Advertisement putting model training method, advertisement putting method, device and electronic equipment

Abstract

Embodiments of the present application provide an advertisement delivery model training method, an advertisement delivery method, an apparatus and an electronic device. The advertisement delivery model training method comprises: obtaining a plurality of training samples; the training sample comprises operation data of a sample advertisement in a sample device associated with an account, a user type of a sample device, a device identifier of the sample device, and a delivery result of the sample advertisement in the sample device; the label of the training sample comprises: an actual conversion result of a product corresponding to the sample advertisement in the sample device; extracting first features of the plurality of training samples and second features of the training samples corresponding to each user type through an advertisement delivery model; fusing the first features and the second features, obtaining a predicted conversion result of the sample advertisement in the sample device based on the fused features; adjusting the model parameters according to the difference between the actual conversion result and the predicted conversion result to obtain a trained advertisement delivery model.

Description

Technical Field

[0001] This invention relates to the field of advertising processing technology, and in particular to advertising placement model training methods, advertising placement methods, devices, and electronic equipment. Background Technology

[0002] With the continuous development of multimedia platforms, the demand for placing advertisements (e.g., ads) on these platforms is also increasing. For example, ads can be placed through the Real-Time API (RTA) provided by the multimedia platform.

[0003] The lack of a method in related technologies to predict ad conversion results makes it impossible to target ads based on conversion outcomes. Summary of the Invention

[0004] The purpose of this invention is to provide an advertising delivery model training method, advertising delivery method, apparatus, and electronic device to address the problem in related technologies of lacking a method to predict advertising conversion results, which prevents targeted advertising delivery based on conversion outcomes. The specific technical solution of this application is as follows: In a first aspect of this application, an advertising delivery model training method is provided, comprising: acquiring multiple sets of training samples; the training samples include: operation data of sample advertisements in sample devices associated with an account, user types logged into the sample devices, device identifiers of the sample devices, and delivery results of the sample advertisements in the sample devices; the training samples have labels, the labels including: the actual conversion results of the products corresponding to the sample advertisements in the sample devices; extracting first features of the multiple sets of training samples and second features of training samples corresponding to each user type through an advertising delivery model; the first features are features extracted based on training samples corresponding to multiple user types, and the second features are features extracted based on training samples corresponding to one user type, the first features and the second features are obtained through different networks in the advertising delivery model; fusing the first features and the second features to obtain fused features; determining the predicted conversion results of the sample advertisements in the sample devices based on the fused features; adjusting the model parameters of the advertising delivery model according to the difference between the actual conversion results and the predicted conversion results to obtain a trained advertising delivery model.

[0005] In a second aspect of the present invention, an advertising delivery method is also provided, comprising: acquiring multiple sets of data to be processed; the data to be processed includes: historical operation data of the advertisement on a device associated with an account, the user type logged into the device, the device identifier of the device, and preset delivery results of the advertisement on the device; inputting the multiple sets of data to be processed into a trained advertising delivery model to obtain conversion results corresponding to each device; the advertising delivery model is generated by any of the methods described in the first aspect above; and determining the device to be delivered to the advertisement from the multiple devices according to the conversion results corresponding to each device.

[0006] In a third aspect of this invention, an advertising delivery model training device is also provided, comprising: a first acquisition module for acquiring multiple sets of training samples; the training samples include: operation data of sample advertisements in sample devices associated with an account, user types logged into the sample devices, device identifiers of the sample devices, and delivery results of the sample advertisements in the sample devices; the training samples have labels, the labels including: actual conversion results of the products corresponding to the sample advertisements in the sample devices; a feature extraction module for extracting first features of the multiple sets of training samples and second features of training samples corresponding to each user type through an advertising delivery model; the first features are features extracted based on training samples corresponding to multiple user types, and the second features are features extracted based on training samples corresponding to one user type, the first features and the second features are obtained through different networks in the advertising delivery model; a feature fusion module for fusing the first features and the second features to obtain fused features; a prediction model for determining the predicted conversion results of the sample advertisements in the sample devices based on the fused features; and a parameter adjustment module for adjusting the model parameters of the advertising delivery model according to the difference between the actual conversion results and the predicted conversion results to obtain a trained advertising delivery model.

[0007] In a fourth aspect of the present invention, an advertising delivery device is also provided, comprising: a second acquisition module, configured to acquire multiple sets of data to be processed; the data to be processed includes: historical operation data of the advertisement on a device associated with an account, the user type logged into the device, the device identifier of the device, and a preset delivery result of the advertisement on the device; a third acquisition module, configured to input the multiple sets of data to be processed into a trained advertising delivery model to obtain a conversion result corresponding to each device; the advertising delivery model is generated by any of the methods described in the first aspect; and a first determination module, configured to determine a device to be delivered to for the advertisement from among the multiple devices based on the conversion result corresponding to each device.

[0008] In a fifth aspect of the present invention, 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; and the processor is used to implement any of the steps of the method described in the first aspect when executing the program stored in the memory.

[0009] In a sixth aspect of the invention, a computer-readable storage medium is also provided, wherein a computer program is stored therein, and the computer program, when executed by a processor, implements the steps of any of the methods described above.

[0010] In a seventh aspect of the present invention, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the above-described advertising delivery model training methods or advertising delivery methods.

[0011] In this embodiment, multiple sets of training samples are obtained. These training samples include: operation data of sample advertisements on sample devices associated with the account, user types logged into the sample devices, device identifiers of the sample devices, and the delivery results of sample advertisements on the sample devices. The labels of the training samples include: the actual conversion results of the sample advertisements on the sample devices. Through the advertising delivery model, first features of the multiple sets of training samples and second features corresponding to each user type are extracted. These features reflect the commonalities of different user types, while the second features reflect the unique characteristics corresponding to each user type. The first features are extracted from training samples corresponding to multiple user types, and the second features are extracted from training samples corresponding to one user type. The first and second features are obtained through different networks in the advertising delivery model. The first and second features are fused to obtain fused features, which reflect both the first features of different user types and the second features corresponding to each user type. Based on the fused features, the predicted conversion results of the sample advertisements on the sample devices are determined. According to the difference between the actual conversion results and the predicted conversion results, the model parameters of the advertising delivery model are adjusted to obtain the trained advertising delivery model. This advertising delivery model can predict the conversion results of ads on devices, providing a basis for ad delivery and enabling targeted ad delivery based on the conversion results. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0013] Figure 1 A flowchart illustrating the steps of an advertising placement model training method provided in this application embodiment; Figure 2 A flowchart illustrating the steps of another advertising delivery model training method provided in this application embodiment; Figure 3 This is a schematic diagram of the structure of an advertising delivery model provided in an embodiment of this application; Figure 4 A flowchart illustrating the steps of an advertising delivery method provided in this application embodiment; Figure 5 A flowchart illustrating the steps of another advertising delivery method provided in this application embodiment; Figure 6 This is a schematic diagram of the structure of an advertising delivery model in related technologies; Figure 7 This is a schematic diagram of another advertising delivery model in related technologies; Figure 8 This is a schematic diagram of another advertising delivery model provided in an embodiment of this application; Figure 9 A structural block diagram of an advertising placement model training device provided in an embodiment of this application; Figure 10 A structural block diagram of an advertising delivery device provided in an embodiment of this application; Figure 11 This is a schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0014] The technical solutions of the present invention will now be described with reference to the accompanying drawings in the embodiments of the present invention.

[0015] Figure 1 This is a flowchart illustrating the steps of an advertising placement model training method provided in this application, for reference only. Figure 1 The method may include the following steps: Step 101: Obtain multiple sets of training samples.

[0016] The training samples include: operation data of sample ads on sample devices associated with the account, user types logged into the sample devices, device identifiers of the sample devices, and the delivery results of sample ads on the sample devices; the training samples are labeled, and the labels include: the actual conversion results of sample ads on the sample devices.

[0017] For example, training samples are collected from sample devices. The device identifier of the sample device may include the type of sample device and / or a device identifier. The sample device may be a mobile phone, tablet, or other electronic device that can be used to deliver advertisements. The sample device associated with an account may be the sample device logged into that account.

[0018] For example, a sample advertisement can be a file slice of a sample multimedia file; for instance, a sample advertisement can include: a file slice obtained by slicing a sample multimedia file (such as a movie or TV series) in a video playback platform, and promote the sample multimedia file corresponding to the file slice to users by placing the file slice on the video playback platform or on other multimedia platforms; in response to the user's click on the file slice, the file slice display interface can jump to the video playback platform, and then in response to the user's click on the sample video in the video playback platform, play the video clicked by the user.

[0019] For example, sample ads can also be product ads; for instance, shopping platforms can place product ads on their platforms or other multimedia platforms to promote the products associated with those ads to users. In response to a user clicking on a product ad, the user can be redirected from the product ad display interface to the product details page corresponding to the product ad. Then, in response to the user placing an order on the product details page, a product order corresponding to the product ad can be generated.

[0020] For example, the user types logging into the sample device can include high-value users, potential users, and churned users. For instance, user types can be categorized based on user behavior on the sample device. For example, if a user frequently repurchases the product corresponding to the sample advertisement, they can be identified as a high-value user; if a user browses the product details page corresponding to the sample advertisement several times, they can be identified as a potential user; if a user browses the product details page corresponding to the sample advertisement or purchases the product, but does not browse the product details page or purchase the product within a preset time period, they can be identified as a churned user.

[0021] For example, the operational data of sample ads on sample devices associated with an account may include at least one of the following: the number of sample videos corresponding to the sample ads played, the number of times the sample videos corresponding to the sample ads were searched, and the number of times the sample ads were exposed on the sample devices within a historical time period. For example, the historical time period may include 3 days, 7 days, 14 days, and 30 days prior to the current time.

[0022] For example, the result of serving a sample ad on a sample device may include serving the sample ad on the sample device or not serving the sample ad. The training sample includes the result of serving the sample ad on the device or not serving the sample ad. Based on the training sample, the ad serving model is trained. The trained ad serving model can accurately predict the conversion result of the product corresponding to the ad when the ad is served, and the conversion result of the product corresponding to the ad when the ad is not served.

[0023] For example, if an advertisement is a file slice of a sample multimedia file, the conversion result of the product corresponding to the advertisement can include: the playback duration of the target multimedia file corresponding to the file slice on the device. Based on this embodiment, the trained advertisement delivery model can accurately predict: the playback duration of the target multimedia file on the device when the file slice is delivered, and the playback duration of the target multimedia file on the device when the file slice is not delivered.

[0024] For example, the results of sample ad delivery on sample devices may also include the number of impressions of the sample ad on the sample devices.

[0025] For example, training samples may also include: user attributes of users who log in to sample devices, model of sample devices, type of sample advertisements, and platform information of the multimedia platforms on which sample advertisements are placed.

[0026] For example, the labels corresponding to the training samples include the actual conversion results of the sample ads on the sample devices; there can be one or more actual conversion results. For instance, if the sample ad is a product ad, the actual conversion results could include: the quantity and amount of products purchased by sample users through the sample devices that correspond to the product ad.

[0027] Step 102: Using the advertising delivery model, extract the first features of multiple training samples and the second features of training samples corresponding to each user type.

[0028] For example, the first feature is a feature extracted from training samples corresponding to multiple user types, and the second feature is a feature extracted from training samples corresponding to one user type. The first and second features are obtained through different networks in the ad delivery model, that is, the first and second features are obtained separately through independent networks of the ad delivery model.

[0029] For example, the first feature includes at least one of the following: a feature for characterizing operational data, a feature for characterizing user type, a feature for characterizing device identifier, and a feature for characterizing delivery result; the second feature includes at least one of the following: a feature for characterizing operational data, a feature for characterizing user type, a feature for characterizing device identifier, and a feature for characterizing delivery result.

[0030] The first feature represents the common features of multiple sets of advertising information samples; the second feature represents the unique features of advertising information samples corresponding to user types. For example, the advertising delivery model can be a multi-task network model, such as a Multi-gate Mixture-of-Experts (MMoE) network, a Progressive Layered Extraction (PLE) network, or other multi-task network models. Alternatively, the advertising delivery model can also be a single-task network model. For example, the feature extraction sub-model in the advertising delivery model extracts the first feature of multiple sets of training samples and the second feature of the training samples corresponding to each user type.

[0031] For example, training samples may include multiple feature data. These could include first feature data f1, second feature data f2, and so on, up to the nth feature data fn. Furthermore, these n feature data (f1 to fn) are aligned. Vectorizing these n feature data yields n feature vectors. These feature vectors are then input into the feature extraction sub-model of the ad delivery model. The feature extraction sub-model learns the first feature of all training samples and outputs the first feature in vector format. Similarly, for each user type, the feature vectors of the n feature data from the training samples corresponding to that user type are input into the feature extraction sub-model of the ad delivery model. The feature extraction sub-model learns the second feature of the training samples corresponding to that user type and outputs the second feature in vector format.

[0032] Step 103: Fuse the first feature and the second feature to obtain the fused feature.

[0033] For example, the first feature and the second feature each have corresponding weights. Based on the weights of the first feature and the second feature, a weighted summation is performed on the first feature and the second feature, and the weighted summation result is determined as the fusion feature of the first feature and the second feature.

[0034] Step 104: Based on the fusion features, determine the predicted conversion results of the sample advertisements on the sample devices.

[0035] In this case, the predicted conversion result and the actual conversion result are of the same type; for example, if the sample ad is a sample slice of a sample multimedia file, and the actual conversion result includes the duration of playing the sample multimedia file corresponding to the sample ad, then the predicted conversion result also includes the duration of playing the sample multimedia file corresponding to the file slice.

[0036] Step 105: Based on the difference between the actual conversion result and the predicted conversion result, adjust the model parameters of the advertising delivery model to obtain the trained advertising delivery model.

[0037] For example, the loss function value is obtained based on the difference between the actual conversion result and the predicted conversion result. If the loss function value is less than or equal to the preset loss function value, the advertising model is considered to have converged, training is stopped, and the advertising model corresponding to the loss function value less than the preset loss function value is identified as the completed advertising model. If the loss function value is greater than the preset loss function value, the model parameters of the advertising model are adjusted, and the model is trained again until the loss function value is less than or equal to the preset loss function value.

[0038] For example, model parameters include the first weight corresponding to the first feature and the second weight corresponding to the second feature. By weighted summing of the first and second features, a fused feature is obtained. This fused feature is input into the prediction sub-model within the ad delivery model to obtain the predicted conversion result. A loss function value is calculated based on the actual and predicted conversion results. If the loss function value is greater than a preset loss function value, the first weight corresponding to the first feature and the second weight corresponding to the second feature are adjusted, and the ad delivery model undergoes another round of training until the loss function value is less than or equal to the preset loss function value.

[0039] For example, after obtaining the trained ad delivery model, you can obtain data to be processed. There are multiple sets of data to be processed. Each set of data to be processed includes: historical operation data of ads on devices associated with the account, user type of logged-in device, device representation, and ad delivery results on the device. By inputting the data to be processed into the trained ad delivery model, you can obtain the predicted conversion results. Based on the conversion results, you can determine which devices to deliver ads on.

[0040] In this embodiment, multiple sets of training samples are obtained. These training samples include: operation data of sample advertisements on sample devices associated with the account, user types logged into the sample devices, device identifiers of the sample devices, and the delivery results of sample advertisements on the sample devices. The labels of the training samples include: the actual conversion results of the sample advertisements on the sample devices. Through the advertising delivery model, first features of multiple sets of training samples and second features corresponding to each user type are extracted. These features reflect the commonalities of different user types, while the second features reflect the unique characteristics corresponding to each user type. The first features are extracted from training samples corresponding to multiple user types, and the second features are extracted from training samples corresponding to one user type. The first and second features are obtained through different networks in the advertising delivery model. The first and second features are fused to obtain fused features, which reflect both the first features of different user types and the second features corresponding to each user type. Based on the fused features, the predicted conversion results of the sample advertisements on the sample devices are determined. According to the difference between the actual conversion results and the predicted conversion results, the model parameters of the advertising delivery model are adjusted to obtain the trained advertising delivery model. This advertising delivery model can predict the conversion results of ads on devices, providing a basis for ad delivery and enabling targeted ad delivery based on the conversion results.

[0041] Figure 2 This is a flowchart illustrating the steps of another advertising placement model training method provided in this application embodiment, see below. Figure 2 The method may include the following steps: Step 201: Obtain multiple sets of training samples.

[0042] The training samples include: operation data of sample ads on sample devices associated with the account, user types logged into the sample devices, device identifiers of the sample devices, and the delivery results of sample ads on the sample devices; the training samples are labeled, and the labels include: the actual conversion results of sample ads on the sample devices.

[0043] It can acquire characteristic data related to ad delivery, such as feature data and other multimedia platform identifiers, of sample devices on advertisers' own multimedia platforms and other multimedia platforms, and construct training samples based on this data. For example, training samples may include operation data of sample ads on sample devices associated with accounts, user types logged into sample devices, device identifiers of sample devices, and the delivery results of sample ads on sample devices.

[0044] Training samples may also include: the ad type of the sample ad, and / or the ad identifier of the sample ad. For example, a sample ad may be a file slice of a sample multimedia file, and the ad type may include video types such as comedy or suspense; or the ad identifier may be the name of the sample multimedia file corresponding to the file slice. For example, a sample ad may be a product ad, and the ad type may include product types such as food or clothing; or the ad identifier may be the product's brand name and product model.

[0045] For example, the sample multimedia file can be in video, audio, or other formats.

[0046] In this embodiment, the training samples also include: the ad type of the sample ad, and / or the ad identifier of the sample ad. Therefore, during the training of the ad delivery model, the model can accurately learn the features of the training samples belonging to the ad type and / or ad identifier, and thus accurately predict the conversion results of ads belonging to the ad type and / or ad identifier. Based on the accurately predicted conversion results, precise ad delivery can be performed. For example, if the sample ad is a file slice of a sample multimedia file, and the training samples include the file type of the file slice (e.g., suspense, drama, etc.), then the ad delivery model can be trained based on the training samples including this file type. The trained ad delivery model can accurately predict the conversion results of suspense file slices, and then precisely deliver suspense multimedia files based on the accurately predicted conversion results.

[0047] In some embodiments of this application, the sample advertisement includes a file slice of a sample multimedia file; the actual conversion result includes at least one of the following: whether the user clicks on the sample multimedia file on the sample device; the playback duration of the sample multimedia file on the sample device; and the revenue generated by the sample multimedia file on the sample device.

[0048] For example, the sample multimedia file can be a video from an advertiser's video platform, and the file slice can be several video segments cut from the video; or, for another example, the sample multimedia file can be audio, and the file slice can be a slice of audio.

[0049] The training samples include operational data for sample ads on sample devices associated with the account; the tags include the actual conversion results of the sample ads on the sample devices. In this embodiment, the sample ads include file slices of sample multimedia files. When the actual conversion result includes whether the user clicks on the sample multimedia file on the sample device, the trained ad delivery model can accurately predict whether the user will play the target multimedia file corresponding to the file slice on the device; when the actual conversion result includes the playback duration of the sample multimedia file on the sample device, the trained ad delivery model can accurately predict the playback duration of the target multimedia file corresponding to the sample slice on the device; when the actual conversion result includes the revenue generated through the sample multimedia file on the sample device, the trained ad delivery model can accurately predict the revenue generated through the target multimedia file corresponding to the file slice on the device.

[0050] Furthermore, when conducting RTA advertising, advertisers typically need to consider the duration of the target multimedia file (e.g., target video) corresponding to the ad, whether the target multimedia file is launched by clicking the ad, and the revenue generated by the target multimedia file corresponding to the ad. Based on these conversion results, the advertisers will conduct advertising. In this embodiment, these conversion results are used as tags to train the advertising model. The trained advertising model can accurately predict these conversion results. Based on these conversion results, the advertising can be precisely targeted to file slices.

[0051] In some embodiments of this application, the advertising delivery model includes a shared network and multiple expert networks, with each expert network corresponding to a user type. Correspondingly, following step 201, the method further includes the following steps: Step 202: Input multiple sets of training samples into the shared network, and use the features extracted by the shared network as the first feature.

[0052] For example, there can be one or more Shared Experts, such as... Figure 3 The advertising delivery model shown includes two shared networks, specifically shared network i and shared network j.

[0053] The first feature is a feature extracted from training samples corresponding to multiple user types; the first feature includes at least one of the following: a feature for characterizing operational data, a feature for characterizing user type, a feature for characterizing device identifier, and a feature for characterizing delivery result.

[0054] Step 203: For any set of training samples, input the training samples into an expert network corresponding to the user type in the training samples, and use the features extracted by the expert network as the second features of the training samples corresponding to the user type.

[0055] In this embodiment of the invention, by inputting training samples with the same dimensions to both the shared network and the expert network, the dimensions of the first and second features output by the shared network and the expert network are made consistent. This facilitates subsequent fusion by weighted summation of the first and second features. Simultaneously, the training samples input to the shared network also include the user type of the logged-in sample device, ensuring that information in the user type dimension contributes to the training of the shared network in the model, thus guaranteeing the training effectiveness of the model.

[0056] The second feature is a feature extracted from training samples corresponding to a user type. In other words, the second feature is a feature extracted from training samples corresponding to a user type, and this feature corresponds to the user type. The second feature includes at least one of the following: a feature for characterizing operational data, a feature for characterizing user type, a feature for characterizing device identification, and a feature for characterizing delivery results.

[0057] For example, both the shared network and the group expert network can be deep neural network (DNN). The shared network is used for training samples across all user types to learn the first features of different user types, while the group expert network is used for its corresponding user type to learn the second features specific to that user type.

[0058] For example, refer to Figure 3 If there are two types of users, then the advertising model includes two expert networks. Expert network i is used to obtain the second feature of the training sample corresponding to user type i, and expert network j is used to obtain the second feature of the training sample corresponding to user type j.

[0059] When training samples are input into the shared network, the features of multiple training samples are aligned. After these training samples are vectorized and entered into the advertising delivery model, all training samples will pass through the advertising delivery model shared network so that the shared network can learn the first features of all training samples and output them in vector form. The vector output by the shared network is the first feature of multiple training samples.

[0060] Shared networks can use all training samples, while expert networks specific to user types can only use training samples specific to user types. Because expert networks can only use training samples specific to user types, expert networks specific to user types can also be called population-specific networks.

[0061] In one embodiment, the training samples include three sets of data: the first set: device 1, user type i, f11, f12, f13, ..., f1n, label 11, label 12, label 13; the second set: device 2, user type j, f21, f22, f23, ..., f2n, label 21, label 22, label 23; the third set: device 3, user type k, f31, f32, f33, ..., f3n, label 31, label 32, label 33.

[0062] By combining professional knowledge, users can be divided into different user types. For example, user type i, user type j, and user type k can represent high-value users, potential users, and churned users, respectively.

[0063] Training samples can include various types of sample data. For example, sample advertisements can be file slices of multimedia files. Training samples can also include operation data on sample advertisements, user attributes and device models, file slice types, and feature data such as the number of days since the user's last active activity. For example, in the aforementioned embodiment, the first type of feature data (f11, f21, and f31) can be user attributes or device models, such as membership attributes or device models; the second type of feature data (f12, f22, and f32) can be the number of times the sample multimedia file (e.g., sample video) corresponding to the sample advertisement was played in the past 3, 7, 14, and 30 days; the third type of feature data (f13, f23, and f33) can be the number of times the sample multimedia file corresponding to the sample advertisement was searched in the past 3, 7, 14, and 30 days; the fourth type of feature data (f14) can be... The first feature data (f15, f25, and f34) can be the number of times the sample advertisement was exposed on the multimedia platform of the sample device in the past 3, 7, 14, and 30 days; the second feature data (f16, f26, and f36) can be the number of times the sample advertisement was clicked on the multimedia platform of the sample device in the past 3, 7, 14, and 30 days; the third feature data (f17, f27, and f37) can be whether the user prefers comedy or suspense videos; the fourth feature data (f17, f27, and f37) can be the number of days since the user's last active activity.

[0064] There can be multiple tags. For example, the first type of tag (tag 11, tag 21 and tag 31) can indicate whether the sample multimedia file corresponding to the sample advertisement is launched on the sample device; the second type of tag (tag 12, tag 22 and tag 32) can indicate the playback duration of the sample multimedia file corresponding to the sample advertisement on the sample device; and the third type of tag (tag 13, tag 23 and tag 33) can indicate the revenue generated by the sample multimedia file corresponding to the sample advertisement on the sample device (e.g., daily revenue).

[0065] By acquiring features in two dimensions—first feature and second feature—and training the ad delivery model based on these two features, the model's prediction performance can be improved.

[0066] In this embodiment, multiple sets of training samples are input into a shared network, and the features extracted by the shared network are used as the first feature. For any set of training samples, the training samples are input into an expert network corresponding to the user type in the training samples, and the features extracted by the expert network are used as the second feature of the training sample corresponding to the user type. Through the shared network and the expert network, the first feature of all training samples and the second feature of the training samples corresponding to each user type can be explicitly separated, improving the accuracy of extracting the first and second features. The advertising delivery model includes a shared network and an expert network. The first feature is obtained through the shared network, and the second feature is obtained through the expert network. The advertising delivery model only needs one forward computation to obtain the first and second features, which can improve the processing efficiency of the advertising delivery model.

[0067] Furthermore, the shared network ensures the effective learning of the advertising delivery model. By using an expert network to obtain second features corresponding to user types, which accurately reflect the differences between user types, the expert network can learn the unique features specific to each user type. Based on the acquired first and second features, more refined and accurate conversion predictions can be achieved. Using these accurate predictions, the precision and overall return on investment of advertising can be effectively improved.

[0068] For example, in an advertising delivery model, the shared network and the expert network run in parallel. By using the shared network and the expert network in parallel, the number of models can be reduced and the generalization ability of the models can be improved.

[0069] In some embodiments of this application, there is at least one actual conversion result, and each actual conversion result has a corresponding conversion result type. The advertising delivery model includes multiple gating networks, and each gating network corresponds one-to-one with a conversion result type. Correspondingly, following step 203, the method may further include the following steps: Step 204: For each type of conversion result, input multiple sets of training samples into the gating network corresponding to the type of conversion result to obtain the first weight value corresponding to the first feature and the second weight value corresponding to the second feature.

[0070] For example, when there are multiple sets of the first feature, each set of the first feature has a corresponding first weight value; when there are multiple sets of the second feature, each set of the second feature has a corresponding second weight value. For instance, the shared network in an ad delivery model may have multiple (e.g., including...) Figure 3In the shared network (i and j), the input data for each shared network consists of multiple sets of training samples, and the output data consists of the first feature; each shared network's corresponding first feature has its own corresponding first weight. There are multiple expert networks (e.g., including...). Figure 3 In the example, expert networks i and j are used. The input data of each expert network is the training samples of the user type corresponding to the shared network, and the output data is the second feature corresponding to the user type. Each second feature has a corresponding second weight.

[0071] There can be multiple gating networks, each corresponding one-to-one with a conversion result type. For example, conversion result types include two categories: whether to launch the sample multimedia file corresponding to the sample ad through the sample ad, and the playback duration of the sample multimedia file corresponding to the sample ad. (See reference...) Figure 3 There are also two gating networks. The conversion result type corresponding to gating network m is whether the sample multimedia file corresponding to the sample ad is launched through the sample ad; the conversion result type corresponding to gating network n is the playback duration of the sample multimedia file corresponding to the sample ad.

[0072] The input data to the gating network consists of all training samples, for example, reference... Figure 3 The training samples for the input gating network include Figure 3 The sparsity and density features of devices, and the user types corresponding to the sample data (i.e., Figure 3 The Group ID in the sample, and whether the sample ad was placed on the sample device.

[0073] Furthermore, the training samples are input into gating networks m and n. The set of weights output by gating network m includes a first weight and a second weight. Specifically, the first weight includes a first weight assigned to the first feature output by shared network i and a first weight assigned to the first feature output by shared network j; the second weight includes a second weight assigned to the second feature output by expert network i and a second weight assigned to the second feature output by expert network j.

[0074] Step 205: Based on the first weight value and the second weight value, perform a weighted summation on the first feature and the second feature to obtain the fusion feature corresponding to the transformation result type.

[0075] For example, refer to Figure 3Based on the first weight assigned to the first feature output by the gating network m, the first weight assigned to the first feature output by the shared network j, the second weight assigned to the second feature output by the expert network i, and the second weight assigned to the second feature output by the expert network j, the first feature output by the shared network i, the first feature output by the shared network j, the second feature output by the expert network i, and the second feature output by the expert network j are weighted and summed to obtain the fusion feature corresponding to the conversion result type (e.g., whether the video corresponding to the sample ad is launched through the sample ad) corresponding to the gating network m.

[0076] In this embodiment, by using a gating network corresponding to the type of transformation result, the first weight value corresponding to the first feature and the second weight value corresponding to the second feature can be adaptively obtained. Compared with the method of manually setting the first and second weight values, the accuracy of the first and second weights learned by the gating network is higher. By weighting and summing the first and second features according to the first and second weights, a more accurate fusion feature can be obtained.

[0077] In some embodiments of this application, the advertising delivery model includes multiple task towers, each corresponding one-to-one with a conversion result type. Correspondingly, following step 205, the method may further include the following steps: Step 206: For each type of conversion result, input the fusion feature corresponding to the conversion result type into the task tower corresponding to the conversion result type to obtain the predicted conversion result corresponding to the conversion result type.

[0078] For example, conversion result types include two categories: whether the sample ad triggers the corresponding sample multimedia file, and the playback duration of the sample ad's corresponding multimedia file. (See reference...) Figure 3 There are also two task towers. The conversion result type corresponding to task tower m is whether the sample multimedia file corresponding to the sample ad is launched through the sample ad; the conversion result type corresponding to task tower n is the playback duration of the sample multimedia file corresponding to the sample ad.

[0079] For example, the task tower can be a deep neural network (DNN), or other network models. (See reference...) Figure 3 The conversion result types corresponding to task tower m and gated network m are the same; the conversion result types corresponding to task tower n and gated network n are also the same. For example, the conversion result type corresponding to gated network m and task tower m is: whether to launch the video corresponding to the sample ad through the sample ad.

[0080] Furthermore, the first weights corresponding to the first features output by shared network m and shared network i, the first weights corresponding to the first features output by shared network j, the second weights corresponding to the second features output by expert network i, and the second weights corresponding to the second features output by expert network j are weighted and summed according to the first and second weights to obtain the fused features. Then, the fused features are input into task tower m, and the conversion result is predicted based on the output of task tower m. The conversion result type of the predicted conversion result is: whether to launch the video corresponding to the sample advertisement through the sample advertisement.

[0081] In this embodiment, the ad delivery model includes multiple task towers, each corresponding to a conversion result type. For each conversion result type, the fusion feature corresponding to that type is input into the corresponding task tower to obtain the predicted conversion result. Thus, through a single processing step, the ad delivery model can obtain predicted conversion results corresponding to multiple conversion result types, achieving multi-task prediction for ad delivery and improving the processing efficiency of the ad delivery model.

[0082] In some embodiments of this application, the method may further include: extracting sparse features from multiple sets of training samples.

[0083] For example, training samples can include various features. In the training samples constructed from collected data, if some features are not collected, or the collected features have a value of zero, and the number of these training samples accounts for a proportion exceeding a preset proportion in the total number of training samples, these features can be called sparse features. After these features are converted into feature vectors, the feature vectors are 0. For example, if the training samples include user membership levels and sample device models, and in 100 sets of collected training samples, most (e.g., 80 sets) of training samples do not include membership levels and sample device models, then membership levels and sample device models are sparse features.

[0084] Features other than sparse features in the training samples are dense features. For example, in 100 sets of training data, if most (e.g., 80 sets) of the training samples have the following features collected and their values ​​are not zero, then these features are dense features in the training samples: the number of sample multimedia files corresponding to the sample advertisements played in the past 3 days, 7 days, 14 days, and 30 days.

[0085] Furthermore, step 206 may include sub-steps A1 to A2: Sub-step A1: Input the sparse features into the factorization machine network, and input the fusion features corresponding to the transformation result type into the task tower corresponding to the transformation result type.

[0086] For example, refer to Figure 3 The sparse features (i.e., device sparse features) are input into the Factorization Machine (FM) network; the fused features are input into the task stack m used to predict whether to launch the video corresponding to the advertisement through the advertisement.

[0087] Sub-step A2: For each type of transformation result, the output of the factorization machine network and the output of the task tower corresponding to the transformation result type are combined to obtain the predicted transformation result corresponding to the transformation result type.

[0088] For example, the conversion result type corresponding to gated network m is whether to launch the sample multimedia file corresponding to the sample ad; the conversion result type corresponding to gated network n is the playback duration of the sample multimedia file corresponding to the sample ad. (See reference...) Figure 3 By fusing the outputs of the FM network and the task tower m, the predicted conversion result is: whether to start the sample multimedia file corresponding to the sample advertisement; by fusing the outputs of the FM network and the task tower n, the predicted conversion result is: the playback duration of the sample multimedia file corresponding to the sample advertisement.

[0089] In this embodiment, inputting sparse features into the FM network can save storage and computing resources during the training process of the advertising delivery model. Furthermore, sparse features are typically features with strong business logic meaning. Inputting sparse features into the FM network and then fusing the output of the FM network with the data results from the task tower allows the advertising delivery model to more directly learn the correlation between sparse features and predicted conversion results.

[0090] In some embodiments of this application, there are multiple actual conversion results, each with a corresponding conversion result type. Correspondingly, following step 206, the method may further include the following steps: Step 207: For each type of conversion result, obtain the loss function value corresponding to the type of conversion result based on the actual conversion result and the predicted conversion result.

[0091] For example, for each conversion result type, the loss function value between the actual conversion result and the predicted conversion result corresponding to the conversion result type is calculated based on a preset loss function.

[0092] Step 208: Combine the loss function values ​​corresponding to each type of transformation result to obtain the combined loss function value.

[0093] For example, the loss function values ​​corresponding to each type of transformation result can be summed directly to obtain the fusion loss function value.

[0094] For example, there are two conversion result types, specifically: whether the sample multimedia file corresponding to the sample advertisement is launched on the sample device, and the playback duration of the sample multimedia file corresponding to the sample advertisement. The loss function values ​​corresponding to these two conversion result types are Loss launch and Loss playback, respectively. The two loss function values ​​are summed to obtain the fused loss function value TotalLoss = Loss launch + Loss playback.

[0095] Step 209: Adjust the model parameters of the advertising delivery model based on the difference between the fusion loss function value and the preset loss function threshold to obtain the trained advertising delivery model.

[0096] For example, training an ad delivery model involves multiple iterations. If, after the current iteration, the fusion loss function value is less than or equal to a preset loss function threshold, the iteration stops, and the ad delivery model obtained in the current iteration is identified as the trained ad delivery model. If, after the current iteration, the fusion loss function value is greater than the preset loss function threshold, the next iteration begins, and the model parameters of the ad delivery model are adjusted until the fusion loss function value is less than or equal to the preset loss function value. Then, the iteration stops, and the ad delivery model that satisfies the condition that the fusion loss function value is less than or equal to the preset loss function value is identified as the trained ad delivery model.

[0097] In this embodiment, for each conversion result type, a loss function value corresponding to the actual conversion result and the predicted conversion result is obtained. These loss function values ​​are then fused together to obtain a fused loss function value. The fused loss function value accurately reflects the prediction accuracy of the predicted conversion results for multiple conversion result types. Based on the difference between the fused loss function value and a preset loss function threshold, the model parameters of the advertising delivery model are adjusted to obtain a trained advertising delivery model. This trained model can accurately predict various conversion results for advertising.

[0098] Reference Figure 4 This application also provides an advertising delivery method, which may include the following steps: Step 301: Obtain multiple sets of data to be processed.

[0099] The data to be processed includes: historical data on advertising operations on devices associated with the account, user types of logged-in devices, device identifiers of devices, and preset advertising delivery results on devices.

[0100] The historical operation data has the same data type as the operation data in the training samples used to train the ad delivery model in the aforementioned embodiments. For example, the operation data in the training samples includes the number of times a user has viewed the sample multimedia file corresponding to the sample ad on the sample device in the past 3, 7, 14, and 30 days, while the historical operation data includes the number of times a user has viewed the target multimedia file corresponding to the ad on the device in the past 3, 7, 14, and 30 days.

[0101] Step 302: Input multiple sets of data to be processed into the trained advertising delivery model to obtain the conversion results corresponding to each device.

[0102] The trained advertising delivery model can be generated using any of the advertising delivery model training methods described in the preceding embodiments, which will not be elaborated here.

[0103] In some embodiments of this application, the advertisement includes a file slice of the target multimedia file; the conversion result includes at least one of the following: whether the user clicks on the target multimedia file on the device; the playback duration of the target multimedia file on the device; and the revenue generated by the target multimedia file on the device.

[0104] In practical applications, when an advertisement includes a file slice of a target multimedia file, advertisers are more concerned with whether users click on the target multimedia file on their devices; the playback duration of the target multimedia file on the user's device; and the revenue generated by the target multimedia file on the device. Based on this embodiment, the conversion results of the aforementioned conversion result types are obtained. Based on these conversion results, the delivery of the target multimedia file corresponding to the file slice can be accurately realized.

[0105] Step 303: Based on the conversion results corresponding to each device, determine the devices to be deployed for the advertisement from multiple devices.

[0106] For example, conversion results are positively correlated with the effectiveness of ad placement. From multiple devices, several devices with higher conversion results can be identified as the devices to be targeted, and the ads can be placed on these devices.

[0107] Specifically, a mapping relationship can be established between advertisements and selected devices for delivery of those advertisements. In one implementation, upon receiving an advertisement retrieval request from a device, the advertisement corresponding to that device in the mapping relationship is sent to that device, thereby delivering the advertisement to the corresponding device. In another implementation, the advertisement can be directly pushed to each device, or the advertisement can be pushed to an edge device associated with the device (e.g., an edge node of a distribution network), which then pushes the advertisement to the device when push conditions are met. The push conditions can be set as needed, and this embodiment of the invention does not limit this. For example, push conditions could include the current time reaching a preset time, receiving an advertisement retrieval request from a device, or the multimedia platform on the device being launched, etc.

[0108] For example, if a device to be advertised has multiple multimedia platform software installed, the advertisement can be advertised to all the multimedia platform software on the device, or it can be advertised to some of the multimedia platform software.

[0109] For example, if a sample device has multiple multimedia platforms installed, the training samples used to train the advertising delivery model include: platform information of the multimedia platforms that deliver sample ads on the sample device; then, the advertising delivery model trained based on the training samples can predict the conversion results corresponding to each multimedia software platform on the device; after determining the device to be delivered, the target multimedia platform can be determined from multiple multimedia platforms based on the comparison results of the conversion results of each multimedia platform on the device to be delivered, so as to deliver ads on the target multimedia platform.

[0110] In this embodiment, multiple sets of data to be processed are acquired. These data include: historical operation data of advertisements on devices associated with the account, the user type of the logged-in device, the device identifier of the device, and the preset delivery results of the advertisement on the device. The multiple sets of data are input into a trained advertisement delivery model to obtain the conversion results corresponding to each device. The trained advertisement delivery model is obtained through the first features of multiple training samples and the second features of training samples corresponding to each user type. The first and second features are obtained through different networks in the advertisement delivery model. The fused features accurately reflect the first features of all training samples and the second features of training samples corresponding to each user type. Based on the fused features, the trained advertisement delivery model is obtained, and the trained advertisement delivery model can predict the conversion results corresponding to each device. Using the conversion results corresponding to each device as a basis, the devices to be delivered to for advertisements can be specifically determined from multiple devices, thereby improving the accuracy of advertisement delivery to a certain extent.

[0111] In some embodiments of this application, reference is made to Figure 5 Advertising placement methods may include the following steps: Step 401: Obtain multiple sets of data to be processed.

[0112] The data to be processed includes: historical operation data of advertisements on devices associated with the account, user type of logged-in device, device identifier of device, and preset delivery results of advertisements on device.

[0113] In some embodiments of this application, the data to be processed further includes: the advertisement type of the advertisement; and / or, the advertisement identifier of the advertisement.

[0114] In this embodiment, if the data to be processed also includes the ad type of the advertisement, the data to be processed is input into the trained ad delivery model. The trained ad delivery model can predict the conversion result corresponding to the ad type of the advertisement, and then achieve precise ad delivery based on the conversion result. If the data to be processed also includes the ad identifier of the advertisement, the data to be processed is input into the trained ad delivery model. The trained ad delivery model can predict the conversion result corresponding to the ad identifier of the advertisement, and then achieve precise ad delivery based on the conversion result.

[0115] Step 402: Set the preset delivery result to deliver the advertisement on the device, and input the data to be processed into the trained advertisement delivery model to obtain the first conversion result corresponding to each device.

[0116] In this step, the data to be processed includes: historical operation data of advertisements on the device associated with the account, user type of the logged-in device, device identifier of the device, and preset delivery results of advertisements on the device; and the preset delivery result is: to deliver advertisements on the device.

[0117] Step 403: Set the preset delivery result to not deliver ads on the device, and input the data to be processed into the trained ad delivery model to obtain the second conversion result corresponding to each device.

[0118] In this step, the advertisements to be processed include: historical operation data of advertisements on the device associated with the account, user type of the logged-in device, device identifier of the device, and preset delivery results of the advertisements on the device; and the preset delivery result is: no advertisements are delivered on the device.

[0119] The difference between the pending data used to obtain the first conversion result and the pending data used to obtain the second conversion result is as follows: In the pending data used to obtain the first conversion result, the preset delivery result is: to deliver the advertisement on the device; in the pending data used to obtain the second conversion result, the preset delivery result is: not to deliver the advertisement on the device.

[0120] Step 404: For each device, obtain the difference between the first conversion result and the second conversion result corresponding to the device.

[0121] The first conversion result reflects the conversion outcome after advertising is placed on the device, while the second conversion result reflects the conversion outcome without advertising on the device. The difference between the first and second conversion results reflects the difference in conversion outcomes between advertising on and without advertising on the device, and this difference can characterize the conversion gain value score generated by the advertising.

[0122] For example, if there are multiple first transformation results and multiple second transformation results, step 404 may include the following sub-steps: Sub-step B1: For each device, fuse multiple first transformation results corresponding to the device to obtain a first fused transformation result corresponding to the device, and fuse multiple second transformation results corresponding to the device to obtain a second fused transformation result corresponding to the device.

[0123] For example, the first conversion result includes: whether the user clicked on the target multimedia file on the device, and the playback duration of the target multimedia file on the device. These two first conversion results can be weighted and summed to obtain the first fused conversion result corresponding to the device, or the two first conversion results can be directly summed to obtain the first fused conversion result corresponding to the device.

[0124] For example, the conversion result types of the first conversion result include: the playback duration of the target multimedia file on the device, and the revenue generated by the target multimedia file on the device.

[0125] Multiple primary conversion results can be normalized and then merged; alternatively, multiple primary conversion results can be merged directly. In direct merging, the playback duration level (e.g., level 1, level 2, level 3, level 4, level 5) can be obtained based on the playback duration and the correspondence between preset playback duration ranges and preset duration levels. The revenue level (e.g., level 1, level 2, level 3, level 4, level 5) can also be obtained based on the revenue and the correspondence between preset revenue ranges and preset revenue levels. Then, the playback duration level and revenue level can be directly merged.

[0126] The method for fusing multiple second transformation results can refer to the method for fusing multiple first transformation results described above, and will not be repeated here.

[0127] Sub-step B2: The difference between the first fusion conversion result and the second fusion conversion result corresponding to the device is determined as the difference between the first conversion result and the second conversion result corresponding to the device.

[0128] In this embodiment, the first fusion conversion result reflects the overall conversion result when advertising is placed on the device, and the second fusion conversion result reflects the overall conversion result when advertising is not placed on the device. The difference between the first and second fusion conversion results accurately reflects the gain value between placing advertising on the device and not placing advertising. The difference between the first and second fusion conversion results corresponding to the device is determined as the difference between the first and second conversion results corresponding to the device. Subsequently, based on the difference between the first and second conversion results corresponding to the device, the device to be used for advertising can be accurately determined.

[0129] Step 405: Based on the difference, determine the devices to be used for advertising from multiple devices.

[0130] In this embodiment, the trained ad delivery model includes a shared network, multiple expert networks, multiple gating networks, and multiple task towers. Each expert network corresponds one-to-one with a user type, each gating network corresponds one-to-one with a conversion result type, and each task tower corresponds one-to-one with a conversion result type.

[0131] The data to be processed includes the preset delivery results of the advertisement. By setting the preset delivery results to either "delivering advertisements on the device" or "not delivering advertisements", two types of data to be processed with different preset delivery results can be obtained.

[0132] For each preset delivery result, the data to be processed is input into the trained ad delivery model. The shared network extracts the first feature of all data to be processed, and the expert network extracts the second feature corresponding to each user type. For each conversion result type, an adaptive gating network fuses the features output by the shared network and the expert network. The fused features are then input into the corresponding task tower, and the training samples are input into the FM network. The outputs of the task tower and the FM network are fused to obtain the predicted conversion result. There are multiple task towers, each corresponding to a conversion result type. Based on multiple task towers, multi-task prediction can be achieved, resulting in multiple conversion results. Multiple conversion results are fused to obtain a fused conversion result corresponding to the preset delivery result. Specifically, this includes a first fused conversion result when ads are delivered on the device, and a second fused conversion result when ads are not delivered on the device. The difference between the first and second fused conversion results is obtained. This difference reflects the ad gain value score. The results are sorted according to the gain value score, and the sorting results accurately determine the devices to be delivered to.

[0133] In this embodiment, the preset delivery result is set to deliver ads on devices. Then, the data to be processed is input into the trained ad delivery model to obtain a first conversion result corresponding to each device. Alternatively, the preset delivery result is set to not deliver ads on devices. Then, the data to be processed is input into the trained ad delivery model to obtain a second conversion result corresponding to each device. For each device, the difference between the first and second conversion results is obtained. The first conversion result is the conversion result when ads are delivered on the device, and the second conversion result is the conversion result when ads are not delivered on the device. The difference between the first and second conversion results characterizes the difference in conversion results between delivering ads on devices and not delivering ads, accurately reflecting the gain value score of delivering ads on devices. When delivering ads, advertisers focus more on the gain value score generated by the ad delivery than on the conversion result of the product corresponding to the ad after delivery. Therefore, based on the difference between the first and second conversion results, the devices to be delivered to the ad can be accurately determined from multiple devices.

[0134] For example, step 405, which determines the devices to be used for advertising from multiple devices based on the difference, may include the following sub-steps: Sub-step C1: Sort the differences to obtain the sorted results.

[0135] Obtain the difference between the first and second conversion results corresponding to each device. This difference represents the gain value score generated by the advertisement on the device. Sort the differences, which is equivalent to sorting the ad gain value score generated by the advertisement on each device.

[0136] Sub-step C2: Based on the sorting results, determine the devices to be used for advertising from multiple devices.

[0137] The difference between the first conversion result and the second conversion result can accurately reflect the gain value score generated by the advertisement on the device. The higher the gain value score, the more obvious the effect of the advertisement. Therefore, the difference between the first conversion result and the second conversion result corresponding to the device is positively correlated with the effect of the advertisement.

[0138] Furthermore, based on the sorting results, several target differences with larger values ​​can be selected from the differences, and the devices corresponding to the target differences can be identified as the devices to be deployed.

[0139] For example, in one embodiment, the advertisement is a file slice of a multimedia file. Both the first and second conversion results are the duration of playback of the target multimedia file corresponding to the advertisement on the device. Specifically, after the advertisement is placed on device 1, the first conversion result is 60 minutes, while the second conversion result is 30 minutes if the advertisement is not placed on device 1. Similarly, after the advertisement is placed on device 2, the first conversion result is 80 minutes, while the second conversion result is 70 minutes if the advertisement is not placed on device 1. In this embodiment, after the advertisements are placed on devices 1 and 2, the playback duration of the target multimedia file on device 1 (60 minutes) is shorter than that on device 2 (80 minutes). However, the difference in playback duration between placing and not placing an advertisement on device 1 is 30 minutes, while the difference is 10 minutes. That is, placing an advertisement on device 1 yields a greater gain value; therefore, in subsequent processing, device 1 is identified as the device to be advertised.

[0140] In this embodiment, the differences between the first conversion result and the second conversion result corresponding to the device are sorted to obtain a sorting result. The difference between the first conversion result and the second conversion result can accurately reflect the gain value score generated by advertising on the device. The sorting result of the difference is equivalent to the sorting result of the gain value score generated by advertising on the device, and also equivalent to the conversion result generated by advertising. In practical applications, advertisers pay more attention to the gain value score generated by advertising. Therefore, based on the sorting result, the device to be advertised can be accurately determined from multiple devices to improve the accuracy of advertising.

[0141] When ads are file slices of multimedia files (such as video, audio, etc.), advertisers need to consider the conversion results when placing Real-Time Application Programming Interface (RTA) ads, as there are differences in conversion results between different user types.

[0142] An advertising evaluation model can be established for each user group. Based on this model, it can be used to assess whether to target ads to users of that specific user type. For example, refer to... Figure 6 This approach involves building model a for user group 'a' and model b for user group 'b'. Training samples from user group 'a' are input into model a to obtain predicted conversion results for user group 'a', and training samples from user group 'b' are input into model b to obtain predicted conversion results for user group 'b'. Ads are then targeted based on these predicted conversion results. However, this method ignores the commonalities between different user groups, leading to low accuracy in ad targeting based on predicted conversion results. Furthermore, this method requires building a complete predictive model for each user group, resulting in high maintenance costs for the ad models.

[0143] Alternatively, it can be used without differentiating user groups, employing a general advertising evaluation model to assess whether to target ads to users. For example, referring to... Figure 7 A shared model is built for user group A and user group B to predict conversion results for both groups. However, this method struggles to capture the deep, unique attributes between different groups and limits the model's generalization ability. The model's conversion prediction accuracy is low, resulting in low precision in ad targeting based on conversion results.

[0144] In this embodiment, due to the significant differences between different user groups, refer to Figure 8 In this embodiment, an advertising delivery model is constructed that includes a shared network and an expert network, wherein the shared network model and the expert network operate in parallel. The expert network is designed as a user group-specific network for each user type, based on the shared network that enables multi-task sharing.

[0145] For example, the ad delivery model includes a multi-task network, such as an MMOE network or a PLE network, which converts training samples into vectors and inputs them into the multi-task network to obtain the output results of the multi-task network.

[0146] It can acquire feature data related to sample ads on the advertiser's own multimedia platform within the sample device, as well as feature data related to sample ads on other multimedia platforms, platform identifiers of other multimedia platforms, and the placement scenarios of sample ads, and construct training samples based on this data. For example, it can use labels such as whether the sample multimedia file corresponding to the sample ad is launched on the sample device, the duration of playback of the sample multimedia file corresponding to the sample ad on the device, and the revenue generated by the sample multimedia file corresponding to the sample ad on the sample device.

[0147] For example, we can obtain conversion result characteristics from multiple multimedia platforms on sample devices, as well as feature data such as exposure rate and click-through rate of sample ads on other multimedia platforms besides the advertiser's. Based on this feature data, we can construct behavioral characteristics of sample devices on the advertiser's multimedia platforms over short and long periods (e.g., 3 days, 7 days, 14 days, and 30 days), and conversion results of clicking on sample ads and the corresponding products on other multimedia platforms. Training samples can then be constructed based on this acquired data.

[0148] For example, samples can be collected starting from the ad delivery dimension. User exposure can be used as training samples, and user deep conversion behavior can be used as labels. Deep conversion behavior refers to the actual conversion results in the aforementioned embodiments. After obtaining the training samples, the ad delivery model can be trained using the training samples to obtain the trained ad delivery model.

[0149] For example, in the field of long videos, there are many video and text features. These video and text features can be used as part of the training samples to introduce them into the advertising delivery model of this embodiment, thereby increasing the accuracy of the model's prediction results.

[0150] Significant behavioral differences exist between different user types (e.g., high-value users, potential users, churned users, etc.). In this embodiment, users can be categorized into different user types based on business knowledge. Each user type corresponds to a core user group. A shared network learns the common attributes of training samples for all user types, while an expert network focuses on learning the deeper and more specific information of training samples for its corresponding user type, thus paying attention to the common characteristics and individual differences among different user types. For example, expert network i learns the deeper and more specific information of training samples for user type i, and expert network j learns the deeper and more specific information of training samples for user type j.

[0151] Reference Figure 3The advertising delivery model in this embodiment also includes an adaptive gating network (including gating network m and gating network n). During model training, after the training samples are converted into feature vectors, the feature vectors are input into the gating network. Through the gating network, the first weights corresponding to each first feature and the second weights corresponding to each second feature can be dynamically output. The advertising delivery model can then fuse the first and second features output by the shared network and the expert network through the first and second weights. The fused features are then input into the multi-task tower to predict various conversion results. The model parameters of the advertising delivery model are adjusted according to the predicted conversion results to obtain the trained advertising delivery model.

[0152] Furthermore, in this embodiment, the output of the adaptive gating network is the input of the multi-task tower. Based on the output of the multi-task tower and the output of the FM network, a conversion result of the type corresponding to the task tower is obtained. Further, a gain value score is calculated based on the first conversion result when advertising is being performed and the second conversion result when advertising is not being performed.

[0153] For example, a graph neural network can be introduced into the multi-scenario, multi-task advertising delivery model of this embodiment to increase the ability of the advertising delivery model to extract information from training samples, increase the accuracy of obtaining the first feature and the second feature, and thus increase the prediction accuracy of the advertising delivery model.

[0154] In this embodiment, the trained advertising delivery model is a multi-task gain value scoring model for different user groups. During the advertising delivery decision-making process using this model, the model predicts a first conversion result (delivering an ad on the device) and a second conversion result (not delivering an ad). The difference between the first and second conversion results is then calculated; this difference is the gain value score corresponding to the device. Specifically, according to steps 402 to 404 and sub-steps B1 to B2, the difference between the first and second conversion results corresponding to the device is obtained, and this difference is the gain value score corresponding to the device; further details are omitted here.

[0155] After obtaining the gain value score, the devices to be targeted for advertising are determined from multiple devices based on the ranking results of the gain value score. Specifically, the RTA advertising targeting decision is made according to the aforementioned step 405, as well as sub-steps C1 and C2, which will not be elaborated here.

[0156] Reference Figure 9This application embodiment also provides an advertising delivery model training device. The device 40 includes: a first acquisition module 501, used to acquire multiple sets of training samples; the training samples include: operation data of sample advertisements in sample devices associated with an account, user type of the logged-in sample device, device identifier of the sample device, and delivery results of sample advertisements in the sample device; the training samples have labels, the labels include: the actual conversion result of the product corresponding to the sample advertisement in the sample device; a feature extraction module 502, used to extract first features of multiple sets of training samples and second features of training samples corresponding to each user type through the advertising delivery model; the first features are features extracted based on training samples corresponding to multiple user types, and the second features are features extracted based on training samples corresponding to one user type, the first features and the second features are obtained through different networks in the advertising delivery model; a feature fusion module 503, used to fuse the first features and the second features to obtain fused features; a prediction model 504, used to determine the predicted conversion result of the sample advertisement in the sample device based on the fused features; and a parameter adjustment module 505, used to adjust the model parameters of the advertising delivery model according to the difference between the actual conversion result and the predicted conversion result to obtain the trained advertising delivery model.

[0157] In some embodiments of this application, the advertising delivery model includes a shared network and multiple expert networks, with each expert network corresponding to a user type. The feature extraction module 502 may include: a first acquisition submodule, used to input multiple sets of training samples into the shared network and use the features extracted by the shared network as a first feature; and a second acquisition submodule, used to input any set of training samples into the expert network corresponding to the user type in the training samples and use the features extracted by the expert network as a second feature corresponding to the user type.

[0158] Optionally, the actual conversion result includes at least one type, and each actual conversion result has a corresponding conversion result type; the advertising delivery model includes multiple gated networks, and the gated networks correspond one-to-one with the conversion result types; the feature fusion module 503 may include: a third acquisition submodule, used to input multiple sets of training samples into the gated network corresponding to the conversion result type for each conversion result type to obtain a first weight value corresponding to the first feature and a second weight value corresponding to the second feature; a fourth acquisition submodule, used to perform a weighted summation of the first feature and the second feature based on the first weight value and the second weight value to obtain a fused feature corresponding to the conversion result type.

[0159] Optionally, the advertising delivery model includes multiple task towers, each corresponding to a conversion result type; the prediction model 504 may include: a fifth acquisition submodule, used to input the fusion features corresponding to each conversion result type into the task tower corresponding to the conversion result type to obtain the predicted conversion result corresponding to the conversion result type.

[0160] Optionally, the apparatus further includes: an extraction module for extracting sparse features from multiple sets of training samples; the fifth acquisition submodule may include: a first processing unit for inputting sparse features into a factorization machine network and inputting fused features corresponding to the transformation result type into a task tower corresponding to the transformation result type; and a fusion unit for fusing the output of the factorization machine network and the output of the task tower corresponding to the transformation result type for each transformation result type to obtain a predicted transformation result corresponding to the transformation result type.

[0161] Optionally, there are multiple actual conversion results, each with a corresponding conversion result type. The parameter adjustment module may include: a sixth acquisition submodule, used to obtain the loss function value corresponding to each conversion result type based on the actual conversion result and the predicted conversion result; a seventh acquisition submodule, used to fuse the loss function values ​​corresponding to each conversion result type to obtain a fused loss function value; and a parameter adjustment submodule, used to adjust the model parameters of the advertising delivery model based on the difference between the fused loss function value and the preset loss function threshold to obtain the trained advertising delivery model.

[0162] Optionally, the sample advertisement includes a file slice of the sample multimedia file; the actual conversion result includes at least one of the following: whether the user clicks on the sample multimedia file on the sample device; the playback duration of the sample multimedia file on the sample device; and the revenue generated by the sample multimedia file on the sample device.

[0163] Based on this embodiment, by obtaining the first feature and the second feature, and fusing the first feature and the second feature to obtain the fused feature, the trained advertising delivery model is obtained based on the fused feature. The trained advertising delivery model can predict the conversion results of advertisements on devices, providing a basis for advertising delivery, so that subsequent targeted advertising can be carried out based on the advertising conversion results.

[0164] Reference Figure 10This embodiment also provides an advertising delivery device. The device 60 includes: a second acquisition module 601, used to acquire multiple sets of data to be processed; the data to be processed includes: historical operation data of advertisements in devices associated with the account, user type of the logged-in device, device identifier of the device, and preset delivery results of the advertisement in the device; a third acquisition module 602, used to input the multiple sets of data to be processed into a trained advertising delivery model to obtain conversion results corresponding to each device; the advertising delivery model is generated by the advertising delivery model training method of any of the above embodiments; and a first determination module 603, used to determine the device to be delivered for the advertisement from multiple devices according to the conversion results corresponding to each device.

[0165] The third acquisition module 602 may include: an eighth acquisition submodule, used to set the preset delivery result to deliver the advertisement on the device, and input the data to be processed into the trained advertisement delivery model to obtain a first conversion result corresponding to each device; a ninth acquisition submodule, used to set the preset delivery result to not deliver the advertisement on the device, and input the data to be processed into the trained advertisement delivery model to obtain a second conversion result corresponding to each device; the first determination module 603 may include: a tenth acquisition submodule, used to acquire the difference between the first conversion result and the second conversion result corresponding to each device; and a determination submodule, used to determine the device to be delivered for the advertisement from multiple devices based on the difference.

[0166] In some embodiments of this application, when there are multiple first conversion results and multiple second conversion results, the tenth acquisition submodule may include: a fusion unit, configured to fuse multiple first conversion results corresponding to each device to obtain a first fused conversion result corresponding to the device, and fuse multiple second conversion results corresponding to the device to obtain a second fused conversion result corresponding to the device; and a first determination unit, configured to determine the difference between the first fused conversion result and the second fused conversion result corresponding to the device as the difference between the first conversion result and the second conversion result corresponding to the device.

[0167] In some embodiments of this application, the first determining submodule may include: a sorting unit for sorting the differences to obtain a sorting result; and a second determining unit for determining the device to be deployed for the advertisement from multiple devices based on the sorting result.

[0168] In some embodiments of this application, the advertisement includes a file slice of the target multimedia file; the conversion result includes at least one of the following: whether the user clicks on the target multimedia file on the device; the playback duration of the target multimedia file on the device; and the revenue generated by the target multimedia file on the device.

[0169] Based on this embodiment, the conversion result corresponding to each device can be predicted, and then the devices to be used for advertising can be determined in a targeted manner according to the conversion result corresponding to each device. The method has the advantage of high advertising accuracy.

[0170] This invention also provides an electronic device, such as... Figure 11 As shown, it includes a processor 701, a communication interface 702, a memory 703, and a communication bus 704, wherein the processor 701, the communication interface 702, and the memory 703 communicate with each other through the communication bus 704. Memory 703 is used to store computer programs; When processor 701 executes a program stored in memory 703, it performs the following steps: Multiple training samples are obtained; the training samples include: operation data of sample ads on sample devices associated with the account, user types logged into the sample devices, device identifiers of the sample devices, and the delivery results of sample ads on the sample devices; the training samples are labeled, and the labels include: the actual conversion results of the product corresponding to the sample ad on the sample device; through the ad delivery model, the first feature of the multiple training samples and the second feature of the training samples corresponding to each user type are extracted; the first feature is the feature extracted based on the training samples corresponding to multiple user types, and the second feature is the feature extracted based on the training samples corresponding to one user type. The first feature and the second feature are obtained through different networks in the ad delivery model; the first feature and the second feature are fused to obtain the fused feature; based on the fused feature, the predicted conversion result of the sample ad on the sample device is determined; according to the difference between the actual conversion result and the predicted conversion result, the model parameters of the ad delivery model are adjusted to obtain the trained ad delivery model.

[0171] In some embodiments of this application, multiple sets of data to be processed are obtained; the data to be processed includes: historical operation data of advertisements on devices associated with the account, user type of the logged-in device, device identifier of the device, and preset delivery results of advertisements on the device; the multiple sets of data to be processed are input into the trained advertisement delivery model to obtain the conversion results corresponding to each device; the advertisement delivery model is generated by the advertisement delivery model training method in any of the above embodiments; based on the conversion results corresponding to each device, the device to be delivered to the advertisement is determined from multiple devices.

[0172] 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.

[0173] The communication interface is used for communication between the aforementioned terminal and other devices.

[0174] 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.

[0175] 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.

[0176] In another embodiment of the present invention, a computer-readable storage medium is also provided, which stores instructions that, when executed on a computer, cause the computer to perform any of the advertising delivery model training methods or advertising delivery methods described in the above embodiments.

[0177] In another embodiment of the present invention, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the advertising delivery model training methods or advertising delivery methods in the above embodiments.

[0178] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A 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 flow or function according to the embodiments of the present invention is 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, 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., a solid-state disk (SSD)).

[0179] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, 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. Without further limitations, 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 the element.

[0180] 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.

[0181] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for training an advertising delivery model, characterized in that, include: Obtain multiple sets of training samples; The training samples include: operation data of sample advertisements in sample devices associated with the account, user types logged into the sample devices, device identifiers of the sample devices, and the delivery results of the sample advertisements in the sample devices; the training samples have tags, and the tags include: the actual conversion results of the products corresponding to the sample advertisements in the sample devices; The advertising delivery model extracts first features from multiple sets of training samples and second features from training samples corresponding to each user type. The first features are features extracted from training samples corresponding to multiple user types, and the second features are features extracted from training samples corresponding to one user type. The first features and the second features are obtained through different networks in the advertising delivery model. The first feature and the second feature are combined to obtain the fused feature; Based on the fusion features, the predicted conversion result of the sample advertisement in the sample device is determined; Based on the difference between the actual conversion result and the predicted conversion result, the model parameters of the advertising delivery model are adjusted to obtain the trained advertising delivery model.

2. The method according to claim 1, characterized in that, The advertising delivery model includes a shared network and multiple expert networks, each expert network corresponding one-to-one with a user type; the step of extracting first features from multiple sets of training samples and second features from training samples corresponding to each user type through the advertising delivery model includes: Multiple sets of training samples are input into the shared network, and the features extracted by the shared network are used as the first feature. For any set of training samples, the training samples are input into an expert network corresponding to the user type in the training samples, and the features extracted by the expert network are used as the second features of the training samples corresponding to the user type.

3. The method according to claim 1, characterized in that, The actual conversion result includes at least one type, and each actual conversion result has a corresponding conversion result type; the advertising delivery model includes multiple gate networks, and each gate network corresponds one-to-one with the conversion result type; The fusion of the first feature and the second feature to obtain the fused feature includes: For each type of conversion result, multiple sets of training samples are input into the gating network corresponding to the type of conversion result to obtain a first weight value corresponding to the first feature and a second weight value corresponding to the second feature; Based on the first weight value and the second weight value, the first feature and the second feature are weighted and summed to obtain the fusion feature corresponding to the transformation result type.

4. The method according to claim 3, characterized in that, The advertising delivery model includes multiple task towers, each corresponding one-to-one with a conversion result type; determining the predicted conversion result of the sample advertisement on the sample device based on the fusion features includes: For each type of conversion result, the fusion feature corresponding to the conversion result type is input into the task tower corresponding to the conversion result type to obtain the predicted conversion result corresponding to the conversion result type.

5. The method according to claim 4, characterized in that, The method further includes: Extract sparse features from multiple sets of training samples; For each of the conversion result types, the fusion feature corresponding to the conversion result type is input into the task tower corresponding to the conversion result type to obtain the predicted conversion result corresponding to the conversion result type, including: The sparse features are input into the factorization machine network, and the fusion features corresponding to the transformation result type are input into the task tower corresponding to the transformation result type. For each type of transformation result, the output of the factorization machine network and the output of the task tower corresponding to the transformation result type are fused to obtain the predicted transformation result corresponding to the transformation result type.

6. The method according to claim 1, characterized in that, There are multiple actual conversion results, each with a corresponding conversion result type; adjusting the model parameters of the advertising delivery model based on the difference between the actual conversion results and the predicted conversion results to obtain the trained advertising delivery model includes: For each type of conversion result, the loss function value corresponding to the conversion result type is obtained based on the actual conversion result and the predicted conversion result corresponding to the conversion result type; The loss function values ​​corresponding to each of the aforementioned transformation result types are combined to obtain the fusion loss function value; Based on the difference between the fusion loss function value and the preset loss function threshold, the model parameters of the advertising delivery model are adjusted to obtain the trained advertising delivery model.

7. The method according to any one of claims 1 to 6, characterized in that, The sample advertisement includes file slices of sample multimedia files; The actual conversion results include at least one of the following: whether the user clicks on the sample multimedia file on the sample device; the playback duration of the sample multimedia file on the sample device; and the revenue generated by the sample multimedia file on the sample device.

8. An advertising placement method, characterized in that, include: Acquire multiple sets of data to be processed; The data to be processed includes: historical operation data of the advertisement on the device associated with the account, the user type logged into the device, the device identifier of the device, and the preset delivery results of the advertisement on the device; Multiple sets of the data to be processed are input into the trained advertising delivery model to obtain conversion results corresponding to each device; the advertising delivery model is generated by the advertising delivery model training method described in any one of claims 1-7. Based on the conversion results corresponding to each of the devices, the devices to be deployed for the advertisement are determined from among the multiple devices.

9. The method according to claim 8, characterized in that, The step of inputting multiple sets of the data to be processed into the trained advertising delivery model to obtain the conversion result corresponding to each device includes: The preset delivery result is set to deliver the advertisement on the device, and the data to be processed is input into the trained advertisement delivery model to obtain the first conversion result corresponding to each device. The preset delivery result is set to not deliver the advertisement on the device, and the data to be processed is input into the trained advertisement delivery model to obtain a second conversion result corresponding to each device. The step of determining the device to be deployed for the advertisement from a plurality of devices based on the conversion results corresponding to each of the devices includes: For each of the devices, the difference between the first conversion result and the second conversion result corresponding to the device is obtained; Based on the difference, the device to be used for the advertisement is determined from among the multiple devices.

10. The method according to claim 9, characterized in that, When there are multiple first conversion results and multiple second conversion results, obtaining the difference between the first conversion result and the second conversion result corresponding to each device includes: For each device, multiple first conversion results corresponding to the device are fused to obtain a first fused conversion result corresponding to the device, and multiple second conversion results corresponding to the device are fused to obtain a second fused conversion result corresponding to the device; The difference between the first fusion conversion result and the second fusion conversion result corresponding to the device is determined as the difference between the first conversion result and the second conversion result corresponding to the device.

11. The method according to claim 9, characterized in that, The step of determining the device to be used for advertising from among the multiple devices based on the difference includes: The differences are sorted to obtain the sorting results; Based on the sorting results, the devices to be used for the advertisement are determined from among the multiple devices.

12. The method according to any one of claims 8 to 11, characterized in that, The advertisement includes file slices of the target multimedia file; The conversion result includes at least one of the following: whether the user clicks on the target multimedia file on the device; the playback duration of the target multimedia file on the device; and the revenue generated by the target multimedia file on the device.

13. An advertising placement model training device, characterized in that, include: The first acquisition module is used to acquire multiple sets of training samples; The training samples include: operation data of sample advertisements in sample devices associated with the account, user types logged into the sample devices, device identifiers of the sample devices, and the delivery results of the sample advertisements in the sample devices; the training samples have tags, and the tags include: the actual conversion results of the products corresponding to the sample advertisements in the sample devices; The feature extraction module is used to extract first features of multiple sets of training samples and second features of training samples corresponding to each user type through an advertising delivery model; the first features are features extracted based on training samples corresponding to multiple user types, and the second features are features extracted based on training samples corresponding to one user type; the first features and the second features are obtained through different networks in the advertising delivery model. The feature fusion module is used to fuse the first feature and the second feature to obtain a fused feature; A predictive model is used to determine the predicted conversion result of the sample advertisement in the sample device based on the fusion features; The parameter adjustment module is used to adjust the model parameters of the advertising delivery model based on the difference between the actual conversion result and the predicted conversion result, so as to obtain the trained advertising delivery model.

14. An advertising delivery device, characterized in that, include: The second acquisition module is used to acquire multiple sets of data to be processed. The data to be processed includes: historical operation data of the advertisement on the device associated with the account, the user type logged into the device, the device identifier of the device, and the preset delivery results of the advertisement on the device; The third acquisition module is used to input multiple sets of the data to be processed into the trained advertising delivery model to obtain the conversion result corresponding to each of the devices; the advertising delivery model is generated by the method described in any one of claims 1-7. The first determining module is used to determine the device to be deployed for the advertisement from among the multiple devices based on the conversion results corresponding to each of the devices.

15. 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; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the steps of the method described in any one of claims 1-12.

16. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method as described in any one of claims 1-12.

Images