Method and device for determining media resources to be placed and media resource processing model

Abstract

The present disclosure provides a method and device for determining media resources to be delivered and a media resource processing model, and relates to the field of Internet technology, in order to at least improve the effect of advertising delivery. The determination method includes: for each candidate media resource group in the media resource candidate set, using the media resource processing model to perform a processing operation to obtain the contents of at least two indicators of each candidate media resource group; the processing operation includes: obtaining the characteristics of each candidate media resource group; inputting the characteristics into at least one expert network to obtain a learning feature related to at least one indicator output by each expert network; determining the content of each indicator of each candidate media resource group based on the obtained learning feature and the weight set corresponding to each indicator of at least two indicators; determining the target candidate media resource group based on the contents of at least two indicators of each candidate media resource group, and using the candidate media resources in the target candidate media resource group as the media resources to be delivered.

Description

Technical Field

[0001] The present disclosure relates to the field of Internet technology, and in particular to a method and device for determining media resources to be delivered and a media resource processing model. Background Art

[0002] Usually, advertisers will expand and copy the ads that have good online performance, and continue to run them in other scenarios to attract more users. In the prior art, it is usually determined whether the ad needs to be extended based on a certain performance of the ad during the delivery process. However, the single performance of an ad during the delivery process cannot fully reflect the effectiveness of the ad. Therefore, when the extended delivery is used, the delivery effect is not good. Summary of the invention

[0003] The present disclosure provides a method and device for determining media resources to be delivered and a media resource processing model, so as to at least improve the effect of advertising delivery. The technical solution of the present disclosure is as follows:

[0004] According to the first aspect of the present disclosure, a method for determining media resources to be delivered is provided, and the determination method includes: first, for each candidate media resource group in the media resource candidate set, a pre-trained media resource processing model is used to perform a processing operation to obtain the content of at least two indicators of each candidate media resource group. The processing operation includes: obtaining the characteristics of each candidate media resource group; inputting the characteristics into at least one expert network to obtain a learning feature related to at least one indicator output by each expert network; determining the content of each indicator of each candidate media resource group based on the obtained learning feature and the weight set corresponding to each indicator of the at least two indicators; the weight set includes the weight of at least one expert network. Then, based on the content of at least two indicators of each candidate media resource group, the target candidate media resource group is determined, and the candidate media resources in the target candidate media resource group are used as the media resources to be delivered.

[0005] Each candidate media resource group includes at least one candidate media resource.

[0006] Optionally, the above-mentioned method of "determining the content of each indicator of each candidate media resource group based on the obtained learning features and the weight set corresponding to each indicator of at least two indicators" includes: based on the weight set, performing weighted summation on the obtained learning features to obtain the summed features; predicting each indicator on the summed features to obtain the content of each indicator of each candidate media resource group.

[0007] Optionally, the above-mentioned method of "determining the target candidate media resource group based on the contents of at least two indicators of each candidate media resource group" includes: determining the candidate media resource group whose contents of at least two indicators meet preset conditions as the target candidate media resource group; or, determining the comprehensive indicator value of each first candidate media resource group, the first candidate media resource group being the candidate media resource group whose contents of at least two indicators meet preset conditions; sorting all first candidate media resource groups according to the determined comprehensive indicator value to obtain sorted first candidate media resource groups; and determining the target candidate media resource group from the sorted first candidate media resource groups.

[0008] The preset condition includes that the content of each indicator of at least two indicators is within a preset range corresponding to each indicator.

[0009] Optionally, the above-mentioned method of "determining the comprehensive index value of each first candidate media resource group" includes: normalizing the content of each indicator of each first candidate media resource group respectively to obtain at least two numerical values; weighted summing the at least two numerical values ​​and at least two preset indicator weights to obtain a comprehensive index value.

[0010] Optionally, the features of each candidate media resource group are represented by a vector. Optionally, the above-mentioned method of "obtaining the features of each candidate media resource group" includes: obtaining multiple initial features corresponding to each candidate media resource group; embedding each initial feature corresponding to each candidate media resource group to obtain an initial feature vector corresponding to each initial feature; performing dot product operation on every two initial feature vectors of all initial feature vectors of each candidate media resource group to obtain an interactive feature vector of each candidate media resource group; splicing the interactive feature vectors of each candidate media resource group, and using the spliced ​​vector as the feature of each candidate media resource.

[0011] Optionally, the above-mentioned method of "obtaining multiple initial features corresponding to each candidate media resource group" includes: obtaining feature data of each candidate media resource in each candidate media resource group within a preset time period; aggregating the feature data corresponding to each candidate media resource group to obtain multiple initial features corresponding to each candidate media resource group.

[0012] Optionally, the multiple initial features corresponding to each candidate media resource group include at least three types of initial features. The at least three types of initial features include numerical features, categorical sparse features, and sequence features. The numerical features include the sum of the first consumption data of each candidate media resource group within a preset time period. The categorical sparse features include the category of each candidate media resource group within a preset time period. The sequence features include the sum of the second consumption data of all candidate media resources in each candidate media resource group within each unit time in the preset time period.

[0013] According to a second aspect of the present disclosure, a method for determining a media resource processing model is provided, the method comprising: obtaining features of at least one historical media resource group within a first preset time period, and contents of at least two indicators of at least one historical media resource group within a second preset time period; inputting features of each historical media resource group into at least one expert network in an initial network model; extracting initial learning features through each expert network; obtaining association information of each indicator based on a weight set corresponding to each indicator and all initial learning features extracted by at least one expert network; predicting the indicator content of the association information of each indicator through a prediction model corresponding to each indicator in the initial network model to obtain the predicted content of each indicator; determining a loss value between the predicted contents of at least two indicators and the contents of at least two indicators of each historical media resource group; iteratively training the initial network model based on the loss value until the loss value is less than or equal to a preset threshold, and determining the trained initial network model as the media resource processing model.

[0014] The time of the first preset time period is earlier than the time of the second preset time period. Each historical media resource group includes at least one historical media resource. The weight set includes the weight of at least one expert network.

[0015] Optionally, each training input sample is represented by a vector. The above-mentioned "obtaining at least one training input sample" includes: obtaining multiple initial features corresponding to the historical media resource group within the first preset time period; inputting the multiple initial features corresponding to the historical media resource group into the representation layer in the initial network model; embedding each initial feature corresponding to the historical media resource group through the representation layer to obtain the initial feature vector corresponding to each initial feature; performing dot product operation on every two initial feature vectors of all initial feature vectors corresponding to the historical media resource group through the interaction layer in the initial network model to obtain the interaction feature vector corresponding to the historical media resource group; splicing the interaction feature vectors corresponding to the historical media resource group through the connection layer in the initial network model, and using the spliced ​​vector as each training input sample.

[0016] Optionally, the weight sets corresponding to at least two indicators are different.

[0017] According to a third aspect of the present disclosure, a device for determining media resources to be delivered is provided, and the determination device includes a processing module and a determination module. The processing module is configured to execute a processing operation for each candidate media resource group in the candidate set of media resources using a pre-trained media resource processing model to obtain the content of at least two indicators of each candidate media resource group. Each candidate media resource group includes at least one candidate media resource. The processing operation includes: obtaining features of each candidate media resource group; inputting the features into at least one expert network to obtain learning features related to at least one indicator output by each expert network; determining the content of each indicator of each candidate media resource group based on the obtained learning features and the weight set corresponding to each indicator of the at least two indicators; the weight set includes the weight of at least one expert network.

[0018] The determination module is configured to determine a target candidate media resource group according to the contents of at least two indicators of each candidate media resource group, and use the candidate media resources in the target candidate media resource group as the media resources to be delivered.

[0019] Optionally, the processing module is specifically configured to execute: based on a weight set, weighted summing of the obtained learning features to obtain the summed features; predicting each indicator of the summed features to obtain the content of each indicator of each candidate media resource group.

[0020] Optionally, the determination module is specifically configured to execute: determining a candidate media resource group whose contents of at least two indicators meet preset conditions as a target candidate media resource group; or, determining a comprehensive indicator value of each first candidate media resource group, the first candidate media resource group being a candidate media resource group whose contents of at least two indicators meet preset conditions; sorting all first candidate media resource groups according to the determined comprehensive indicator value to obtain sorted first candidate media resource groups; and determining the target candidate media resource group from the sorted first candidate media resource groups.

[0021] The preset condition includes that the content of each indicator of at least two indicators is within a preset range corresponding to each indicator.

[0022] Optionally, the determination module is specifically configured to execute: normalizing the content of each indicator of each first candidate media resource group respectively to obtain at least two numerical values; and performing weighted summation of the at least two numerical values ​​and at least two preset indicator weights to obtain a comprehensive indicator value.

[0023] Optionally, the features of each candidate media resource group are represented by vectors. The processing module is specifically configured to execute: obtaining multiple initial features corresponding to each candidate media resource group; embedding each initial feature corresponding to each candidate media resource group to obtain an initial feature vector corresponding to each initial feature; performing dot product operation on every two initial feature vectors of all initial feature vectors of each candidate media resource group to obtain an interactive feature vector of each candidate media resource group; splicing the interactive feature vectors of each candidate media resource group, and using the spliced ​​vector as the feature of each candidate media resource.

[0024] Optionally, the processing module is specifically configured to execute: obtaining feature data of each candidate media resource in each candidate media resource group within a preset time period; aggregating the feature data corresponding to each candidate media resource group to obtain multiple initial features corresponding to each candidate media resource group.

[0025] Optionally, the multiple initial features corresponding to each candidate media resource group include at least three types of initial features. The at least three types of initial features include numerical features, categorical sparse features, and sequence features. The numerical features include the sum of the first consumption data of each candidate media resource group within a preset time period. The categorical sparse features include the category of each candidate media resource group within a preset time period. The sequence features include the sum of the second consumption data of all candidate media resources in each candidate media resource group within each unit time in the preset time period.

[0026] According to a fourth aspect of the present disclosure, a device for determining a media resource processing model is provided, the device comprising an acquisition module and a training module. The acquisition module is configured to execute acquisition of the characteristics of each historical media resource in at least one historical media resource group within a first preset time period, and the contents of at least two indicators of each historical media resource group within a second preset time period. The time of the first preset time period is earlier than the time of the second preset time period; each historical media resource group includes at least one historical media resource.

[0027] The training module is configured to execute: inputting the features of each historical media resource group into at least one expert network in the initial network model; extracting initial learning features through each expert network; obtaining the associated information of each indicator according to the weight set corresponding to each indicator and all the initial learning features extracted by at least one expert network; wherein the weight set includes the weight of at least one expert network; predicting the indicator content of the associated information of each indicator through the prediction model corresponding to each indicator in the initial network model to obtain the predicted content of each indicator; determining the loss value between the predicted content of at least two indicators and the content of at least two indicators of each historical media resource group; iteratively training the initial network model according to the loss value until the loss value is less than or equal to a preset threshold, and determining the trained initial network model as the media resource processing model.

[0028] Optionally, the feature of each historical media resource is represented by a vector. The acquisition module is specifically configured to execute: acquiring multiple initial features corresponding to each historical media resource group within a first preset time period; inputting the multiple initial features corresponding to each historical media resource group into the representation layer in the initial network model; embedding each initial feature corresponding to each historical media resource group through the representation layer to obtain an initial feature vector corresponding to each initial feature; performing dot product operation on every two initial feature vectors of all initial feature vectors corresponding to each historical media resource group through the interaction layer in the initial network model to obtain an interaction feature vector corresponding to each historical media resource group; splicing the interaction feature vectors corresponding to each historical media resource group through the connection layer in the initial network model, and using the spliced ​​vector as the feature of each historical media resource.

[0029] Optionally, the weight sets corresponding to at least two indicators are different.

[0030] According to a fifth aspect of the present disclosure, an electronic device is provided, comprising: a processor and a memory for storing instructions executed by the processor. The processor is configured to execute the instructions to implement any optional method for determining a to-be-delivered media resource in the first aspect; or the processor is configured to execute the instructions to implement any optional method for determining a media resource processing model in the second aspect.

[0031] According to a sixth aspect of the present disclosure, a computer-readable storage medium is provided, on which computer instructions are stored. When the computer instructions are executed by an electronic device, the electronic device executes any optional method for determining a to-be-delivered media resource in the first aspect; or, when the computer instructions are executed by an electronic device, the electronic device executes any optional method for determining a media resource processing model in the second aspect.

[0032] According to a seventh aspect of the present disclosure, a computer program product is provided, the computer program product comprising computer instructions. When the computer instructions are executed by an electronic device, the method for determining a media resource to be delivered as any one of the optional items in the first aspect above is implemented; or, when the computer instructions are executed by an electronic device, the method for determining a media resource processing model as any one of the optional items in the second aspect above is implemented.

[0033] The technical solution provided by the embodiments of the present disclosure brings at least the following beneficial effects:

[0034] In the above scheme, the electronic device uses each expert network to extract learning features related to different indicators for the features of each candidate media resource group, and then uses the weight set corresponding to each indicator and all the learning features extracted by at least one expert network to predict the content of the indicator. In this way, for different indicators, at least one expert network is shared. Then, through the weight set corresponding to each indicator (including the weight of the at least one expert network), the importance of all learning features extracted by at least one expert network in predicting the content of any indicator can be controlled. For example, the expert network whose extracted learning features have a greater correlation with an indicator has a greater weight in the weight set corresponding to the indicator, then the learning feature has a greater impact on predicting the content of the indicator, and the predicted content of the indicator is more accurate. In this way, it is ensured that the learning features with a smaller correlation with an indicator interfere with the prediction of the content of the indicator, that is, the independence of the electronic device in predicting the content of different indicators is guaranteed.

[0035] Secondly, the electronic device can simultaneously predict the contents of at least two indicators for each candidate media resource group, thereby avoiding determining the media resources to be delivered based on a single performance of each candidate media resource group during the delivery process, thereby improving the delivery effect.

[0036] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute improper limitations on the present disclosure.

[0038] Figure 1 is a schematic diagram of an implementation environment of a method for determining media resources to be delivered according to an exemplary embodiment;

[0039] Figure 2 is a flow chart showing a method for determining media resources to be delivered according to an exemplary embodiment;

[0040] Figure 3is one of the structural schematic diagrams of a media resource processing model according to an exemplary embodiment;

[0041] Figure 4 is a second structural diagram of a media resource processing model according to an exemplary embodiment;

[0042] Figure 5A A method for determining a media resource processing model according to an exemplary embodiment is shown in FIG. Figure 1 ;

[0043] Figure 5B A method for determining a media resource processing model according to an exemplary embodiment is shown in FIG. Figure 2 ;

[0044] Figure 6 is a structural schematic diagram of a device for determining media resources to be delivered according to an exemplary embodiment;

[0045] Figure 7 is a schematic diagram of the hardware structure of a device for determining a media resource processing model according to an exemplary embodiment;

[0046] Figure 8 The figure is a schematic diagram of the hardware structure of a device for determining media resources to be delivered according to an exemplary embodiment. DETAILED DESCRIPTION

[0047] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings.

[0048] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0049] First, the application scenarios of the embodiments of the present disclosure are introduced.

[0050] The method for determining the media resources to be delivered in the embodiment of the present disclosure is applied to the scene of delivering advertisements (i.e., the media resources in the present disclosure). For example, the scene of delivering game advertisements, the scene of delivering video advertisements, etc. In the process of delivering advertisements, advertisers usually extend and copy the advertisements with good online performance, and continue to deliver them in other scenes to obtain more users. In the related art, it is usually determined whether the advertisement needs to be extended based on a certain performance of the advertisement during the delivery process.

[0051] However, in the actual process of advertising, advertisers usually have various goals, such as focusing on multiple goals such as activation conversion rate, cumulative return on investment (ROI), and retention rate during the advertising process. Therefore, the method of determining the advertising information that needs to be extended based on a certain performance of the advertisement during the delivery process cannot meet the advertising needs of advertisers. In addition, a single performance of an advertisement during the delivery process cannot fully reflect the effectiveness of the advertisement. Therefore, when the delivery is extended, the delivery effect is not good.

[0052] In order to solve the above problems, the embodiments of the present disclosure provide a method for determining media resources to be delivered, which can determine the contents of at least two indicators of each candidate media resource group based on the characteristics of each candidate media resource group, and determine the media resources to be delivered based on the contents of at least two indicators, thereby avoiding determining the media resources to be delivered based on a single performance of the candidate media resources (such as candidate advertisements) during the delivery process, and improving the delivery effect.

[0053] Next, the implementation environment of the embodiment of the present disclosure is introduced.

[0054] Figure 1 FIG. 1 is a schematic diagram of an implementation environment of a method for determining a media resource to be delivered or a method for determining a media resource processing model according to an exemplary embodiment. Figure 1 As shown, the method for determining the media resources to be delivered and the method for determining the media resource processing model in the present disclosure can be applied to the implementation environment. The implementation environment includes a terminal 01 and a server 02. The terminal 01 and the server 02 can be interconnected and communicated through a network.

[0055] The terminal 01 may be a mobile phone, tablet computer, desktop, laptop, handheld computer, notebook computer, vehicle-mounted device, ultra-mobile personal computer (UMPC), netbook, etc. The disclosed embodiment does not impose any special restrictions on the specific form of the terminal 01.

[0056] The server 02 may be a single server, a server cluster consisting of multiple servers, or a cloud computing service center. The server 02 may include a processor, a memory, a network interface, and the like.

[0057] It should be noted that the method for determining the media resources to be placed and the method for determining the media resource processing model provided in the embodiment of the present disclosure can be applied to the above-mentioned server 02, and can also be applied to the above-mentioned terminal 01. The server 02 and the terminal 01 can be collectively referred to as electronic devices. The execution subject of the method for determining the media resources to be placed provided in the embodiment of the present disclosure can be a device for determining the media resources to be placed, and the execution subject of the method for determining the media resource processing model can be a device for determining the media resource processing model. The device for determining the media resources to be placed can be the above-mentioned electronic device (such as the server 02 or the terminal 01); or, it can be an application (application, APP) installed in the electronic device that provides the function of determining the media resources to be placed; or, it can be a central processing unit (Central Processing Unit, CPU) in the electronic device; or, it can be a control module in the electronic device for executing the method for determining the media resources to be placed. Similarly, the device for determining the media resource processing model can be the above-mentioned electronic device; or, it can be an APP installed in the electronic device that provides the function of determining the media resource processing model; or, it can be a CPU in the electronic device; or, it can be a control module in the electronic device for executing the method for determining the media resource processing model.

[0058] Taking an electronic device as an example, the method for determining the media resources to be delivered and the method for determining the media resource processing model provided in the embodiment of the present application are described in detail below.

[0059] Those skilled in the art should understand that the above-mentioned terminals and servers are only examples, and other existing or future terminals or servers that are applicable to the present disclosure should also be included in the protection scope of the present disclosure and are included here by reference.

[0060] Figure 2 FIG. 1 is a flow chart showing a method for determining a media resource to be delivered according to an exemplary embodiment. Figure 2 As shown, the method may include the following steps.

[0061] Step 21: The electronic device uses a pre-trained media resource processing model to perform processing operations on each candidate media resource group in the media resource candidate set to obtain contents of at least two indicators of each candidate media resource group.

[0062] The media resource candidate set may include at least one candidate media resource group. Each candidate media resource group may include one candidate media resource, or may include multiple candidate media resources. The multiple candidate media resources may be divided according to the same candidate media resource type (e.g., creative type), or may be divided according to other requirements. The media resource processing model has the ability to predict the content of at least two indicators of each candidate media resource group based on the characteristics of each candidate media resource group.

[0063] Specifically, the processing operation includes the following steps.

[0064] Step A1: The electronic device obtains the characteristics of each candidate media resource group.

[0065] The feature of each candidate media resource group may represent the interaction relationship between multiple initial features of each candidate media resource group. The feature of each candidate media resource group may be represented by a vector.

[0066] First, the electronic device obtains a plurality of initial features corresponding to each candidate media resource group. The plurality of initial features corresponding to each candidate media resource group may include at least three types of initial features. Each type of initial feature may include one initial feature or multiple initial features.

[0067] Specifically, the electronic device first obtains the characteristic data of each candidate media resource in each candidate media resource group within a preset time period. Specifically, when the electronic device determines that it is necessary to determine the media resource to be delivered in the candidate set of media resources, it obtains the characteristic data of each candidate media resource in each candidate media resource group. Among them, the characteristic data of each candidate media resource at least includes the type of the candidate media resource and the performance of the candidate media resource during the delivery process. The characteristic data can be called delivery data. For example, the type of the candidate media resource includes the creative type of the candidate media resource, the identity document (ID) of the candidate media resource, etc.; the performance of the candidate media resource during the delivery process within the preset time period includes the historical cost sequence of the candidate media resource, the completion rate of the candidate media resource, etc.

[0068] Then, after determining the characteristic data of each candidate media resource in each candidate media resource group within a preset time period, the electronic device aggregates the characteristic data corresponding to each candidate media resource group to obtain a plurality of initial characteristics corresponding to each candidate media resource group. Specifically, the electronic device can determine the delivery target (i.e., indicator) and determine the aggregation method according to the delivery target. For example, if the delivery target is to pursue the activation conversion rate of the candidate media resource group and the cumulative return on investment (ROI) of the candidate media resource group, the characteristic data corresponding to each candidate media resource group can be aggregated according to the specified time range (e.g., aggregated by day, or aggregated by week).

[0069] Among them, the multiple initial features corresponding to each candidate media resource group may include at least three types of initial features; the at least three types of initial features may include numerical features (such as the total cost of each candidate media resource group, the total activation number of each candidate media resource group, the total click-through rate of the cover of each candidate media resource group), category-type sparse features (such as the delivery platform type of each candidate media resource group, the creative type of each candidate media resource group) and sequence features (the historical total activation number sequence of each candidate media resource group, the historical total cost sequence of candidate media resource creatives).

[0070] In the above scheme, the electronic device can obtain the characteristic data of each candidate media resource in each candidate media resource group. Then, in order to predict the content of at least two indicators of each candidate media resource group, the electronic device can aggregate the characteristic data of all candidate media resources included in the candidate media resource group to obtain multiple initial features of the candidate media resource group. Therefore, the content of at least two indicators of the candidate media resource group can be determined based on the multiple initial features of the candidate media resource group, so as to determine the target candidate media resource group from all candidate media resource groups in the media resource candidate set, and the candidate media resources in the target candidate media resource group are the media resources to be delivered. Secondly, the electronic device can also determine the time range of aggregation based on the content of which indicators are the delivery target, so that the characteristic data within different time ranges can be flexibly used for the content of different indicators.

[0071] For example, when the weekly aggregation process is determined, the electronic device may obtain the feature data of each candidate media resource in each candidate media resource group in the week closest to the current time within the preset time period. Then, the electronic device may aggregate the same feature data of all candidate media resources in each candidate media resource group in the week to obtain each initial feature corresponding to each candidate media resource group.

[0072] Further, assuming that each candidate media resource group includes multiple candidate media resources of the same creative type, the characteristic data corresponding to each candidate media resource group within the preset time period may include: the identifier of each candidate media resource among the multiple candidate media resources of the same creative type, the historical cost sequence of each candidate media resource, and the completion rate of each candidate media resource. The historical cost sequence of each candidate media resource may include the historical cost of the candidate media resource in each unit time within the preset time period; the completion rate of each candidate media resource may include the completion rate in each unit time within the preset time period. The above-mentioned aggregation of the characteristic data corresponding to each candidate media resource group within the week closest to the current moment within the preset time period may include: the electronic device sums the historical costs of all candidate media resources in each candidate media resource group in each unit time period within the week to obtain the total cost in each unit time period, and then obtains the total cost sequence composed of the total costs in multiple unit time ends within the week; the electronic device may also sum the completion rates of all candidate media resources in each candidate media resource group in the week to obtain the total completion rate of each candidate media resource group in the week.

[0073] Alternatively, assuming that each candidate media resource group includes multiple candidate media resources with different creative types, the characteristic data corresponding to each candidate media resource group within the preset time period may include: the creative type of each candidate media resource among the multiple candidate media resources with different creative types, the identifier of each candidate media resource, the historical cost sequence of each candidate media resource, and the completion rate of each candidate media resource. The above-mentioned aggregation of the characteristic data corresponding to each candidate media resource group within the preset time period and the week closest to the current moment may include: the electronic device obtains the creative type of all candidate media resources in each candidate media resource group; the electronic device sums the historical costs of all candidate media resources in each candidate media resource group in each unit time period in the week to obtain the total cost in each unit time period, and then obtains the total cost sequence composed of the total costs in multiple unit time ends in the week; the electronic device can also sum the completion rates of all candidate media resources in each candidate media resource group in the week to obtain the total completion rate of each candidate media resource group in the week.

[0074] Afterwards, when the feature of each candidate media resource group is represented by a vector, the electronic device may perform embedding processing on each of the multiple initial features corresponding to each candidate media resource group to obtain an initial feature vector corresponding to each initial feature.

[0075] Specifically, embedding processing is a processing method for vectorizing features, including specific implementation methods such as word embedding, sentence embedding, and entity embedding. Different embedding processing can be used for different types of initial features. For example, after the electronic device normalizes the numerical feature (such as the above-mentioned total completion rate), it uses a multi-layer perceptron (MLP) network to embed it to obtain a numerical feature vector; uses the Embedding layer to embed the categorical sparse features (such as the creative types of all candidate media resources in each candidate media resource group) to obtain a categorical sparse feature vector; uses the Transformer layer to embed the sequence features (such as the above-mentioned total cost sequence) to obtain a sequence feature vector.

[0076] Then, the electronic device performs a dot product operation on every two initial feature vectors of all initial feature vectors of each candidate media resource group to obtain an interactive feature vector of each candidate media resource group, wherein an interactive feature vector of each candidate media resource group represents the interactive relationship between every two initial features.

[0077] Finally, the electronic device performs splicing processing on the interactive feature vectors of each candidate media resource group, and uses the spliced ​​vectors as the features of each candidate media resource group. The spliced ​​vectors can be called the full vector corresponding to each candidate media resource group.

[0078] In the above scheme, the electronic device embeds each initial feature of each candidate media resource group, which can realize the vectorization of the initial features. For electronic devices, the initial feature vector belongs to a readable machine language, which improves the readability of the initial features of each candidate media resource group. Then, the electronic device performs a dot product operation on each two initial feature vectors, which can realize the interaction between different initial features of each candidate media resource group and enhance the connection between different initial features. Moreover, if the two initial features that affect the content of any indicator are associated to form an interactive feature vector, the content of the indicator is predicted according to the interactive feature vector, which is more accurate than predicting the content of the indicator according to an initial feature that affects the content of the indicator. For example, any indicator is the activation conversion rate of each candidate media resource group, and the two initial features corresponding to the candidate media resource group include the creative type and the historical total activation number sequence. These two initial features will affect the activation conversion rate of the candidate media resource group; the activation conversion rate of the candidate media resource group is predicted by the interactive feature vector composed of these two initial features, which is more accurate than predicting the activation conversion rate of the candidate media resource group by using any of the two initial features. Therefore, it can be known that the electronic device can improve the accuracy of predicting the content of at least two indicators by using the interactive feature vector composed of the association of every two initial feature vectors.

[0079] In addition, the electronic device can concatenate all interactive feature vectors of each candidate media resource group to obtain a concatenated vector as the feature of each candidate media resource group. Predicting the content of at least two indicators based on the feature of the candidate media resource group is to predict the content of at least two indicators based on all interactive feature vectors of each candidate media resource group.

[0080] A2: The electronic device inputs the features of each candidate media resource group into at least one expert network to obtain a learning feature output by each expert network that is related to at least one indicator.

[0081] Each expert network is used to extract learning features related to at least one indicator from the features of each candidate media resource group. Extracting learning features related to at least one indicator is to map the features of each candidate media resource group to the feature space of the expert network, so as to facilitate the subsequent determination of the content of the indicator of each candidate media resource group.

[0082] Each expert network corresponds to at least one indicator. Each expert network extracts learning features related to at least one indicator corresponding to each candidate media resource group. The number of all indicators corresponding to all expert networks is equal to the total number of at least two indicators. At least one indicator corresponding to any two expert networks may be the same or different.

[0083] A3: The electronic device determines the content of each indicator of each candidate media resource group according to the obtained learning features and a weight set corresponding to each indicator of at least two indicators; the weight set includes a weight of at least one expert network.

[0084] The learning features that can be obtained by the electronic device include at least one learning feature that corresponds to at least one expert network. The weight sets corresponding to different indicators may be the same or different. Any two weight sets being the same means that the weights of all expert networks in the two weight sets are the same. Any two weight sets being different may mean that the weights of all expert networks in the two weight sets are different, or the weights of some expert networks in the two weight sets are different.

[0085] Among them, at least one expert network includes n expert networks, and the weight set corresponding to each indicator includes n weights of n expert networks, and the value range of n can be 1-6. At least two indicators include m indicators. n and m are both positive integers. n≥m. The larger m is, the larger n is.

[0086] For example, n=4, m=3. The weight set corresponding to the first indicator is A, the weight set corresponding to the second indicator is B, the weight set corresponding to the third indicator is C, and the weight set corresponding to the fourth indicator is D. The weight set includes 4 weights of 4 expert networks. A, B, C and D can be the same or different.

[0087] The electronic device performs weighted summation on the obtained learning features based on the weight set to obtain the summed features; then, the summed features are predicted for each indicator to obtain the content of each indicator of each candidate media resource group. Specifically, the electronic device multiplies each learning feature by the weight of the expert network corresponding to each learning feature in the weight set corresponding to each indicator, and then sums all the products to obtain the summed features.

[0088] In which, when at least two indicators correspond to the same weight set, the electronic device multiplies each learning feature by the weight of the expert network corresponding to each learning feature in a weight set and then sums the results to obtain a summed feature, and predicts at least two indicators for the summed feature respectively to obtain the content of at least two indicators for each candidate media resource group.

[0089] In the case where at least two indicators correspond to different weight sets, the electronic device multiplies at least one learning feature by the weight of at least one expert network in the weight set corresponding to the indicator for each indicator, and then sums the results to obtain the summed feature corresponding to the indicator, and predicts the indicator for the summed feature corresponding to the indicator to obtain the content of the indicator for each candidate media resource group. The at least one learning feature and the weight of at least one expert network in the weight set corresponding to the indicator are multiplied one by one.

[0090] In the above scheme, each expert network extracts learning features related to at least one indicator. It can be seen that the learning features output by different expert networks have different correlations with each indicator. Furthermore, the electronic device can set a larger weight for the expert network whose output learning features have a greater correlation with the indicator, and a larger weight for the expert network whose output learning features have a smaller correlation with the indicator in the weight set corresponding to each indicator. Then, based on the weight set corresponding to the indicator, the electronic device performs weighted summation on all learning features extracted by at least one expert network to obtain the summed feature. The summed feature has a greater correlation with the indicator and is more accurate in predicting the content of the indicator.

[0091] This allows for subsequent determination of media resources to be delivered based on the contents of at least two indicators.

[0092] Optionally, the multiple initial features corresponding to each candidate media resource group may include at least three types of initial features; the at least three types of initial features may include numerical features, categorical sparse features, and sequence features. Correspondingly, the feature data of each candidate media resource may also include at least three types of feature data; the at least three types of feature data include: numerical data, categorical sparse data, and sequence data.

[0093] Among them, the numerical data of each candidate media resource may refer to the first consumption data in each unit time in a preset time period. The first consumption data may include at least one of the following: cost, click-through rate, exposure, activation number, video completion rate, and conversion rate. For example, the numerical data of a candidate media resource may include the cost of the candidate media resource creative, the income of the candidate media resource, the age of the candidate media resource creative, the activation number of the candidate media resource, the activation unit price of the candidate media resource, the number of exposures of the candidate media resource cover, the click-through rate of the candidate media resource cover, the exposure number of the candidate media resource video, the completion rate of the candidate media resource video, the conversion rate of the candidate media resource, etc.

[0094] Furthermore, the numerical features of each candidate media resource group may include the sum of the first consumption data of each candidate media resource group within a preset time period, that is, the sum of the first consumption data of all candidate media resources in each candidate media resource group. For example, the numerical features of each candidate media resource group may include: the sum of the costs of all candidate media resource creatives in each candidate media resource group within a preset time period (that is, the above-mentioned total costs), the sum of the activation numbers of all candidate media resources (that is, the above-mentioned total activation numbers), the sum of the cover click-through rates of all candidate media resources (that is, the above-mentioned total cover click-through rates), and the like.

[0095] The category-type sparse data may be the category of each candidate media resource. For example, the category-type sparse data may include the type of the creative of the candidate media resource, the ID of the product in the candidate media resource, the cover ID of the candidate media resource, the cover type of the candidate media resource, the video ID, the video category, the delivery channel ID of the candidate media resource, the delivery platform type of the candidate media resource, etc.

[0096] Furthermore, the category-type sparse feature of each candidate media resource group refers to the category of each candidate media resource group. The feature of each candidate media resource group may be a category shared by all candidate media resources in each candidate media resource group. For example, the creative type of all candidate media resources in each candidate media resource group, the delivery platform type of all candidate media resources are the same, and the product ID of all candidate media resources is the same, then the category of each candidate media resource group may include: the creative type, the delivery platform type, and the product ID.

[0097] The sequence data of each candidate media resource may refer to the second consumption data in multiple unit times in a preset time period, that is, the change of the second consumption data in the preset time period. The second consumption data may include at least one of the following: activation number, historical consumption value (which may be called historical cost). For example, the sequence data may include the historical activation number sequence of the candidate media resource, the historical cost sequence of the creative of the candidate media resource, etc.

[0098] Furthermore, the sequence feature (which may be referred to as a time series feature) of each candidate media resource group includes the sum of the second consumption data of all candidate media resources in each candidate media resource group in each unit time in a preset time period. The sequence feature may characterize the change in the sum of the second consumption data of all candidate media resources in each candidate media resource group in a preset time period. For example, the sequence feature of each candidate media resource group may include: a historical total activation number sequence of each candidate media resource group, a historical total expenditure sequence of candidate media resource creatives, etc.

[0099] It can be understood that the sequence characteristics of each candidate media resource group can reflect the information fluctuation of each candidate media resource group over a period of time (such as fluctuations in the number of activations, fluctuations in spending). Furthermore, by using any information fluctuation trend reflected by a sequence characteristic, it is possible to predict the information that the candidate media resource group may display in the future. Secondly, the information fluctuations of different categories of candidate media resources are different, so it can be seen that the category-type sparse features will affect the sequence characteristics, and thus affect the information that the candidate media resource group may display in the future. In summary, the multiple types of initial features (including sequence features and category-type sparse features) of each candidate media resource group used by the electronic device are factors that affect the content of at least two indicators of the candidate media resource group. So. Multiple types of initial features are used to predict the content of at least two indicators of the candidate media resource group, and the prediction accuracy is higher.

[0100] In addition, the electronic device divides the features of the candidate media resources into numerical features, categorical sparse features, and sequence features, and can distinguish between numerical features, type features, and time series features, so as to facilitate the subsequent vectorization processing of features through different implementation methods.

[0101] Optionally, the content of at least two indicators may include a regression value of each candidate media resource group and / or a value corresponding to a classification label of each candidate media resource group. Specifically, the regression value is a value that can intuitively reflect the delivery effect of each candidate media resource group, for example, the revenue of each candidate media resource group. The value corresponding to the classification label of each candidate media resource group is a value that can assist relevant personnel in distinguishing the delivery effect, for example, the activation conversion rate, cumulative ROI, total life cycle value and retention rate of each candidate media resource group.

[0102] In the above scheme, the regression value can directly reflect the delivery effect of each candidate media resource group, and the value corresponding to the classification label of each candidate media resource group can assist relevant personnel in distinguishing the delivery effect. In this way, the media resources to be delivered determined based on the content of at least two indicators can avoid determining the media resources to be delivered based on the single performance of each candidate media resource group during the delivery process, thereby improving the delivery effect.

[0103] Step 22: The electronic device determines a target candidate media resource group according to the contents of at least two indicators of each candidate media resource group, and uses the candidate media resources in the target candidate media resource group as the media resources to be delivered.

[0104] Specifically, the electronic device determines the candidate media resource group whose contents of at least two indicators meet the preset conditions as the target candidate media resource group. Alternatively, the electronic device determines the comprehensive indicator value of each first candidate media resource group, and the first candidate media resource group is the candidate media resource group whose contents of at least two indicators meet the preset conditions; then, the electronic device sorts all the first candidate media resource groups according to the determined comprehensive indicator value to obtain sorted first candidate media resource groups; and then determines the target candidate media resource group from the sorted first candidate media resource groups.

[0105] Among them, the preset condition includes that the content of each indicator of at least two indicators is greater than or equal to the preset threshold value corresponding to each indicator within the preset range corresponding to each indicator. The at least two indicators can be divided into: a first indicator for which the user requires the value to be as large as possible, and a second indicator for which the user requires the value to be as small as possible. The preset range corresponding to the first indicator can be an indicator threshold value corresponding to the indicator and the range of values ​​above it. The preset range corresponding to the second indicator can be an indicator threshold value corresponding to the indicator and the range of values ​​below it. The indicator threshold value corresponding to each indicator can be a default value, or a value written by relevant personnel according to actual conditions, or a value calculated by the processor of the electronic device according to a preset formula.

[0106] Specifically, the electronic device performs an indicator summing operation for each first candidate media resource group to obtain a comprehensive indicator value for each first candidate media resource group. The indicator summing operation includes: normalizing the content of each indicator of each first candidate media resource group to obtain at least two values. Performing weighted summation of the at least two values ​​and at least two preset indicator weights to obtain a comprehensive indicator value.

[0107] Among them, at least two preset indicator weights, at least two indicators, and at least two values ​​are all in one-to-one correspondence. If the user requires that the preset indicator weight corresponding to the first indicator with a larger value is a positive number (or a negative number), then the preset indicator weight corresponding to the second indicator with a smaller value is a negative number (or a positive number). The at least two preset indicator weights can be set according to the importance of the at least two indicators. The more important any indicator is, the greater the absolute value of the preset indicator weight corresponding to any indicator is.

[0108] Exemplarily, when the preset indicator weight corresponding to the first indicator is a positive number and the preset indicator weight corresponding to the second indicator is a negative number, the electronic device sorts all first candidate media resource groups from large to small according to the comprehensive indicator value, and then determines the top K candidate media resource groups as target candidate media resource groups.

[0109] Exemplarily, when the preset indicator weight corresponding to the first indicator is a negative number and the preset indicator weight corresponding to the second indicator is a positive number, the electronic device sorts all first candidate media resource groups from small to large according to the comprehensive indicator value, and then determines the top K candidate media resource groups as target candidate media resource groups.

[0110] It is understandable that the electronic device determines the media resources to be delivered according to the preset range corresponding to each indicator, or determines the media resources to be delivered according to the preset range corresponding to each indicator and the comprehensive indicator value, so that relevant personnel can deliver media resources according to the determined media resources to be delivered and the delivery plan.

[0111] Among them, if there are multiple indicators of multiple first candidate media resource groups whose contents all meet the preset conditions, and each candidate media resource group has the contents of multiple indicators, it is impossible to simply select the target candidate media resource group based on the best indicator. Therefore, the electronic device can normalize the content of each indicator of each first candidate media resource group to obtain at least two numerical values. There is comparability between the at least two normalized numerical values. Then, the electronic device performs weighted summation of the at least two numerical values ​​and at least two preset indicator weights to obtain a comprehensive indicator value. Since the importance of different indicators is different, at least two preset indicator weights can be set according to the importance of at least two indicators, so as to achieve that the normalized values ​​corresponding to the more important indicators account for a larger proportion in the comprehensive indicator value. Finally, the electronic device can quickly determine the target candidate media resource group with the best comprehensive indicator value based only on the comprehensive indicator value.

[0112] Optionally, after determining the media resources to be delivered, the electronic device delivers the online media resources according to the delivery plan.

[0113] In the above scheme, the electronic device uses at least one expert network to extract learning features related to different indicators from the features of each candidate media resource group, and then uses the weight set corresponding to each indicator and all the learning features extracted by at least one expert network to predict the content of the indicator. In this way, for different indicators, at least one expert network is shared. Then, through the weight set corresponding to each indicator (including the weight of the at least one expert network), the importance of all learning features extracted by at least one expert network in predicting the content of any indicator can be controlled. For example, the expert network whose extracted learning features have a greater correlation with an indicator has a greater weight in the weight set corresponding to the indicator, then the greater the influence of the learning features on predicting the content of the indicator, the more accurate the predicted content of the indicator. In this way, it is ensured that the learning features with a smaller correlation with an indicator interfere with the prediction of the content of the indicator, that is, the independence of the electronic device in predicting the content of different indicators is guaranteed.

[0114] Secondly, the electronic device can simultaneously predict the contents of at least two indicators for each candidate media resource group, thereby avoiding determining the media resources to be delivered based on a single performance of each candidate media resource group during the delivery process, thereby improving the delivery effect.

[0115] Optionally, the electronic device may include a pre-trained media resource processing model for predicting the contents of at least two indicators of each candidate media resource group based on the characteristics of each candidate media resource group.

[0116] Specifically, the media resource processing model may include a multi-task model (task tower) and at least one expert network (expert). The multi-task model (which may be called a multi-prediction model) includes at least two parallel tasks corresponding to at least two indicators. The structures of the tasks in the expert network and the multi-task model are both MLP+Relu. The number of expert networks is determined according to the amount of parallel tasks in the multi-task model. For example, the multi-task model includes 2 parallel tasks, and the media resource processing model includes 3 expert networks; the multi-task model includes 4 parallel tasks, and the media resource processing model includes 6 expert networks.

[0117] Each expert network is used to extract learning features related to at least one indicator from the features of each candidate media resource group. The multi-task model is used to predict the content of at least two indicators.

[0118] For example, Figure 3 FIG. 1 is a structural diagram of a media resource processing model shown in an embodiment of the present disclosure. Figure 3 As shown, the media resource processing model includes an input layer 34, a control layer 33, a multi-task model layer 32, and an output layer 31. The multi-task model layer 32 includes parallel tasks 1 to m. The control layer 33 includes expert networks 1 to n. The control layer 33 also includes 1 gate control unit or m gate control units (gate) ( Figure 3 (not shown), the structure of each gate control unit is MLP+Softmax.

[0119] In this way, when the features of each candidate media resource group (such as the full vector corresponding to the candidate media resource) are input into the input layer 34, expert networks 1 to expert networks n respectively extract learning features related to at least one indicator for the features of each candidate media resource group, thereby obtaining n learning features. Afterwards, the gate control unit corresponding to each indicator multiplies each learning feature by the weight of an expert network corresponding to the learning feature in the weight set corresponding to each indicator and sums the results. Finally, tasks 1 to m respectively predict the indicators of the summed vectors to obtain indicator content 1 to indicator content m. If there are m gate control units, m summed vectors are obtained, and tasks 1 to m respectively predict the indicators of the m summed vectors. Tasks 1 to m correspond one-to-one to the m summed vectors.

[0120] Optionally, the above-mentioned media resource processing model may also include a processing function for multiple initial features of each candidate media resource group (i.e., embedding processing, dot product operation processing, and splicing processing of multiple initial features of each candidate media resource group are implemented in the media resource processing model). In this way, after obtaining multiple initial features of each candidate media resource group, the electronic device directly inputs them into the media resource processing model to obtain the content of at least two indicators of each candidate media resource group.

[0121] For example, Figure 4 FIG. 1 is a structural diagram of a media resource processing model shown in an embodiment of the present disclosure. Figure 4 As shown in the figure, the media resource processing model includes an input layer 34, a representation layer 37, an interaction layer 36, a connection layer 35, a control layer 33, a multi-task model layer 32, and an output layer 31. The multi-task model layer 32 includes parallel tasks 1 to m; the control layer 33 includes expert networks 1 to n. The control layer 33 also includes one gate control unit or m gate control units (gates) (not shown in the figure), and the structure of each gate control unit is MLP+Softmax.

[0122] In this way, when the input layer 34 inputs multiple initial features of each candidate media resource group, the representation layer 37 embeds each of the multiple initial features of the candidate media resource to obtain the initial feature vector corresponding to each initial feature. The interaction layer 36 performs a dot product operation on every two initial feature vectors of all the initial feature vectors of each candidate media resource group to obtain the interactive feature vector of each candidate media resource group. The connection layer 35 concatenates all the interactive feature vectors of each candidate media resource group and uses the concatenated vector as the feature of each candidate media resource group (i.e., the full vector corresponding to each candidate media resource group). Afterwards, the expert network 1 to the expert network n respectively extracts learning features related to at least one indicator from the features of each candidate media resource group to obtain n learning features. Afterwards, the gate control unit corresponding to each indicator multiplies each learning feature by the weight of an expert network corresponding to the learning feature in the weight set corresponding to each indicator and then sums them. Finally, tasks 1 to m respectively predict the indicators of the summed vectors to obtain the content of indicators 1 to the content of indicators m.

[0123] In the above scheme, the gate control unit in the media resource processing model can make different tasks (i.e. different indicators) use the learning features output by the expert network in a diversified manner. In this way, even when the correlation between different indicators is not high, the learning features output by the partial expert network related to each indicator can be used to predict the content of the corresponding indicator, thereby improving the flexibility of the media resource processing model. The multi-task model in the media resource processing model includes at least two parallel tasks, which can predict the content of at least two indicators, avoiding the determination of the media resources to be delivered based on a single performance during the media resource delivery process, thereby improving the delivery effect. At least two parallel tasks and at least two indicators are in one-to-one correspondence.

[0124] Optionally, in the case where the electronic device includes a media resource processing model, the embodiment of the present disclosure further provides a method for determining the media resource processing model. Figure 5A As shown, the method for determining the media resource processing model includes the following steps.

[0125] Step 41: The electronic device obtains the characteristics of each historical media resource group in at least one historical media resource group within a first preset time period, and the contents of at least two indicators of each historical media resource group within a second preset time period.

[0126] The time of the first preset time period is earlier than the time of the second preset time period. Each historical media resource group includes one historical media resource or multiple historical media resources.

[0127] Specifically, the electronic device may obtain feature data (which may be referred to as delivery data) of each historical media resource group within a first preset time period, aggregate the feature data corresponding to each historical media resource group, and obtain multiple initial features corresponding to each historical media resource group.

[0128] Among them, the features of at least one historical media resource group are used as training samples of the initial network model. In order to ensure the balance of the training samples, if there is a long tail phenomenon in the feature data of a certain historical media resource group (a large number of media resource creatives have low costs), the features of the historical media resource group and the contents of at least two indicators are also obtained for training the initial network model. The electronic device can perform negative sampling on the feature data of the historical media resource group to obtain negative sample data (the negative sample data can also include at least three types of feature data); and then aggregate the negative sample data to obtain multiple initial features corresponding to the historical media resource group.

[0129] Afterwards, the electronic device sequentially performs embedding processing, dot product operation processing, and splicing processing on the multiple initial features of each historical media resource group collected to obtain features corresponding to each historical media resource group.

[0130] Finally, the electronic device obtains the contents of at least two indicators corresponding to the characteristic data of each historical media resource group. For example, the first preset time period is the N days before a certain moment, and the second preset time period is a continuous period after a certain moment. The characteristic data of each historical media resource group is the characteristic data of a certain media resource N days before, and the contents of at least two indicators corresponding to each historical media resource group are the activation conversion rate of each historical media resource group in the classification label and the cumulative ROI of each historical media resource group. Then, the electronic device obtains the activation conversion rate and cumulative ROI of each historical media resource group in a continuous period after N days. In this way, the electronic device can determine the characteristics of each historical media resource group based on the characteristic data of each historical media resource group N days before, and determine the activation conversion rate and cumulative ROI of each historical media resource group in a continuous period after N days as the contents of the two indicators of each historical media resource group.

[0131] Optionally, the feature of each historical media resource is represented by a vector. The electronic device may first obtain a plurality of initial features corresponding to each historical media resource group within a first preset time period. Then, the electronic device inputs the plurality of initial features corresponding to each historical media resource group into the representation layer in the initial network model; each initial feature corresponding to each historical media resource group is embedded through the representation layer to obtain an initial feature vector corresponding to each initial feature. The electronic device then performs a dot product operation on every two initial feature vectors of all the initial feature vectors corresponding to each historical media resource group through the interaction layer in the initial network model to obtain an interaction feature vector corresponding to each historical media resource group. Finally, the electronic device splices the interaction feature vectors corresponding to each historical media resource group through the connection layer in the initial network model, and uses the spliced ​​vector as the feature of each historical media resource.

[0132] It is understandable that the electronic device embeds each initial feature of each historical media resource group to realize the vectorization of the initial features. For electronic devices, the initial feature vector belongs to a readable machine language, which improves the readability of the initial features of each historical media resource group. Then, the electronic device performs a dot product operation on every two initial feature vectors, which can realize the interaction between different initial features of each historical media resource group and enhance the connection between different initial features. Moreover, if two initial features that affect the content of any indicator are associated to form an interactive feature vector, the content of the indicator is predicted based on the interactive feature vector, which is more accurate than predicting the content of the indicator based on an initial feature that affects the content of the indicator. Therefore, it can be seen that the electronic device can improve the accuracy of predicting the content of at least two indicators by using the interactive feature vector composed of every two initial feature vectors.

[0133] In addition, the electronic device can splice all interactive feature vectors of each historical media resource group to obtain a spliced ​​vector as the feature of each historical media resource group. Predicting the content of at least two indicators based on the feature of the historical media resource group is to predict the content of at least two indicators based on all interactive feature vectors of each historical media resource group.

[0134] Step 42: The electronic device trains the initial network model according to the characteristics of at least one historical media resource group and the contents of at least two indicators of at least one historical media resource group to obtain a media resource processing model.

[0135] Specifically, the electronic device trains the initial network model M times according to the characteristics of each historical media resource group in at least one historical media resource group and the contents of at least two indicators until a preset convergence condition is met to obtain a media resource processing model. The initial network module may be a neural network model.

[0136] Among them, the steps of the i-th training in M ​​trainings include: inputting the characteristics of a historical media resource group into the initial network model obtained in the i-1th training for processing to obtain the predicted content of at least two indicators; then, adjusting the parameters of the initial network model obtained in the i-1th training according to the loss value between the predicted content of at least two indicators and the content of at least two indicators of the historical media resource group.

[0137] The preset convergence condition is that the loss value is less than or equal to the first threshold, 1≤i≤M. The first threshold is preset, for example, it can be a default value, or a value written by relevant personnel according to actual conditions, or a value calculated by the processor of the electronic device according to a preset formula.

[0138] Optionally, when the media resource processing model in the present disclosure includes a processing function for multiple initial features of each candidate media resource group, the electronic device trains the initial network model based on the multiple initial features of each historical media resource group in at least one historical media resource group and the contents of at least two indicators to obtain the media resource processing model.

[0139] Specifically, Figure 5B As shown, step 42 may specifically include the following steps.

[0140] Step 421: The electronic device inputs the features of each historical media resource group into at least one expert network in the initial network model; and extracts initial learning features through each expert network.

[0141] Step 422: The electronic device obtains association information of each indicator based on a weight set corresponding to each indicator and all initial learning features extracted by at least one expert network; wherein the weight set includes the weight of at least one expert network.

[0142] The electronic device can obtain the association information of each indicator through the gate control unit corresponding to each indicator in the initial network model according to the weight set corresponding to each indicator and all initial learning features extracted by at least one expert network. Among them, at least two indicators can correspond to the same gate control unit, or correspond to different gate control units.

[0143] Optionally, the weight sets corresponding to at least two indicators are different.

[0144] It is understandable that different indicators have different degrees of correlation with the learning features extracted by each expert network; therefore, each indicator uses an exclusive set of weights (including the weight of at least one expert network) to ensure that the learning features extracted by the expert network that are less correlated with the indicator have less impact on predicting the content of the indicator; the learning features extracted by the expert network that are more correlated with the indicator have a greater impact on predicting the content of the indicator; thereby improving the prediction accuracy of the content of each indicator.

[0145] Step 423: The electronic device predicts the indicator content of the associated information of each indicator through the prediction model corresponding to each indicator in the initial network model to obtain the predicted content of each indicator.

[0146] Step 424: The electronic device determines a loss value between the predicted contents of the at least two indicators and the contents of the at least two indicators of each historical media resource group.

[0147] The electronic device may calculate a difference (ie, a loss value) between the predicted content of each indicator and the content of each indicator.

[0148] Step 425: The electronic device iteratively trains the initial network model according to the loss value until the loss value is less than or equal to a preset threshold, and determines the trained initial network model as the media resource processing model.

[0149] After the electronic device obtains the loss value based on the characteristics of a historical media resource group and the contents of at least two indicators, it adjusts the parameters of the initial network model according to the loss value; then the characteristics of the next historical media resource group and the contents of at least two indicators are used to train the adjusted initial network model until the loss value is less than a preset threshold.

[0150] It is understandable that the electronic device predicts different indicator contents of each historical media resource group, and uses the same at least one expert network, thereby achieving the sharing of at least one expert network. Secondly, the association information of each indicator is obtained by using the weight set corresponding to each indicator and all learning features extracted by at least one expert network. By setting the weight set corresponding to each indicator (including the weight of at least one expert network), the weight of the expert network with a greater correlation between the extracted learning features and each indicator can be amplified, and the weight of the expert network with a lesser correlation between the extracted learning features and each indicator can be reduced. In this way, the proportion of the learning features with a greater correlation with each indicator in the association information of each indicator can be amplified, and the proportion of the learning features with a lesser correlation with each indicator in the association information of each indicator can be reduced, so that the content of each indicator can be more accurately predicted based on the association information of each indicator.

[0151] It is understandable that the above-mentioned method for determining the media resources to be delivered can be implemented by a device for determining the media resources to be delivered. In order to realize the above-mentioned functions, the device for determining the media resources to be delivered includes hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art should easily realize that, in combination with the modules and algorithm steps of each example described in the embodiments disclosed herein, the embodiments of the present disclosure can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in the form of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the embodiments of the present disclosure.

[0152] The embodiments of the present disclosure may divide the functional modules of the above-mentioned determination device for the media resources to be delivered, etc. according to the above-mentioned method example. For example, each functional module may be divided corresponding to each function, or two or more functions may be integrated into one processing module. The above-mentioned integrated modules may be implemented in the form of hardware or in the form of software functional modules. It should be noted that the division of modules in the embodiments of the present disclosure is schematic and is only a logical functional division. There may be other division methods in actual implementation.

[0153] The embodiment of the present disclosure also provides a device for determining a media resource to be delivered. The device can be an electronic device, a chip in an electronic device, or a control module in an electronic device for determining an advertisement.

[0154] Figure 6 FIG. 1 is a schematic diagram showing a structure of a device for determining a media resource to be delivered according to an exemplary embodiment. Figure 6 As shown, the determination device includes a processing module 51 and a determination module 52 .

[0155] The processing module 51 is configured to execute a processing operation for each candidate media resource group in the media resource candidate set using a pre-trained media resource processing model to obtain the contents of at least two indicators of each candidate media resource group. Each candidate media resource group includes at least one candidate media resource. The processing operation includes: obtaining the features of each candidate media resource group; inputting the features into at least one expert network to obtain a learning feature related to at least one indicator output by each expert network; determining the content of each indicator of each candidate media resource group based on the obtained learning features and the weight set corresponding to each indicator of the at least two indicators. The weight set includes the weight of at least one expert network.

[0156] The determination module 52 is configured to determine a target candidate media resource group according to the contents of at least two indicators of each candidate media resource group, and use the candidate media resources in the target candidate media resource group as the media resources to be delivered.

[0157] Optionally, the processing module 51 is specifically configured to execute: based on the weight set, weighted summing of the obtained learning features to obtain the summed features; predicting each indicator of the summed features to obtain the content of each indicator of each candidate media resource group.

[0158] Optionally, the determination module 52 is specifically configured to execute: determining a candidate media resource group whose contents of at least two indicators meet preset conditions as a target candidate media resource group; or determining a comprehensive indicator value of each first candidate media resource group, the first candidate media resource group being a candidate media resource group whose contents of at least two indicators meet preset conditions; sorting all first candidate media resource groups according to the determined comprehensive indicator value to obtain sorted first candidate media resource groups; and determining the target candidate media resource group from the sorted first candidate media resource groups.

[0159] The preset condition includes that the content of each indicator of at least two indicators is within a preset range corresponding to each indicator.

[0160] Optionally, the determination module 52 is specifically configured to execute: normalizing the content of each indicator of each first candidate media resource group respectively to obtain at least two numerical values; and performing weighted summation of the at least two numerical values ​​and at least two preset indicator weights to obtain a comprehensive indicator value.

[0161] Optionally, the feature of each candidate media resource group is represented by a vector. The processing module 51 is specifically configured to execute: obtaining multiple initial features corresponding to each candidate media resource group; embedding each initial feature corresponding to each candidate media resource group to obtain an initial feature vector corresponding to each initial feature; performing a dot product operation on every two initial feature vectors of all initial feature vectors of each candidate media resource group to obtain an interactive feature vector of each candidate media resource group; splicing the interactive feature vectors of each candidate media resource group, and using the spliced ​​vector as the feature of each candidate media resource.

[0162] Optionally, the processing module 51 is specifically configured to execute: obtaining feature data of each candidate media resource in each candidate media resource group within a preset time period; aggregating the feature data corresponding to each candidate media resource group to obtain multiple initial features corresponding to each candidate media resource group.

[0163] Optionally, the multiple initial features corresponding to each candidate media resource group include at least three types of initial features. The at least three types of initial features include numerical features, category-type sparse features, and sequence features. The numerical features include the sum of the first consumption data of each candidate media resource group within a preset time period. The category-type sparse features include the category of each candidate media resource group within a preset time period. The sequence features include the sum of the second consumption data of all candidate media resources in each candidate media resource group within each unit time in the preset time period.

[0164] Figure 7 FIG. 1 is a schematic diagram of a structure of a device for determining a media resource processing model according to an exemplary embodiment. Figure 7 As shown, the determination device includes an acquisition module 61 and a training module 62 .

[0165] The acquisition module 61 is configured to execute the acquisition of the characteristics of each historical media resource in at least one historical media resource group within a first preset time period, and the contents of at least two indicators of each historical media resource group within a second preset time period; the time of the first preset time period is earlier than the time of the second preset time period; each historical media resource group includes at least one historical media resource.

[0166] The training module 62 is configured to execute: inputting the features of each historical media resource group into at least one expert network in the initial network model; extracting initial learning features through each expert network; obtaining the associated information of each indicator according to the weight set corresponding to each indicator and all the initial learning features extracted by at least one expert network; wherein the weight set includes the weight of at least one expert network; predicting the indicator content of the associated information of each indicator through the prediction model corresponding to each indicator in the initial network model to obtain the predicted content of each indicator; determining the loss value between the predicted content of at least two indicators and the content of at least two indicators of each historical media resource group; iteratively training the initial network model according to the loss value until the loss value is less than or equal to a preset threshold, and determining the trained initial network model as the media resource processing model.

[0167] Optionally, the feature of each historical media resource is represented by a vector. The acquisition module 61 is specifically configured to execute: acquiring multiple initial features corresponding to each historical media resource group within the first preset time period; inputting the multiple initial features corresponding to each historical media resource group into the representation layer in the initial network model; embedding each initial feature corresponding to each historical media resource group through the representation layer to obtain an initial feature vector corresponding to each initial feature; performing dot product operation on every two initial feature vectors of all initial feature vectors corresponding to each historical media resource group through the interaction layer in the initial network model to obtain an interaction feature vector corresponding to each historical media resource group; splicing the interaction feature vectors corresponding to each historical media resource group through the connection layer in the initial network model, and using the spliced ​​vector as the feature of each historical media resource.

[0168] Optionally, the weight sets corresponding to at least two indicators are different.

[0169] Figure 8 7 is a schematic diagram of a hardware structure of an electronic device according to an exemplary embodiment. The electronic device may include a processor 702. The processor 702 is used to execute application code to implement the method for determining the media resources to be delivered in the present disclosure; or, the processor 702 is used to execute application code to implement the method for determining the media resource processing model in the present disclosure.

[0170] The processor 702 may be a central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits for controlling the execution of the program of the disclosed solution.

[0171] like Figure 8 As shown, the electronic device may further include a memory 703. The memory 703 is used to store application code for executing the solution of the present disclosure, and is controlled to be executed by the processor 702.

[0172] The memory 703 may be a read-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM) or other types of dynamic storage devices that can store information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical disc, laser disc, optical disc, digital versatile disc, Blu-ray disc, etc.), a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory 703 may exist independently and be connected to the processor 702 via the bus 704. The memory 703 may also be integrated with the processor 702.

[0173] like Figure 8 As shown, the electronic device may further include a communication interface 701, wherein the communication interface 701, the processor 702, and the memory 703 may be coupled to each other, for example, via a bus 704. The communication interface 701 is used to exchange information with other devices, for example, to support information exchange between the electronic device and other devices.

[0174] It should be pointed out that Figure 8 The device structure shown in the figure does not constitute a limitation on the electronic device, except Figure 8 In addition to the components shown, the electronic device may include more or fewer components than shown, or combine certain components, or arrange the components differently.

[0175] In actual implementation, the functions implemented by the processing module 51 and the determining module 52 can be Figure 8 The processor 702 shown is implemented by calling the program code in the memory 703. The specific execution process can refer to the description of the method for determining the media resources to be delivered in the above method embodiment, which will not be repeated here.

[0176] In actual implementation, the functions implemented by the acquisition module 61 and the training module 62 can be Figure 8The processor 702 shown is implemented by calling the program code in the memory 703. The specific execution process can refer to the description of the determination method of the media resource processing model in the above method embodiment, which will not be repeated here.

[0177] Another embodiment of the present disclosure further provides a computer-readable storage medium, in which computer instructions are stored. When the computer instructions are executed on the electronic device, the electronic device executes the method for determining the media resources to be delivered in the method embodiment; or, when the computer instructions are executed on the electronic device, the electronic device executes the method for determining the media resource processing model in the method embodiment.

[0178] For example, the computer readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc (CD-ROM), a magnetic tape, a floppy disk, an optical data storage device, and the like.

[0179] In another embodiment of the present disclosure, a computer program product is also provided, the computer program product comprising computer instructions. When the computer instructions are executed on the electronic device, the electronic device executes the method for determining the media resource to be delivered in the method embodiment; or, when the computer instructions are executed on the electronic device, the electronic device executes the method for determining the media resource processing model in the method embodiment.

[0180] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0181] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. The present disclosure is intended to cover any variations, uses or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The description and examples are to be considered exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0182] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A method for determining media resources to be delivered, characterized in that: The determination method comprises: For each candidate media resource group in the media resource candidate set, a pre-trained media resource processing model is used to perform a processing operation to obtain the contents of at least two indicators of each candidate media resource group; the contents of the indicators are used to represent the numerical value of the delivery effect of the candidate media resource group; Wherein, each candidate media resource group includes at least one candidate media resource; The processing operation includes: obtaining the characteristics of each candidate media resource group; the characteristics include one or more of the creative type of each candidate media resource, the identifier of each candidate media resource, the historical spending sequence of each candidate media resource, and the completion rate of each candidate media resource; Input the features into multiple expert networks to obtain learning features related to at least one indicator output by each expert network; determine the content of each indicator of each candidate media resource group according to the obtained learning features and a weight set corresponding to each indicator of the at least two indicators; the weight set includes the weights of the multiple expert networks; Determine a target candidate media resource group according to the contents of at least two indicators of each candidate media resource group, and use the candidate media resources in the target candidate media resource group as the media resources to be delivered; The determining, according to the obtained learning feature and the weight set corresponding to each indicator of the at least two indicators, the content of each indicator of each candidate media resource group includes: Based on the weight set, weighted summing is performed on the obtained learning features to obtain summed features; Predicting each indicator of the summed features to obtain the content of each indicator of each candidate media resource group; The characteristics of each candidate media resource group are represented by a vector; The obtaining the feature of each candidate media resource group includes: Acquire a plurality of initial features corresponding to each candidate media resource group; Performing embedding processing on each initial feature corresponding to each candidate media resource group to obtain an initial feature vector corresponding to each initial feature; Performing a dot product operation on every two initial feature vectors of all the initial feature vectors of each candidate media resource group to obtain an interactive feature vector of each candidate media resource group; Performing splicing processing on the interactive feature vectors of each candidate media resource group, and using the spliced ​​vectors as the features of each candidate media resource; The obtaining of a plurality of initial features corresponding to each candidate media resource group includes: Acquire feature data of each candidate media resource in each candidate media resource group within a preset time period; Aggregating the feature data corresponding to each candidate media resource group to obtain a plurality of initial features corresponding to each candidate media resource group; The multiple initial features corresponding to each candidate media resource group include at least three types of initial features; The at least three types of initial features include numerical features, categorical sparse features and sequence features; the numerical features include the sum of the first consumption data of each candidate media resource group within the preset time period, the categorical sparse features include the category of each candidate media resource group within the preset time period, and the sequence features include the sum of the second consumption data of all candidate media resources in each candidate media resource group within each unit time in the preset time period.

2. The determination method according to claim 1, characterized in that: The step of determining a target candidate media resource group according to the contents of the at least two indicators of each candidate media resource group includes: Determine a candidate media resource group whose contents of at least two indicators satisfy a preset condition as the target candidate media resource group; wherein the preset condition includes that the contents of each indicator of the at least two indicators are within a preset range corresponding to each indicator; or, Determine a comprehensive index value for each first candidate media resource group, where the first candidate media resource group is a candidate media resource group whose contents of at least two indicators satisfy the preset conditions; sort all first candidate media resource groups according to the determined comprehensive index value to obtain sorted first candidate media resource groups; and determine the target candidate media resource group from the sorted first candidate media resource groups.

3. The determination method according to claim 2, characterized in that: The step of determining the comprehensive index value of each first candidate media resource group includes: Normalizing the content of each indicator of each first candidate media resource group respectively to obtain at least two values; The at least two numerical values ​​and at least two preset indicator weights are weighted and summed to obtain the comprehensive indicator value.

4. A method for determining a media resource processing model, characterized in that: The determination method comprises: Acquire the characteristics of each historical media resource in at least one historical media resource group within a first preset time period, and the contents of at least two indicators of each historical media resource group within a second preset time period; the characteristics of each historical media resource group include one or more of the creative type of each historical media resource, the identifier of each historical media resource, the historical cost sequence of each historical media resource, and the completion rate of each historical media resource; The content of the indicator is used to represent the value of the delivery effect of the historical media resource group; The time of the first preset time period is earlier than the time of the second preset time period; each of the historical media resource groups includes at least one historical media resource; Inputting the features of each historical media resource group into multiple expert networks in the initial network model; extracting initial learning features through each expert network; Obtaining association information of each indicator according to a weight set corresponding to each indicator and all initial learning features extracted by the at least one expert network; wherein the weight set includes weights of the multiple expert networks; Predicting the indicator content of the associated information of each indicator by using the prediction model corresponding to each indicator in the initial network model to obtain the predicted content of each indicator; Determine a loss value between the obtained predicted contents of the at least two indicators and the contents of the at least two indicators of each historical media resource group; Iteratively training the initial network model according to the loss value until the loss value is less than or equal to a preset threshold, and determining the trained initial network model as a media resource processing model; The characteristics of each historical media resource are represented by a vector; The obtaining of the characteristics of each historical media resource includes: Acquire a plurality of initial features corresponding to each of the historical media resource groups within the first preset time period; Inputting a plurality of initial features corresponding to each historical media resource group into a representation layer in the initial network model; Embedding each initial feature corresponding to each historical media resource group through the representation layer to obtain an initial feature vector corresponding to each initial feature; Performing a dot product operation on every two initial feature vectors of all initial feature vectors corresponding to each historical media resource group through the interaction layer in the initial network model to obtain an interaction feature vector corresponding to each historical media resource group; Performing splicing processing on the interaction feature vectors corresponding to each historical media resource group through the connection layer in the initial network model, and using the spliced ​​vectors as the features of each historical media resource; The obtaining of a plurality of initial features corresponding to each candidate media resource group includes: Acquire feature data of each candidate media resource in each candidate media resource group within a preset time period; Aggregating the feature data corresponding to each candidate media resource group to obtain a plurality of initial features corresponding to each candidate media resource group; The multiple initial features corresponding to each candidate media resource group include at least three types of initial features; The at least three types of initial features include numerical features, categorical sparse features and sequence features; the numerical features include the sum of the first consumption data of each candidate media resource group within the preset time period, the categorical sparse features include the category of each candidate media resource group within the preset time period, and the sequence features include the sum of the second consumption data of all candidate media resources in each candidate media resource group within each unit time in the preset time period.

5. The determination method according to claim 4, characterized in that: The weight sets corresponding to the at least two indicators are all different.

6. A device for determining media resources to be delivered, characterized in that: The determining device comprises: The processing module is configured to execute a processing operation for each candidate media resource group in the media resource candidate set using a pre-trained media resource processing model to obtain the contents of at least two indicators of each candidate media resource group; the contents of the indicators are used to represent the numerical value of the delivery effect of the candidate media resource group; Each candidate media resource group includes at least one candidate media resource; the processing operation includes: obtaining the characteristics of each candidate media resource group; the characteristics include one or more of the creative type of each candidate media resource, the identifier of each candidate media resource, the historical spending sequence of each candidate media resource, and the completion rate of each candidate media resource; Input the features into multiple expert networks to obtain learning features related to at least one indicator output by each expert network; determine the content of each indicator of each candidate media resource group according to the obtained learning features and a weight set corresponding to each indicator of the at least two indicators; the weight set includes the weights of the multiple expert networks; A determination module is configured to determine a target candidate media resource group according to the contents of at least two indicators of each candidate media resource group, and use the candidate media resources in the target candidate media resource group as the media resources to be delivered; The processing module is specifically configured to execute: Based on the weight set, weighted summing is performed on the obtained learning features to obtain summed features; Predicting each indicator of the summed features to obtain the content of each indicator of each candidate media resource group; The characteristics of each candidate media resource group are represented by a vector; The processing module is specifically configured to execute: Acquire a plurality of initial features corresponding to each candidate media resource group; Performing embedding processing on each initial feature corresponding to each candidate media resource group to obtain an initial feature vector corresponding to each initial feature; Performing a dot product operation on every two initial feature vectors of all the initial feature vectors of each candidate media resource group to obtain an interactive feature vector of each candidate media resource group; Performing splicing processing on the interactive feature vectors of each candidate media resource group, and using the spliced ​​vectors as the features of each candidate media resource; The processing module is specifically configured to execute: Acquire feature data of each candidate media resource in each candidate media resource group within a preset time period; Aggregating the feature data corresponding to each candidate media resource group to obtain a plurality of initial features corresponding to each candidate media resource group; The multiple initial features corresponding to each candidate media resource group include at least three types of initial features; The at least three types of initial features include numerical features, categorical sparse features and sequence features; the numerical features include the sum of the first consumption data of each candidate media resource group within the preset time period, the categorical sparse features include the category of each candidate media resource group within the preset time period, and the sequence features include the sum of the second consumption data of all candidate media resources in each candidate media resource group within each unit time in the preset time period.

7. The determination device according to claim 6, characterized in that: The determining module is specifically configured to execute: Determine a candidate media resource group whose contents of at least two indicators satisfy a preset condition as the target candidate media resource group; wherein the preset condition includes that the contents of each indicator of the at least two indicators are within a preset range corresponding to each indicator; or, Determine a comprehensive index value for each first candidate media resource group, where the first candidate media resource group is a candidate media resource group whose contents of at least two indicators satisfy the preset conditions; sort all first candidate media resource groups according to the determined comprehensive index value to obtain sorted first candidate media resource groups; and determine the target candidate media resource group from the sorted first candidate media resource groups.

8. The determination device according to claim 7, characterized in that: The determining module is specifically configured to execute: Normalizing the content of each indicator of each first candidate media resource group respectively to obtain at least two values; The at least two numerical values ​​and at least two preset indicator weights are weighted and summed to obtain the comprehensive indicator value.

9. A device for determining a media resource processing model, characterized in that: The determining device comprises: An acquisition module is configured to acquire the characteristics of each historical media resource in at least one historical media resource group within a first preset time period, and the contents of at least two indicators of each historical media resource group within a second preset time period; The characteristics of each historical media resource group include one or more of a creative type of each historical media resource, an identifier of each historical media resource, a historical spending sequence of each historical media resource, and a completion rate of each historical media resource; The content of the indicator is used to represent the value of the delivery effect of the historical media resource group; The time of the first preset time period is earlier than the time of the second preset time period; each of the historical media resource groups includes at least one historical media resource; The training module is configured to execute: Inputting the features of each historical media resource group into multiple expert networks in the initial network model; extracting initial learning features through each expert network; According to the weight set corresponding to each indicator and all the initial learning features extracted by the multiple expert networks, the association information of each indicator is obtained; wherein the weight set includes the weight of the at least one expert network; Predicting the indicator content of the associated information of each indicator by using the prediction model corresponding to each indicator in the initial network model to obtain the predicted content of each indicator; Determine a loss value between the obtained predicted contents of the at least two indicators and the contents of the at least two indicators of each historical media resource group; Iteratively training the initial network model according to the loss value until the loss value is less than or equal to a preset threshold, and determining the trained initial network model as a media resource processing model; The characteristics of each historical media resource are represented by a vector; The acquisition module is specifically configured to execute: Acquire a plurality of initial features corresponding to each of the historical media resource groups within the first preset time period; Inputting a plurality of initial features corresponding to each historical media resource group into a representation layer in the initial network model; Embedding each initial feature corresponding to each historical media resource group through the representation layer to obtain an initial feature vector corresponding to each initial feature; Performing a dot product operation on every two initial feature vectors of all initial feature vectors corresponding to each historical media resource group through the interaction layer in the initial network model to obtain an interaction feature vector corresponding to each historical media resource group; Performing splicing processing on the interaction feature vectors corresponding to each historical media resource group through the connection layer in the initial network model, and using the spliced ​​vectors as the features of each historical media resource; The processing module is specifically configured to execute: Acquire feature data of each candidate media resource in each candidate media resource group within a preset time period; Aggregating the feature data corresponding to each candidate media resource group to obtain a plurality of initial features corresponding to each candidate media resource group; The multiple initial features corresponding to each candidate media resource group include at least three types of initial features; The at least three types of initial features include numerical features, categorical sparse features and sequence features; the numerical features include the sum of the first consumption data of each candidate media resource group within the preset time period, the categorical sparse features include the category of each candidate media resource group within the preset time period, and the sequence features include the sum of the second consumption data of all candidate media resources in each candidate media resource group within each unit time in the preset time period.

10. The determination device according to claim 9, characterized in that: The weight sets corresponding to the at least two indicators are all different.

11. An electronic device, characterized in that: The electronic device comprises: a processor and a memory for storing instructions executed by the processor; The processor is configured to execute the instruction so that the electronic device executes the method for determining the media resource to be delivered according to any one of claims 1 to 3; Alternatively, the processor is configured to execute the instructions so that the electronic device performs the method for determining the media resource processing model as described in claim 4 or 5.

12. A computer-readable storage medium, characterized in that: The computer readable storage medium has computer instructions stored thereon; When the computer instruction is executed by an electronic device, the electronic device executes the method for determining the media resource to be delivered as described in any one of claims 1 to 3; Alternatively, when the computer instructions are executed by the electronic device, the electronic device executes the method for determining the media resource processing model as described in claim 4 or 5.

13. A computer program product, characterized in that The computer program product comprises computer instructions; When the computer instruction is executed by the electronic device, the method for determining the media resources to be delivered as described in any one of claims 1 to 3 is implemented; Alternatively, when the computer instructions are executed by the electronic device, the method for determining the media resource processing model as described in claim 4 or 5 is implemented.

Images