[0016] On the other hand, a click-through rate prediction model processing device is provided, the device comprising:

[0017] The training data acquisition module is used to acquire training data corresponding to the sample content. The training data includes location features, non-location features, exposure location tags, and actual click-through rate. The non-location features include content features and user profile features. The location features are the features of the actual exposure location of the sample content, and the exposure location tags are tags constructed based on the actual exposure location of the sample content.

[0018] The prediction information acquisition module is used to input the location features of the sample content and the non-location features of the sample content into the click-rate-location prediction model to obtain the predicted click-rate of the sample content and the predicted exposure position of the sample content.

[0019] The loss function value calculation module is used to calculate the loss function value based on the exposure location label of the sample content, the actual click-through rate, the predicted click-through rate of the sample content, and the predicted location of the sample content, using the cross-entropy loss function.

[0020] The model training module is used to update the model parameters of the click-rate-location prediction model based on the loss function value, so as to train the click-rate-location prediction model.

[0021] The model building module is used to build a click-through rate prediction model based on the trained click-through rate-location prediction model; the click-through rate prediction model is used to process the non-location features of the content to be recommended to obtain the predicted click-through rate of the content to be recommended.

[0022] In one possible implementation, the click-rate-location prediction model includes a generative adversarial network, a click-rate prediction network, and a bypass network;

[0023] The prediction information acquisition module includes:

[0024] The exposure feature acquisition submodule is used to input the position features of the sample content into the bypass network to obtain the exposure features of the sample content at the actual exposure position;

[0025] The hidden layer representation vector acquisition submodule is used to input the non-positional features of the sample content into the generative network in the generative adversarial network to obtain the hidden layer representation vector of the non-positional features;

[0026] The predicted click-through rate acquisition submodule is used to process the exposure features of the sample content at the actual exposure position and the hidden layer representation vector of the non-position features through the click-through rate prediction network to obtain the predicted click-through rate of the sample content.

[0027] The predicted exposure location acquisition module is used to process the hidden layer representation vector of the non-positional features through the discriminant network in the generative adversarial network to obtain the predicted exposure location of the sample content.

[0028] In one possible implementation, the loss function value calculation module includes:

[0029] The first loss function value calculation submodule is used to calculate the first cross-entropy loss function value based on the exposure position label of the sample content and the predicted exposure position of the sample content.

[0030] The second loss function value calculation submodule is used to calculate the second cross-entropy loss function value based on the actual click-through rate of the sample content and the predicted click-through rate of the sample content.

[0031] The third loss function value calculation submodule is used to calculate the third cross-entropy loss function value based on the predicted exposure position of the sample content and the uniform distribution information of the sample content; the uniform distribution information is the specified exposure probability of the sample content at at least one of the predicted positions.

[0032] In one possible implementation, the model training module includes:

[0033] The first parameter update submodule is used to update the parameters in the discriminant network based on the first cross-entropy loss function value;

[0034] The second parameter update submodule is used to update the parameters in the generator network, the click-through rate prediction network, and the bypass network based on the second cross-entropy loss function value.

[0035] The third parameter update submodule is used to update the parameters in the generator network based on the value of the third cross-entropy loss function.

[0036] In one possible implementation, the generative network has a neural network-based factorization machine structure.

[0037] In one possible implementation, the model building module is used to obtain the model composed of the generator network, the click-rate prediction network, and the bypass network in the trained click-rate-location prediction model as the click-rate prediction model.

[0038] In one possible implementation, the exposure location label is a soft label generated based on the actual exposure location of the sample content, wherein the exposure probability corresponding to the actual exposure location in the soft label is greater than the exposure probability of other exposure locations; and the sum of the exposure probabilities of each exposure location in the soft label is equal to 1.

[0039] On the other hand, a content recommendation device is provided, the device comprising:

[0040] The feature acquisition module is used to acquire non-location features of the content to be recommended in response to receiving a content recommendation request from a target user. The non-location features include content features and user profile features.

[0041] The click-through rate (CTR) prediction module is used to input the non-location features into the CTR prediction model to obtain the predicted CTR of the content to be recommended, output by the CTR prediction model. The CTR prediction model is built based on a trained CTR-location prediction model. The training method of the CTR-location prediction model includes: acquiring training data corresponding to the sample content, the training data including location features, non-location features, exposure location tags, and the actual CTR; the non-location features including content features and user profile features; the location features being the features of the actual exposure location of the sample content, and the exposure location tags being tags constructed based on the actual exposure location of the sample content; inputting the location features and non-location features of the sample content into the CTR-location prediction model to obtain the predicted CTR and predicted exposure location of the sample content; calculating a loss function value using a cross-entropy loss function based on the exposure location tags, the actual CTR, the predicted CTR, and the predicted location of the sample content; and updating the model parameters of the CTR-location prediction model based on the loss function value.

[0042] The content recommendation module is used to recommend content to the target user based on the predicted click-through rate of the content to be recommended.

[0043] In one possible implementation, the click-through rate prediction model includes a generator network, a click-through rate prediction network, and a bypass network; the apparatus further includes:

[0044] The input setting module is used to set the input of the bypass network to the position feature corresponding to a specified position.

[0045] In one possible implementation, the content recommendation module includes:

[0046] The target content acquisition submodule is used to acquire target content based on the predicted click-through rate of the content to be recommended;

[0047] The content recommendation submodule is used to recommend the target content to the target user.

[0048] In one possible implementation, the target content satisfies at least one of the following conditions:

[0049] The predicted click-through rate is higher than the click-through rate threshold.

[0050] The top n are ranked according to their predicted click-through rates from highest to lowest; n is a positive integer.

[0051] On the other hand, a computer device is provided, the computer device including a processor and a memory, the memory storing at least one instruction, at least one program, code set or instruction set, the at least one instruction, the at least one program, the code set or instruction set being loaded and executed by the processor to implement the click-through rate prediction model processing method or the content recommendation method as described above.

[0052] On the other hand, a computer-readable storage medium is provided, wherein at least one computer program is stored in the computer program, which is loaded and executed by a processor to implement the click-through rate prediction model processing method or the content recommendation method as described above.

[0053] On the other hand, a computer program product or computer program is provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the click-through rate prediction model processing method or the content recommendation method described above in the various optional implementations.

[0054] The technical solution provided in this application may include the following beneficial effects:

[0055] By using training data corresponding to sample content, the loss function value between the model's predicted value and the label value is calculated based on the cross-entropy loss function. The click-through rate-location prediction model is trained based on this loss function value, and the click-through rate prediction model is determined based on the trained click-through rate-location prediction model. This achieves accurate click-through rate prediction based on non-location information. The calculation of the loss function value is optimized during the training process, which improves the prediction accuracy of the trained model. This allows the obtained click-through rate prediction model to improve the accuracy of click-through rate prediction while eliminating the influence of location bias on click-through rate prediction.

[0056] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0057] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0058] Figure 1 This is a schematic diagram of the structure of a system used in a content recommendation method according to an exemplary embodiment;

[0059] Figure 2A flowchart illustrating a click-through rate prediction model processing method provided in an exemplary embodiment of this application is shown.

[0060] Figure 3 This is a framework diagram illustrating click-rate-location prediction model training and content recommendation according to an exemplary embodiment;

[0061] Figure 4 A flowchart illustrating a content recommendation method provided in an exemplary embodiment of this application is shown;

[0062] Figure 5 This illustration shows a schematic diagram of a terminal push interface for pushing content to a user, as shown in an exemplary embodiment of this application.

[0063] Figure 6 A flowchart illustrating the click-through rate prediction model training and content recommendation method provided in an exemplary embodiment of this application is shown.

[0064] Figure 7 This illustration shows a schematic diagram of the click-rate-location prediction model provided in an exemplary embodiment of this application;

[0065] Figure 8 A schematic diagram of the Deep-FM structure provided in an exemplary embodiment of this application is shown;

[0066] Figure 9 A block diagram of a click-through rate prediction model processing apparatus provided in an exemplary embodiment of this application is shown;

[0067] Figure 10 A block diagram of a content recommendation apparatus provided in an exemplary embodiment of this application is shown;

[0068] Figure 11 This is a structural block diagram of a computer device according to an exemplary embodiment;

[0069] Figure 12 A structural block diagram of a computer device provided in an exemplary embodiment of this application is shown. Detailed Implementation

[0070] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0071] This application proposes a method for generating a click-through rate (CTR) prediction model and recommending content. This method, based on a CTR prediction model trained with artificial intelligence (AI), predicts the CTR of content to be recommended, thereby improving the accuracy of CTR prediction and ultimately enhancing the effectiveness of content recommendation. For ease of understanding, the terms used in this application are explained below.

[0072] 1) Artificial Intelligence

[0073] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0074] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0075] 2) Machine Learning (ML)

[0076] Machine learning is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory, among others. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and learn-by-doing.

[0077] 3) Feed stream recommendations

[0078] Feed recommendation is a type of content recommendation that aggregates information; it's a continuously updated information stream presented to users. Feeds disseminate dynamic information to subscribers in real time, and the continuous updating of information within a feed is an effective way for users to access information. For example, a status update in WeChat Moments is a feed, a post on Weibo is a feed, and each user's WeChat Moments and Weibo following pages constitute a feed stream.

[0079] To illustrate, Timeline is a feed type that sorts feeds chronologically, with earlier posts seen first and later posts at the top. Rank is another feed type that sorts feeds based on a non-time factor, such as user preference, with favorites at the top and less favorites at the bottom. Aggregate represents an aggregation type, such as when several users watched a movie, this can be aggregated into a single feed: A, B, and C watched movie X. A typical Aggregate type is Timeline combined with client-side aggregation. Notice is a notification type, a functional type, generally used for various notifications and private messages in different apps, and also includes Timeline and Aggregate types.

[0080] 4) CTR (Click-Through-Rate)

[0081] CTR, short for click-through rate, is a commonly used term in internet advertising and an important indicator for measuring the effectiveness of internet advertising. CTR refers to the click-through rate of online ads (image ads / text ads / keyword ads / ranking ads / video ads, etc.), which is the actual number of clicks on the ad (strictly speaking, the number of times the ad reaches the target page) divided by the number of ad impressions.

[0082] 5) Position Bias

[0083] In recommendation systems, the attention each item receives is affected by its display position. Items displayed earlier are usually more likely to be noticed and clicked by users than items displayed later. This can lead to a bias in the model's perception of user preferences and inaccurate CTR predictions.

[0084] 6) Generative Adversarial Networks (GANs)

[0085] Generative Adversarial Networks (GANs) are deep learning models that consist of two modules: a generative module and a discriminative module. These two modules learn to produce fairly good outputs through mutual interaction.

[0086] Figure 1 This is a schematic diagram illustrating the structure of a system used in a content recommendation method according to an exemplary embodiment. For example... Figure 1 As shown, the system includes: server 110 and terminal 120.

[0087] Among them, the aforementioned server 110 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.

[0088] The aforementioned terminal 120 is a terminal with content recommendation function. For example, terminal 120 can be a smartphone, tablet computer, e-book reader, smart glasses, smart watch, smart TV, smart in-vehicle device, MP3 player (Moving Picture Experts Group Audio Layer III), MP4 player (Moving Picture Experts Group Audio Layer IV), laptop computer, and desktop computer, etc.

[0089] Optionally, the system described above includes one or more servers 110 and multiple terminals 120. This application embodiment does not limit the number of servers 110 and terminals 120.

[0090] The terminal and server are connected via a communication network. Optionally, the communication network can be a wired network or a wireless network.

[0091] Optionally, the aforementioned wireless or wired networks use standard communication technologies and / or protocols. The network is typically the Internet, but can also be any network, including but not limited to Local Area Networks (LANs), Metropolitan Area Networks (MANs), Wide Area Networks (WANs), mobile, wired or wireless networks, private networks, or any combination of virtual private networks. In some embodiments, technologies and / or formats including Hyper Text Markup Language (HTML), Extensible Markup Language (XML), etc., are used to represent data exchanged over the network. Furthermore, conventional encryption technologies such as Secure Socket Layer (SSL), Transport Layer Security (TLS), Virtual Private Networks (VPNs), and Internet Protocol Security (IPsec) can be used to encrypt all or some links. In other embodiments, customized and / or dedicated data communication technologies can be used to replace or supplement the aforementioned data communication technologies. This application does not impose any limitations.

[0092] Figure 2 This application illustrates a flowchart of a click-through rate prediction model processing method provided in an exemplary embodiment. The method is executed by a computer device, which can be implemented as a terminal or a server. Figure 1 The terminal or server shown, such as Figure 2 As shown, the click-through rate prediction model processing method includes the following steps:

[0093] Step 210: Obtain training data corresponding to the sample content. The training data includes location features, non-location features, exposure location labels, and actual click-through rate. The non-location features include content features and user profile features. The location features are the features of the actual exposure location of the sample content, and the exposure location labels are labels constructed based on the actual exposure location of the sample content.

[0094] The content features can include content size, content category, content author, content exposure, content sharing rate, number of times the content is viewed, number of times the content is completed, content popularity, etc. Taking the content as a video as an example, the video features can include video length, video category, video author, video resolution and clarity, video exposure, video interaction penetration rate, video completion rate, number of times the video is played, number of times the video is viewed, number of likes, number of favorites, number of comments, etc.

[0095] User profile features can include user interests, login frequency, content browsing duration, content browsing frequency, content interaction frequency, probability of completing content viewing, probability of quickly scrolling through content, and so on.

[0096] It should be noted that the content features and user profile features in the embodiments of this application may include all or part of the content listed above, or may include other related features not listed. The content features and user profile features may be obtained according to actual needs. This application does not limit the content features and user profile features used in training.

[0097] In one possible implementation, the sample content can be in various forms such as images, articles, public accounts, videos, news, advertisements, and product links.

[0098] Step 220: Input the location features and non-location features of the sample content into the click-through rate-location prediction model to obtain the predicted click-through rate and predicted exposure location of the sample content.

[0099] Step 230: Based on the exposure location label, actual click-through rate, predicted click-through rate, and predicted location of the sample content, calculate the loss function value using the cross-entropy loss function.

[0100] The Cross Entropy Error Function (CEC) is commonly used in classification problems. In the case of binary classification, its expression is:

[0101]

[0102] Among them, y i This represents the label of sample i, with 1 for positive class and 0 for negative class, p i This represents the probability that sample i is predicted to be positive.

[0103] In the case of multiple classifications, its expression is:

[0104]

[0105] Where M represents the number of categories, y icThis indicates an indicator variable (0 or 1), where 1 represents the category and 0 represents the category of sample i. ic The predicted probability that observed sample i belongs to category c.

[0106] Step 240: Update the model parameters of the click-rate-location prediction model based on the loss function value in order to train the click-rate-location prediction model.

[0107] Step 250: Based on the trained click-rate-location prediction model, construct a click-rate prediction model.

[0108] This click-through rate prediction model is used to process the non-location features of the content to be recommended in order to obtain the predicted click-through rate of the content.

[0109] In summary, the click-through rate (CTR) prediction model training method provided in this application uses training data corresponding to sample content, calculates the loss function value between the model's predicted value and the label value based on the cross-entropy loss function, trains the CTR-location prediction model based on this loss function value, and determines the CTR prediction model based on the trained CTR-location prediction model. This achieves accurate CTR prediction based on non-location information. The calculation of the loss function value is optimized during the training process, improving the prediction accuracy of the trained model. This allows the obtained CTR prediction model to improve the accuracy of CTR prediction while eliminating the influence of location bias on CTR prediction.

[0110] The scheme described in this application embodiment obtains a click-through rate (CTR)-location prediction model by training a cross-entropy loss function based on the location and non-location features of two sample contents. Then, a CTR prediction model is constructed using this CTR-location prediction model. By improving the loss function during model training, the prediction accuracy of the obtained CTR prediction model is enhanced. The above CTR prediction model can be applied to, but is not limited to, the following scenarios:

[0111] 1. Image and text information push scenario

[0112] For example, in scenarios where there is a need to push text and image information, such as Weibo, Toutiao, and browser homepages, in order to make the pushed text and image information more in line with user needs, or to push text and image information that is likely to become the most popular trending information to users, a click-through rate (CTR) prediction model can be used to predict the CTR of the text and image information to be released, based on the content characteristics of the text and image information to be released and the user profile characteristics of the logged-in user. Based on the predicted CTR, the text and image information can be pushed and displayed in an appropriate position. For example, text and image information with a high predicted CTR can be displayed first in the recommendation order, and text and image information with a low predicted CTR can be displayed last.

[0113] 2. Video recommendation scenario

[0114] Video recommendation scenarios can include short video scenarios and long video scenarios. In this scenario, for newly released videos, the click-through rate prediction model generated by the click-through rate prediction model processing method provided in this application can predict the click-through rate of the video based on the content features of the video content and user profile features, so as to determine the recommendation position of the video. For example, when the predicted click-through rate of the video is high, the video can be pushed on the homepage of the video application, or when the predicted click-through rate of the video is low, the video can not be pushed.

[0115] 3. Advertising placement scenarios

[0116] In advertising scenarios, the predicted click-through rate (CTR) and the advertiser's bid often determine the ad ranking. However, the samples used to train CTR prediction models often contain biased information. Many features used in training these models are statistical features, such as the relatively rich and reliable CTR of search keywords and ads. However, for keyword / ad pairs that are rarely or never displayed, this information is very sparse and unreliable. The model cannot fully learn from this information and struggles to advance the display position of these keyword / ad pairs. Therefore, the data used to train the model contains biased information, which is mostly manifested as positional bias—that is, different display positions have different impacts on CTR. Generally speaking, ads displayed in earlier positions tend to have higher CTRs. The click-through rate prediction model training method provided in this application can eliminate the influence of location information in the sample on the generated model's click-through rate prediction. This improves the accuracy of click-through rate prediction when using the generated click-through rate prediction model to determine the click-through rate of ads that are not frequently displayed or have never been displayed. This allows for the reasonable arrangement of the ad's display position, thereby maximizing the effectiveness of ad placement.

[0117] The solution involved in this application includes the training phase of the click-rate-location prediction model and the content recommendation phase. Figure 3 This is a framework diagram illustrating click-rate-location prediction model training and content recommendation according to an exemplary embodiment, such as... Figure 3As shown, in the click-rate-location prediction model training phase, the click-rate-location prediction model training device 310 obtains the click-rate-location prediction model using pre-set training data corresponding to sample content (including at least two sample contents and their respective location features, non-location features, exposure location labels, and actual click-rates), and determines the click-rate model based on this model. In the content recommendation phase, the content recommendation device 320 predicts the click-rate of the input content to be recommended based on the determined click-rate prediction model, obtains the predicted click-rate of the content to be recommended, and thus determines the recommendation position or recommendation order of the content to be recommended based on the predicted click-rate.

[0118] The click-rate-location prediction model training device 310 and the content recommendation device 320 can be computer devices, such as personal computers, servers, or other fixed computer devices, or tablet computers, e-book readers, or other mobile computer devices.

[0119] Optionally, the click-rate-location prediction model training device 310 and the content recommendation device 320 can be the same device, or they can be different devices. Furthermore, when the click-rate-location prediction model training device 310 and the content recommendation device 320 are different devices, they can be of the same type, such as both being servers; or they can be of different types, such as the content recommendation device 320 being a personal computer or terminal, while the click-rate-location prediction model training device 310 can be a server, etc. This application embodiment does not limit the specific types of the click-rate-location prediction model training device 310 and the content recommendation device 320.

[0120] Figure 4 The flowchart shown is an exemplary embodiment of the content recommendation method provided in this application. This face recommendation method can be executed by a computer device, which can be implemented as a terminal or a server, or a terminal and a server. The terminal or server can be implemented as... Figure 1 For the terminal or server shown, the recommended method for this content includes the following steps:

[0121] Step 410: In response to receiving a content recommendation request from a target user, obtain the non-location features of the content to be recommended, which include content features and user profile features.

[0122] In one possible implementation, the content recommendation request for the target user can be an automatic recommendation request sent when the user logs in or visits a specified page. For example, it could be a proactive recommendation request sent when the user opens a specified interface of the application. For instance, when the user opens the "Discover" module in WeChat, a content recommendation request is automatically sent. Alternatively, the user's content recommendation request can also be a recommendation request generated based on the user's search operation. For example, if the user searches for "landscape pictures" in the search box, a content recommendation request for "landscape pictures" is generated based on this search operation. This allows the computer device to obtain the predicted click-through rate of each piece of content to be recommended based on the user's user profile characteristics and the content characteristics of each piece of content to be recommended, and then recommend relevant content to the user based on the predicted click-through rate.

[0123] Step 420: Input the non-location features into the click-through rate (CTR) prediction model to obtain the predicted CTR corresponding to the content to be recommended. This CTR prediction model is built based on the trained CTR-location prediction model. The training method of the CTR-location prediction model includes: acquiring training data corresponding to the sample content, which includes location features, non-location features, exposure location labels, and the actual CTR; the non-location features include content features and user profile features; the location features are the features of the actual exposure location of the sample content, and the exposure location labels are labels constructed based on the actual exposure location of the sample content; input the location features and non-location features of the sample content into the CTR-location prediction model to obtain the predicted CTR and predicted exposure location of the sample content; calculate the loss function value using the cross-entropy loss function based on the exposure location labels, actual CTR, predicted CTR, and predicted location of the sample content; and update the model parameters of the CTR-location prediction model based on the loss function value.

[0124] Step 430: Recommend content to the target user based on the predicted click-through rate of the content to be recommended.

[0125] In one possible implementation, when the above content recommendation method is executed by the server, the server recommends content to the target user through the terminal based on the predicted click-through rate of the content to be recommended. That is, the server first sends the determined recommended content to the terminal, and then the terminal displays the content recommendation to the user through the interface.

[0126] Figure 5 This illustration shows a schematic diagram of a terminal push interface for pushing content to a user, as shown in an exemplary embodiment of this application. Figure 5As shown, the terminal interface can sort and push various push content based on the predicted click-through rate (CTR) obtained from the CTR prediction model. This push content can include different types of content such as images, articles, WeChat official accounts, videos, news, advertisements, and product links. Figure 5 As shown, the push content is presented as a public account with text and images. Among them, the push content 510 with a predicted click-through rate is ranked higher in the terminal push interface, while the push content 520 with a predicted click-through rate is ranked lower in the terminal push interface.

[0127] In summary, the content recommendation method provided in this application, upon receiving a content recommendation request from a target user, inputs the non-location features corresponding to the content to be recommended into a click-through rate (CTR) model determined by a CTR-location prediction model trained using a cross-entropy loss function, thereby obtaining the predicted CTR of the content to be recommended output by the CTR prediction model. This CTR prediction model can improve the accuracy of CTR prediction while eliminating the influence of location bias on CTR prediction, thus improving the content push effect.

[0128] In one possible implementation, the click-rate-location prediction model includes a generative adversarial network, a click-rate prediction network, and a bypass network. Figure 6 This application illustrates a flowchart of a click-through rate prediction model training and content recommendation method provided in an exemplary embodiment. The method is executed by a computer device, which can be implemented as a terminal or a server, or a terminal and a server. Figure 1 The terminal or server shown, such as Figure 6 As shown, the click-through rate prediction model training and content recommendation method include the following steps:

[0129] Step 610: Obtain training data corresponding to the sample content. The training data includes location features, non-location features, exposure location labels, and actual click-through rate. The non-location features include content features and user profile features. The location features are the features of the actual exposure location of the sample content, and the exposure location labels are labels constructed based on the actual exposure location of the sample content.

[0130] The positional feature of the sample content represents the actual exposure position of the sample content. Taking a total of 5 exposure positions as an example, if the exposure position of the sample content is the 2nd position, the positional feature of the sample content can be represented as (0, 1, 0, 0, 0).

[0131] In one possible implementation, the click-rate-location prediction model in this application is a model built on generative adversarial networks. During model training, in order to reduce the training difficulty and prevent overfitting to the discriminant network in the generative adversarial network, the exposure location label set in this embodiment is a soft label generated based on the real exposure location of each sample content. The exposure probability corresponding to the real exposure location in the soft label is greater than the exposure probability of other exposure locations, and the sum of the exposure probabilities of each exposure location in the soft label is equal to 1.

[0132] Taking the exposure position of the sample content as the second position, with a total of 5 exposure positions as an example, its corresponding one-hot label is (0, 1, 0, 0, 0). That is, the actual exposure position of the sample content is set to 1, and the other positions are all set to 0. In this application, the actual exposure position is represented in the form of a soft label, which can be (0.1, 0.8, 0.1, 0, 0), or (0.25, 0.5, 0.25, 0, 0). This application does not restrict the setting method of the soft label, but in the soft label, the exposure probability corresponding to the actual exposure position is greater than the exposure probability of other exposure positions.

[0133] In one possible implementation, the soft tag can be a soft tag set by relevant personnel based on actual needs and the actual exposure location, or it can be a soft tag generated by computer equipment according to a specified rule set in the computer equipment. This application does not limit the generation method of the soft tag.

[0134] Step 620: Input the location features and non-location features of the sample content into the click-through rate-location prediction model to obtain the predicted click-through rate and predicted exposure location of the sample content.

[0135] In one possible implementation, the above process is implemented as follows:

[0136] The location features of the sample content are input into the bypass network to obtain the exposure features of the sample content at the actual exposure location.

[0137] The non-positional features of the sample content are input into the generative network in the generative adversarial network to obtain the hidden layer representation vector of the non-positional features;

[0138] The click-through rate prediction network processes the exposure features of sample content at the actual exposure location and the hidden layer representation vector of non-location features to obtain the predicted click-through rate of the sample content.

[0139] By processing the hidden layer representation vectors of non-positional features through the discriminant network in a generative adversarial network, the predicted exposure position of the sample content can be obtained.

[0140] In this generative adversarial network, the discriminant network determines the predicted exposure position of the sample content by judging whether there is positional information in the hidden layer representation vector of non-positional features.

[0141] In one possible implementation, the predicted exposure position can be indicated by the predicted exposure position probability distribution. For example, if the predicted exposure position of a sample content is (0.2, 0.6, 0.1, 0.1, 0), it means that the probability of the sample content falling on the first exposure position is 0.2, the probability of falling on the second exposure position is 0.6, the probability of falling on the third exposure position is 0.1, the probability of falling on the fourth exposure position is 0.1, and the probability of falling on the first exposure position is 0.

[0142] In one possible implementation, the click-through rate prediction network obtains the predicted click-through rate of the sample content by fusing the exposure features of the sample content at the actual exposure location with the hidden layer representation vectors of the non-location features.

[0143] Figure 7 This application shows a schematic diagram of the structure of a click-rate-location prediction model provided in an exemplary embodiment, as shown below. Figure 7 As shown, the click-through rate-location prediction model 700 includes a generative adversarial network 710, a click-through rate prediction network (Predictnet) 720, and a bypass network (Bypass net) 730. The generative adversarial network 710 includes a generator network (Main net) 711 and a discriminator network (Bias net) 712. The bypass network 730 is used to generate exposure features of each sample content at the corresponding real exposure location based on the location features of each input sample. The generator network 711 is used to process the non-location features of each input sample and output the hidden layer representation vector of the non-location features. The click-through rate prediction network 720 is used to generate the predicted click-through rate of each sample based on the hidden layer representation vector of the non-location features and the exposure features of each sample at the corresponding real exposure location. The discriminator network 712 is used to obtain the predicted exposure location of each sample from the hidden layer representation vector of the non-location features output by the generator network 711.

[0144] In one possible implementation, the generative network has a neural network-based factorization machine (Deep-FM) structure.

[0145] The generative network in this generative adversarial network is used in a neural network-based factorization machine structure to process non-positional features in the training dataset and generate hidden layer representation vectors of the non-positional features.

[0146] In one possible implementation, the inner product unit in the Deep-FM structure performs second-order feature crossing on the content features and user profile features of the input content to obtain second-order combined features. Then, the deep neural network in the Deep-FM structure performs deep feature comparison on the content features and user profile features to obtain higher-order combined features. Figure 8 A schematic diagram of the Deep-FM structure provided in an exemplary embodiment of this application is shown, as follows: Figure 8 As shown, the Deep-FM structure includes two parts: a deep neural network 810 (Hidden layer) and a factorization machine 820 (FM layer). These two parts share the same embedding layer 830 features. Low-order feature extraction and high-order feature extraction are performed on the embedding layer features, respectively. In this embodiment, the embedding layer features include content features and user profile features. The inner product unit in the factorization machine is used to extract the second-order combination features between the content features and the user profile features. The deep neural network is used to extract the high-order combination features between the content features and the user profile features. Finally, the hidden layer representation vector of the non-positional features of each sample content at at least two predicted exposure positions is obtained through the output unit 840 (OutputUnits).

[0147] In one possible implementation, in order to obtain a more comprehensive hidden layer representation vector of non-positional features, a Wide layer 850 is also set in the Deep-FM structure to extract the first-order features from the content features and user profile features. Based on the connection layer 860, the first-order features of the content, the first-order features of the user profile, the second-order combined features, and the higher-order combined features are concatenated to obtain the hidden layer representation vector of the non-positional features. The output unit 840 outputs the hidden layer representation vector of the non-positional features.

[0148] Step 630: Based on the exposure location label of the sample content, the actual click-through rate, the predicted click-through rate of the sample content, and the predicted location of the sample content, calculate the loss function value using the cross-entropy loss function.

[0149] The loss function value includes the first cross-entropy loss function value, the second cross-entropy loss function value, and the third cross-entropy loss function value.

[0150] In one possible implementation, the above-mentioned exposure location label, actual click-through rate, predicted click-through rate of sample content, and predicted location of sample content are used to calculate the loss function value through the cross-entropy loss function, as follows:

[0151] Based on the exposure location labels of the sample content and the predicted exposure location of the sample content, the first cross-entropy loss function value is calculated; that is:

[0152]

[0153] Where b represents the exposure location label, b' represents the predicted exposure location probability, c represents the predicted exposure location, M represents the number of predicted exposure locations, and N represents the number of sample contents.

[0154] Based on the actual click-through rate (CTR) and the predicted CTR of the sample content, calculate the value of the second cross-entropy loss function; that is:

[0155]

[0156] Where y represents the actual click-through rate and y' represents the predicted click-through rate.

[0157] Based on the predicted exposure positions of the sample content and the uniform distribution information of the sample content, the third cross-entropy loss function value is calculated. This uniform distribution information represents the specified exposure probability of the sample content at at least one predicted position. For example, the exposure probability of the sample content at each predicted exposure position can be set to 0 or 1, i.e., the uniform distribution information is (0, 0, 0, 0, 0) or (1, 1, 1, 1, 1). The calculation of the third cross-entropy loss function value is implemented as follows:

[0158]

[0159] Where B represents a uniform distribution and b' represents the probability of the predicted exposure location.

[0160] Step 640: Update the model parameters of the click-through rate-location prediction model based on the loss function value.

[0161] Corresponding to the fact that the loss function values ​​in step 630 include three cross-entropy loss function values, the above steps can be implemented as follows:

[0162] The parameters in the discriminant network are updated based on the first cross-entropy loss function value;

[0163] Based on the value of the second cross-entropy loss function, the parameters in the generator network, click-through rate prediction network, and bypass network are updated;

[0164] The parameters in the generator network are updated based on the value of the third cross-entropy loss function.

[0165] Specifically, when updating the parameters in the corresponding network using the corresponding loss function value, the parameters in other networks remain unchanged. In other words:

[0166] Keeping the parameters in the generator network, the click-through rate prediction network, and the bypass network unchanged, the parameters in the discriminator network are updated based on the first cross-entropy loss function value.

[0167] Keeping the parameters in the discriminant network, the click-through rate prediction network, and the bypass network unchanged, the parameters in the generator network are updated based on the value of the third cross-entropy loss function.

[0168] Based on the second cross-entropy loss function value, the parameters of the generator network, click-through rate prediction network, and bypass network in the generative adversarial network are updated.

[0169] It should be noted that this application does not restrict the order in which the three loss functions update the network parameters during model training.

[0170] During the parameter update process, since both the second cross-entropy loss function LossY and the third cross-entropy loss function LossG will generate parameters in the generator network of the adversarial network for updating, in order to balance the influence of the two on the parameters of the generator network, different weights are set for the two loss functions, which is expressed as: λLossG(B, b')+LossY(y, y').

[0171] Step 650: Obtain the model composed of the generator network, the click-rate prediction network, and the bypass network in the trained click-rate-location prediction model as the click-rate prediction model.

[0172] The click-through rate model is used to predict the click-through rate of candidate content based on its content features and the user profile features of the target user.

[0173] Step 660: In response to receiving a content recommendation request from a target user, obtain the non-location features of the content to be recommended, which include content features and user profile features of the target user.

[0174] Step 670: Input the non-location features into the click-through rate prediction model to obtain the predicted click-through rate corresponding to the content to be recommended.

[0175] The click-through rate (CTR) prediction model is obtained based on the CTR model generation process shown in steps 610 to 650. Specifically, it is built upon the trained CTR-location prediction model. The training method for this model includes: acquiring training data corresponding to the sample content, which includes location features, non-location features, exposure location labels, and the actual CTR; the non-location features include content features and user profile features; the location features are the characteristics of the actual exposure location of the sample content, and the exposure location labels are labels constructed based on the actual exposure location of the sample content; inputting the location features and non-location features of the sample content into the CTR-location prediction model to obtain the predicted CTR and predicted exposure location of the sample content; calculating the loss function value using the cross-entropy loss function based on the exposure location labels, actual CTR, predicted CTR, and predicted location of the sample content; and updating the model parameters of the CTR-location prediction model based on the loss function value.

[0176] The click-through rate prediction model includes a generator network, a click-through rate prediction network, and a bypass network.

[0177] When applying the click-through rate model, the input to the bypass network is set to the location feature corresponding to a specified location. For example, it can be set to the location feature of location 1, or it can be set to the location feature corresponding to other specified locations; this application does not impose any restrictions on this.

[0178] Step 680: Recommend content to the target user based on the predicted click-through rate of the content to be recommended.

[0179] In one possible implementation, the target content is obtained based on the predicted click-through rate of the content to be recommended;

[0180] Recommend target content to target users.

[0181] The target content satisfies at least one of the following conditions:

[0182] The predicted click-through rate is higher than the click-through rate threshold.

[0183] The top n are ranked according to their predicted click-through rates from highest to lowest; n is a positive integer.

[0184] To illustrate, there are 1000 pieces of content to be recommended. These 1000 pieces of content are input into the click-through rate (CTR) prediction model to obtain the predicted CTR for each of the 1000 pieces of content. The computer device is set with a CTR threshold, for example, a CTR threshold of 50%. Then, the content to be recommended with a predicted CTR higher than 50% is selected as the target content and recommended to the user.

[0185] In one possible implementation, the computer device has a recommendation quantity threshold. If the number of target content exceeds the recommendation quantity threshold set in the computer device, the target content can be randomly selected from the recommended quantity threshold and recommended to the user. Alternatively, the target content can be sorted in descending order of predicted click-through rate, and the target content ranked within the recommendation quantity threshold range can be recommended to the user.

[0186] Alternatively, the content to be recommended can be sorted in descending order of its predicted click-through rate.

[0187] The top n ranked content to be recommended is selected as the target content and recommended to the target users, where n is a positive integer.

[0188] Where n can be equal to the aforementioned threshold for the number of recommendations.

[0189] Experimental results showed that the AUC (Area Under the Curve) of the offline click-through rate prediction model was improved by 0.3% compared to models in related technologies. After going live, content was pushed based on the click-through rate predicted by this click-through rate prediction model, which increased the click-through rate of viewed content by 1.8%, the average viewing time per person by 0.58%, and the average number of interactions per person by 1.5%.

[0190] In summary, the click-through rate (CTR) prediction model training and content recommendation method proposed in this application uses a CTR-location prediction model trained with a cross-entropy loss function based on training data from sample content to determine the CTR model. By improving the loss function, the influence of location bias on the model's predicted CTR is reduced, thereby improving the accuracy of the CTR prediction model. Subsequently, when a content recommendation request is received from a target user, the content features of the content to be recommended and the user profile features of the target user are input into the CTR prediction model. The CTR prediction model is then used to predict the CTR, and content is pushed based on the predicted CTR, thereby improving the content push effect.

[0191] Figure 9 A block diagram of a click-through rate prediction model processing apparatus provided in an exemplary embodiment of this application is shown, as follows: Figure 9 As shown, the device includes:

[0192] The training data acquisition module 910 is used to acquire training data corresponding to the sample content. The training data includes location features, non-location features, exposure location labels, and actual click-through rate. The non-location features include content features and user profile features. The location features are the features of the actual exposure location of the sample content, and the exposure location labels are labels constructed based on the actual exposure location of the sample content.

[0193] The prediction information acquisition module 920 is used to input the location features of the sample content and the non-location features of the sample content into the click-rate-location prediction model to obtain the predicted click-rate of the sample content and the predicted exposure position of the sample content.

[0194] The loss function value calculation module 930 is used to calculate the loss function value based on the exposure position label of the sample content, the actual click-through rate, the predicted click-through rate of the sample content, and the predicted position of the sample content, using the cross-entropy loss function.

[0195] The model training module 940 is used to update the model parameters of the click-rate-location prediction model based on the loss function value, so as to train the click-rate-location prediction model.

[0196] The model building module 950 is used to build a click-through rate prediction model based on the trained click-through rate-location prediction model; the click-through rate prediction model is used to process the non-location features of the content to be recommended to obtain the predicted click-through rate of the content to be recommended.

[0197] In one possible implementation, the click-rate-location prediction model includes a generative adversarial network, a click-rate prediction network, and a bypass network;

[0198] The prediction information acquisition module 920 includes:

[0199] The exposure feature acquisition submodule is used to input the position features of the sample content into the bypass network to obtain the exposure features of the sample content at the actual exposure position;

[0200] The hidden layer representation vector acquisition submodule is used to input the non-positional features of the sample content into the generative network in the generative adversarial network to obtain the hidden layer representation vector of the non-positional features;

[0201] The predicted click-through rate acquisition submodule is used to process the exposure features of the sample content at the actual exposure position and the hidden layer representation vector of the non-position features through the click-through rate prediction network to obtain the predicted click-through rate of the sample content.

[0202] The predicted exposure location acquisition module is used to process the hidden layer representation vector of the non-positional features through the discriminant network in the generative adversarial network to obtain the predicted exposure location of the sample content.

[0203] In one possible implementation, the loss function value calculation module 930 includes:

[0204] The first loss function value calculation submodule is used to calculate the first cross-entropy loss function value based on the exposure position label of the sample content and the predicted exposure position of the sample content.

[0205] The second loss function value calculation submodule is used to calculate the second cross-entropy loss function value based on the actual click-through rate of the sample content and the predicted click-through rate of the sample content.

[0206] The third loss function value calculation submodule is used to calculate the third cross-entropy loss function value based on the predicted exposure position of the sample content and the uniform distribution information of the sample content; the uniform distribution information is the specified exposure probability of the sample content at at least one of the predicted positions.

[0207] In one possible implementation, the model training module 940 includes:

[0208] The first parameter update submodule is used to update the parameters in the discriminant network based on the first cross-entropy loss function value;

[0209] The second parameter update submodule is used to update the parameters in the generator network, the click-through rate prediction network, and the bypass network based on the second cross-entropy loss function value.

[0210] The third parameter update submodule is used to update the parameters in the generator network based on the value of the third cross-entropy loss function.

[0211] In one possible implementation, the generative network has a neural network-based factorization machine structure.

[0212] In one possible implementation, the model building module 950 is used to obtain the model composed of the generator network, the click-rate prediction network, and the bypass network in the trained click-rate-location prediction model as the click-rate prediction model.

[0213] In one possible implementation, the exposure location label is a soft label generated based on the actual exposure location of the sample content, wherein the exposure probability corresponding to the actual exposure location in the soft label is greater than the exposure probability of other exposure locations; and the sum of the exposure probabilities of each exposure location in the soft label is equal to 1.

[0214] In summary, the click-through rate (CTR) prediction model training device provided in this application uses training data corresponding to sample content, calculates the loss function value between the model's predicted value and the label value based on the cross-entropy loss function, trains the CTR-location prediction model based on this loss function value, and determines the CTR prediction model based on the trained CTR-location prediction model. This achieves accurate prediction of CTR based on non-location information. The calculation of the loss function value is optimized during the training process, improving the prediction accuracy of the trained model. This allows the obtained CTR prediction model to improve the accuracy of CTR prediction while eliminating the influence of location bias on CTR prediction.

[0215] Figure 10 A block diagram of a content recommendation apparatus provided in an exemplary embodiment of this application is shown, such as Figure 10 As shown, the device includes:

[0216] The feature acquisition module 1010 is used to acquire non-location features of the content to be recommended in response to receiving a content recommendation request from a target user. The non-location features include content features and user profile features.

[0217] Click-through rate (CTR) prediction module 1020 is used to input the non-location features into a CTR prediction model to obtain the predicted CTR of the content to be recommended, output by the CTR prediction model. The CTR prediction model is a model built based on a trained CTR-location prediction model. The training method of the CTR-location prediction model includes: acquiring training data corresponding to the sample content, the training data including location features, non-location features, exposure location tags, and the actual CTR; the non-location features including content features and user profile features; the location features being the features of the actual exposure location of the sample content, and the exposure location tags being tags constructed based on the actual exposure location of the sample content; inputting the location features and non-location features of the sample content into the CTR-location prediction model to obtain the predicted CTR and predicted exposure location of the sample content; calculating a loss function value using a cross-entropy loss function based on the exposure location tags, the actual CTR, the predicted CTR, and the predicted location of the sample content; and updating the model parameters of the CTR-location prediction model based on the loss function value.

[0218] The content recommendation module 1030 is used to recommend content to the target user based on the predicted click-through rate of the content to be recommended.

[0219] In one possible implementation, the click-through rate prediction model includes a generator network, a click-through rate prediction network, and a bypass network; the apparatus further includes:

[0220] The input setting module is used to set the input of the bypass network to the position feature corresponding to a specified position.

[0221] In one possible implementation, the content recommendation module 1030 includes:

[0222] The target content acquisition submodule is used to acquire target content based on the predicted click-through rate of the content to be recommended;

[0223] The content recommendation submodule is used to recommend the target content to the target user.

[0224] In one possible implementation, the target content satisfies at least one of the following conditions:

[0225] The predicted click-through rate is higher than the click-through rate threshold.

[0226] The top n are ranked according to their predicted click-through rates from highest to lowest; n is a positive integer.

[0227] In summary, the content recommendation device provided in this application, upon receiving a content recommendation request from a target user, inputs the non-location features corresponding to the content to be recommended into a click-through rate (CTR) model determined by a CTR-location prediction model trained using a cross-entropy loss function, thereby obtaining the predicted CTR of the content to be recommended output by the CTR prediction model. This CTR prediction model can improve the accuracy of CTR prediction by eliminating the influence of location bias on CTR prediction, thus improving the content push effect.

[0228] Figure 11 This is a structural block diagram of a computer device 1100 according to an exemplary embodiment. This computer device can be implemented as a server as described in the above-described scheme of this application. The computer device 1100 includes a Central Processing Unit (CPU) 1101, a system memory 1104 including Random Access Memory (RAM) 1102 and Read-Only Memory (ROM) 1103, and a system bus 1105 connecting the system memory 1104 and the CPU 1101. The computer device 1100 also includes a basic input / output system (I / O system) 1106 to facilitate information transfer between various devices within the computer, and a mass storage device 1107 for storing the operating system 1113, application programs 1114, and other program modules 1115.

[0229] The basic input / output system 1106 includes a display 1108 for displaying information and an input device 1109 for user input, such as a mouse or keyboard. Both the display 1108 and the input device 1109 are connected to the central processing unit 1101 via an input / output controller 1110 connected to the system bus 1105. The basic input / output system 1106 may also include the input / output controller 1110 for receiving and processing input from multiple other devices such as a keyboard, mouse, or electronic stylus. Similarly, the input / output controller 1110 also provides output to a display screen, printer, or other types of output devices.

[0230] The mass storage device 1107 is connected to the central processing unit 1101 via a mass storage controller (not shown) connected to the system bus 1105. The mass storage device 1107 and its associated computer-readable media provide non-volatile storage for the computer device 1100. That is, the mass storage device 1107 may include computer-readable media (not shown) such as a hard disk or a compact disc read-only memory (CD-ROM) drive.

[0231] Without loss of generality, the computer-readable medium may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer storage media include RAM, ROM, erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other solid-state storage technologies, CD-ROM, digital versatile disc (DVD) or other optical storage, magnetic tape cassettes, magnetic tape, disk storage, or other magnetic storage devices. Of course, those skilled in the art will recognize that the computer storage media are not limited to the above-mentioned types. The system memory 1104 and mass storage device 1107 described above can be collectively referred to as memory.

[0232] According to various embodiments of this disclosure, the computer device 1100 can also be connected to a remote computer on a network, such as the Internet. That is, the computer device 1100 can be connected to the network 1112 via the network interface unit 1111 connected to the system bus 1105, or the network interface unit 1111 can be used to connect to other types of networks or remote computer systems (not shown).

[0233] The memory further includes at least one instruction, at least one program, code set, or instruction set, which are stored in the memory. The central processing unit 1101 executes the at least one instruction, at least one program, code set, or instruction set to implement all or part of the steps in the click-through rate prediction model processing method and / or content recommendation method shown in the above embodiments.

[0234] Figure 12 A structural block diagram of a computer device 1200 provided in an exemplary embodiment of this application is shown. The computer device 1200 can be implemented as a click-rate-location prediction model training device and / or content recommendation device, such as a smartphone, tablet, laptop, or desktop computer. The computer device 1200 may also be referred to as a user device, portable terminal, laptop terminal, desktop terminal, or other names.

[0235] Typically, computer device 1200 includes a processor 1201 and a memory 1202.

[0236] Processor 1201 may include one or more processing cores, such as a quad-core processor or a twelfth-core processor. Processor 1201 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). Processor 1201 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 1201 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, processor 1201 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0237] The memory 1202 may include one or more computer-readable storage media, which may be non-transitory. The memory 1202 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in the memory 1202 is used to store at least one instruction, which is executed by the processor 1201 to implement the click-through rate prediction model training method and / or content recommendation method provided in the method embodiments of this application.

[0238] In some embodiments, the computer device 1200 may optionally include a peripheral device interface 1203 and at least one peripheral device. The processor 1201, memory 1202, and peripheral device interface 1203 can be connected via a bus or signal line. Each peripheral device can be connected to the peripheral device interface 1203 via a bus, signal line, or circuit board. Specifically, the peripheral device includes at least one of the following: a radio frequency circuit 1204, a display screen 1205, a camera assembly 1206, an audio circuit 1207, a positioning assembly 1208, and a power supply 1209.

[0239] Peripheral device interface 1203 can be used to connect at least one I / O (Input / Output) related peripheral device to processor 1201 and memory 1202. In some embodiments, processor 1201, memory 1202 and peripheral device interface 1203 are integrated on the same chip or circuit board; in some other embodiments, any one or two of processor 1201, memory 1202 and peripheral device interface 1203 can be implemented on separate chips or circuit boards, which is not limited in this embodiment.

[0240] In some embodiments, the computer device 1200 further includes one or more sensors 1210. The one or more sensors 1210 include, but are not limited to: an accelerometer 1211, a gyroscope 1212, a pressure sensor 1213, a fingerprint sensor 1214, an optical sensor 1215, and a proximity sensor 1216.

[0241] Those skilled in the art will understand that Figure 12 The structure shown does not constitute a limitation on the computer device 1200 and may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0242] In one exemplary embodiment, a computer-readable storage medium is also provided for storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement all or part of the steps in the click-through rate prediction model training and / or content recommendation method described above. For example, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, floppy disk, or optical data storage device, etc.

[0243] In one exemplary embodiment, a computer program product or computer program is also provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the aforementioned actions. Figure 2 , Figure 4 or Figure 6 All or part of the steps of the method shown in any embodiment.

[0244] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0245] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A click-through rate prediction model processing method, characterized in that, The method includes: Acquire training data corresponding to the sample content. The training data includes location features, non-location features, exposure location tags, and actual click-through rate. The non-location features include content features and user profile features. The location features are the features of the actual exposure location of the sample content, and the exposure location tags are tags constructed based on the actual exposure location of the sample content. The location features of the sample content are input into the bypass network of the click-through rate-location prediction model to obtain the exposure features of the sample content at the actual exposure location. The click-through rate-location prediction model also includes a generative adversarial network (GAN) and a click-through rate prediction network. The non-location features of the sample content are input into the generative network of the GAN to obtain the hidden layer representation vectors of the non-location features. The click-through rate prediction network processes the exposure features of the sample content at the actual exposure location and the hidden layer representation vectors of the non-location features to obtain the predicted click-through rate of the sample content. The discriminative network in the GAN processes the hidden layer representation vectors of the non-location features to obtain the predicted exposure location of the sample content. Based on the exposure location label of the sample content, the actual click-through rate, the predicted click-through rate of the sample content, and the predicted exposure location of the sample content, the loss function value is calculated using the cross-entropy loss function; The model parameters of the click-rate-location prediction model are updated based on the loss function value in order to train the click-rate-location prediction model. Based on the trained click-rate-location prediction model, a click-rate prediction model is constructed; the click-rate prediction model is used to process the non-location features of the content to be recommended to obtain the predicted click-rate of the content to be recommended.

2. The method according to claim 1, characterized in that, The loss function value is calculated using the cross-entropy loss function based on the exposure location label, the actual click-through rate, the predicted click-through rate, and the predicted exposure location of the sample content. This includes: Based on the exposure location label of the sample content and the predicted exposure location of the sample content, calculate the value of the first cross-entropy loss function; Based on the actual click-through rate of the sample content and the predicted click-through rate of the sample content, calculate the value of the second cross-entropy loss function; Based on the predicted exposure position of the sample content and the uniform distribution information of the sample content, a third cross-entropy loss function value is calculated; the uniform distribution information is the specified exposure probability of the sample content at at least one of the predicted positions.

3. The method according to claim 2, characterized in that, The step of updating the model parameters of the click-through rate-location prediction model based on the loss function value includes: Based on the first cross-entropy loss function value, the parameters in the discriminant network are updated; Based on the second cross-entropy loss function value, the parameters in the generator network, the click-through rate prediction network, and the bypass network are updated; The parameters in the generator network are updated based on the value of the third cross-entropy loss function.

4. The method according to claim 1, characterized in that, The generating network has a neural network-based factorization machine structure.

5. The method according to claim 1, characterized in that, The click-rate prediction model is constructed based on the trained click-rate-location prediction model, including: The model consisting of the generator network, the click-rate prediction network, and the bypass network in the trained click-rate-location prediction model is obtained as the click-rate prediction model.

6. The method according to claim 1, characterized in that, The exposure location label is a soft label generated based on the actual exposure location of the sample content, and the exposure probability corresponding to the actual exposure location in the soft label is greater than the exposure probability of other exposure locations; Furthermore, the sum of the exposure probabilities of each exposure position in the soft label is equal to 1.

7. A content recommendation method, characterized in that, The method includes: In response to receiving a content recommendation request from a target user, the non-location features of the content to be recommended are obtained, including content features and user profile features of the target user. The non-location features are input into the click-through rate (CTR) prediction model to obtain the predicted CTR of the content to be recommended, output by the CTR prediction model. The CTR prediction model is built upon a trained CTR-location prediction model, which includes a generative adversarial network (GAN), a CTR prediction network, and a side-channel network. The training method of the CTR-location prediction model includes: acquiring training data corresponding to the sample content, the training data including location features, non-location features, exposure location labels, and the actual CTR; the non-location features include content features and user profile features; the location features are the features of the actual exposure location of the sample content, and the exposure location labels are labels constructed based on the actual exposure location of the sample content; the location features of the sample content are input into the side-channel network of the CTR-location prediction model. A generative adversarial network (GAN) is used to obtain the exposure features of the sample content at the actual exposure location. The non-positional features of the sample content are input into the generative network of the GAN to obtain the hidden layer representation vectors of the non-positional features. A click-through rate (CTR) prediction network processes the exposure features of the sample content at the actual exposure location and the hidden layer representation vectors of the non-positional features to obtain the predicted CTR of the sample content. A discriminant network in the GAN processes the hidden layer representation vectors of the non-positional features to obtain the predicted exposure location of the sample content. Based on the exposure location label of the sample content, the actual CTR, the predicted CTR of the sample content, and the predicted exposure location of the sample content, a loss function value is calculated using the cross-entropy loss function. The model parameters of the CTR-location prediction model are updated based on the loss function value. Content recommendations are made to the target user based on the predicted click-through rate of the content to be recommended.

8. The method according to claim 7, characterized in that, The click-through rate prediction model includes a generator network, a click-through rate prediction network, and a bypass network; the method further includes: The input of the bypass network is set to the location feature corresponding to the specified location.

9. The method according to claim 7, characterized in that, The step of recommending content to the target user based on the predicted click-through rate of the content to be recommended includes: Target content is obtained based on the predicted click-through rate of the content to be recommended; Recommend the target content to the target user.

10. The method according to claim 9, characterized in that, The target content satisfies at least one of the following conditions: The predicted click-through rate is higher than the click-through rate threshold. The top n are ranked according to their predicted click-through rates from highest to lowest; n is a positive integer.

11. A click-through rate prediction model processing device, characterized in that, The device comprises: The training data acquisition module is used to acquire training data corresponding to the sample content. The training data includes location features, non-location features, exposure location tags, and actual click-through rate. The non-location features include content features and user profile features. The location features are the features of the actual exposure location of the sample content, and the exposure location tags are tags constructed based on the actual exposure location of the sample content. The predicted exposure location acquisition module is used to input the location features of the sample content into the bypass network of the click-rate-location prediction model to obtain the exposure features of the sample content at the actual exposure location. The click-rate-location prediction model also includes a generative adversarial network (GAN) and a click-rate prediction network. The non-location features of the sample content are input into the generative network of the GAN to obtain the hidden layer representation vector of the non-location features. The click-rate prediction network processes the exposure features of the sample content at the actual exposure location and the hidden layer representation vector of the non-location features to obtain the predicted click-rate of the sample content. The discriminative network in the GAN processes the hidden layer representation vector of the non-location features to obtain the predicted exposure location of the sample content. The loss function value calculation module is used to calculate the loss function value based on the exposure position label of the sample content, the actual click-through rate, the predicted click-through rate of the sample content, and the predicted exposure position of the sample content, using the cross-entropy loss function. The model training module is used to update the model parameters of the click-rate-location prediction model based on the loss function value, so as to train the click-rate-location prediction model. The model building module is used to build a click-through rate prediction model based on the trained click-through rate-location prediction model; the click-through rate prediction model is used to process the non-location features of the content to be recommended to obtain the predicted click-through rate of the content to be recommended.

12. A content recommendation device, characterized in that, The device comprises: The feature acquisition module is used to acquire non-location features of the content to be recommended in response to receiving a content recommendation request from a target user. The non-location features include content features and user profile features. The click-through rate (CTR) prediction module is used to input the non-location features into the CTR prediction model to obtain the predicted CTR of the content to be recommended, output by the CTR prediction model. The CTR prediction model is a model built based on a trained CTR-location prediction model, which includes a generative adversarial network (GAN), a CTR prediction network, and a bypass network. The training method of the CTR-location prediction model includes: acquiring training data corresponding to the sample content, the training data including location features, non-location features, exposure location labels, and the actual CTR; the non-location features include content features and user profile features; the location features are the features of the actual exposure location of the sample content, and the exposure location labels are labels constructed based on the actual exposure location of the sample content; and inputting the location features of the sample content into the CTR-location prediction model. The model uses a bypass network to obtain the exposure features of the sample content at the actual exposure location; inputs the non-positional features of the sample content into the generative network of the generative adversarial network to obtain the hidden layer representation vector of the non-positional features; processes the exposure features of the sample content at the actual exposure location and the hidden layer representation vector of the non-positional features through a click-through rate prediction network to obtain the predicted click-through rate of the sample content; processes the hidden layer representation vector of the non-positional features through a discriminative network of the generative adversarial network to obtain the predicted exposure location of the sample content; calculates the loss function value using the cross-entropy loss function based on the exposure location label of the sample content, the actual click-through rate, the predicted click-through rate of the sample content, and the predicted exposure location of the sample content; and updates the model parameters of the click-through rate-location prediction model based on the loss function value. The content recommendation module is used to recommend content to the target user based on the predicted click-through rate of the content to be recommended.

13. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing at least one instruction, at least one program, a code set, or an instruction set, the at least one instruction, the at least one program, the code set, or the instruction set being loaded and executed by the processor to implement the click-through rate prediction model processing method as described in any one of claims 1 to 6 or the content recommendation method as described in any one of claims 7 to 10.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one computer program, which is loaded and executed by a processor to implement the click-through rate prediction model processing method as described in any one of claims 1 to 6 or the content recommendation method as described in any one of claims 7 to 10.

Images