Model training method, advertisement putting method and related equipment

Abstract

The invention provides a model training method. The method comprises the following steps: acquiring a first training user feature, a training scene feature, an advertisement putting service training data feature, an advertisement putting service target training label, a second training user feature, an advertisement non-putting service training data feature and an advertisement non-putting service target training label; training scene vectors are obtained through the training scene features; obtaining a first training service natural prediction vector through the first training user features and the advertisement putting service training data features, and obtaining a second training service natural prediction vector through the second training user features and the advertisement non-putting service training data features; obtaining a training service delivery gain prediction value through the training scene vector and the first training service natural prediction vector; obtaining a training service natural prediction value through the second training service natural prediction vector; and adjusting model parameters of the to-be-trained model through the training service natural predicted value and the training service delivery gain predicted value to obtain an advertisement gain model.

Description

Technical Field

[0001] This application relates to the field of advertising technology, and in particular to model training methods, advertising delivery methods and related equipment. Background Technology

[0002] In related technologies, advertisers can use RTA (Real-time Application Programming Interface) systems to place ads in application software, websites and other advertising scenarios, thereby improving certain business objectives of advertisers.

[0003] Advertisers typically have multiple ad placement scenarios to choose from, and they also usually have multiple business objectives. To evaluate the value gained by each ad placement scenario for each business objective, advertisers need to train an ad placement model and an ad-free model for each combination of "ad placement scenario + business objective." Using these models, advertisers can determine the estimated value of ad placement and the estimated value of ad-free placement for each business objective, thus obtaining the value gained by each ad placement scenario. Advertisers can select ad placement scenarios with high value from among multiple scenarios; however, this method suffers from the problem of utilizing a large number of models. Summary of the Invention

[0004] The purpose of this application is to provide a model training method, an advertising delivery method, and related equipment, so as to at least solve the problem that, in order to evaluate the value gained by each advertising delivery scenario for each business objective, advertisers need to train an advertising delivery model and an advertising non-delivery model for each combination of "advertising delivery scenario + business objective," resulting in a large number of models; the specific technical solution is as follows: In a first aspect of this application, a model training method is provided, comprising: Multiple batches of training samples are obtained. Each batch of training samples includes: training samples with ads and training samples without ads. The training samples with ads include first training user features, training scenario features, advertising business training data features, and advertising business target training labels. The training samples without ads include second training user features, advertising non-ad training data features, and advertising non-ad training target training labels. The model to be trained consists of a scenario processing sub-model, a business target prediction sub-model, a first causal inference sub-model, and a second causal inference sub-model. The training scene features are input into the scene processing sub-model, which outputs a training scene vector. Input the first training user features and advertising delivery business training data features into the business target prediction sub-model to obtain the first training business natural prediction vector; Input the second training user features and the advertising non-delivery business training data features into the business target prediction sub-model to obtain the second training business natural prediction vector; The training scenario vector and the first training service natural prediction vector are input into the first causal inference sub-model to obtain the training service delivery gain prediction value. Input the second training business natural prediction vector into the second causal inference sub-model to obtain the training business natural prediction value; A first loss function value is obtained by using the natural prediction value of the training business and the target training label of the advertising-not-delivered business; a second loss function value is obtained by using the predicted value of the training business delivery gain and the target training label of the advertising delivery business. The model parameters of different sub-models constituting the model to be trained are adjusted by the first loss function value and the second loss function value until the model to be trained after adjusting the model parameters meets the preset convergence condition, thus obtaining the advertising gain model.

[0005] In a second aspect of this application, an advertising delivery method is also provided, applied to a business server of an advertising gain model, comprising: Obtain the scene characteristics of at least one preset advertising scenario, and the user characteristics of the users in the advertising scenario; The scene features are input into the preset scene feature processing sub-model in the advertising gain model obtained by any of the model training methods described in the first aspect, to obtain the scene vector of the advertising delivery scene; The user features are input into the preset business prediction sub-model of the advertising gain model obtained by any of the model training methods described in the first aspect, to obtain the natural business prediction vector of the user's contribution to the preset business objective; the business objective is the preset business objective corresponding to the advertisement to be delivered. The business natural prediction vector and the scenario vector are input into the pre-set gain prediction sub-model in the advertising gain model obtained by any of the model training methods described in the first aspect to obtain the business delivery gain prediction value of the user's contribution to the business objective; the business delivery gain prediction value is the business delivery gain prediction value of the user to the business objective when the advertisement to be delivered is delivered in the advertising delivery scenario. Based on the business delivery gain prediction value, a target advertising delivery scenario is determined in at least one of the advertising delivery scenarios, and the advertisement to be delivered is delivered to the target advertising delivery scenario.

[0006] In a third aspect of this application, a model training apparatus is also provided, comprising: The first acquisition module is used to acquire multiple batches of training samples. Each batch of training samples includes: training samples for ads that have been placed and training samples for ads that have not been placed. The training samples for ads that have been placed include first training user features, training scenario features, advertising placement business training data features, and advertising placement business target training labels. The training samples for ads that have not been placed include second training user features, advertising non-placement business training data features, and advertising non-placement business target training labels. The model to be trained consists of a scenario processing sub-model, a business target prediction sub-model, a first causal inference sub-model, and a second causal inference sub-model. The first model processing module is used to input the training scene features into the scene processing sub-model and output the training scene vector. The second model processing module is used to input the first training user features and advertising placement business training data features into the business target prediction sub-model to obtain the first training business natural prediction vector. The third model processing module is used to input the second training user features and the advertising non-delivery business training data features into the business target prediction sub-model to obtain the second training business natural prediction vector. The fourth model processing module is used to input the training scene vector and the first training service natural prediction vector into the first causal inference sub-model to obtain the training service delivery gain prediction value. The fifth model processing module is used to input the second training business natural prediction vector into the second causal inference sub-model to obtain the training business natural prediction value; The function value calculation module is used to obtain a first loss function value through the natural prediction value of the training business and the target training label of the advertising non-delivery business, and to obtain a second loss function value through the predicted value of the training business delivery gain and the target training label of the advertising delivery business; The adjustment module is used to adjust the model parameters of different sub-models constituting the model to be trained by using the first loss function value and the second loss function value, until the model to be trained after adjusting the model parameters meets the preset convergence condition, thus obtaining the advertising gain model.

[0007] In a fourth aspect of this application, an advertising delivery device is also provided, applied to a business server of an advertising gain model, comprising: The feature acquisition module is used to acquire scene features of at least one preset advertising scenario, and user features of users in the advertising scenario. The scene vector acquisition module is used to input the scene features into the preset scene feature processing sub-model in the advertising gain model obtained by any of the model training methods described in the first aspect, so as to obtain the scene vector of the advertising delivery scene. The business natural prediction vector acquisition module is used to input the user features into a preset business prediction sub-model in the advertising gain model obtained by any of the model training methods described in the first aspect, and obtain the business natural prediction vector of the user's contribution to the preset business objective; the business objective is the preset business objective corresponding to the advertisement to be delivered. The business delivery gain prediction module is used to input the natural prediction vector of the business and the scenario vector into the pre-set gain prediction sub-model in the advertising gain model obtained by any of the model training methods described in the first aspect, and obtain the business delivery gain prediction value of the user's contribution to the business objective; the business delivery gain prediction value is the business delivery gain prediction value of the user to the business objective when the advertisement to be delivered is delivered in the advertising delivery scenario. The advertising delivery module is used to determine a target advertising delivery scenario in at least one of the advertising delivery scenarios based on the business delivery gain prediction value, and to deliver the advertisement to be delivered to the target advertising delivery scenario.

[0008] In another aspect of this application, an electronic device is also provided, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; The processor is used to execute the program stored in the memory to implement the model training method described above, or any of the advertising delivery methods described above.

[0009] In another aspect of this application, a computer-readable storage medium is also provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform the model training method described above or any of the advertising delivery methods described above.

[0010] This application provides a model training method that acquires multiple batches of training samples. Each batch of training samples includes: advertising training samples and non-advertising training samples. The advertising training samples include first training user features, training scenario features, advertising placement business training data features, and advertising placement business target training labels. The non-advertising training samples include second training user features, advertising non-placement business training data features, and advertising non-placement business target training labels. The model to be trained consists of a scenario processing sub-model, a business target prediction sub-model, a first causal inference sub-model, and a second causal inference sub-model. The training scenario features are input into the scenario processing sub-model, which outputs a training scenario vector. The first training user features and advertising placement business training data features are input into the business target prediction sub-model to obtain a first training business natural prediction vector. The second training user features and advertising non-placement business training data features are input into the business target prediction sub-model to obtain a second training business natural prediction vector. The training scenario vector and... The first training business natural prediction vector is input into the first causal inference sub-model to obtain the training business delivery gain prediction value; the second training business natural prediction vector is input into the second causal inference sub-model to obtain the training business natural prediction value; the first loss function value is obtained through the training business natural prediction value and the training label of the advertising non-delivery business target; the second loss function value is obtained through the training business delivery gain prediction value and the advertising delivery business target training label; the model parameters of the different sub-models constituting the model to be trained are adjusted using the first and second loss function values ​​until the model to be trained after adjusting the model parameters meets the preset convergence condition, thus obtaining the advertising gain model. This realizes the integration of a model that can handle multiple training scenarios, a model that can predict multiple business targets corresponding to advertising delivery and non-delivery, a model that can predict the business gain brought by advertising delivery in the training scenario, and a model that can predict the business prediction value when no advertising is delivered in the training scenario into a single model, resulting in the advertising gain model—a general model. This general model only needs to make one prediction to obtain the business delivery gain prediction value, business natural prediction value, and business natural prediction value for advertising delivery and non-delivery under multiple scenarios and multiple business targets, reducing the number of models, improving model utilization efficiency, and improving model prediction performance. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0012] Figure 1 This is a flowchart of the steps of a model training method provided in the embodiments of this application; Figure 2 This is a flowchart illustrating the steps of an advertising delivery method provided in this application embodiment; Figure 3 This is a schematic diagram of an advertising gain model provided in an embodiment of this application; Figure 4 This is a structural block diagram of a model training device provided in the embodiments of this application; Figure 5 This is a structural block diagram of an advertising delivery device provided in the embodiments of this application; Figure 6 This is one of the block diagrams of an electronic device provided in the embodiments of this application; Figure 7 This is a second block diagram of an electronic device provided in the embodiments of this application; Figure 8 This is a schematic diagram of a computer-readable medium provided in an embodiment of this application. Detailed Implementation

[0013] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0014] To facilitate understanding of the technical solutions and effects of the embodiments of this application, the relevant technologies of this application will be briefly described below.

[0015] In related technologies, advertisers can use RTA (Real-time Application Programming Interface) systems to place ads in application software, websites and other advertising scenarios, thereby improving certain business objectives of advertisers.

[0016] Specifically, advertisers can use ad delivery models to estimate the expected value of their business objectives when advertising is performed in a particular ad delivery scenario; simultaneously, using an ad-not-delivered model, they can estimate the expected value of their business objectives when advertising is not performed in that same ad delivery scenario. Advertisers can then use these estimated values ​​to determine the added value of that ad delivery scenario to their business objectives. Based on this, advertisers can prioritize ad delivery scenarios with higher added value among multiple ad delivery scenarios.

[0017] However, advertisers typically have multiple ad placement scenarios to choose from, and they also usually have multiple business objectives. In order to evaluate the value gained by each ad placement scenario for each business objective, advertisers need to train an ad placement model and an ad-free model for each combination of "ad placement scenario + business objective," resulting in a large number of models.

[0018] Furthermore, when training the advertising delivery model and the non-delivery model based on the combination of "ad delivery scenario + business objective", only sample data related to this combination of "ad delivery scenario + business objective" can be used, and sample data from other advertising delivery scenarios or other business objectives cannot be utilized, resulting in insufficient use of sample data.

[0019] Reference Figure 1 The diagram illustrates a flowchart of a model training method provided in an embodiment of this application, which may specifically include the following steps: Step 101: Obtain multiple batches of training samples. Each batch of training samples includes: training samples for ads placed and training samples for ads not placed. The training samples for ads placed include first training user features, training scenario features, advertising placement business training data features, and advertising placement business target training labels. The training samples for ads not placed include second training user features, advertising non-placement business training data features, and advertising non-placement business target training labels. The model to be trained consists of a scenario processing sub-model, a business target prediction sub-model, a first causal inference sub-model, and a second causal inference sub-model.

[0020] Step 102: Input the training scene features into the scene processing sub-model and output the training scene vector.

[0021] Step 103: Input the first training user features and advertising placement business training data features into the business target prediction sub-model to obtain the first training business natural prediction vector.

[0022] Step 104: Input the second training user features and the advertising-unserved business training data features into the business target prediction sub-model to obtain the second training business natural prediction vector.

[0023] Step 105: Input the training scene vector and the first training business natural prediction vector into the first causal inference sub-model to obtain the training business delivery gain prediction value.

[0024] Step 106: Input the second training business natural prediction vector into the second causal inference sub-model to obtain the training business natural prediction value.

[0025] Step 107: Obtain the first loss function value by training the natural prediction value of the business and the training label of the business target of the advertising not being delivered, and obtain the second loss function value by training the prediction value of the business delivery gain and the training label of the business target of the advertising delivery.

[0026] Step 108: Adjust the model parameters of the different sub-models constituting the model to be trained by using the first loss function value and the second loss function value until the model to be trained after adjusting the model parameters meets the preset convergence condition, and obtain the advertising gain model.

[0027] In step 101, when training the model using multiple batches of training samples, each batch is trained once, and the model parameters are updated once. The training scenario features of this application include features from multiple training scenarios. The advertising placement business training data features include training data features for various advertising placement business objectives. These objectives can be different business objectives when advertising is placed in different training scenarios. For example, if the training scenario is software A and software B owned by the advertiser, the advertising placement business objectives could be: the number of times software A is launched, its usage duration, and its economic benefits after the advertisement is placed in software A; or the number of times software B is launched, its usage duration, and its economic benefits after the advertisement is placed in software B. The non-advertising business training data features include training data features for various non-advertising business objectives. These objectives can be different business objectives when the advertisement is not placed in different training scenarios. For example, when the advertisement is not placed in software A, the number of times software A is launched, its usage duration, and its economic benefits; or when the advertisement is not placed in software B, the number of times software B is launched, its usage duration, and its economic benefits.

[0028] In this application, the training labels for both advertising placement and non-ad placement business objectives are set according to the corresponding business objectives, so they can be set consistently or inconsistently. However, the training labels for advertising placement business objectives need to be set based on the characteristics of the advertising placement business training data, and these labels can be defined as standard business gain placement prediction values. Similarly, the training labels for non-ad placement business objectives need to be set based on the characteristics of the non-ad placement business training data, and these labels can be defined as standard business natural prediction values. These can include tags for launch counts, usage durations, and economic benefits in the training scenario under both advertising placement and non-ad placement conditions. The advertising placement training samples correspond to the training sample data under the advertising placement scenario, while the non-ad placement training samples correspond to the training sample data under the non-ad placement scenario.

[0029] The scene processing sub-model can be a multi-scene PPNET (Parameter Personalized Network), which can encode scene information into vectors through a scene encoder.

[0030] The business objective prediction sub-model can be a multi-gate mixed expert model (MMOE). Multi-gate mixed expert models can predict multiple related but different advertising objectives at the same time, such as the click-through rate and registration rate of advertiser software.

[0031] Both the first and second causal inference sub-models can be Explicit Feature Interaction-aware Uplift Networks (EFINs). EFINs can estimate the causal effect of advertisements. This application integrates multi-scenario PPNETs, ​​multi-objective MMOEs, and two EFINs into a single model to obtain the model to be trained. Therefore, in this embodiment, a training set can be used to train multiple sub-models of the model to be trained. Sample data from all advertising scenarios and all business objectives in this training set can be used to train the model, fully utilizing the sample data. Furthermore, using a single training set only requires one training iteration, improving training efficiency. The advertising-served training samples are used to train the scenario processing sub-model, the business objective prediction sub-model, and the first causal inference sub-model. The non-advertising-served training samples are used to train the business objective prediction sub-model and the second causal inference sub-model.

[0032] In step 102, the training scenario features under the advertising placement scenario may include: historical training scenario features of the advertising placement scenario. Historical training scenario features include: historical training scenario interaction basic features and historical training scenario interaction bidding features. Historical training scenario interaction basic features can refer to historical training scenario basic features, and historical training scenario interaction bidding features can refer to historical training scenario bidding features. Among these, the training scenario basic features may include: the name of the advertising placement scenario, the conversion rate of the advertising placement scenario to the advertiser's business objectives, etc. The training scenario bidding features may include the price information of the advertiser placing ads in the advertising placement scenario.

[0033] In step 103, the first training user features may include user information and user preferences for advertising scenarios. User information may include: user's age, gender, occupation, etc. The advertising business training data features include training data features for various advertising business objectives. Therefore, different first training business natural prediction vectors can be obtained based on different advertising business objectives. Therefore, in one possible embodiment, step 103 specifically includes the following sub-steps: Sub-step 1031: Input the first training user features and training data features including multiple advertising business objectives into the business objective prediction sub-model to obtain the first training business natural prediction vector corresponding to the training data features of different advertising business objectives.

[0034] It should be noted that the first training user feature in this application embodiment includes a first sparse training user feature and / or a first dense training user feature. Therefore, inputting the first training user feature and the advertising delivery business training data feature into the business target prediction sub-model to obtain the first training business natural prediction vector further includes: Extract the first dense training user feature from the first training user feature, and input the first dense training user feature and the advertising placement business training data feature into the business target prediction sub-model to obtain the first training business natural prediction vector. The method includes: Extract the first sparse training user feature from the first training user feature, and process the first sparse training user feature using the preset factorization machine in the advertising gain model to obtain the training embedding vector of the first sparse training user feature.

[0035] In the embodiments of this application, Figure 3 Feature 1, Feature 2, Feature 3, Feature 4, Feature 5, and Feature N can all be represented as the first training user features. The first training user features include the first sparse training user features and / or the first dense training user features. Figure 3 The dense features in the text correspond to the first dense training user features, and the sparse features correspond to the first sparse training user features. Sparse training user features refer to high-dimensional, discrete features with a large value space but sparse actual observations, while dense training user features refer to low-dimensional, continuous, and densely valued numerical features.

[0036] Reference Figure 3 It can extract the first dense training user features from the first training user features, and input the first dense training user features into the business target prediction sub-model to obtain the first training business natural prediction vector of the advertiser placing ads in the training scenario.

[0037] The advertising gain model also includes Factorization Machines (FM). In this embodiment, a first sparse training user feature can be extracted from the first training user feature, and the first sparse training user feature can be processed using FM to obtain the training embedding vector of the first sparse training user feature.

[0038] In step 104, the second training user features may also include user information, such as the user's age, gender, occupation, etc. It may also include business features related to the advertiser's business objectives when no ads are being placed in the advertising scenario. For example, if the advertising scenario is software A, and the advertiser's business objective is the number of times the advertiser's software D is launched, the user features may also include the number of times users of software A launch software D when no ads are placed in software A. The training data features for non-advertising business objectives include training data features for various non-advertising business objectives. Therefore, different second training business natural prediction vectors can be obtained based on different non-advertising business objectives. Therefore, in one possible embodiment, step 104 specifically includes the following sub-steps: Sub-step 1041: Input the second training user features and training data features including multiple non-advertising business objectives into the business objective prediction sub-model to obtain the second training business natural prediction vector corresponding to the training data features of different non-advertising business objectives.

[0039] The first and second training user features of this application are both types of user features, therefore, in the embodiments of this application, Figure 3 Features 1, 2, 3, 4, 5 and N in the dataset can also be represented as second training user features, which include second sparse training user features and / or second dense training user features.

[0040] In step 105, the first training business natural prediction vector and the training scenario vector are input into the first causal inference sub-model to determine the training business delivery gain prediction value of the training user's contribution to the business objective when advertising is delivered in the training scenario. The training business delivery gain prediction value is the training business delivery gain prediction value of the first training user to the business objective when advertising is delivered in the training scenario.

[0041] Furthermore, as can be seen from the above, in this embodiment, the first training user feature includes a first dense training user feature and a first sparse training user feature. The first sparse training user feature is extracted from the first training user feature, and FM is used to process the first sparse training user feature to obtain the embedding vector of the first sparse training user feature. When the advertising gain model predicts the business placement gain prediction value when placing ads in the advertising placement training scenario, the embedding vector of the first sparse training user feature is introduced to improve the prediction accuracy of the subsequent training business placement gain prediction value. Therefore, step 105 specifically includes the following sub-steps: Sub-step 1051: Input the first training service natural prediction vector, the training scenario vector, and the training embedding vector into the first causal inference sub-model to obtain the training service delivery gain prediction value of the first training user's contribution to the service objective.

[0042] In this embodiment of the application, the predicted value of the training business delivery gain when advertising is delivered in the training scenario is obtained. By introducing the training embedding vector of the first sparse training user features, the prediction accuracy of the predicted value of the training business delivery gain is improved.

[0043] Step 106: Input the second training business natural prediction vector into the second causal inference sub-model to obtain the training business natural prediction value of the training user for the business objective.

[0044] In some embodiments of this application, the predicted value of the training service delivery contribution of the first training user to the service objective can be determined by the natural predicted value of the training service and the predicted value of the training service delivery gain, including: Calculate the difference between the natural predicted value of the training service and the predicted value of the training service delivery gain; The difference is used to determine the predicted value of the training service deployment based on the contribution of the first training user to the business objective.

[0045] In this embodiment, the training scene vector, the first training service natural prediction vector, and the training embedding vector can be input into the first causal inference sub-model to obtain the training service delivery gain prediction value for the first training user towards the service target. Alternatively, the second training service natural prediction vector can be input into the second causal inference sub-model to obtain the training service natural prediction value for the second training user towards the service target.

[0046] Then, based on the difference between the predicted value of the training service delivery gain and the natural predicted value of the training service, the predicted value of the training service delivery is determined.

[0047] In step 107, the first loss function value can be calculated using the training labels for the advertising-not-delivered business objective and the natural predicted value of the training business. The second loss function value can be calculated using the predicted value of the delivery gain of the training business and the training labels for the advertising delivery business objective. The calculation can employ methods such as cross-entropy and mean squared error. Because the business objective training labels include the number of launches, usage duration, and economic benefit labels for the training scenario under both advertising delivery and non-delivery conditions, and the natural predicted value of the training business and the predicted value of the delivery gain of the training business are also calculated based on training data features that include multiple business objectives, the loss function value can be calculated for different business objectives. Therefore, step 107 specifically includes the following sub-steps: Sub-step 1071: Identify from the natural predicted values ​​of the number of launches, usage duration, and economic benefits of the training scenario when no advertisements are placed, the natural predicted values ​​of the training business.

[0048] Sub-step 1072: Using the naturally predicted number of launches without ad delivery and the launch number label, obtain the first loss function value for the number of launches without ad delivery.

[0049] Sub-step 1073: Using the naturally predicted usage duration without advertising and the usage duration label, obtain the first loss function value for usage duration without advertising.

[0050] Sub-step 1074: Using the natural predicted value of economic returns without advertising and the economic return label, obtain the first loss function value of economic returns without advertising.

[0051] Sub-step 1075: Identify from the training service delivery gain prediction values ​​the launch count delivery gain prediction value, usage duration delivery gain prediction value, and economic benefit delivery gain prediction value of the training scenario under advertising delivery.

[0052] Sub-step 1076: Using the predicted value of the launch gain under the ad campaign and the launch number label, obtain the second loss function value for the launch number under the ad campaign.

[0053] Sub-step 1077: Obtain the second loss function value for usage duration under advertising by using the predicted value of usage duration under advertising and the usage duration label.

[0054] Sub-step 1078: Using the predicted economic return gain under advertising and the economic return label, obtain the second loss function value for the economic return under advertising.

[0055] Step 108: The gradient can be calculated using the first loss function value. Based on this gradient, the parameters of the scene processing sub-model, the business objective prediction sub-model, and the first causal inference sub-model can be updated. The gradient can also be calculated using the second loss function value. Based on this gradient, the parameters of the business objective prediction sub-model and the second causal inference sub-model can be updated. After the model parameters are updated, they are respectively trained into a scene feature processing sub-model capable of handling multiple advertising scenarios, a business prediction sub-model capable of predicting multiple business objectives corresponding to the advertisements to be placed, a placement gain prediction sub-model capable of predicting the business gains brought by placing advertisements in advertising scenarios, and a natural gain prediction sub-model capable of predicting the business gains brought by not placing advertisements. These trained models are merged into an advertising gain model. Therefore, in one possible embodiment, step 108 specifically includes the following sub-steps: Sub-step 1081: Adjust the model parameters of the different sub-models constituting the model to be trained by using the first loss function value and the second loss function value.

[0056] Sub-step 1082: When the scene processing sub-model of the model to be trained after adjusting the model parameters meets the preset convergence condition, the scene feature processing sub-model is obtained.

[0057] Sub-step 1083: When the business target prediction sub-model of the model to be trained after adjusting the model parameters meets the preset convergence condition, the business prediction sub-model is obtained.

[0058] Sub-step 1084: When the first causal inference sub-model of the model to be trained after adjusting the model parameters satisfies the preset convergence condition, the delivery gain prediction sub-model is obtained.

[0059] Sub-step 1085: When the second causal inference sub-model of the model to be trained after adjusting the model parameters satisfies the preset convergence condition, the natural gain prediction sub-model is obtained.

[0060] Sub-step 1086: The scene feature processing sub-model, the business prediction sub-model, the delivery gain prediction sub-model, and the natural gain prediction sub-model are fused to obtain the advertising gain model.

[0061] Among them, the aforementioned preset convergence conditions can be set as follows: the loss changes of the first loss function value and the second loss function value are less than a certain threshold for multiple consecutive times; the calculated gradient mean / variance changes very little; the adjustment changes of the model parameters are less than a certain value; and the maximum number of iterations is reached.

[0062] This application provides a model training method that acquires multiple batches of training samples. Each batch of training samples includes: advertising training samples and non-advertising training samples. The advertising training samples include first training user features, training scenario features, advertising placement business training data features, and advertising placement business target training labels. The non-advertising training samples include second training user features, advertising non-placement business training data features, and advertising non-placement business target training labels. The model to be trained consists of a scenario processing sub-model, a business target prediction sub-model, a first causal inference sub-model, and a second causal inference sub-model. The training scenario features are input into the scenario processing sub-model, which outputs a training scenario vector. The first training user features and advertising placement business training data features are input into the business target prediction sub-model to obtain a first training business natural prediction vector. The second training user features and advertising non-placement business training data features are input into the business target prediction sub-model to obtain a second training business natural prediction vector. The training scenario vector and... The first training business natural prediction vector is input into the first causal inference sub-model to obtain the training business delivery gain prediction value; the second training business natural prediction vector is input into the second causal inference sub-model to obtain the training business natural prediction value; the first loss function value is obtained through the training business natural prediction value and the training label of the advertising non-delivery business target; the second loss function value is obtained through the training business delivery gain prediction value and the advertising delivery business target training label; the model parameters of the different sub-models constituting the model to be trained are adjusted using the first and second loss function values ​​until the model to be trained after adjusting the model parameters meets the preset convergence condition, thus obtaining the advertising gain model. This realizes the integration of a model that can handle multiple training scenarios, a model that can predict multiple business targets corresponding to advertising delivery and non-delivery, a model that can predict the business gain brought by advertising delivery in the training scenario, and a model that can predict the business prediction value when no advertising is delivered in the training scenario into a single model, resulting in the advertising gain model—a general model. This general model only needs to make one prediction to obtain the business delivery gain prediction value, business natural prediction value, and business natural prediction value for advertising delivery and non-delivery under multiple scenarios and multiple business targets, reducing the number of models, improving model utilization efficiency, and improving model prediction performance.

[0063] Reference Figure 2 This document illustrates a flowchart of an advertising delivery method provided in an embodiment of this application, applied to a business server of an advertising gain model, and specifically includes the following steps: Step 201: Obtain the scene characteristics of at least one preset advertising scenario, and the user characteristics of the user in the advertising scenario; In this embodiment, advertisers can place advertisements in advertising scenarios to improve their business goals. These advertising scenarios can refer to applications, websites, etc. Applications can include those for playing videos, allowing users to watch videos.

[0064] In one example, the advertising scenario includes software A, software B, and software C. The advertiser owns its own software D, and the advertiser's business objectives may include: the number of times software D is launched, the user's usage time on software D, and the economic benefits of software D. If software D is an application for playing videos, the advertiser's business objectives may also include: the amount of time users spend watching videos on software D.

[0065] Advertisers can select advertising scenarios from software A, B, and C, and place related advertisements in software D to improve their business goals.

[0066] In this embodiment, advertisers may have at least one preset advertising scenario to choose from for ad placement; users using an advertising scenario can be considered as users within that scenario. Advertisers can utilize an ad gain model to determine the predicted gain value of the advertising scenario for their business objectives. Then, advertisers can select the advertising scenario with the higher predicted gain value from the at least one advertising scenario for ad placement.

[0067] In this application embodiment, at least one preset advertising scenario's scenario characteristics and the user characteristics of the user in the advertising scenario can be obtained.

[0068] The characteristics of an advertising campaign can include: current and historical scenario characteristics. Current scenario characteristics can include: current basic scenario characteristics and scenario bidding characteristics. Historical scenario characteristics can include: historical basic scenario interaction characteristics and historical scenario interaction bidding characteristics. Historical basic scenario interaction characteristics can refer to historical scenario basic characteristics, and historical scenario interaction bidding characteristics can refer to historical scenario bidding characteristics.

[0069] The basic characteristics of a scenario can include: the name of the advertising scenario and the conversion rate of the advertising scenario to the advertiser's business objectives. Scenario bidding characteristics can include the price information for the advertiser to place ads in the advertising scenario.

[0070] User characteristics in an advertising context can include user information and user preferences for that specific advertising scenario. User information may include: age, gender, occupation, etc.

[0071] In one example, the advertising scenario is software A, and the advertiser's business objective is the number of times their software D is launched. Basic scenario characteristics may include: the name of software A and the conversion rate of software A's launches of software D. The conversion rate of software A's launches of software D can refer to the number of times a user is redirected to software D while using software A after ads related to software D are displayed on software A. User characteristics of software A may include user information and user preferences for software A. If software A is a video application, user preferences for software A may include user viewing preferences when watching videos on software A.

[0072] In this embodiment of the application, the user characteristics of a user in an advertising scenario may also include business characteristics of the user related to the advertiser's business objectives when no ads are being displayed in the advertising scenario. For example, if the advertising scenario is software A, and the advertiser's business objective is the number of times the advertiser's software D is launched, the user characteristics may also include the number of times users of software A launch software D when no ads are being displayed in software A.

[0073] In this embodiment, advertisers can utilize an advertising gain model to determine the predicted gain value of an advertising placement scenario on their business objectives. The advertising gain model is obtained by fusing a scenario feature processing sub-model, a business prediction sub-model, a placement gain prediction sub-model, and a natural gain prediction sub-model.

[0074] Step 202, input the scene features into such a system as... Figure 1 The advertising gain model obtained by the model training method shown uses a preset scene feature processing sub-model to obtain the scene vector of the advertising delivery scene.

[0075] Reference Figure 3 The diagram illustrates an advertising gain model provided in an embodiment of this application. The advertising gain model may include a scene feature processing sub-model, a business prediction sub-model, a delivery gain prediction sub-model, and a natural gain prediction sub-model. Figure 3 The multi-scenario network part is the scenario feature processing sub-model, the multi-objective network (MMoE) is the business prediction sub-model, the delivery network part is the delivery gain prediction sub-model, and the non-delivery network part is the natural gain prediction sub-model.

[0076] The scenario characteristics of an advertising campaign can include: current scenario characteristics and historical scenario interaction characteristics. Current scenario characteristics can include: current basic scenario characteristics and scenario bidding characteristics. Historical scenario interaction characteristics can include: historical basic scenario interaction characteristics and historical scenario interaction bidding characteristics. Scenario characteristics and historical scenario interaction characteristics constitute the personalized scenario characteristics of an advertising campaign.

[0077] In this embodiment, the scene features of the advertising placement scenario can be input into the scene feature processing sub-model in the advertising gain model. The scene feature processing sub-model includes two gate nu1s (gating units 1), where gate nu1 is the scene encoder. The scene feature processing sub-model can use gate nu1 to convert the scene features into scene vectors of the advertising placement scenario.

[0078] In this embodiment, scene features and interaction features between users and the advertising scenario can also be input into the scene feature processing sub-model to obtain a scene vector. The interaction features between users and the advertising scenario may include user preferences for the advertising scenario from the user features.

[0079] Step 203, input the user characteristics into such as Figure 1 The model training method shown obtains a business prediction sub-model in the advertising gain model, which is preset to obtain the natural business prediction vector of the user's contribution to the preset business objective; the business objective is the preset business objective corresponding to the advertisement to be delivered. In this application embodiment, the user characteristics of a user in an advertising scenario may include user information, user preferences for the advertising scenario, and business characteristics of the user related to the advertiser's business objectives when no advertising is displayed in the advertising scenario. Figure 3 Feature 1, Feature 2, Feature 3, Feature 4, Feature 5, and Feature N are the user features.

[0080] In this embodiment, user features can be input into the business prediction sub-model in the advertising gain model to obtain the natural prediction vector of the contribution of users in the advertising scenario to the business objectives corresponding to the advertising to be placed, when the advertiser does not place the advertising to be placed in the advertising scenario.

[0081] The business prediction sub-model can include at least one business objective tower, and can utilize this tower to determine a natural prediction vector of a user's contribution to a business objective based on user characteristics. (Refer to...) Figure 3 The advertiser's business objectives include revenue and video app playback time. The business objective tower can include DNN Layer 1 (revenue tower) and DNN Layer 1 (playback tower). The revenue tower outputs a natural prediction vector of user revenue for the advertiser, while the playback tower outputs a natural prediction vector of user playback time for the advertiser's video app.

[0082] Step 204: Input the business natural prediction vector and the scene vector into... Figure 1The model training method shown obtains a pre-set gain prediction sub-model in the advertising gain model, which yields the predicted business delivery gain value of the user's contribution to the business objective; the predicted business delivery gain value is the predicted business delivery gain value of the user to the business objective when the advertisement to be delivered is delivered in the advertising delivery scenario.

[0083] In this embodiment of the application, the scene vector output by the scene feature processing sub-model and the natural business prediction vector output by the business prediction sub-model are input into the gain prediction sub-model in the advertising gain model to obtain the predicted value of the business delivery gain contributed by the user to the business objective when the advertisement to be delivered is delivered in the advertising delivery scenario.

[0084] In the embodiments of this application, the gain prediction sub-model in the advertising gain model includes the natural gain prediction sub-model and / or the delivery gain prediction sub-model. Figure 3 The non-loaded network part is the natural gain prediction sub-model, and the loaded network part is the loaded gain prediction sub-model.

[0085] In this embodiment of the application, the natural prediction vector of the business forecasting sub-model can be input into the natural gain prediction sub-model to obtain the natural prediction value of the user's business target.

[0086] In this embodiment of the application, the natural prediction vector of the business forecasting sub-model and the embedding vector of the FM output can also be input into the natural gain prediction sub-model to obtain the natural prediction value of the user's business objective.

[0087] Reference Figure 3 The natural gain prediction sub-model includes at least one set of "self-attention network + multilayer perception", and each set of "self-attention network + multilayer perception" corresponds to a business target tower in the business prediction sub-model.

[0088] The Natural Gain Prediction Submodel utilizes a self-attention network to process the natural prediction vector output by the corresponding business target tower, capturing its internal structure or temporal dependencies. This results in a first enhanced representation of the natural prediction vector. The submodel then uses a multilayer perception layer (part of the same self-attention network group) to further process this first enhanced representation, for example, by performing a non-linear transformation, to obtain a second enhanced representation. Finally, the submodel combines this second enhanced representation with the embedding vector output by the FM (Fast Moving Average) to obtain the corresponding natural prediction value. This natural prediction value is more accurate and precise than the natural prediction vector itself. By inputting the embedding vector from the FM into the Natural Gain Prediction Submodel, user features are reintroduced into the determination of the natural prediction value, further improving its accuracy.

[0089] exist Figure 3 In the model, the natural gain prediction sub-model outputs the natural predicted value of economic income and the natural predicted value of playback duration when no ads are displayed in the advertising scenario.

[0090] In this embodiment, the natural prediction vector of the business and the scenario vector can be input into the delivery gain prediction sub-model to obtain the predicted value of the user's contribution to the business objective.

[0091] In this embodiment, the natural prediction vector of the business is input into the natural gain prediction sub-model to obtain the natural prediction value of the user's business for the business target; the difference between the natural prediction value of the business and the business delivery prediction value is calculated, and the business delivery gain prediction value contributed by the user to the business target is determined by the difference. This realizes the determination of the business delivery gain prediction value when placing advertisements in the advertising delivery scenario by using the natural gain prediction sub-model and the delivery gain prediction sub-model in the gain prediction sub-model.

[0092] Reference Figure 3 The delivery gain prediction sub-model includes at least one set of "multilayer perception", each set of "multilayer perception" includes three multilayer perceptions, and each set of "multilayer perception" corresponds to a business target tower in the business prediction sub-model.

[0093] exist Figure 3First, the natural prediction vector output by each business target tower is combined with the scene vector output by the scene feature processing sub-model to obtain a fused scene-aware natural prediction vector, thus incorporating the advertising placement scenario into the determination of business placement prediction values. Then, the placement gain prediction sub-model uses multilayer perception to process the fused scene-aware natural prediction vector of the business target tower corresponding to that multilayer perception, obtaining two identical preliminary gain representations. Next, the placement gain prediction sub-model combines these two preliminary gain representations again with the scene vector output by the scene feature processing sub-model, strengthening the weight of the advertising placement scenario in the determination of business placement prediction values, resulting in a first enhanced gain representation and a second enhanced gain representation corresponding to the two preliminary gain representations. The placement gain prediction sub-model can then use two multilayer perceptions in the same group as that multilayer perception to process the first and second enhanced gain representations respectively, obtaining a third gain corresponding to the first enhanced gain representation and constraints corresponding to the second enhanced gain representation. These constraints are used to prevent the model from overfitting in complex scenarios.

[0094] Finally, the delivery gain prediction sub-model can combine the third gain with the embedding vector output by FM, and reintroduce user features to obtain the business delivery gain prediction value. The delivery gain prediction sub-model can then subtract the business natural prediction value output by the natural gain prediction sub-model to obtain the business delivery prediction value.

[0095] Here, the reason is that when constructing the training set of the first causal inference sub-model and using the training set to train the first causal inference sub-model into a delivery gain prediction sub-model, the constructed training set can only contain the standard business delivery gain prediction value corresponding to the training scenario features, but cannot contain the standard business delivery prediction value corresponding to the training scenario features. Therefore, the delivery gain prediction sub-model can only be trained to obtain the business delivery gain prediction value first, and then determine the business delivery prediction value based on the business delivery gain prediction value.

[0096] Reference Figure 3 The delivery gain prediction sub-model yielded the revenue gain prediction value, revenue constraint, playback gain prediction value, and playback constraint for economic revenue and playback duration. The delivery gain prediction sub-model used the natural revenue prediction value and revenue gain prediction value of economic revenue to obtain the revenue delivery prediction value of economic revenue; and used the natural playback prediction value and playback gain prediction value of playback duration to obtain the playback delivery prediction value of playback duration.

[0097] In the embodiments of this application, during the inference process of the advertising gain model, for any advertising scenario, the gain prediction sub-model can obtain the business placement prediction value, business natural prediction value and business placement gain prediction value of each business objective of the advertisement to be placed through one inference prediction process, which effectively improves inference efficiency and saves inference costs.

[0098] Step 205: Based on the business delivery gain prediction value, determine the target advertising delivery scenario in at least one of the advertising delivery scenarios, and deliver the advertisement to be delivered to the target advertising delivery scenario.

[0099] In this embodiment, the advertising gain model outputs a predicted value of the business delivery gain that a user contributes to the business objectives related to the advertising to be delivered after the advertising to be delivered in an advertising delivery scenario. Based on the predicted value of the business delivery gain corresponding to at least one advertising delivery scenario, the advertising delivery scenario with the higher predicted value of the business delivery gain can be selected as the target advertising delivery scenario, and the advertising to be delivered can be delivered to the target advertising delivery scenario.

[0100] In this embodiment, at least one preset advertising scenario's scenario features and user features within the advertising scenario are obtained. The scenario features are input into a preset scenario feature processing sub-model in the advertising gain model to obtain a scenario vector for the advertising scenario. The user features are input into a preset business prediction sub-model in the advertising gain model to obtain a natural business prediction vector of the user's contribution to a preset business objective. The business objective is the business objective corresponding to the preset ad to be placed. The natural business prediction vector and / or scenario vector are input into a preset gain prediction sub-model in the advertising gain model to obtain a predicted business placement gain value of the user's contribution to the business objective. The predicted business placement gain value is the predicted business placement gain value of the user for the business objective when the ad to be placed is placed in the advertising scenario. Based on the predicted business placement gain value, a target advertising scenario is determined in at least one advertising scenario, and the ad to be placed is placed in the target advertising scenario. This integrates a model that can process multiple advertising scenarios, a model that can predict multiple business objectives corresponding to the ad to be placed, and a model that can predict the business gain brought by placing an ad in an advertising scenario into one model, resulting in the general model of the advertising gain model. This universal model only needs to make one prediction to obtain the predicted business delivery gain for all business objectives in any advertising scenario, reducing the number of models, improving model utilization efficiency, and enhancing model prediction performance.

[0101] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of this application are not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of this application.

[0102] Reference Figure 4 The diagram shows a structural block diagram of a model training device provided in an embodiment of this application, which may specifically include the following modules: The first acquisition module 301 is used to acquire multiple batches of training samples. Each batch of training samples includes: training samples for advertising and training samples for not advertising. The training samples for advertising include first training user features, training scenario features, advertising business training data features, and advertising business target training labels. The training samples for not advertising include second training user features, advertising non-advertising business training data features, and advertising non-advertising business target training labels. The model to be trained consists of a scenario processing sub-model, a business target prediction sub-model, a first causal inference sub-model, and a second causal inference sub-model. The first model processing module 302 is used to input the training scene features into the scene processing sub-model and output the training scene vector. The second model processing module 303 is used to input the first training user features and advertising placement business training data features into the business target prediction sub-model to obtain the first training business natural prediction vector. The third model processing module 304 is used to input the second training user features and the advertising non-delivery business training data features into the business target prediction sub-model to obtain the second training business natural prediction vector. The fourth model processing module 305 is used to input the training scene vector and the first training service natural prediction vector into the first causal inference sub-model to obtain the training service delivery gain prediction value. The fifth model processing module 306 is used to input the second training business natural prediction vector into the second causal inference sub-model to obtain the training business natural prediction value; The function value calculation module 307 is used to obtain a first loss function value through the natural prediction value of the training business and the target training label of the advertising non-delivery business, and to obtain a second loss function value through the prediction value of the training business delivery gain and the target training label of the advertising delivery business; The adjustment module 308 is used to adjust the model parameters of different sub-models constituting the model to be trained by using the first loss function value and the second loss function value, until the model to be trained after adjusting the model parameters meets the preset convergence condition, thus obtaining the advertising gain model.

[0103] In one optional embodiment of this application, the training data features for advertising placement business include training data features for various advertising placement business objectives.

[0104] The second model processing module 303 includes: The first model processing submodule is used to input the first training user features and training data features including multiple advertising business objectives into the business objective prediction submodel to obtain the first training business natural prediction vector corresponding to the training data features of different advertising business objectives.

[0105] In one optional embodiment of this application, the training data features of the advertising non-delivery business include training data features of various advertising non-delivery business objectives; The third model processing module 304 includes: The second model processing submodule is used to input the second training user features and training data features including multiple non-advertising business objectives into the business objective prediction submodel to obtain the second training business natural prediction vector corresponding to the training data features of different non-advertising business objectives.

[0106] In one optional embodiment of this application, the business objective training tags include launch count tags, usage duration tags, and economic benefit tags for the training scenario under both advertising and non-advertising conditions; The function value calculation module 307 includes: The first identification submodule is used to identify, from the natural prediction values ​​of the training business, the natural prediction values ​​of the number of launches, the natural prediction values ​​of the usage duration, and the natural prediction values ​​of the economic benefits of the training scenario when no ads are placed; The first function value calculation submodule is used to obtain the first loss function value of the number of launches without ads by using the naturally predicted value of the number of launches without ads and the launch number label; The second function value calculation submodule is used to obtain the first loss function value of usage time without advertising by using the naturally predicted value of usage time and the usage time tag when no ads are delivered; The third function value calculation submodule is used to obtain the first loss function value of the economic benefits without advertising by using the natural predicted value of economic benefits and the economic benefit label. The second identification submodule is used to identify, from the training service delivery gain prediction value, the launch number delivery gain prediction value, the usage duration delivery gain prediction value, and the economic benefit delivery gain prediction value of the training scenario under advertising delivery; The fourth function value calculation submodule is used to obtain the second loss function value about the number of launches under advertising by using the predicted value of the launch gain and the launch number label. The fifth function value calculation submodule is used to obtain the second loss function value for usage time under advertising by using the predicted value of usage time gain under advertising and the usage time label; The sixth function value calculation submodule is used to obtain the second loss function value of the economic benefits under advertising by using the predicted value of the economic benefit under advertising and the economic benefit label.

[0107] In one optional embodiment of this application, the adjustment module 308 includes: The adjustment submodule is used to adjust the model parameters of different sub-models that constitute the model to be trained by using the first loss function value and the second loss function value; The first condition judgment submodule is used to obtain the scene feature processing submodel when the scene processing submodel of the model to be trained after adjusting the model parameters meets the preset convergence condition. The second condition judgment submodule is used to obtain the business prediction submodel when the business target prediction submodel of the model to be trained after adjusting the model parameters meets the preset convergence condition. The third condition judgment submodule is used to obtain the delivery gain prediction submodel when the first causal inference submodel of the model to be trained after adjusting the model parameters meets the preset convergence condition. The fourth condition judgment submodule is used to obtain the natural gain prediction submodel when the second causal inference submodel of the model to be trained after adjusting the model parameters meets the preset convergence condition. The fifth condition judgment submodule is used to fuse the scene feature processing submodel, the business prediction submodel, the delivery gain prediction submodel and the natural gain prediction submodel to obtain the advertising gain model.

[0108] In one optional embodiment of this application, the training service delivery gain prediction value is the training service delivery gain prediction value of the first training user for the service target when advertising is delivered in the training scenario. The model training device also includes: The difference calculation module is used to calculate the difference between the natural predicted value of the training service and the predicted value of the training service delivery gain. The determination module is used to determine the predicted value of the first training user's contribution to the business objective based on the difference.

[0109] In one optional embodiment of this application, the first training user feature includes a first sparse training user feature and / or a first dense training user feature; The second model processing module 303 also includes: The first feature extraction submodule is used to extract the first dense training user feature from the first training user feature, and input the first dense training user feature and the advertising placement business training data feature into the business target prediction submodel to obtain the first training business natural prediction vector. The second feature extraction submodule is used to extract the first sparse training user feature from the first training user feature, and to process the first sparse training user feature using the preset factorization machine in the advertising gain model to obtain the training embedding vector of the first sparse training user feature.

[0110] In one optional embodiment of this application, the fourth model processing module 305 includes: The third model processing submodule is used to input the first training business natural prediction vector, the training scenario vector, and the training embedding vector into the first causal inference submodel to obtain the training business delivery gain prediction value of the first training user's contribution to the business objective.

[0111] Reference Figure 5 This diagram illustrates a structural block diagram of an advertising delivery device provided in an embodiment of this application. This device is applied to a business server for an advertising gain model and may specifically include the following modules: The feature acquisition module 401 is used to acquire scene features of at least one preset advertising scenario, and user features of users in the advertising scenario. Scene vector acquisition module 402 is used to input the scene features as follows: Figure 1 The advertising gain model obtained by the model training method shown in the figure has a preset scene feature processing sub-model, which obtains the scene vector of the advertising delivery scene. The business natural prediction vector acquisition module 403 is used to input the user features as follows: Figure 1 The model training method shown in the figure obtains a business prediction sub-model in the advertising gain model, which is preset to obtain the natural business prediction vector of the user's contribution to the preset business objective; the business objective is the preset business objective corresponding to the advertisement to be delivered. The service delivery gain prediction value acquisition module 404 is used to input the service natural prediction vector and the scene vector as follows: Figure 1The model training method shown in the figure obtains a pre-set gain prediction sub-model in the advertising gain model, which yields the predicted value of the user's contribution to the business objective; the predicted value of the business gain is the predicted value of the user's contribution to the business objective when the advertisement to be advertised is advertised in the advertising scenario. The advertising delivery module 405 is used to determine a target advertising delivery scenario in at least one of the advertising delivery scenarios based on the business delivery gain prediction value, and to deliver the advertisement to be delivered to the target advertising delivery scenario.

[0112] As the device embodiment is basically similar to the method embodiment, the description is relatively simple, and relevant parts can be found in the description of the method embodiment.

[0113] In addition, embodiments of this application also provide an electronic device, such as... Figure 6 As shown, it includes a processor 501, a communication interface 502, a memory 503, and a communication bus 504, wherein the processor 501, the communication interface 502, and the memory 503 communicate with each other through the communication bus 504. Memory 503 is used to store computer programs; Processor 501 is used to execute the program stored in memory 503, and to perform the following steps: Multiple batches of training samples are obtained. Each batch of training samples includes: training samples with ads and training samples without ads. The training samples with ads include first training user features, training scenario features, advertising business training data features, and advertising business target training labels. The training samples without ads include second training user features, advertising non-ad training data features, and advertising non-ad training target training labels. The model to be trained consists of a scenario processing sub-model, a business target prediction sub-model, a first causal inference sub-model, and a second causal inference sub-model. The training scene features are input into the scene processing sub-model, which outputs a training scene vector. Input the first training user features and advertising delivery business training data features into the business target prediction sub-model to obtain the first training business natural prediction vector; Input the second training user features and the advertising non-delivery business training data features into the business target prediction sub-model to obtain the second training business natural prediction vector; The training scenario vector and the first training service natural prediction vector are input into the first causal inference sub-model to obtain the training service delivery gain prediction value. Input the second training business natural prediction vector into the second causal inference sub-model to obtain the training business natural prediction value; A first loss function value is obtained by using the natural prediction value of the training business and the target training label of the advertising-not-delivered business; a second loss function value is obtained by using the predicted value of the training business delivery gain and the target training label of the advertising delivery business. The model parameters of different sub-models constituting the model to be trained are adjusted by the first loss function value and the second loss function value until the model to be trained after adjusting the model parameters meets the preset convergence condition, thus obtaining the advertising gain model.

[0114] In addition, embodiments of this application also provide an electronic device, such as... Figure 7 As shown, it includes a processor 601, a communication interface 602, a memory 603, and a communication bus 604, wherein the processor 601, the communication interface 602, and the memory 603 communicate with each other through the communication bus 604. Memory 603 is used to store computer programs; Processor 601 is used to execute the program stored in memory 603, and to perform the following steps: Obtain the scene characteristics of at least one preset advertising scenario, and the user characteristics of the users in the advertising scenario; Input the scene features as follows Figure 1 The advertising gain model obtained by the model training method shown uses a preset scene feature processing sub-model to obtain the scene vector of the advertising delivery scene. Input the user characteristics as follows Figure 1 The model training method shown obtains a business prediction sub-model in the advertising gain model, which is preset to obtain the natural business prediction vector of the user's contribution to the preset business objective; the business objective is the preset business objective corresponding to the advertisement to be delivered. Input the business natural prediction vector and the scene vector as follows: Figure 1 The model training method shown in the figure obtains a pre-set gain prediction sub-model in the advertising gain model, which yields the predicted value of the user's contribution to the business objective; the predicted value of the business gain is the predicted value of the user's contribution to the business objective when the advertisement to be advertised is advertised in the advertising scenario. Based on the business delivery gain prediction value, a target advertising delivery scenario is determined in at least one of the advertising delivery scenarios, and the advertisement to be delivered is delivered to the target advertising delivery scenario.

[0115] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0116] The communication interface is used for communication between the aforementioned terminal and other devices.

[0117] The memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0118] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0119] like Figure 8 As shown, in another embodiment provided in this application, a computer-readable medium 701 is also provided, which stores instructions that, when run on a computer, cause the computer to execute a model training method or an advertising delivery method as described in the above embodiments.

[0120] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute a model training method or an advertising delivery method as described in the above embodiments.

[0121] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available 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 disk (SSD)).

[0122] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0123] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0124] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the scope of protection of this application.

Claims

1. A model training method, characterized in that, The method includes: Multiple batches of training samples are obtained. Each batch of training samples includes: training samples with ads and training samples without ads. The training samples with ads include first training user features, training scenario features, advertising business training data features, and advertising business target training labels. The training samples without ads include second training user features, advertising non-ad training data features, and advertising non-ad training target training labels. The model to be trained consists of a scenario processing sub-model, a business target prediction sub-model, a first causal inference sub-model, and a second causal inference sub-model. The training scene features are input into the scene processing sub-model, which outputs a training scene vector. Input the first training user features and advertising delivery business training data features into the business target prediction sub-model to obtain the first training business natural prediction vector; Input the second training user features and the advertising non-delivery business training data features into the business target prediction sub-model to obtain the second training business natural prediction vector; The training scenario vector and the first training service natural prediction vector are input into the first causal inference sub-model to obtain the training service delivery gain prediction value. Input the second training business natural prediction vector into the second causal inference sub-model to obtain the training business natural prediction value; A first loss function value is obtained by using the natural prediction value of the training business and the target training label of the advertising-not-delivered business; a second loss function value is obtained by using the predicted value of the training business delivery gain and the target training label of the advertising delivery business. The model parameters of different sub-models constituting the model to be trained are adjusted by the first loss function value and the second loss function value until the model to be trained after adjusting the model parameters meets the preset convergence condition, thus obtaining the advertising gain model.

2. The method according to claim 1, characterized in that, The training data features for the advertising placement business include training data features for various advertising placement business objectives; The step of inputting the first training user features and advertising delivery business training data features into the business target prediction sub-model to obtain the first training business natural prediction vector includes: The first training user features and training data features including multiple advertising business objectives are input into the business objective prediction sub-model to obtain the first training business natural prediction vector corresponding to the training data features of different advertising business objectives.

3. The method according to claim 1, characterized in that, The training data features for the non-advertising business include training data features for various non-advertising business objectives; The step of inputting the second training user features and the advertising-unserved business training data features into the business target prediction sub-model to obtain the second training business natural prediction vector includes: The second training user features and training data features including various non-advertising business objectives are input into the business objective prediction sub-model to obtain the second training business natural prediction vector corresponding to the training data features of different non-advertising business objectives.

4. The method according to claim 1, characterized in that, Business objective training tags include launch count tags, usage duration tags, and economic benefit tags for training scenarios under both advertising and non-advertising conditions; The step of obtaining a first loss function value using the natural prediction value of the training business and the target training label of the advertising-not-delivered business, and obtaining a second loss function value using the predicted gain value of the training business and the target training label of the advertising-delivered business, includes: From the natural predicted values ​​of the training business, we can identify the natural predicted values ​​of the number of times the training scene is launched, the natural predicted value of the usage time, and the natural predicted value of the economic benefits when the advertisement is not placed. By using the naturally predicted number of launches without ads and the launch number label, we obtain the first loss function value for the number of launches without ads. By using the naturally predicted usage duration and usage duration tags when no ads were delivered, we obtain the first loss function value for usage duration when no ads were delivered. By using the natural predicted value of economic benefits without advertising and the economic benefit label, we obtain the first loss function value for economic benefits without advertising. From the predicted gain values ​​of the training service delivery, the predicted gain values ​​of the number of launches, the predicted gain values ​​of the usage time, and the predicted gain values ​​of the economic benefits of the training scenario under the advertising delivery are identified. By using the predicted value of the launch gain and the launch number label under the ad campaign, we obtain the second loss function value for the launch number under the ad campaign. By using the predicted value of usage time gain under advertising and the usage time label, we obtain the second loss function value for usage time under advertising. By using the predicted economic gain from advertising and the economic gain label, we obtain the second loss function value for the economic returns under advertising.

5. The method according to claim 1, characterized in that, The step of adjusting the model parameters of different sub-models constituting the model to be trained using the first loss function value and the second loss function value until the model to be trained after adjusting the model parameters meets the preset convergence condition to obtain the advertising gain model includes: The model parameters of the different sub-models constituting the model to be trained are adjusted by using the first loss function value and the second loss function value; When the scene processing sub-model of the model to be trained after adjusting the model parameters meets the preset convergence condition, the scene feature processing sub-model is obtained. When the business objective prediction sub-model of the training model after adjusting the model parameters meets the preset convergence condition, the business prediction sub-model is obtained. When the first causal inference sub-model of the model to be trained after adjusting the model parameters satisfies the preset convergence condition, the delivery gain prediction sub-model is obtained. When the second causal inference sub-model of the model to be trained after adjusting the model parameters satisfies the preset convergence condition, the natural gain prediction sub-model is obtained. The scene feature processing sub-model, the business prediction sub-model, the delivery gain prediction sub-model, and the natural gain prediction sub-model are fused together to obtain the advertising gain model.

6. The method according to claim 1, characterized in that, The training service delivery gain prediction value is the training service delivery gain prediction value of the first training user for the service target when advertising is delivered in the training scenario. After inputting the second training business natural prediction vector into the second causal inference sub-model to obtain the training business natural prediction value, the method further includes: Calculate the difference between the natural predicted value of the training service and the predicted value of the training service delivery gain; The difference is used to determine the predicted value of the training service deployment based on the contribution of the first training user to the business objective.

7. The method according to claim 1, characterized in that, The first training user features include a first sparse training user feature and / or a first dense training user feature; the step of inputting the first training user features and advertising delivery business training data features into the business target prediction sub-model to obtain the first training business natural prediction vector further includes: Extract the first dense training user feature from the first training user feature, and input the first dense training user feature and the advertising placement business training data feature into the business target prediction sub-model to obtain the first training business natural prediction vector. The method includes: Extract the first sparse training user feature from the first training user feature, and process the first sparse training user feature using the preset factorization machine in the advertising gain model to obtain the training embedding vector of the first sparse training user feature.

8. The method according to claim 7, characterized in that, The step of inputting the training scene vector and the first training service natural prediction vector into the first causal inference sub-model to obtain the training service delivery gain prediction value includes: The first training service natural prediction vector, the training scenario vector, and the training embedding vector are input into the first causal inference sub-model to obtain the training service delivery gain prediction value of the first training user's contribution to the service objective.

9. An advertising placement method, characterized in that, The business servers used to deploy the ad gain model include: Obtain the scene characteristics of at least one preset advertising scenario, and the user characteristics of the users in the advertising scenario; The scene features are input into the preset scene feature processing sub-model in the advertising gain model obtained by any of the model training methods described in claims 1 to 8 to obtain the scene vector of the advertising delivery scene; The user features are input into the preset business prediction sub-model of the advertising gain model obtained by any of the model training methods described in claims 1 to 8 to obtain the natural business prediction vector of the user's contribution to the preset business objective; the business objective is the preset business objective corresponding to the advertisement to be delivered. The business natural prediction vector and the scenario vector are input into the preset gain prediction sub-model in the advertising gain model obtained by any of the model training methods described in claims 1 to 8 to obtain the business delivery gain prediction value of the user's contribution to the business objective; the business delivery gain prediction value is the business delivery gain prediction value of the user to the business objective when the advertisement to be delivered is delivered in the advertising delivery scenario. Based on the business delivery gain prediction value, a target advertising delivery scenario is determined in at least one of the advertising delivery scenarios, and the advertisement to be delivered is delivered to the target advertising delivery scenario.

10. A model training device, characterized in that, The device includes: The first acquisition module is used to acquire multiple batches of training samples. Each batch of training samples includes: training samples for ads that have been placed and training samples for ads that have not been placed. The training samples for ads that have been placed include first training user features, training scenario features, advertising placement business training data features, and advertising placement business target training labels. The training samples for ads that have not been placed include second training user features, advertising non-placement business training data features, and advertising non-placement business target training labels. The model to be trained consists of a scenario processing sub-model, a business target prediction sub-model, a first causal inference sub-model, and a second causal inference sub-model. The first model processing module is used to input the training scene features into the scene processing sub-model and output the training scene vector. The second model processing module is used to input the first training user features and advertising placement business training data features into the business target prediction sub-model to obtain the first training business natural prediction vector. The third model processing module is used to input the second training user features and the advertising non-delivery business training data features into the business target prediction sub-model to obtain the second training business natural prediction vector. The fourth model processing module is used to input the training scene vector and the first training service natural prediction vector into the first causal inference sub-model to obtain the training service delivery gain prediction value. The fifth model processing module is used to input the second training business natural prediction vector into the second causal inference sub-model to obtain the training business natural prediction value; The function value calculation module is used to obtain a first loss function value through the natural prediction value of the training business and the target training label of the advertising non-delivery business, and to obtain a second loss function value through the predicted value of the training business delivery gain and the target training label of the advertising delivery business; The adjustment module is used to adjust the model parameters of different sub-models constituting the model to be trained by using the first loss function value and the second loss function value, until the model to be trained after adjusting the model parameters meets the preset convergence condition, thus obtaining the advertising gain model.

11. An advertising delivery device, characterized in that, The device, which is used in a business server for an advertising gain model, includes: The feature acquisition module is used to acquire scene features of at least one preset advertising scenario, and user features of users in the advertising scenario. The scene vector acquisition module is used to input the scene features into the preset scene feature processing sub-model in the advertising gain model obtained by any of the model training methods described in claims 1 to 8, so as to obtain the scene vector of the advertising delivery scene; The business natural prediction vector acquisition module is used to input the user features into a preset business prediction sub-model in the advertising gain model obtained by any of the model training methods described in claims 1 to 8, and obtain the business natural prediction vector of the user's contribution to the preset business objective; the business objective is the preset business objective corresponding to the advertisement to be delivered; The business delivery gain prediction module is used to input the natural prediction vector of the business and the scenario vector into the pre-set gain prediction sub-model in the advertising gain model obtained by any of the model training methods described in claims 1 to 8, and obtain the business delivery gain prediction value of the user's contribution to the business objective; the business delivery gain prediction value is the business delivery gain prediction value of the user to the business objective when the advertisement to be delivered is delivered in the advertising delivery scenario. The advertising delivery module is used to determine a target advertising delivery scenario in at least one of the advertising delivery scenarios based on the business delivery gain prediction value, and to deliver the advertisement to be delivered to the target advertising delivery scenario.

12. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; When the processor executes a program stored in the memory, it implements the method as described in any one of claims 1-9.

13. A computer-readable medium having instructions stored thereon that, when executed by one or more processors, cause the processors to perform the method as described in any one of claims 1-9.

Images