Model training method and related product

Abstract

The invention discloses a model training method and a related product. The method comprises the steps of obtaining a training sample; the training sample comprises at least two pieces of first user data of the user, at least two pieces of first content data of the content, a first label used for representing whether the user selects the content or not and a second label used for representing whether the user interacts with the content or not; inputting the training sample into a first module and a second module of the first model to obtain a target selection rate and a target conversion rate representing the interaction probability of the content under the condition that the content is selected by the user; the target selection rate is obtained based on first feature interaction and second feature interaction of the training sample and represents the probability that the content is selected by the user; obtaining a target interaction rate based on the target selection rate and the target conversion rate, wherein the target interaction rate refers to the probability of interaction between the user and the content; and adjusting parameters of the first model by using the target selection rate, the first label, the target interaction rate and the second label to obtain a prediction model capable of accurately predicting the probability that the content is selected by the user.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a model training method and related products. Background Technology

[0002] As applications, web pages, and other platforms diversify their functions, users are increasingly accustomed to browsing, clicking, and viewing videos, images, and text, or interacting with such content. The content covers a wide range of fields and is vast in quantity. Personalized content recommendations not only satisfy user preferences and enhance user experience but also improve user retention and platform stickiness. Therefore, how to develop a model that can predict the probability of users selecting content is a problem that requires ongoing exploration in this field. Summary of the Invention

[0003] Therefore, it is necessary to provide a model training method and related products to address the aforementioned technical problems. The model training method trains a predictive model that can accurately predict the probability of content being selected by a user. The related products include a training sample acquisition method, a content recommendation method, a model training device, a training sample acquisition device, a content recommendation device, computer equipment, computer-readable storage media, and computer program products.

[0004] Firstly, this application provides a model training method, the method comprising:

[0005] Obtain training samples; the training samples include at least two first user data points of the user, at least two first content data points of the content, a first tag indicating whether the user has selected the content, and a second tag indicating whether the user has interacted with the content;

[0006] The training samples are respectively input into the first module and the second module of the first model to obtain the target selection rate determined by the first module and the target conversion rate determined by the second module. The target selection rate is obtained based on the first feature interaction and the second feature interaction of the training samples, representing the probability that the content is selected by the user. The target conversion rate refers to the interaction probability when the content is selected by the user. The first feature interaction refers to the influence of the combined data on the prediction result of whether the content is selected by the user when any combination of first user data and first content data is less than a first threshold number. The second feature interaction refers to the influence of the combined data on the prediction result of whether the content is selected by the user when any combination of first user data and first content data is greater than or equal to the first threshold number.

[0007] The target interaction rate is obtained based on the target selection rate and the target conversion rate; the target interaction rate refers to the probability that a user interacts with the content.

[0008] By using the target selection rate and the first label, the target interaction rate and the second label, the parameters of the first model are adjusted to obtain the prediction model.

[0009] Using the model training method provided in the first aspect, the target interaction rate is obtained based on the target selection rate and the target conversion rate. When adjusting the parameters of the first model using the target interaction rate and the second label, backpropagation can optimize the first module used to predict the target selection rate. This allows the first model to learn not only which type of content users with a certain characteristic would choose during parameter tuning using the target selection rate and the first label, but also which type of content users with a certain characteristic are more likely to interact with. Based on this, it is beneficial to improve the prediction accuracy of the selection probability. Furthermore, it can be seen that implicitly modeling through the second module to predict the target conversion rate to obtain the target interaction rate also helps to improve the prediction accuracy of the selection probability.

[0010] Taking a first threshold of three as an example, the first module predicts the probability of content selection by the user based on the impact of any combination of two first user data and first content data in the training sample on the prediction result of whether the content is selected by the user, and the impact of any combination of at least three first user data and first content data in the training sample on the prediction result of whether the content is selected by the user. That is, it predicts the target selection rate. Through explicit modeling in the first module, on the one hand, it can consider not only the interaction strength between the higher-order feature data of at least three feature data in the training sample, but also the interaction strength between the lower-order feature data of two feature data, which is conducive to a more comprehensive prediction of the content selection probability. That is, it is conducive to improving the prediction accuracy of the selection probability. Here, the interaction strength refers to the impact of feature data combination on the prediction result. On the other hand, in the training method of this model, the target interaction rate is obtained based on the target selection rate and the target conversion rate. By improving the prediction accuracy of the target selection rate, through forward propagation, it is conducive to improving the prediction accuracy of the target interaction rate. Furthermore, when adjusting the parameters of the first model based on the target interaction rate and the second label, through backpropagation, it is also conducive to making the prediction model better and enabling the prediction model to achieve accurate prediction.

[0011] Based on the above analysis, it can be seen that the model training method provided in the first aspect can train a prediction model that can accurately predict the probability of content being selected by the user.

[0012] In one embodiment, the training samples are training samples from at least two content recommendation scenarios of the content recommendation platform, and the prediction model is used to make predictions on the content recommendation platform.

[0013] In one embodiment, the training sample further includes first context data, which is feature data of the context feature type when the user corresponding to the first user data browses the content corresponding to the first content data.

[0014] In one embodiment, the step of inputting the training samples into a first module and a second module of a first model, respectively, to obtain the target selection rate determined by the first module and the target conversion rate determined by the second module, includes:

[0015] The training samples are converted into embedding vectors using the first model;

[0016] The embedding vector is input into the first module and the second module in the first model respectively to obtain the target selection rate determined by the first module and the target conversion rate determined by the second module.

[0017] In one embodiment, adjusting the parameters of the first model using the target selection rate and the first label, the target interaction rate and the second label to obtain a prediction model includes:

[0018] Using the first difference between the target selection rate and the first label, a first loss of the first module is determined, wherein the first loss is positively correlated with the first difference;

[0019] Using the second difference between the target interaction rate and the second tag, a second loss for the first module and the second module is determined, wherein the second loss is positively correlated with the second difference;

[0020] Based on the first loss and the second loss, the parameters of the first model are adjusted to obtain the prediction model.

[0021] Secondly, this application also provides a method for obtaining training samples, wherein the training samples are used to train a model that predicts the probability of content being selected by a user in a content recommendation platform, the content recommendation platform including at least two content recommendation scenarios; the method includes:

[0022] Obtain historical data from the at least two content recommendation scenarios; the historical data includes user feature data, content feature data, a first tag indicating whether the user selected the content, and a second tag indicating whether the user interacted with the content.

[0023] Obtain a first user feature and a first content feature between at least two scene feature sets; the at least two scene feature sets correspond one-to-one with the at least two content recommendation scenarios, and the scene feature sets in the at least two scene feature sets include user feature types and content feature types existing in the corresponding content recommendation scenarios; the first user feature refers to the same user feature type between the at least two scene feature sets; the first content feature refers to the same content feature type between the at least two scene feature sets;

[0024] The historical data is filtered according to the first user feature and the first content feature to obtain a training sample; the training sample includes at least two first user data, at least two first content data, the first tag and the second tag; the first user data is the feature data belonging to the first user feature in the user feature data; the first content data is the feature data belonging to the first content feature in the content feature data.

[0025] Thirdly, this application also provides a content recommendation method, the method comprising:

[0026] Acquire target feature data; the target feature data includes target user data of the first user and target content data of the first content;

[0027] The target feature data is input into the prediction model to obtain a first selection rate; the first selection rate refers to the probability that the first content is selected by the first user.

[0028] The prediction model is trained by the method described in the first aspect or any embodiment of the first aspect.

[0029] In one embodiment, the target feature data further includes target context data; the target context data is feature data belonging to the context features when the first user browses the first content.

[0030] In one embodiment, after inputting the target feature data into the prediction model to obtain a first selection rate, the method further includes:

[0031] Obtain at least one second selection rate for at least one second content; the at least one second content is different from the first content; the at least one second content corresponds one-to-one with the at least one second selection rate, and the second selection rate refers to the probability that the corresponding second content is selected by the first user;

[0032] Based on the first selection rate and the at least one second selection rate, at least one recommended content is determined from the first content and the at least one second content.

[0033] In one embodiment, determining at least one recommended content from the first content and the at least one second content based on the first selection rate and the at least one second selection rate includes:

[0034] Based on the magnitude of the first selection rate and the at least one second selection rate, at least two recommended contents are determined from the first content and the at least one second content, and the recommendation order of the at least two recommended contents is determined.

[0035] In one embodiment, inputting the target feature data into the prediction model to obtain a first selection rate includes:

[0036] The target feature data is input into the prediction model to obtain the first selection rate and the first interaction rate; the first interaction rate refers to the probability that the first user interacts with the first content.

[0037] In one embodiment, after inputting the target feature data into the prediction model to obtain the first selection rate and the first interaction rate, the method further includes:

[0038] Obtain at least one second selection rate and at least one second interaction rate for at least one second content; the at least one second content is different from the first content; the at least one second content, the at least one second selection rate, and the at least one second interaction rate correspond one-to-one, the second selection rate refers to the probability that the corresponding second content is selected by the first user, and the second interaction rate refers to the probability that the first user interacts with the corresponding second content;

[0039] Based on a first probability and at least one second probability, at least one recommended content is determined from the first content and the at least one second content; the first probability includes the first selection rate and the first interaction rate of the first content; the second probability in the at least one second probability includes a second selection rate and a second interaction rate corresponding to a second content, wherein the second content is one of the at least one second content.

[0040] In one embodiment, the first user refers to a user requesting entry into a target recommendation scenario, the target recommendation scenario being one of at least two content recommendation scenarios, and the prediction model being trained using training samples from the at least two content recommendation scenarios; the method further includes:

[0041] Based on the target feature data, a corrected selection rate is determined; the corrected selection rate refers to the probability that the first content will be selected by the first user in the target recommendation scenario, provided that the first content has been selected by the first user.

[0042] The scene selection rate is obtained by multiplying the first selection rate and the corrected selection rate; the scene selection rate refers to the probability that the first content is selected by the first user in the target recommendation scene.

[0043] In one embodiment, determining the corrected selection rate based on the target feature data includes:

[0044] The target feature data is input into the correction model to obtain the correction selection rate;

[0045] The correction model is trained using training samples of the target in the at least two content recommendation scenarios. The training samples of the target refer to training samples of content selected by the user. The training samples of the target include at least two first user data, at least two first content data, a first tag indicating whether the user has selected the content, and a scenario tag. The scenario tag is used to indicate whether the training sample of the target is a training sample in the target recommendation scenario.

[0046] In one embodiment, the correction model is obtained by adjusting the parameters of the second model using a reference selection rate and the first label and the scene label; the reference selection rate is obtained by inputting the training samples of the target into the second model; the reference selection rate refers to the probability that the content is selected by the user in the target recommendation scene, provided that the content corresponding to at least two first content data is selected by the user corresponding to at least two first user data.

[0047] In one embodiment, after obtaining the scene selection rate based on the product of the first selection rate and the modified selection rate, the method further includes:

[0048] Obtain at least one third selection rate for at least one third content; the at least one third content is different from the first content; the at least one third content corresponds one-to-one with the at least one third selection rate, and the third selection rate refers to the probability that the corresponding third content is selected by the first user in the target recommendation scenario;

[0049] Based on the scene selection rate and the at least one third selection rate, at least one recommended content in the target recommended scene is determined from the first content and the at least one third content.

[0050] In one embodiment, determining at least one recommended content in the target recommended scenario from the first content and the at least one third content based on the scene selection rate and the at least one third selection rate includes:

[0051] Based on the magnitude of the scene selection rate and the at least one third selection rate, at least two recommended contents in the target recommended scene are determined from the first content and the at least one third content, and the recommendation order of the at least two recommended contents is determined.

[0052] Fourthly, this application also provides a model training apparatus, the apparatus comprising:

[0053] The sample acquisition module is used to acquire training samples; the training samples include at least two first user data of the user, at least two first content data of the content, a first tag indicating whether the user has selected the content, and a second tag indicating whether the user has interacted with the content.

[0054] The sample input module is used to input the training samples into the first module and the second module of the first model respectively, to obtain the target selection rate determined by the first module and the target conversion rate determined by the second module. The target selection rate is obtained based on the first feature interaction and the second feature interaction of the training samples, representing the probability that the content is selected by the user. The target conversion rate refers to the interaction probability when the content is selected by the user. The first feature interaction refers to the influence of the combined data on the prediction result of whether the content is selected by the user when any combination of first user data and first content data is less than a first threshold number. The second feature interaction refers to the influence of the combined data on the prediction result of whether the content is selected by the user when any combination of first user data and first content data is greater than or equal to the first threshold number.

[0055] The probability calculation module is used to obtain the target interaction rate based on the target selection rate and the target conversion rate;

[0056] The parameter adjustment module is used to adjust the parameters of the first model using the target selection rate and the first label, the target interaction rate and the second label, to obtain the prediction model.

[0057] Fifthly, this application also provides a training sample acquisition device, wherein the training samples are used to train a model that predicts the probability of content being selected by a user in a content recommendation platform, the content recommendation platform including at least two content recommendation scenarios; the device includes:

[0058] The history acquisition module is used to acquire historical data in the at least two content recommendation scenarios; the historical data includes user feature data, content feature data, a first tag indicating whether the user has selected the content, and a second tag indicating whether the user has interacted with the content.

[0059] The feature acquisition module is used to acquire a first user feature and a first content feature between at least two scene feature sets; the at least two scene feature sets correspond one-to-one with the at least two content recommendation scenarios, and the scene feature sets in the at least two scene feature sets include user feature types and content feature types existing in the corresponding content recommendation scenarios; the first user feature refers to the same user feature type between the at least two scene feature sets; the first content feature refers to the same content feature type between the at least two scene feature sets;

[0060] The sample determination module is used to filter the historical data according to the first user feature and the first content feature to obtain training samples; the training samples include at least two first user data, at least two first content data, the first tag and the second tag; the first user data is the feature data belonging to the first user feature in the user feature data; the first content data is the feature data belonging to the first content feature in the content feature data.

[0061] Sixthly, this application also provides a content recommendation device, the device comprising:

[0062] The data acquisition module is used to acquire target feature data; the target feature data includes target user data of the first user and target content data of the first content.

[0063] The data input module is used to input the target feature data into the prediction model to obtain a first selection rate; the first selection rate refers to the probability that the first content is selected by the first user.

[0064] The prediction model is obtained by training the method described in the first aspect or any embodiment of the first aspect.

[0065] In a seventh aspect, this application also provides a computer device, including: a memory and a processor, wherein the memory stores computer program instructions; when the computer program instructions are executed by the processor, the processor causes the processor to perform the method as described in the first aspect or any embodiment of the first aspect, or to perform the method as described in the second aspect or any embodiment of the second aspect, or to perform the method as described in the third aspect or any embodiment of the third aspect.

[0066] Eighthly, this application also provides a computer-readable storage medium storing computer program instructions; when the computer program instructions are executed on one or more processors, they perform the method as described in the first aspect or any embodiment of the first aspect, or perform the method as described in the second aspect or any embodiment of the second aspect, or perform the method as described in the third aspect or any embodiment of the third aspect.

[0067] Ninthly, this application also provides a computer program product comprising computer program instructions; when the computer program instructions are executed on a computer, the computer causes the computer to perform the method as described in the first aspect or any embodiment of the first aspect, or to perform the method as described in the second aspect or any embodiment of the second aspect, or to perform the method as described in the third aspect or any embodiment of the third aspect.

[0068] The beneficial effects that can be achieved by the training sample acquisition method provided in the second aspect, the content recommendation method provided in the third aspect, the model training device provided in the fourth aspect, the training sample acquisition device provided in the fifth aspect, the content recommendation device provided in the sixth aspect, the computer equipment provided in the seventh aspect, the computer-readable storage medium provided in the eighth aspect, and the computer program product provided in the ninth aspect can refer to the beneficial effects of the model training method provided in the first aspect. Attached Figure Description

[0069] To clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the embodiments will be described below.

[0070] Figure 1 A schematic flowchart illustrating the model training method provided in this application embodiment;

[0071] Figure 2 A schematic diagram illustrating the process of obtaining training samples provided in an embodiment of this application;

[0072] Figure 3 A schematic diagram of the model framework provided for embodiments of this application;

[0073] Figure 4 A flowchart illustrating the content recommendation method provided in this application embodiment;

[0074] Figure 5 A flowchart illustrating the content recommendation method provided in this application embodiment;

[0075] Figure 6 A schematic diagram of the model framework provided for embodiments of this application;

[0076] Figure 7A schematic diagram illustrating the process of determining the scene selection rate provided in an embodiment of this application;

[0077] Figure 8 A schematic diagram illustrating the process of obtaining training samples provided in an embodiment of this application;

[0078] Figure 9 A flowchart illustrating the training sample acquisition method provided in this application embodiment;

[0079] Figure 10 This is a schematic diagram of the structure of the model training device provided in the embodiments of this application;

[0080] Figure 11 This is a schematic diagram of the structure of the training sample acquisition device provided in the embodiments of this application;

[0081] Figure 12 A schematic diagram of the structure of the content recommendation device provided in the embodiments of this application;

[0082] Figure 13 A schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0083] To facilitate understanding of the embodiments of this application, a more comprehensive description of the embodiments of this application will be provided below with reference to the accompanying drawings. The drawings illustrate preferred embodiments of the embodiments of this application. However, the embodiments of this application can be implemented in many different forms and are not limited to the embodiments described herein. These embodiments are provided to make the disclosure of the embodiments of this application more thorough and comprehensive.

[0084] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which embodiments of this application pertain.

[0085] The terms “first,” “second,” etc., used in this document are used to distinguish different objects, not to describe a specific order. It should also be understood that the term “including / contains” means the presence of the stated features, wholes, steps, operations, parts, or combinations thereof, but does not exclude the possibility of the presence of one or more other features, wholes, steps, operations, parts, or combinations thereof.

[0086] This application provides a model training method. This method enables the training of a predictive model that can accurately predict the probability of a user clicking on content. The predictive model obtained through this method helps in providing personalized content recommendations to users. Figure 1 As shown, the model training method includes the following steps S101 to S104.

[0087] S101, Obtain training samples. The training samples contain at least two sets of first user data, at least two sets of first content data, a first label indicating whether the user selected the content, and a second label indicating whether the user interacted with the content. The training samples represent the user's behavior towards the content corresponding to the first user data, and the behavior includes at least one of the following: selection, no selection, interaction, and no interaction.

[0088] The content refers to media data such as videos, images and text, notes including videos or images and text, and product details including videos or images and text. The first user data is the feature data corresponding to the user's user characteristic type. For example, if the user characteristic type is age, the feature data corresponding to the user's age characteristic is 40 years old; if the user characteristic type is gender, the feature data corresponding to the user's gender characteristic is male. The first content data is the feature data corresponding to the content's content feature type. For example, if the content feature type is theme, the feature data corresponding to the content's theme characteristic is digital camera. Behaviors include at least one of selection, non-selection, interaction, and no interaction. This application embodiment aims to represent the user's journey from browsing content entry to selecting and viewing content details, and does not limit the specific method of selection. Optionally, the user performs a click operation on the content entry to make a selection, such as selecting by displaying content details through voice control. It should be noted that the user browsing content mentioned in this application embodiment refers to the user browsing content entry, not viewing content details. This application embodiment does not distinguish the specific interaction methods between the user and the content. Optionally, the interaction can be liking, commenting, collecting, sharing, etc. The training sample is a behavioral sample, representing user behavior towards content. Users are represented by the first user data, content by the first content data, and behavior by a first tag indicating whether a selection was made and a second tag indicating whether an interaction was performed.

[0089] In one embodiment, the training samples are training samples from at least two content recommendation scenarios of the content recommendation platform, and the prediction model is used to make predictions on the content recommendation platform.

[0090] A content recommendation platform refers to a platform with content recommendation functionality, allowing users to browse, click to view, or interact with content. At least two content recommendation scenarios refer to all content recommendation scenarios within the content recommendation platform. These scenarios include homepage recommendations and search scenarios.

[0091] Compared to training samples from a single content recommendation scenario, training samples from at least two content recommendation scenarios can expand the sample size, provide richer user behavior information, and alleviate the problem of sample sparsity in single content recommendation scenarios. Furthermore, training samples from at least two content recommendation scenarios have complementary value in terms of user preferences, providing richer and more comprehensive user profile information, preference information, and potential interest information. For new users, training samples from at least two content recommendation scenarios play a crucial role in enabling the model to accurately capture and comprehensively learn new user profile information, accurately predict new user behavior based on the learned profile information, and provide personalized recommendations based on the prediction results. Training samples from at least two content recommendation scenarios integrate user behavior information from different content recommendation scenarios, as well as content interaction information from different content recommendation scenarios. Training the model using an ensemble of training samples from multiple content recommendation scenarios, compared to training the model using training samples from a single content recommendation scenario, helps the model learn the correlations between user behavior patterns and features across multiple content recommendation scenarios, resulting in a more accurate and comprehensive learning of user profiles and preferences. Furthermore, it helps the model learn the interactive relationships between content across multiple content recommendation scenarios. For example, learning the relationship between content with the theme "women's clothing" and the fact that it won't be selected by male users in the homepage scenario or by male users in the search scenario. This, in turn, helps train a model that can more accurately predict the probability of content being selected by users. Based on more accurate predictions, content recommendation platforms can then recommend content that matches users' preferences, thereby improving the user experience.

[0092] In one embodiment, the training samples further include first context data, which is feature data of the context feature type when the user corresponding to the first user data browses the content corresponding to the first content data.

[0093] The first context data consists of feature data corresponding to the context feature types when a user browses content. Context feature types include time, location, device, and operating system. For example, the feature data corresponding to the time feature when a user browses content is the commuting period from 7:00 to 9:00; the feature data corresponding to the location feature when a user browses content is a first-tier city; and the feature data corresponding to the device feature when a user browses content is a computer.

[0094] When the training samples also include contextual data, training the model with these samples helps it not only capture and learn real user profiles and preferences, but also learn the influence of the environment on user behavior. For example, the model learns that users aged 25 to 45 prefer short news articles during their commute; that users located in first-tier cities prefer content related to local cuisine; and that users logging in on a computer prefer longer articles. This, in turn, allows the trained model to more accurately predict the probability of users selecting content based on the environment.

[0095] In one embodiment, step S101 above, obtaining training samples, includes the following steps S201 to S203.

[0096] S201, Obtain historical data from at least two content recommendation scenarios. Historical data includes user characteristic data, content content characteristic data, contextual characteristic data, first tag, and second tag.

[0097] Among them, user feature data refers to the feature data corresponding to user feature types; content feature data refers to the feature data corresponding to content feature types; and context feature data refers to the feature data corresponding to context feature types. A single historical data point represents the behavior of a user corresponding to the user feature data towards the content corresponding to the content feature data within the context of the context feature data. For example, during a commute, the behavior of a user with the age characteristic of 30 years old towards content with the topic characteristic of current events includes selection and interaction.

[0098] When users access content online through different content recommendation scenarios, the server records every behavioral event of the user in each scenario in real time and persists these behavioral events as a sample. By integrating historical behavioral events from various content recommendation scenarios, cross-scenario samples from at least two content recommendation scenarios can be obtained. In other words, historical data from at least two content recommendation scenarios can be obtained. Figure 2 This is a flowchart illustrating the process of obtaining training samples. It takes at least two content recommendation scenarios, including the homepage scenario and other scenarios, as an example. Figure 2 The cross-scenario sample generation process shown in the left-middle border refers to the process of acquiring historical data from at least two content recommendation scenarios. Specifically, Figure 2The online features shown represent potential feature types within a pre-defined content recommendation scenario. These online features include at least user feature types, content feature types, and context feature types. When a user requests access to the homepage, the server extracts feature data corresponding to the online features based on the user, the content in the initial ranking results, and the real-time environment. This feature data is then input into the online model for content recommendation. Furthermore, the server records each user's behavioral event in the homepage scenario in real-time, based on the user's browsing process. A behavioral event represents the user's behavior towards a piece of content in the real-time environment. Feature data is extracted from the online features for each behavioral event, resulting in historical data corresponding to each event. The historical data includes user feature data, content feature data, context feature data, a first tag, and a second tag that match the corresponding behavioral event. Figure 2 The extracted homepage scene data shown in the diagram refers to the integration of historical data from the obtained homepage scene. Similar to the above process, the server also integrates historical data from other scenes. Furthermore, based on the homepage scene data and other scene data, a cross-scene sample set including historical data from at least two content recommendation scenarios can be obtained. One historical data point constitutes one cross-scene sample.

[0099] S202, Obtain the first user feature, first content feature, and first context feature among at least two scene feature sets. The at least two scene feature sets correspond one-to-one with at least two content recommendation scenarios. The scene feature sets in the at least two scene feature sets include the user feature type, content feature type, and context feature type present in the corresponding content recommendation scenario. The first user feature refers to the same user feature type among the at least two scene feature sets; the first content feature refers to the same content feature type among the at least two scene feature sets; the first context feature refers to the same context feature type among the at least two scene feature sets.

[0100] The scenario feature set includes all user feature types, content feature types, and context feature types mined in the corresponding content recommendation scenario. For example... Figure 2 As shown, online features in content recommendation scenarios are provided by corresponding scenario feature sets. For example, offline mining might reveal content location features in a homepage scenario with two columns. These location features can then be used as online features within the homepage scenario, allowing for the extraction of corresponding feature data to be input into the online model.

[0101] Because online features differ across content recommendation scenarios, the feature data contained in historical data will also vary. After constructing the historical data, further optimization is needed to unify the feature dimensions and feature semantics of the feature data contained in each historical dataset. Feature dimension refers to the number of feature types, and feature semantics refers to the specific meaning of each feature type. First, the first user feature, first content feature, and first context feature are obtained from at least two scenario feature sets. Specifically, taking at least two content recommendation scenarios, including the homepage scenario and other scenarios, as an example... Figure 2 The process of generating cross-scene feature sets, as shown in the right-hand box, is the process of obtaining the first user feature, the first content feature, and the first context feature. The homepage scene feature set, i.e., the scene feature set of the homepage scene, includes user feature types, content feature types, and context feature types existing in the homepage scene: feature a, feature b…feature n. Other scene feature sets, i.e., scene feature sets of other scenarios, include user feature types, content feature types, and context feature types existing in other scenarios: feature b, feature c…feature n. The cross-scene feature set is obtained from the homepage scene feature set and other scene feature sets, i.e., at least two scene feature sets share the same user feature types, content feature types, and context feature types: feature b, feature c…feature n. In this embodiment, the user feature types, content feature types, and context feature types shared by at least two scene feature sets are sequentially referred to as the first user feature, the first content feature, and the first context feature. Through the above process, the first user feature, the first content feature, and the first context feature: feature b, feature c…feature n, are obtained from at least two scene feature sets. Understandably, each historical data obtained from at least two content recommendation scenarios contains feature data corresponding to the first user feature, the first content feature, and the first context feature between at least two scenario feature sets.

[0102] S203, the historical data is filtered according to the first user feature, the first content feature, and the first context feature to obtain training samples. The first user data is the feature data belonging to the first user feature in the user feature data; the first content data is the feature data belonging to the first content feature in the content feature data; and the first context data is the feature data belonging to the first context feature in the context feature data.

[0103] If at least two scene feature sets share the same first user feature, first content feature, and first context feature, then filtering historical data according to the same feature type can unify the feature dimensions and semantics of the feature data contained in each historical data set. For example, consider six feature types: two user feature types, two content feature types, and two context feature types. The specific meanings of these six feature types are age, gender; theme, length; time, and location, respectively. Filtering historical data according to these six feature types filters out the differentiated feature data contained in each historical data set, resulting in training samples. Each training sample contains feature data from the six feature types: age, gender, theme, length, time, and location. The feature dimensions and semantics of the feature data contained in each training sample are identical. Figure 2 As shown, historical data are filtered according to the same feature type: feature b, feature c... feature n, to obtain training samples. Each training sample contains feature data of feature b, feature c... feature n. Among them, the first user data, first content data, and first context data contained in the training samples are, in order, the feature data of the first user feature, the feature data of the first content feature, and the feature data of the first context feature.

[0104] The method in this embodiment obtains and unifies the feature dimensions and semantics of the feature data contained in each historical data set, resulting in training samples with identical feature dimensions and semantics. Unifying feature dimensions and semantics yields effective, information-rich, and high-quality training samples, enabling effective model training and improving training performance. Training the model with these training samples from at least two content recommendation scenarios enhances the model's understanding of data from different content recommendation scenarios, improves feature utilization efficiency, and helps the model better capture cross-scenario behavioral patterns during training, thereby improving prediction accuracy and consistency.

[0105] S102, the training samples are input into the first module and the second module of the first model respectively to obtain the target selection rate determined by the first module and the target conversion rate determined by the second module. The target selection rate is obtained based on the first feature interaction and the second feature interaction of the training samples, representing the probability that the content is selected by the user; the target conversion rate refers to the interaction probability when the content is selected by the user; the first feature interaction refers to the influence of the combined data on the prediction result of whether the content is selected by the user when any combination of first user data and first content data is less than a first threshold number; the second feature interaction refers to the influence of the combined data on the prediction result of whether the content is selected by the user when any combination of first user data and first content data is greater than or equal to the first threshold number.

[0106] After generating historical data and constructing the same feature types, the original behavioral events are transformed into valid training samples. These training samples are then input into the first model for training, resulting in a predictive model capable of predicting the probability of content being selected by the user.

[0107] Figure 3 This is a schematic diagram of the model framework. Exemplarily, a feasible model framework for the first model in this application embodiment is as follows: Figure 3As shown. The first model includes at least a first module and a second module. The first module includes a Factorization Machine (FM) component and a Deep Neural Network (DNN) consisting of at least concatenated layers and multiple fully connected layers. In the first module, FM is primarily responsible for learning low-order feature interactions in the training samples, while the DNN is primarily responsible for learning high-order feature interactions in the training samples. Both components aim to determine whether content will be selected by the user. The low-order feature interactions are the first feature interactions described in this embodiment, and the high-order feature interactions are the second feature interactions described in this embodiment. Specifically, taking a first threshold of three as an example, FM primarily learns the influence of any two combinations of first user data and first content data in the training samples on the prediction result of whether the content will be selected by the user. For example, the user's gender characteristic is male, and the content's theme characteristic is basketball. When these two feature data are combined, they will have a positive impact on whether the content will be selected by the user. By learning the influence of these two feature data combinations, FM can learn that male users have a preference for basketball-themed content. The DNN primarily learns the impact of any combination of at least three primary user data and primary content data in the training samples on the prediction results. For example, a user's age is 30 years old, their gender is female, and the content's theme is clothing matching. This combination of three features positively influences whether the user will select the content. By learning the influence of this combination, the DNN can learn that young female users are interested in creating clothing matching guides. When the training samples also include primary context data, the DNN primarily learns the impact of any combination of at least three primary user data, primary content data, and primary context data in the training samples on the prediction results. For example, a user's age is 30 years old, the content's theme is clothing matching, and the environment's time is 8 PM. This combination of three features positively influences whether the user will select the content. By learning the influence of this combination, the DNN can learn that young users are interested in creating clothing matching guides at night. FM and DNN output based on the learned information. The first module, containing FM and DNN, couples the outputs of the two components to determine the target selection rate. That is, determining the probability that content will be selected by the user. The second module contains an independent Multilayer Perceptron (MLP). In other words, the second module also contains a Deep Neural Network (DNN) that includes at least concatenated layers and multiple fully connected layers.DNNs primarily learn high-order feature interactions from training samples and output the learned information. Unlike the DNN in the first module, the DNN in the second module aims to determine whether a user will interact with content given that the content has been selected. Based on the DNN's output, the second module determines the target conversion rate; that is, it determines the probability of interaction when the content is selected by the user.

[0108] In one embodiment, step S102 above, which involves inputting training samples into the first module and the second module of the first model respectively, to obtain the target selection rate determined by the first module and the target conversion rate determined by the second module, includes the following steps S301 to S302.

[0109] S301 uses the first model to convert training samples into embedding vectors.

[0110] Specifically, such as Figure 3 As shown, the first module also includes an embedding layer. Training samples are input into the embedding layer of the first model, and the embedding layer uses feature engineering methods such as bucketing and hashing to transform the training samples containing feature data into dense embedding vectors.

[0111] S302, the embedding vectors are input into the first module and the second module in the first model respectively to obtain the target selection rate determined by the first module and the target conversion rate determined by the second module.

[0112] Furthermore, using the first model, the embedding vectors output from the embedding layer are input into the first module and the second module respectively, to obtain the target selection rate determined by the first module and the target conversion rate determined by the second module. The first module and the second module share the embedding layer and training samples.

[0113] Through transformation, the embedding vectors not only retain the original feature data information but also possess a format recognizable by the first model, which is beneficial for improving model learning efficiency. When training samples come from at least two content recommendation scenarios, converting each training sample into embedding vectors through embedding technology also has the following benefits: Each embedding vector can model user behavior patterns in different content recommendation scenarios, as well as the correlations between content interactions in different content recommendation scenarios. This processing method allows the first model to learn user behavior information for a certain feature in different content recommendation scenarios, and to learn the interactive correlations of content for a certain feature in different content recommendation scenarios, when training the model based on each embedding vector. In the first module, FM explicitly models low-order feature interactions in cross-scenario embedding vectors through vector inner product; DNN learns high-order feature interactions in cross-scenario embedding vectors through a multilayer perceptron. Embedding technology can provide richer information for model learning.

[0114] S103, the target interaction rate is obtained based on the target selection rate and the target conversion rate.

[0115] When we obtain the probability that content is selected by a user and the probability of interaction with the content given that it has been selected, we can obtain the probability that the user will interact with the content based on these two factors. That is, we obtain the target interaction rate. For example, the target interaction rate is obtained by multiplying the target selection rate and the target conversion rate. For example, depending on the actual situation, the target interaction rate can be obtained by multiplying the target selection rate, the target conversion rate, and a first coefficient. For example, as... Figure 3 As shown, the first model outputs the target selection rate, target conversion rate, and target interaction rate.

[0116] In this embodiment, the second module implicitly models the transformation process from selection to interaction, outputting the target conversion rate, while the first module explicitly models and outputs the target selection rate. This design allows the task of predicting whether content will be selected by the user to fully learn feature interactions through joint optimization of FM and DNN, while the task of predicting whether content will be interacted with by the user focuses on modeling sequential interaction behaviors. The first and second modules share the same underlying feature representations. While preserving task independence, the joint training mechanism models the sequential dependency relationship of "selection-interaction," effectively improving the prediction accuracy of user selection behavior.

[0117] S104. Using the target selection rate, the first label, the target interaction rate, and the second label, adjust the parameters of the first model to obtain the prediction model.

[0118] In one embodiment, step S104 above, which uses the target selection rate and the first label, the target interaction rate and the second label to adjust the parameters of the first model to obtain the prediction model, includes the following steps S401 to S403.

[0119] S401, using the first difference between the target selection rate and the first label, determine the first loss of the first module. The first loss is positively correlated with the first difference.

[0120] S402, using the second difference between the target interaction rate and the second label, determine the second loss of the first module and the second module. The second loss is positively correlated with the second difference.

[0121] When the first model outputs the target selection rate corresponding to the training samples, a first difference is calculated between the target selection rate and the first label in the training samples indicating whether or not a selection has been made. Then, using the first difference and a preset first loss function, a first loss for the first module is determined. The first loss is positively correlated with the first difference. When the first model outputs the target interaction rate corresponding to the training samples, a second difference is calculated between the target selection rate and the second label in the training samples indicating whether or not an interaction has been made. Then, using the second difference and a preset second loss function, a second loss for both the first and second modules is determined. The second loss is positively correlated with the second difference.

[0122] S403, based on the first loss and the second loss, adjust the parameters of the first model to obtain the prediction model.

[0123] Then, the total loss is obtained based on the first and second losses. Through backpropagation, the gradients of each parameter of the first model are calculated based on the total loss. The parameters of the first model are adjusted based on these gradients to obtain the prediction model.

[0124] The prediction model is obtained through the model training method in this application embodiment. The prediction model effectively learns user profiles and captures user preferences, accurately predicting the probability of content being selected by users. Specifically, the target interaction rate is obtained based on the target selection rate and the target conversion rate. When adjusting the parameters of the first model using the target interaction rate and the second label, backpropagation can optimize the first module used to predict the target selection rate. This allows the first model to learn not only which type of content users with a certain characteristic would choose during parameter tuning using the target selection rate and the first label, but also which type of content users with a certain characteristic are more likely to interact with during parameter tuning using the target interaction rate and the second label. The model includes a first module and a second module, and the overall model is optimized based on the outputs of the first module and the second module. This design makes the selection probability prediction and interaction behavior modeling both task-independent and interrelated. Based on this, it is beneficial to improve the prediction accuracy of the selection probability. Furthermore, it can be seen that implicit modeling through the second module to predict the target conversion rate to obtain the target interaction rate also helps to improve the prediction accuracy of the selection probability.

[0125] Furthermore, the first module in the first model predicts the probability of content being selected by the user by learning the impact of arbitrary low-order feature data combinations on the prediction results, and learning the impact of arbitrary high-order feature data combinations on the prediction results. That is, it predicts the target selection rate. Explicit modeling in the first module considers not only the interaction strength between high-order feature data in the training samples but also the interaction strength between low-order feature data, which is beneficial for a more comprehensive prediction of the content selection probability. That is, it helps improve the accuracy of the selection probability prediction. The FM in the first module utilizes low-rank matrix factorization and shared embedding mechanisms to efficiently process sparse data while explicitly modeling the complex relationships between features, providing rich semantic information, which is beneficial for improving the model's performance in prediction, recommendation, and other tasks. Here, interaction strength refers to the impact of feature data combinations on the prediction results. On the other hand, in the training method of this model, the target interaction rate is obtained based on the target selection rate and the target conversion rate. By improving the prediction accuracy of the target selection rate, based on forward propagation, it is beneficial to improve the prediction accuracy of the target interaction rate. Furthermore, when adjusting the parameters of the first model based on the target interaction rate and the second label, backpropagation can also help to improve the prediction model and enable it to make accurate predictions.

[0126] As demonstrated by the aforementioned model training method embodiments, the prediction model obtained through offline training possesses the ability to accurately and comprehensively learn user profiles and preferences, thereby accurately predicting the probability of content being selected by the user. Based on this, this application embodiment also provides a content recommendation method to recommend content that meets user preferences based on the selection probability predicted by the prediction model. Figure 4 As shown, the content recommendation method includes the following steps S501 to S502.

[0127] S501, Obtain target feature data. The target feature data includes the target user data of the first user and the target content data of the first content.

[0128] Specifically, an offline-trained prediction model is deployed in the content recommendation platform. Target feature data, including target user data for the first user and target content data for the first content, is acquired. Here, target user data is the feature data of the first user feature, and target content data is the feature data of the first content feature. The first user feature and the first content feature are, respectively, user feature types and content feature types that are identical across at least two scene feature sets. One content recommendation scenario in the content recommendation platform corresponds to one scene feature set. For a detailed discussion of the first user feature and the first content feature, please refer to the corresponding discussion in the aforementioned model training method embodiments.

[0129] For example, feature data of the first user accessing the content recommendation platform is obtained according to the first user characteristics, and feature data of the first content in the content recommendation platform is obtained according to the first content characteristics, thus obtaining target feature data.

[0130] S502, Input the target feature data into the prediction model to obtain the first selection rate. The first selection rate refers to the probability that the first content is selected by the first user.

[0131] The prediction model is trained using the model training method described in any of the aforementioned model training method embodiments.

[0132] By inputting the target feature data into the prediction model, a first selection rate representing the probability that the first content is selected by the first user can be obtained. The method in this embodiment can accurately predict the probability that content will be selected by the user.

[0133] In one embodiment, the target feature data further includes target context data. The target context data is feature data of the first context feature; the feature data of the first context feature belongs to the feature data of the context feature type when the first user browses the first content.

[0134] In this embodiment, where the training samples used to train the prediction model also include first context data, the target feature data input to the prediction model also includes target context data. The first context data belongs to the feature data of the context feature type represented by the training samples when the user browses the content. The first context feature is the same context feature type across at least two scene feature sets. For a detailed discussion of the first context data and the first context feature, please refer to the corresponding discussion in the foregoing model training method embodiments. For example, this embodiment obtains the feature data of the first context feature from the feature data of the context feature type when the first user browses the first content, based on the first context feature. That is, it obtains the target context data.

[0135] In one embodiment, after inputting the target feature data into the prediction model in step S502 above to obtain the first selection rate, the content recommendation method further includes the following steps S601 to S602.

[0136] S601, obtain at least one second selection rate for at least one second content. The at least one second content is different from the first content; there is a one-to-one correspondence between the at least one second content and the at least one second selection rate, where the second selection rate refers to the probability that the corresponding second content is selected by the first user.

[0137] For example, referring to the method for obtaining the first selection rate in the aforementioned content recommendation method embodiments, the target user data of the first user and the feature data of the first content feature of a second content are used as inputs to the prediction model to obtain a second selection rate. The second selection rate refers to the probability that the corresponding second content is selected by the first user. Based on this, in addition to obtaining the first selection rate of the first content, at least one second selection rate of at least one second content in the content recommendation platform can also be obtained.

[0138] S602, based on a first selection rate and at least one second selection rate, determine at least one recommended content from the first content and at least one second content.

[0139] At least one piece of recommended content refers to content displayed on the content recommendation platform in response to a first user's access, for the first user to browse. When the first user enters the recommended content section, they can select recommended content to view and interact with. For example, they can click on recommended content to view it, or like or comment on it.

[0140] For example, a first selection rate threshold is set. The magnitudes of the first selection rate and at least one second selection rate are compared with the first selection rate threshold to obtain at least one fourth selection rate greater than the first selection rate threshold; the fourth selection rate among the at least one fourth selection rate is either the first selection rate or the second selection rate. At least one fourth piece of content corresponding to the at least one fourth selection rate is designated as at least one recommended piece of content; the fourth piece of content among the at least one fourth piece of content is either the first content or the second content.

[0141] For example, a first quantity threshold is set. The first selection rate and at least one second selection rate are sorted in descending order. A fourth selection rate is determined based on the number of first quantity thresholds that rank highest in the sorting; the fourth selection rate is either the first selection rate or the second selection rate. The fourth content corresponding to each of the fourth selection rates corresponding to the number of first quantity thresholds is determined as recommended content; the fourth content is either the first content or the second content.

[0142] The method in this embodiment, in response to a first user's access to the content recommendation platform, utilizes a prediction model deployed on the content recommendation platform to predict the probability of the first user selecting first content and at least one second piece of content, respectively. Then, based on the prediction results of the prediction model, at least one recommended piece of content that satisfies the first user's preferences is determined from the first content and at least one second piece of content. This content recommendation method facilitates personalized recommendations, thereby improving the user's platform experience and increasing user retention and platform stickiness.

[0143] In one embodiment, step S602 above, which determines at least one recommended content from the first content and at least one second content based on a first selection rate and at least one second selection rate, includes the following steps:

[0144] Based on the magnitude of a first selection rate and at least one second selection rate, at least two recommended contents are determined from the first content and at least one second content, and the recommendation order of the at least two recommended contents is determined.

[0145] This embodiment can not only determine at least one recommended content from first content and at least one second content based on a first selection rate and at least one second selection rate, but also determine the recommendation order of the at least one recommended content. Specifically, referring to the method in the foregoing embodiment, a first selection rate threshold or a first quantity threshold is set. Then, based on the first selection rate and at least one second selection rate, at least two recommended content are determined from the first content and at least one second content. Furthermore, the order of the at least two fourth selection rates corresponding to the at least two recommended content is used as the recommendation order of the at least two recommended content. Here, the fourth selection rate is either the first selection rate or the second selection rate.

[0146] The higher the probability of being selected by the first user, the more the corresponding first or second content satisfies the user's preferences. Prioritizing the first or second content that best meets these preferences, or displaying it in a position that attracts the first user's attention, can significantly improve the user's platform experience, as well as increase user retention and platform stickiness.

[0147] In one embodiment, step S502 above, which involves inputting the target feature data into the prediction model to obtain a first selection rate, includes the following steps:

[0148] The target feature data is input into the prediction model to obtain the first selection rate and the first interaction rate; the first interaction rate refers to the probability that the first user interacts with the first content.

[0149] As can be seen from the aforementioned model training method embodiments, the prediction model can output not only the probability that content is selected by the user, but also the probability that content is interacted with by the user. Therefore, inputting target feature data into the prediction model yields not only a first selection rate but also a first interaction rate. The first interaction rate refers to the probability that the first content is interacted with by the first user, that is, the probability that the first user interacts with the first content. The method in this embodiment can accurately predict both the probability that content is selected by the user and the probability that content is interacted with by the user.

[0150] In one embodiment, after the above steps: inputting target feature data into the prediction model to obtain a first selection rate and a first interaction rate; the first interaction rate refers to the probability that a first user interacts with the first content, the content recommendation method further includes the following steps S701 to S702.

[0151] S701, obtain at least one second selection rate and at least one second interaction rate of at least one second content. The at least one second content is different from the first content; the at least one second content, at least one second selection rate, and at least one second interaction rate correspond one-to-one. The second selection rate refers to the probability that the corresponding second content is selected by the first user, and the second interaction rate refers to the probability that the first user interacts with the corresponding second content.

[0152] For example, referring to the method for obtaining the first selection rate and the first interaction rate in the aforementioned content recommendation method embodiments, the target user data of the first user and the feature data of the first content feature of a second content are used as inputs to the prediction model to obtain a second selection rate and a second interaction rate. The second selection rate refers to the probability that the corresponding second content is selected by the first user, and the second interaction rate refers to the probability that the corresponding second content is interacted with by the first user. Based on this, in addition to obtaining the first selection rate and the first interaction rate of the first content, at least one second selection rate and at least one second interaction rate of at least one second content can also be obtained.

[0153] S702, based on a first probability and at least one second probability, determine at least one recommended content from the first content and at least one second content. The first probability includes a first selection rate and a first interaction rate of the first content; the second probability in the at least one second probability includes a second selection rate and a second interaction rate corresponding to a second content, and the second content is one of the at least one second content.

[0154] Specifically, a recommendation value is obtained based on the selection rate and the interaction rate. For example, the recommendation value is obtained by multiplying the sum of the selection rate and the interaction rate by a second coefficient. For example, the recommendation value is obtained by multiplying the sum of a first multiple of the selection rate and a second multiple of the interaction rate. For example, the recommendation value is obtained by multiplying the sum of a first multiple of the selection rate and a second multiple of the interaction rate by a second coefficient. Based on this, a first recommendation value for the first content is determined using the first selection rate and the first interaction rate. A second recommendation value for each second content is determined using the second selection rate and the second interaction rate.

[0155] For example, a probability threshold is set. The first recommendation value and at least one second recommendation value are compared with the probability threshold to obtain at least one fourth recommendation value greater than the probability threshold; the fourth recommendation value among the at least one fourth recommendation value is either the first recommendation value or the second recommendation value. At least one fourth content corresponding to the at least one fourth recommendation value is taken as at least one recommended content; the fourth content among the at least one fourth content is either the first content or the second content.

[0156] For example, a second quantity threshold is set. The first recommended value and at least one second recommended value are sorted in descending order. A fourth recommended value is determined based on the number of second quantity thresholds that appear at the top of the sorting; the fourth recommended value is either the first or second recommended value. The fourth content corresponding to each of the fourth recommended values ​​based on the number of second quantity thresholds is determined as the recommended content; the fourth content is either the first or second content.

[0157] The method in this embodiment, in response to a first user's access to a content recommendation platform, utilizes a prediction model deployed on the content recommendation platform to predict the probability of the first user selecting first content and at least one second piece of content, as well as the probability of the first user interacting with the first content and at least one second piece of content. Then, based on the probability of selection and the probability of interaction, at least one recommended piece of content that satisfies the first user's preferences is determined from the first content and at least one second piece of content. The recommended content determined by this method not only has a high probability of being selected but also a high probability of converting selection into interaction, making the recommended content more closely aligned with the user's preferences.

[0158] The prediction model is trained using training samples from at least two content recommendation scenarios. Under this premise, when applied to a content recommendation platform, the prediction model can accurately predict the probability of content selection. However, if the prediction model is specifically deployed in a single content recommendation scenario within the platform, the accuracy of the prediction results has room for improvement due to the difference in distribution between the training and prediction samples. Specifically, the training samples contain user behavior information across multiple content recommendation scenarios, and user behavior patterns differ across these scenarios. For example, users typically have a higher probability of selecting content in a search scenario, while their probability of selecting content in a homepage scenario is relatively lower. If the prediction model is deployed in a specific homepage scenario, the difference in distribution between the training and prediction samples may cause the features learned by the model during training to become disconnected from the actual prediction scenario, thus affecting the accuracy and generalization ability of the prediction results. To address this issue, this application innovatively improves the probability prediction process and content recommendation process in a specific content recommendation scenario. The improved content recommendation method can adaptively adjust from the probability across the entire content recommendation platform to the probability of the target content recommendation scenario, so as to accurately predict the probability of the content being selected by the user in the target content recommendation scenario.

[0159] Specifically, in one embodiment, the first user refers to the user requesting entry into the target recommendation scenario, which is one of at least two content recommendation scenarios. The prediction model is trained using training samples from at least two content recommendation scenarios. Figure 5As shown, the content recommendation method includes the following steps S801 to S804. The specific descriptions of steps S801 to S802 can be found in the descriptions of S501 to S502 in the previous embodiments, and will not be repeated in this embodiment.

[0160] S801, Obtain target feature data. The target feature data includes the target user data of the first user and the target content data of the first content.

[0161] S802, the target feature data is input into the prediction model to obtain the first selection rate. The first selection rate refers to the probability that the first content is selected by the first user. The prediction model is trained using the model training method described in any of the aforementioned model training method embodiments.

[0162] S803, determine the corrected selection rate based on the target feature data. The corrected selection rate refers to the probability that the first content will be selected by the first user in the target recommendation scenario, assuming that the first content has been selected by the first user.

[0163] The target recommendation scenario is one of at least two content recommendation scenarios on the content recommendation platform, i.e., the desired content recommendation scenario. Unlike applying a prediction model directly to the content recommendation platform, when deploying a prediction model within a specific target recommendation scenario on the content recommendation platform to predict probabilities within that scenario, a revised selection rate is determined based on target feature data. The revised selection rate refers to the probability that the first content will be selected by the first user in the target recommendation scenario, assuming the first content is selected by the first user. The revised selection rate is used to adjust the initial selection rate to obtain a more accurate prediction result for the target recommendation scenario.

[0164] In one embodiment, step S803 above, which determines the corrected selection rate based on the target feature data, includes the following steps: inputting the target feature data into the correction model to obtain the corrected selection rate.

[0165] The corrected model is trained using training samples from at least two content recommendation scenarios. The training samples for the target are those where the content corresponding to the first content data is selected by the user corresponding to the first user data; in other words, the user has selected the content. The training samples for the target include at least two sets of first user data, at least two sets of first content data, a first tag indicating whether the user selected the content, and a scenario tag. The scenario tag indicates whether the training sample for the target is a training sample within the target recommendation scenario.

[0166] In this embodiment, the correction model and the prediction model share training samples. The training samples for the prediction model are training samples from at least two content recommendation scenarios, representing the user's behavior towards the content corresponding to the first user data. Behaviors include at least one of selection, non-selection, interaction, and no interaction. The correction model is trained using a subset of these training samples; in this embodiment, this subset is referred to as the target's training samples. The target's training samples are those from the training samples where the user's behavior towards the content includes selection. For example, if selection is a click, the target's training samples are those where the user clicked on the content. The target's training samples include training samples where the user clicked without interaction and training samples where the user clicked and interacted. For instance, if the training samples used to train the prediction model contain first context data, the target's training samples used to train the correction model in this embodiment also contain first context data. Sharing training samples to train both the prediction and correction models ensures the synergy of the two models in the feature space.

[0167] In this embodiment, in addition to the first user data, the first content data, and the first tag indicating whether a selection is made, the training sample also includes a scene tag indicating whether the training sample of the target is a training sample in the target recommendation scenario.

[0168] When acquiring historical data from at least two content recommendation scenarios, each historical data point includes not only the feature data of the corresponding online characteristics but also the scenario identifier. For example... Figure 2 As shown, the cross-scenario sample generation technology integrates historical data from multiple content recommendation scenarios to construct a cross-scenario sample set with scenario tag mapping. Each cross-scenario sample in the cross-scenario sample set, i.e., each historical data point, retains a clear scenario identifier. This scenario identifier identifies which content recommendation scenario the cross-scenario sample originates from, such as the homepage scenario or other scenarios. Each cross-scenario sample contains feature data from its corresponding content recommendation scenario and can trace its origin through the scenario identifier. By forming scenario-distributed samples, feature learning can accurately cover various scenario needs that users may encounter, providing a precise sample foundation for probability adjustments for target recommendation scenarios. This, in turn, helps improve the accuracy of probability prediction for target recommendation scenarios, enabling precise content recommendations to users within those scenarios.

[0169] This embodiment aims to identify different content recommendation scenarios to clarify the scenario source of historical data, without restricting the method of scenario identification. For example, historical data scenarios are identified according to a predefined scenario list. Specifically, taking historical data originating from the homepage scenario as an example, its scenario identification is [Homepage scenario: True, Other scenario 1: False, Other scenario 2: False, ...]. Taking historical data originating from other scenario 1 as an example, its scenario identification is [Homepage scenario: False, Other scenario 1: True, Other scenario 2: False, ...]. Here, True indicates belonging to that scenario, and False indicates not belonging to that scenario. For example, historical data scenarios are identified using predefined scenario codes. Specifically, numbers can be used as codes, with the homepage scenario as code 0, other scenario 1 as code 1, other scenario 2 as code 2, and so on. Alternatively, the scenario name can be directly used as the scenario code.

[0170] Based on scene identification of each historical data point, feature data filtering is performed on the historical data to obtain training samples, which also include scene identification. For the target recommendation scenario, the scene identification of the training sample can be converted into a scene label of 0 or 1. Here, 0 indicates that the training sample does not belong to the target recommendation scenario, and 1 indicates that the training sample belongs to the target recommendation scenario.

[0171] In one embodiment, the correction model is obtained by adjusting the parameters of the second model using the reference selection rate and the first label and scene label; the reference selection rate is obtained by inputting the training samples of the target into the second model; the reference selection rate refers to the probability that the content is selected by the user in the target recommendation scenario, provided that the content corresponding to at least two first content data is selected by the user corresponding to at least two first user data.

[0172] Figure 6 This is a schematic diagram of the model framework. Exemplarily, a feasible model framework for the second model in this application embodiment is as follows: Figure 6As shown, taking a training sample containing first user data, first content data, and first context data as an example, the training sample is input into the second model. Specifically, a target training sample is determined from the training samples; the target training sample is then input into the second model to obtain a reference selection rate. Exemplarily, the second model includes at least an embedding layer and a deep neural network (DNN). The DNN includes at least a concatenation layer and multiple fully connected layers. The target training sample is input into the embedding layer, which uses feature engineering methods such as bucketing and hashing to transform the training sample containing feature data into a dense embedding vector. Then, the embedding vector output from the embedding layer is input into the concatenation layer in the DNN. The DNN, assuming the content is selected by the user, determines whether the content was selected in the target recommendation scenario as a task, and learns high-order feature interactions in the cross-scenario embedding vector through a multilayer perceptron. Finally, based on the output of the DNN, the second model outputs a reference selection rate corresponding to the training sample. The reference selection rate refers to the probability that a content will be selected by a user in the target recommendation scenario, assuming that the content corresponding to the first content data is selected by the user corresponding to the first user data. Regarding the role of converting training samples into embedding vectors through embedding technology and the meaning of DNN learning higher-order feature interactions in this embodiment, please refer to the corresponding descriptions in the foregoing embodiments; these will not be repeated here.

[0173] Then, the loss of the second model is determined by using the differences between the reference selection rate and the first label and the scene label. The gradients of each parameter of the second model are calculated through backpropagation. Based on the gradients of each parameter, the parameters of the second model are adjusted to obtain the corrected model.

[0174] For example, the modified model and the prediction model share not only training samples but also an embedding layer. That is, after the training samples are input into the embedding layer to be converted into embedding vectors, these embedding vectors are input into the first and second modules of the first model to train the prediction model. Similarly, the embedding vectors corresponding to the target's training samples are input into the second model to train the modified model. By sharing training samples and the embedding layer, joint training of the prediction model and the modified model is achieved, ensuring synergy between the two models in the feature space.

[0175] The training process described in this embodiment allows for the training of a corrected model. This corrected model predicts the probability that a user will select content within a target recommendation scenario, assuming the content has already been selected. The target recommendation scenario is determined by adjusting the scenario labels. For example, a 1 in the scenario label indicates that the training sample belongs to the homepage scenario, and a 0 indicates that the training sample does not belong to the homepage scenario. These scenario labels are then used to supervise the output of the second model. The corrected model trained in this way is then used to correct the output of the prediction model to determine the probability that the content will be selected within the homepage scenario.

[0176] The prediction model, based on training samples from at least two content recommendation scenarios, predicts the probability of a user selecting content on a content recommendation platform. While the training samples incorporate cross-scenario behavioral information, the lack of explicit modeling of behavioral differences between different content recommendation scenarios leads to probability bias when the prediction model is deployed in a specific target recommendation scenario for scenario-specific probability prediction. This embodiment trains a correction model to model behavioral differences between different content recommendation scenarios, thus correcting for probability bias. The prediction model and the correction model share input and work collaboratively, preserving cross-scenario commonalities while also considering scenario-specific user behavior patterns, which helps to accurately predict the probability of a user selecting content in a specific scenario.

[0177] S804, based on the product of the first selection rate and the modified selection rate, yields the scene selection rate. The scene selection rate refers to the probability that the first piece of content is selected by the first user in the target recommendation scene.

[0178] Figure 7 This is a schematic diagram illustrating the process of determining the scene selection rate. Taking the target feature data as an example, which also includes the first context feature, i.e., taking the data containing the target context data as an example, refer to... Figure 7 The target feature data is input in parallel into the prediction model and the correction model to obtain the first selection rate output by the prediction model and the corrected selection rate output by the correction model. The first selection rate refers to the probability that the first content is selected by the first user in the first context. That is, in the coarse ranking stage, the probability that the first content is selected by the first user across the entire domain. The corrected selection rate refers to the probability that the first content will be selected by the first user in the target recommendation scenario, given that the first content has already been selected by the first user. Figure 7The prediction model and the correction model shown share an embedding layer. Furthermore, in this embodiment, the scene selection rate is obtained based on the product of the first selection rate and the corrected selection rate. The scene selection rate refers to the probability that the first content is selected by the first user in the target recommendation scene. For example, the scene selection rate is obtained based on the product of the first selection rate, the corrected selection rate, and the third coefficient. Through the above method, an adaptive adjustment from the probability of the entire domain to the probability of the target recommendation scene is achieved, resulting in better alignment between the distribution of training data and prediction data, and accurately determining the probability that content is selected by the user in the target recommendation scene.

[0179] For example, the process of determining the scene selection rate in this embodiment can be regarded as the output process of the cross-domain probabilistic model. Specifically, the cross-domain probabilistic model includes two parallel sub-models: a prediction model and a correction model. The prediction model is used to determine the probability that content is selected by the user across the entire content recommendation platform, and the correction model is used to determine the probability that content is selected by the user in the target recommendation scene, given that the content is selected by the user. For example, the two sub-models share the same input and also share the underlying feature embedding layer. The cross-domain probabilistic model combines the outputs of the two sub-models through a probability multiplication mechanism to output the scene selection rate. That is, it outputs the probability that content is selected by the user in the target recommendation scene. The cross-domain probabilistic model not only alleviates the sample bias problem of deploying the prediction model in a specific target recommendation scene, but also accurately matches user preferences to specific scene requirements. While learning the overall user preferences, the cross-domain probabilistic model can also adapt to the differentiated features under different content recommendation scenarios.

[0180] Having trained the prediction model and the correction model, the method for determining the scene selection rate in this embodiment is applied to the target recommendation scene. Figure 8 The flowchart illustrating the process of obtaining training samples is shown. Taking the target recommendation scenario as the homepage scenario, and the scenario selection rate output by the cross-domain probability model as an example, as follows... Figure 8As shown, a cross-domain probabilistic model, including a prediction model and a correction model, is deployed in the homepage scenario. In response to a user's access request to the homepage scenario, feature data of the user, feature data of the content in the coarse-ranking stage, and feature data of the real-time context environment are extracted based on corresponding online features. The content in the coarse-ranking stage consists of candidate content for recommendation display in the homepage scenario. Feature data is filtered based on the same first user feature, first content feature, and first context feature present in at least two content recommendation scenarios to obtain feature data input to the cross-domain probabilistic model. Then, the cross-domain probabilistic model is used to accurately predict the probability that the content in the coarse-ranking stage will be selected by the user in the homepage scenario. The cross-domain probabilistic model can accurately learn user preferences to accurately predict the probability that content will be selected by the user in the homepage scenario. For a detailed discussion of the first user feature, first content feature, and first context feature, please refer to the corresponding discussion in the aforementioned model training method embodiment.

[0181] In one embodiment, after obtaining the scene selection rate based on the product of the first selection rate and the modified selection rate in step S804, the content recommendation method further includes the following steps S901 to S902.

[0182] S901, Obtain at least one third selection rate for at least one third content. The at least one third content is different from the first content; there is a one-to-one correspondence between the at least one third content and the at least one third selection rate, where the third selection rate refers to the probability that the corresponding third content is selected by the first user in the target recommendation scenario.

[0183] For example, referring to the method for obtaining the scene selection rate in the aforementioned content recommendation method embodiments, the target user data of the first user and the feature data of the first content feature of a third content are used as inputs to the prediction model and the correction model to obtain a third selection rate. The third selection rate refers to the probability that the corresponding third content is selected by the first user in the target recommendation scene. Based on this, in addition to obtaining the scene selection rate of the first content, at least one third selection rate of at least one third content can also be obtained.

[0184] S902, based on the scene selection rate and at least one third selection rate, determine at least one recommended content in the target recommendation scene from the first content and at least one third content.

[0185] At least one piece of recommended content refers to content displayed within the target recommendation scenario in response to a first user's access request for that scenario, for the first user to browse. When the first user enters the recommended content section, they can select and interact with the recommended content.

[0186] For example, a second selection rate threshold is set. The scene selection rate and at least one third selection rate are compared with the second selection rate threshold to obtain at least one fifth selection rate greater than the second selection rate threshold; the fifth selection rate among the at least one fifth selection rate is either the scene selection rate or the third selection rate. At least one fifth content corresponding to the at least one fifth selection rate is designated as at least one recommended content; the fifth content among the at least one fifth content is either the first content or the third content.

[0187] For example, a third quantity threshold is set. The scene selection rate and at least one third selection rate are sorted in descending order. The fifth selection rate of the top-ranked third quantity thresholds is determined; the fifth selection rate is either the scene selection rate or the third selection rate. The fifth content corresponding to each of the fifth selection rates of the third quantity thresholds is determined as the recommended content; the fifth content is either the first content or the third content.

[0188] The method in this embodiment, in response to a first user's access request to a target recommendation scenario, predicts the probability that the first user will select a first piece of content and at least one third piece of content in the target recommendation scenario. Then, based on the prediction results, at least one piece of recommended content that satisfies the first user's preferences is determined from the first content and at least one third piece of content. This content recommendation method facilitates personalized recommendations in the target recommendation scenario, thereby improving the user's platform experience and increasing user retention and platform stickiness.

[0189] In one embodiment, step S902 above, which determines at least one recommended content in the target recommended scenario from the first content and at least one third content based on the scene selection rate and at least one third selection rate, includes the following steps:

[0190] Based on the magnitude of the scene selection rate and at least one third selection rate, determine at least two recommended contents in the target recommendation scene from the first content and at least one third content, and determine the recommendation order of the at least two recommended contents.

[0191] This embodiment not only determines at least one recommended content from first content and at least one third content based on scene selection rate and at least one third selection rate, but also determines the recommendation order of at least one recommended content. Specifically, referring to the method in the previous embodiment, a second selection rate threshold or a third quantity threshold is set. Then, based on scene selection rate and at least one third selection rate, at least two recommended content are determined from first content and at least one third content. Furthermore, the order of the magnitude of the at least two fifth selection rates corresponding to the at least two recommended content is used as the recommendation order of the at least two recommended content. Here, the fifth selection rate is either scene selection rate or third selection rate. The higher the probability of being selected by the first user in the target recommendation scene, the more the corresponding first content or third content satisfies the user's preferences. Prioritizing the recommendation of first content or third content that better satisfies preferences, or displaying it in a display position that attracts the first user's attention, can greatly improve the user's platform experience, as well as the user's retention rate and platform stickiness.

[0192] The historical behavioral events of new users in a single content recommendation scenario are sparse, meaning historical data is sparse. This sparse historical data limits the profile information it can provide. Traditional probabilistic prediction methods, based on this sparse historical data, predict the probability of a new user's behavior in a single content recommendation scenario, resulting in low accuracy. Consequently, it's difficult to provide personalized content recommendations to new users. This negatively impacts the new user's platform experience, and also affects user retention and platform stickiness. Taking the homepage scenario as an example, the homepage is the first content recommendation scenario a new user encounters after logging into the platform. The effectiveness of content recommendations on the homepage significantly impacts new user retention. This application provides a content recommendation method capable of predicting the probability of content being selected by a user in a target recommendation scenario. For example, a content recommendation method based on a prediction model and a correction model, where the prediction model and correction model are trained using training samples from at least two content recommendation scenarios within the content recommendation platform. Because the training samples provide richer behavioral information about new users, and the training samples in at least two content recommendation scenarios have complementary value for new user preferences, they can provide richer and more comprehensive new user profile information, preference information, and potential interest information. Both the prediction model and the correction model can learn new user profiles and user preferences more accurately and comprehensively, as well as the interactive correlations of content in multiple content recommendation scenarios. The prediction model can accurately predict the probability that content will be selected by new users. The correction model can accurately predict the probability that content will be selected by new users in the target recommendation scenario, given that the content has been selected by new users. Compared with traditional probability prediction methods, the content recommendation method in this embodiment combines the prediction results of the prediction model and the correction model, which can more accurately predict the probability of new user behavior in target recommendation scenarios such as the homepage scenario. Based on more accurate prediction results, content that meets the preferences of new users can be accurately recommended in target recommendation scenarios such as the homepage scenario, thereby improving the platform experience for new users. When browsing recommended content, new users are interested in clicking to view content and interacting with content. This, in turn, helps to improve the platform's retention rate and platform stickiness for new users.

[0193] This application also provides a method for obtaining training samples. The training samples are used to train a model that predicts the probability of content being selected by a user in a content recommendation platform. The content recommendation platform includes at least two content recommendation scenarios. For example... Figure 9 As shown, the training sample acquisition method includes the following steps S1001 to S1003.

[0194] S1001, Obtain historical data from at least two content recommendation scenarios. The historical data includes user characteristic data, content characteristic data, a first tag indicating whether the user selected the content, and a second tag indicating whether the user interacted with the content.

[0195] S1002, Obtain the first user feature and the first content feature between at least two scene feature sets. The at least two scene feature sets correspond one-to-one with at least two content recommendation scenarios. The scene feature sets in the at least two scene feature sets include the user feature types and content feature types present in the corresponding content recommendation scenarios. The first user feature refers to the same user feature type between the at least two scene feature sets; the first content feature refers to the same content feature type between the at least two scene feature sets.

[0196] S1003, perform feature data filtering on historical data according to the first user feature and the first content feature to obtain training samples. The training samples contain at least two first user data points for users, at least two first content data points for content, a first label, and a second label; the first user data points are feature data belonging to the first user feature in the user feature data; the first content data are feature data belonging to the first content feature in the content feature data.

[0197] In one embodiment, the historical data further includes contextual feature data of the user browsing content corresponding to the user feature data. A first user feature, a first content feature, and a first contextual feature are obtained from at least two scene feature sets. The historical data is then filtered according to the first user feature, the first content feature, and the first contextual feature to obtain training samples. The training samples also include first contextual data, which is feature data belonging to the first contextual feature within the contextual feature data.

[0198] For a detailed discussion of the technical features, implementation process, and technical effects of the training sample acquisition method in this application embodiment, please refer to the detailed discussion of step S101 in the aforementioned model training method embodiment regarding the acquisition of training samples. This will not be repeated in this application embodiment. It should be emphasized that this training sample acquisition method acquires and unifies the feature dimensions and feature semantics of the feature data contained in each historical data set, resulting in training samples with identical feature dimensions and feature semantics. Unifying feature dimensions and feature semantics yields effective, information-rich, and high-quality training samples, which is beneficial for effective model training and improving model training performance, leading to models that can accurately predict the probability of content being selected by users. Training the model using the above-mentioned training samples from at least two content recommendation scenarios enhances the model's understanding of data from different content recommendation scenarios, improves feature utilization efficiency, and helps the model better capture cross-scenario behavioral patterns during training, thereby improving prediction accuracy and consistency.

[0199] like Figure 10 As shown in the illustration, this application also provides a model training device 1000. The model training device 1000 includes a sample acquisition module 1001, a sample input module 1002, a probability calculation module 1003, and a parameter adjustment module 1004. The sample acquisition module 1001 is used to acquire training samples; the training samples include at least two first user data points, at least two first content data points, a first tag indicating whether the user has selected content, and a second tag indicating whether the user has interacted with content. The sample input module 1002 is used to input training samples into the first module and the second module of the first model respectively, to obtain the target selection rate determined by the first module and the target conversion rate determined by the second module. The target selection rate is obtained based on the first feature interaction and the second feature interaction of the training samples, representing the probability that the content is selected by the user. The target conversion rate refers to the interaction probability of the content when it is selected by the user. The first feature interaction refers to the influence of the combined data on the prediction result of whether the content is selected by the user when any combination of first user data and first content data is less than a first threshold number. The second feature interaction refers to the influence of the combined data on the prediction result of whether the content is selected by the user when any combination of first user data and first content data is greater than or equal to the first threshold number. The probability calculation module 1003 is used to obtain the target interaction rate based on the target selection rate and the target conversion rate. The target interaction rate refers to the probability that the user interacts with the content. The parameter adjustment module 1004 is used to adjust the parameters of the first model using the target selection rate, the first label, the target interaction rate, and the second label to obtain the prediction model.

[0200] In one embodiment, the training samples are training samples from at least two content recommendation scenarios of the content recommendation platform, and the prediction model is used to make predictions on the content recommendation platform.

[0201] In one embodiment, the parameter adjustment module 1004 is further configured to determine a first loss of the first module by utilizing a first difference between the target selection rate and the first label, wherein the first loss is positively correlated with the first difference; determine a second loss of the first module and the second module by utilizing a second difference between the target interaction rate and the second label, wherein the second loss is positively correlated with the second difference; and adjust the parameters of the first model based on the first loss and the second loss to obtain a prediction model.

[0202] For explanations of the terms, please refer to the relevant descriptions in the aforementioned model training method embodiments, which will not be elaborated here.

[0203] It should be noted that the specific execution process of the aforementioned model training device 1000 can be found in [reference needed]. Figures 1 to 3 The specific details of the embodiments shown are not elaborated here.

[0204] like Figure 11 As shown in the illustration, this application embodiment also provides a training sample acquisition device 1100. The training samples are used to train a model that predicts the probability of content being selected by a user in a content recommendation platform, which includes at least two content recommendation scenarios. The training sample acquisition device 1100 includes a history acquisition module 1101, a feature acquisition module 1102, and a sample determination module 1103. The history acquisition module 1101 is used to acquire historical data from at least two content recommendation scenarios; the historical data includes user feature data, content feature data, a first tag indicating whether a user has selected content, and a second tag indicating whether a user has interacted with content. The feature acquisition module 1102 is used to acquire a first user feature and a first content feature between at least two scenario feature sets; the at least two scenario feature sets correspond one-to-one with at least two content recommendation scenarios, and the scenario feature sets in the at least two scenario feature sets include user feature types and content feature types existing in the corresponding content recommendation scenarios; the first user feature refers to the same user feature type between at least two scenario feature sets; the first content feature refers to the same content feature type between at least two scenario feature sets. The sample determination module 1103 is used to filter historical data according to the first user feature and the first content feature to obtain training samples; the training samples include at least two first user data, at least two first content data, a first label and a second label; the first user data is the feature data belonging to the first user feature in the user feature data; the first content data is the feature data belonging to the first content feature in the content feature data.

[0205] For explanations of the terms, please refer to the relevant descriptions in the aforementioned training sample acquisition method embodiments, which will not be elaborated here.

[0206] It should be noted that the specific execution process of the aforementioned training sample acquisition device 1100 can be found in [reference needed]. Figure 9 The specific details of the embodiments shown are not elaborated here.

[0207] like Figure 12 As shown in the illustration, this application also provides a content recommendation device 1200. The content recommendation device 1200 includes a data acquisition module 1201 and a data input module 1202. The data acquisition module 1201 is used to acquire target feature data; the target feature data includes target user data of a first user and target content data of the first content. The data input module 1202 is used to input the target feature data into a prediction model to obtain a first selection rate; the first selection rate refers to the probability that the first content is selected by the first user. The prediction model is trained using the model training method described in any of the foregoing model training method embodiments.

[0208] In one embodiment, the content recommendation device 1200 further includes a first content recommendation module, configured to obtain at least one second selection rate of at least one second content; the at least one second content is different from the first content; the at least one second content corresponds one-to-one with the at least one second selection rate, the second selection rate being the probability that the corresponding second content is selected by the first user; and based on the first selection rate and the at least one second selection rate, determine at least one recommended content from the first content and the at least one second content.

[0209] In one embodiment, the first content recommendation module is further configured to determine at least two recommended contents from the first content and at least one second content based on the magnitude of the first selection rate and at least one second selection rate, and to determine the recommendation order of the at least two recommended contents.

[0210] In one embodiment, the data input module 1202 is further configured to input target feature data into a prediction model to obtain a first selection rate and a first interaction rate; the first interaction rate refers to the probability that a first user interacts with the first content.

[0211] In one embodiment, the content recommendation device 1200 further includes a second content recommendation module, configured to acquire at least one second selection rate and at least one second interaction rate of at least one second content; the at least one second content is different from the first content; the at least one second content, at least one second selection rate, and at least one second interaction rate correspond one-to-one, the second selection rate refers to the probability that the corresponding second content is selected by the first user, and the second interaction rate refers to the probability that the first user interacts with the corresponding second content; based on the first probability and at least one second probability, at least one recommended content is determined from the first content and at least one second content; the first probability includes the first selection rate and the first interaction rate of the first content; the second probability in the at least one second probability includes the second selection rate and the second interaction rate corresponding to a second content, and the second content is one of the at least one second content.

[0212] In one embodiment, the first user refers to the user requesting entry into the target recommendation scenario, which is one of at least two content recommendation scenarios. The prediction model is trained using training samples from at least two content recommendation scenarios. The data input module 1202 is further configured to determine a corrected selection rate based on target feature data; the corrected selection rate refers to the probability that the first content will be selected by the first user in the target recommendation scenario, assuming the first content is selected by the first user. The content recommendation device 1200 further includes a probability determination module, configured to obtain a scenario selection rate based on the product of the first selection rate and the corrected selection rate; the scenario selection rate refers to the probability that the first content will be selected by the first user in the target recommendation scenario.

[0213] In one embodiment, the data input module 1202 is further configured to input target feature data into the correction model to obtain a corrected selection rate; wherein the correction model is trained using training samples of the target in at least two content recommendation scenarios, the training samples of the target refer to training samples of content selected by the user, the training samples of the target include at least two first user data of the user, at least two first content data of the content, a first tag for indicating whether the user selects the content, and a scenario tag, the scenario tag being used to indicate whether the training sample of the target is a training sample in the target recommendation scenario.

[0214] In one embodiment, the correction model is obtained by adjusting the parameters of the second model using the reference selection rate and the first label and scene label; the reference selection rate is obtained by inputting the training samples of the target into the second model; the reference selection rate refers to the probability that the content is selected by the user in the target recommendation scenario, provided that the content corresponding to at least two first content data is selected by the user corresponding to at least two first user data.

[0215] In one embodiment, the content recommendation device 1200 further includes a third content recommendation module, used to obtain at least one third selection rate of at least one third content; at least one third content is different from the first content; at least one third content corresponds one-to-one with at least one third selection rate, the third selection rate being the probability that the corresponding third content is selected by the first user in the target recommendation scenario; based on the scenario selection rate and at least one third selection rate, at least one recommended content in the target recommendation scenario is determined from the first content and at least one third content.

[0216] In one embodiment, the aforementioned third content recommendation module is further configured to determine at least two recommended contents in the target recommendation scenario from the first content and at least one third content based on the magnitude of the scene selection rate and at least one third selection rate, and to determine the recommendation order of the at least two recommended contents.

[0217] For explanations of the terms, please refer to the relevant descriptions in the foregoing recommended method embodiments; they will not be elaborated upon here.

[0218] It should be noted that the specific execution process of the above-mentioned recommended device 1200 can be found in [reference needed]. Figures 4 to 8 The specific details of the embodiments shown are not elaborated here.

[0219] like Figure 13 As shown in the illustration, this application also provides a computer device 1300. Exemplarily, the computer device 1300 may include a processor 1301, a communication interface 1302, a communication bus 1303, and a memory 1304. Specifically, the computer device 1300 may include:

[0220] The system includes at least one processor 1301, such as a CPU, at least one communication interface 1302, a memory 1304, and at least one communication bus 1303. The communication bus 1303 is used to enable communication between these components. The communication interface 1302 may optionally include a standard wired interface, a wireless interface (such as a Wi-Fi interface or a Bluetooth interface), etc. The memory 1304 may be high-speed RAM or non-volatile memory, such as at least one disk storage device. Optionally, the memory 1304 may also be at least one storage device located remotely from the aforementioned processor 1301. Figure 13 As shown, the memory 1304, which serves as a computer storage medium, may include an operating system and program instructions.

[0221] For example, processor 1301 can be used to implement the above. Figure 10The steps or methods executed by the sample acquisition module 1001, sample input module 1002, probability calculation module 1003, and parameter adjustment module 1004 in the sample acquisition module 1001, sample input module 1002, probability calculation module 1003, and parameter adjustment module 1004.

[0222] Understandably, the above method is merely an example, and the above can also be executed by the processor 1301 and other modules in the aforementioned computer device 1300. Figure 10 The steps or methods executed by the sample acquisition module 1001, sample input module 1002, probability calculation module 1003 and parameter adjustment module 1004 in this paper are not limited here.

[0223] exist Figure 13 In the computer device 1300 shown, the processor 1301 can be used to load computer program instructions stored in the memory 1304 and specifically perform the following operations:

[0224] Obtain training samples; the training samples contain at least two first user data points for the user, at least two first content data points for the content, a first tag indicating whether the user selected the content, and a second tag indicating whether the user interacted with the content;

[0225] Training samples are input into the first module and the second module of the first model, respectively, to obtain the target selection rate determined by the first module and the target conversion rate determined by the second module. The target selection rate is obtained based on the first feature interaction and the second feature interaction of the training samples, representing the probability that the content is selected by the user. The target conversion rate refers to the interaction probability when the content is selected by the user. The first feature interaction refers to the influence of the combined data on the prediction result of whether the content is selected by the user when any combination of first user data and first content data is less than a first threshold number. The second feature interaction refers to the influence of the combined data on the prediction result of whether the content is selected by the user when any combination of first user data and first content data is greater than or equal to the first threshold number.

[0226] The target engagement rate is derived from the target selection rate and target conversion rate; the target engagement rate refers to the probability that a user interacts with the content.

[0227] By adjusting the parameters of the first model using the target selection rate, the first label, the target interaction rate, and the second label, a prediction model is obtained.

[0228] For explanations of the terms, please refer to the relevant descriptions in the aforementioned model training method embodiments, which will not be elaborated here.

[0229] It should be noted that the specific execution process can be found in [link to relevant documentation]. Figures 1 to 3 The specific details of the embodiments shown are not elaborated here.

[0230] For example, processor 1301 can be used to implement the above. Figure 11 The steps or methods executed by the history acquisition module 1101, feature acquisition module 1102, and sample determination module 1103 in the process.

[0231] Understandably, the above method is merely an example, and the above can also be executed by the processor 1301 and other modules in the aforementioned computer device 1300. Figure 11 The steps or methods executed by the history acquisition module 1101, feature acquisition module 1102, and sample determination module 1103 in this paper are not limited here.

[0232] exist Figure 13 In the computer device 1300 shown, the processor 1301 can be used to load computer program instructions stored in the memory 1304 and specifically perform the following operations:

[0233] Acquire historical data from at least two content recommendation scenarios; the historical data includes user characteristic data, content characteristic data, a first tag indicating whether the user selected the content, and a second tag indicating whether the user interacted with the content;

[0234] Obtain the first user feature and the first content feature between at least two scene feature sets; the at least two scene feature sets correspond one-to-one with at least two content recommendation scenarios, and the scene feature sets in the at least two scene feature sets include the user feature types and content feature types existing in the corresponding content recommendation scenarios; the first user feature refers to the same user feature type between at least two scene feature sets; the first content feature refers to the same content feature type between at least two scene feature sets;

[0235] Historical data is filtered for features based on the first user feature and the first content feature to obtain training samples. The training samples contain at least two first user data points, at least two first content data points, a first tag, and a second tag. The first user data points are the feature data belonging to the first user feature in the user feature data. The first content data are the feature data belonging to the first content feature in the content feature data.

[0236] For explanations of the terms, please refer to the relevant descriptions in the aforementioned training sample acquisition method embodiments, which will not be elaborated here.

[0237] It should be noted that the specific execution process can be found in [link to relevant documentation]. Figure 9 The specific details of the embodiments shown are not elaborated here.

[0238] For example, processor 1301 can be used to implement the above. Figure 12 The steps or methods executed by the data acquisition module 1201 and the data input module 1202 in the middle.

[0239] Understandably, the above method is merely an example, and the above can also be executed by the processor 1301 and other modules in the aforementioned computer device 1300. Figure 12 The steps or methods executed by the data acquisition module 1201 and the data input module 1202 in this document are not limited here.

[0240] exist Figure 13 In the computer device 1300 shown, the processor 1301 can be used to load computer program instructions stored in the memory 1304 and specifically perform the following operations:

[0241] Acquire target feature data; the target feature data includes the target user data of the first user and the target content data of the first content.

[0242] The target feature data is input into the prediction model to obtain the first selection rate; the first selection rate refers to the probability that the first content is selected by the first user.

[0243] The prediction model is trained using the model training method described in any of the aforementioned model training method embodiments. Explanations of the terms can be found in the relevant descriptions in the aforementioned recommended method embodiments, and will not be elaborated upon here.

[0244] It should be noted that the specific execution process can be found in [link to relevant documentation]. Figures 4 to 8 The specific details of the embodiments shown are not elaborated here.

[0245] This application also provides a computer-readable storage medium that can store multiple computer program instructions. These computer program instructions are adapted to be loaded and executed by a processor as described above. Figures 1 to 3 The method steps of the model training method embodiment shown above, or the execution as described above. Figures 4 to 8 The method steps of the content recommendation method embodiment shown above, or the execution as described above. Figure 9 The illustrated example shows the method steps for obtaining training samples. For a detailed execution process, please refer to [link to example]. Figures 1 to 3 For a detailed description of the model training method implementation shown, or refer to... Figures 4 to 8 For a detailed description of the content recommendation method embodiments shown, or refer to... Figure 9 The specific details of the training sample acquisition method embodiment shown are not repeated here.

[0246] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as meaning "if...", or "after...", or "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if (the stated condition or event) is detected" can be interpreted as meaning "if determining...", or "in response to determining...", or "when (the stated condition or event) is detected", or "in response to detecting (the stated condition or event)".

[0247] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. This computer program product includes one or more computer program instructions. When these computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer program instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer program instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0248] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

[0249] The technical features of the above embodiments can be arbitrarily combined. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, they should be considered to be within the scope of this specification.

[0250] The embodiments described above are merely examples of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application.

Claims

1. A model training method, characterized in that, The method includes: Obtain training samples; the training samples include at least two first user data points of the user, at least two first content data points of the content, a first tag indicating whether the user has selected the content, and a second tag indicating whether the user has interacted with the content; The training samples are respectively input into the first module and the second module of the first model to obtain the target selection rate determined by the first module and the target conversion rate determined by the second module. The target selection rate is obtained based on the first feature interaction and the second feature interaction of the training samples, representing the probability that the content is selected by the user. The target conversion rate refers to the interaction probability when the content is selected by the user. The first feature interaction refers to the influence of the combined data on the prediction result of whether the content is selected by the user when any combination of first user data and first content data is less than a first threshold number. The second feature interaction refers to the influence of the combined data on the prediction result of whether the content is selected by the user when any combination of first user data and first content data is greater than or equal to the first threshold number. The target interaction rate is obtained based on the target selection rate and the target conversion rate; the target interaction rate refers to the probability that a user interacts with the content. By using the target selection rate and the first label, the target interaction rate and the second label, the parameters of the first model are adjusted to obtain the prediction model.

2. The method according to claim 1, characterized in that, The training samples are training samples from at least two content recommendation scenarios on the content recommendation platform, and the prediction model is used to make predictions on the content recommendation platform.

3. The method according to claim 1 or 2, characterized in that, The step of adjusting the parameters of the first model using the target selection rate and the first label, the target interaction rate and the second label, to obtain the prediction model includes: Using the first difference between the target selection rate and the first label, a first loss of the first module is determined, wherein the first loss is positively correlated with the first difference; Using the second difference between the target interaction rate and the second tag, a second loss for the first module and the second module is determined, wherein the second loss is positively correlated with the second difference; Based on the first loss and the second loss, the parameters of the first model are adjusted to obtain the prediction model.

4. A method for obtaining training samples, characterized in that, The training samples are used to train a model that predicts the probability of content being selected by a user in a content recommendation platform, wherein the content recommendation platform includes at least two content recommendation scenarios; the method includes: Obtain historical data from the at least two content recommendation scenarios; the historical data includes user feature data, content feature data, a first tag indicating whether the user selected the content, and a second tag indicating whether the user interacted with the content. Obtain a first user feature and a first content feature between at least two scene feature sets; the at least two scene feature sets correspond one-to-one with the at least two content recommendation scenarios, and the scene feature sets in the at least two scene feature sets include user feature types and content feature types existing in the corresponding content recommendation scenarios; the first user feature refers to the same user feature type between the at least two scene feature sets; the first content feature refers to the same content feature type between the at least two scene feature sets; The historical data is filtered according to the first user feature and the first content feature to obtain a training sample; the training sample includes at least two first user data, at least two first content data, the first tag and the second tag; the first user data is the feature data belonging to the first user feature in the user feature data; the first content data is the feature data belonging to the first content feature in the content feature data.

5. A content recommendation method, characterized in that, The method includes: Acquire target feature data; the target feature data includes target user data of the first user and target content data of the first content; The target feature data is input into the prediction model to obtain a first selection rate; the first selection rate refers to the probability that the first content is selected by the first user. The prediction model is obtained by training using the method described in any one of claims 1 to 3.

6. The method according to claim 5, characterized in that, After inputting the target feature data into the prediction model to obtain the first selection rate, the method further includes: Obtain at least one second selection rate for at least one second content; the at least one second content is different from the first content; the at least one second content corresponds one-to-one with the at least one second selection rate, and the second selection rate refers to the probability that the corresponding second content is selected by the first user; Based on the first selection rate and the at least one second selection rate, at least one recommended content is determined from the first content and the at least one second content.

7. The method according to claim 6, characterized in that, The step of determining at least one recommended content from the first content and the at least one second content based on the first selection rate and the at least one second selection rate includes: Based on the magnitude of the first selection rate and the at least one second selection rate, at least two recommended contents are determined from the first content and the at least one second content, and the recommendation order of the at least two recommended contents is determined.

8. The method according to claim 5, characterized in that, The step of inputting the target feature data into the prediction model to obtain the first selection rate includes: The target feature data is input into the prediction model to obtain the first selection rate and the first interaction rate; the first interaction rate refers to the probability that the first user interacts with the first content.

9. The method according to claim 8, characterized in that, After inputting the target feature data into the prediction model to obtain the first selection rate and the first interaction rate, the method further includes: Obtain at least one second selection rate and at least one second interaction rate for at least one second content; the at least one second content is different from the first content; the at least one second content, the at least one second selection rate, and the at least one second interaction rate correspond one-to-one, the second selection rate refers to the probability that the corresponding second content is selected by the first user, and the second interaction rate refers to the probability that the first user interacts with the corresponding second content; Based on a first probability and at least one second probability, at least one recommended content is determined from the first content and the at least one second content; the first probability includes the first selection rate and the first interaction rate of the first content; the second probability in the at least one second probability includes a second selection rate and a second interaction rate corresponding to a second content, wherein the second content is one of the at least one second content.

10. The method according to claim 5, characterized in that, The first user refers to the user who requests to enter the target recommendation scenario, which is one of at least two content recommendation scenarios, and the prediction model is trained using training samples from the at least two content recommendation scenarios; the method further includes: Based on the target feature data, a corrected selection rate is determined; the corrected selection rate refers to the probability that the first content will be selected by the first user in the target recommendation scenario, provided that the first content has been selected by the first user. The scene selection rate is obtained by multiplying the first selection rate and the corrected selection rate; the scene selection rate refers to the probability that the first content is selected by the first user in the target recommendation scene.

11. The method according to claim 10, characterized in that, Determining the corrected selection rate based on the target feature data includes: The target feature data is input into the correction model to obtain the correction selection rate; The correction model is trained using training samples of the target in the at least two content recommendation scenarios. The training samples of the target refer to training samples of content selected by the user. The training samples of the target include at least two first user data, at least two first content data, a first tag indicating whether the user has selected the content, and a scenario tag. The scenario tag is used to indicate whether the training sample of the target is a training sample in the target recommendation scenario.

12. The method according to claim 11, characterized in that, The correction model is obtained by adjusting the parameters of the second model using a reference selection rate, the first label, and the scene label; the reference selection rate is obtained by inputting the training samples of the target into the second model. The reference selection rate refers to the probability that a content is selected by a user in the target recommendation scenario, provided that the content corresponding to at least two first content data is selected by the user corresponding to at least two first user data.

13. The method according to claim 10, characterized in that, After obtaining the scene selection rate based on the product of the first selection rate and the modified selection rate, the method further includes: Obtain at least one third selection rate for at least one third content; the at least one third content is different from the first content; the at least one third content corresponds one-to-one with the at least one third selection rate, and the third selection rate refers to the probability that the corresponding third content is selected by the first user in the target recommendation scenario; Based on the scene selection rate and the at least one third selection rate, at least one recommended content in the target recommended scene is determined from the first content and the at least one third content.

14. The method according to claim 13, characterized in that, The step of determining at least one recommended content in the target recommended scenario from the first content and the at least one third content based on the scenario selection rate and the at least one third selection rate includes: Based on the magnitude of the scene selection rate and the at least one third selection rate, at least two recommended contents in the target recommended scene are determined from the first content and the at least one third content, and the recommendation order of the at least two recommended contents is determined.

15. A model training device, characterized in that, The device includes: The sample acquisition module is used to acquire training samples; the training samples include at least two first user data of the user, at least two first content data of the content, a first tag indicating whether the user has selected the content, and a second tag indicating whether the user has interacted with the content. The sample input module is used to input the training samples into the first module and the second module of the first model respectively, to obtain the target selection rate determined by the first module and the target conversion rate determined by the second module. The target selection rate is obtained based on the first feature interaction and the second feature interaction of the training samples, representing the probability that the content is selected by the user. The target conversion rate refers to the interaction probability when the content is selected by the user. The first feature interaction refers to the influence of the combined data on the prediction result of whether the content is selected by the user when any combination of first user data and first content data is less than a first threshold number. The second feature interaction refers to the influence of the combined data on the prediction result of whether the content is selected by the user when any combination of first user data and first content data is greater than or equal to the first threshold number. The probability calculation module is used to obtain the target interaction rate based on the target selection rate and the target conversion rate; The parameter adjustment module is used to adjust the parameters of the first model using the target selection rate and the first label, the target interaction rate and the second label, to obtain the prediction model.

16. A training sample acquisition device, characterized in that, The training samples are used to train a model that predicts the probability of content being selected by a user in a content recommendation platform, wherein the content recommendation platform includes at least two content recommendation scenarios; the device includes: The history acquisition module is used to acquire historical data in the at least two content recommendation scenarios; the historical data includes user feature data, content feature data, a first tag indicating whether the user has selected the content, and a second tag indicating whether the user has interacted with the content. The feature acquisition module is used to acquire a first user feature and a first content feature between at least two scene feature sets; the at least two scene feature sets correspond one-to-one with the at least two content recommendation scenarios, and the scene feature sets in the at least two scene feature sets include user feature types and content feature types existing in the corresponding content recommendation scenarios; the first user feature refers to the same user feature type between the at least two scene feature sets; the first content feature refers to the same content feature type between the at least two scene feature sets; The sample determination module is used to filter the historical data according to the first user feature and the first content feature to obtain training samples; the training samples include at least two first user data, at least two first content data, the first tag and the second tag; the first user data is the feature data belonging to the first user feature in the user feature data; the first content data is the feature data belonging to the first content feature in the content feature data.

17. A content recommendation device, characterized in that, The device includes: The data acquisition module is used to acquire target feature data; the target feature data includes target user data of the first user and target content data of the first content. The data input module is used to input the target feature data into the prediction model to obtain a first selection rate; the first selection rate refers to the probability that the first content is selected by the first user. The prediction model is obtained by training using the method described in any one of claims 1 to 3.

18. A computer device, characterized in that, include: A memory and a processor, wherein the memory stores computer program instructions; when executed by the processor, the computer program instructions cause the processor to perform the method as described in any one of claims 1 to 3, or the method as described in claim 4, or the method as described in any one of claims 5 to 14.

19. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions; when the computer program instructions are executed on one or more processors, they perform the method as described in any one of claims 1 to 3, or the method as described in claim 4, or the method as described in any one of claims 5 to 14.

20. A computer program product, characterized in that, The computer program product includes computer program instructions; when the computer program instructions are executed on a computer, the computer causes the computer to perform the method as described in any one of claims 1 to 3, or the method as described in claim 4, or the method as described in any one of claims 5 to 14.

Images