Model training method and device for resource delivery configuration and resource delivery configuration method and device

Abstract

The embodiment of the invention provides a model training method and device for resource delivery configuration and a resource delivery configuration method and device. The method comprises the steps of obtaining material data, resource niche data, user data and strategy data; the method comprises the following steps: S1, based on material data, resource niche data and user data, obtaining a plurality of prediction configuration strategies through a to-be-trained model; s2, selecting a prediction configuration strategy, and searching a target configuration strategy corresponding to a target resource niche and a target user in the selected prediction configuration strategy from the strategy data; s3, comparing the selected prediction configuration strategy with the target configuration strategy; s4, adjusting model parameters of the to-be-trained model based on a comparison result of the selected prediction configuration strategy and the target configuration strategy; and S4, repeatedly executing the steps S1 to S4 until a preset iteration ending condition is met, and taking the model parameters meeting the preset iteration ending condition as model parameters of the to-be-trained model to obtain a trained model. And finer resource delivery configuration is realized.

Description

Technical Field

[0001] The present invention relates to the field of Internet application technology, and in particular to a model training and resource delivery configuration method and device for resource delivery configuration. Background Art

[0002] With the growing number of Internet user groups, refined operations for different user groups have become an important development direction of Internet operations, and resource allocation is a key link in refined operations.

[0003] Resource delivery can be understood as: displaying promotional information as resource content on resource positions for set user groups during a set delivery period. For example, promotional information can be information related to Internet products such as games, websites, and platforms. These promotional information to be delivered can also be called materials.

[0004] The configuration of resource allocation is an important process in the resource allocation process, and achieving refined resource allocation has always been a difficult problem in this process. Summary of the invention

[0005] The purpose of the embodiment of the present invention is to provide a model training and resource placement configuration method and device for resource placement configuration, so as to achieve more refined resource placement configuration. The specific technical solution is as follows:

[0006] In a first aspect of the present invention, a model training method for resource placement configuration is provided, comprising:

[0007] Acquire material data, resource location data, user data and policy data, wherein the policy data stores the correspondence between multiple resource locations, multiple users and multiple target historical configuration policies;

[0008] S1, based on the material data, the resource location data and the user data, a plurality of prediction configuration strategies are obtained through a to-be-trained model, each prediction configuration strategy being used to represent a strategy for displaying a target material to a target user at a target resource location;

[0009] S2, selecting a predicted configuration strategy, and searching the strategy data for a target configuration strategy corresponding to the target resource position and the target user in the selected predicted configuration strategy, wherein the target configuration strategy is a configuration strategy in the historical configuration strategy whose benefits meet a preset benefit condition;

[0010] S3, comparing the selected prediction configuration strategy with the target configuration strategy;

[0011] S4, adjusting the model parameters of the model to be trained based on the comparison result between the selected prediction configuration strategy and the target configuration strategy;

[0012] S1 to S4 are repeatedly executed until a preset iteration end condition is met, and the model parameters that meet the preset iteration end condition are used as the model parameters of the model to be trained to obtain a trained model.

[0013] Optionally, obtaining the policy data includes:

[0014] Obtain pre-built policy data; the policy data is built in the following manner:

[0015] Obtain the historical configuration strategies for displaying materials to multiple users on multiple resource locations during historical time, and obtain the corresponding benefits of each historical configuration strategy;

[0016] Find the historical configuration policies for the same target resource location and the same target user;

[0017] A configuration strategy whose benefit satisfies a preset benefit condition among historical configuration strategies for the same target resource location and the same target user is selected as the target configuration strategy corresponding to the target resource location and the target user.

[0018] Optionally, selecting a configuration strategy whose benefit satisfies a preset benefit condition from historical configuration strategies for the same target resource location and the same target user as the target configuration strategy corresponding to the target resource location and the target user includes:

[0019] A historical configuration strategy with the highest benefit among historical configuration strategies for the same target resource location and the same target user is selected as the target configuration strategy corresponding to the target resource location and the target user.

[0020] Optionally, selecting a configuration strategy whose benefit satisfies a preset benefit condition from historical configuration strategies for the same target resource location and the same target user as the target configuration strategy corresponding to the target resource location and the target user includes:

[0021] Compare the benefits corresponding to each historical configuration strategy found for the same target resource location and the same target user with the preset benefits;

[0022] If the benefit corresponding to a historical configuration strategy exceeds the preset benefit, the historical configuration strategy is used as the target configuration strategy corresponding to the target resource position and the target user; or, if the benefit corresponding to a historical configuration strategy does not exceed the preset benefit, the historical configuration strategy is used as the target configuration strategy corresponding to the target resource position and the target user.

[0023] Optionally, obtaining the benefits corresponding to each historical configuration strategy includes:

[0024] According to the revenue type, preset parameters are assigned to the historical configuration strategy; and based on the preset parameters, the revenue of displaying materials to users at the resource position is calculated.

[0025] Optionally, adjusting the model parameters of the to-be-trained model based on the comparison result between the selected prediction configuration strategy and the target configuration strategy includes:

[0026] Based on the difference between the selected prediction configuration strategy and the target configuration strategy, the model parameters of the model to be trained are adjusted.

[0027] In a second aspect of the present invention, a resource placement configuration method is provided, comprising:

[0028] Obtain the materials, resource locations and users to be delivered;

[0029] The material to be delivered, the resource position and the user are input into a pre-trained model for resource delivery configuration, and a configuration strategy for displaying the material to the user at the resource position is obtained, so that resource delivery is performed based on the configuration strategy, wherein the model for resource delivery configuration is trained based on any of the model training methods for resource delivery configuration described in the first aspect above.

[0030] In a third aspect of the present invention, a model training device for resource placement configuration is provided, comprising:

[0031] A first acquisition module is used to acquire material data, resource location data, user data and policy data, wherein the policy data stores a correspondence between multiple resource locations, multiple users and multiple target historical configuration policies;

[0032] The training module is used to perform the following steps: S1, based on the material data, the resource position data and the user data, a plurality of prediction configuration strategies are obtained through the model to be trained, each prediction configuration strategy is used to represent a strategy for displaying the target material to the target user at the target resource position; S2, selecting a prediction configuration strategy, and searching the strategy data for a target configuration strategy corresponding to the target resource position and the target user in the selected prediction configuration strategy, the target configuration strategy being a configuration strategy in the historical configuration strategy whose benefits meet the preset benefit conditions; S3, comparing the selected prediction configuration strategy with the target configuration strategy; S4, adjusting the model parameters of the model to be trained based on the comparison result between the selected prediction configuration strategy and the target configuration strategy; repeatedly executing S1 to S4 until the preset iteration end condition is met, and using the model parameters that meet the preset iteration end condition as the model parameters of the model to be trained to obtain a trained model.

[0033] In a fourth aspect of the present invention, a resource placement configuration device is provided, comprising:

[0034] The second acquisition module is used to acquire the materials, resource positions and users to be delivered;

[0035] A determination module is used to input the material to be delivered, the resource position and the user into a pre-trained model for resource delivery configuration, and obtain a configuration strategy for displaying the material to the user at the resource position, so that resource delivery is performed based on the configuration strategy, wherein the model for resource delivery configuration is trained based on any of the model training methods for resource delivery configuration described in the first aspect above.

[0036] In a fifth aspect of the present invention, there is provided an electronic device, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus;

[0037] Memory, used to store computer programs;

[0038] The processor is used to implement the method steps described in any one of the first aspect or the second aspect when executing the program stored in the memory.

[0039] In another aspect of the implementation of the present invention, a computer-readable storage medium is provided, in which a computer program is stored. When the computer program is executed by a processor, it implements any of the above-mentioned model training methods or resource delivery configuration methods for resource delivery configuration.

[0040] In another aspect of the implementation of the present invention, a computer program product comprising instructions is also provided, which, when executed on a computer, enables the computer to execute any of the above-mentioned model training methods or resource delivery configuration methods for resource delivery configuration.

[0041] In the embodiment of the present invention, by acquiring material data, resource position data, user data and policy data, the corresponding relationship between multiple resource positions, multiple users and multiple target historical configuration strategies is stored in the policy data; the following steps S1 to S4 are repeatedly performed, S1, based on the material data, resource position data and user data, through the model to be trained, multiple prediction configuration strategies are obtained, each prediction configuration strategy is used to represent the strategy of displaying the target material to the target user at the target resource position; S2, select the prediction configuration strategy, and search the target configuration strategy corresponding to the target resource position and the target user in the selected prediction configuration strategy from the policy data, the target configuration strategy is the configuration strategy in the historical configuration strategy whose benefits meet the preset benefit conditions; S3, compare the selected prediction configuration strategy and the target configuration strategy; S4, based on the comparison result of the selected prediction configuration strategy and the target configuration strategy, adjust the model parameters of the model to be trained; when the preset iteration end condition is met, the model parameters under the condition of meeting the preset iteration end condition are used as the model parameters of the model to be trained, and the trained model is obtained. In this way, in the process of resource delivery, the resource delivery configuration can be performed through the trained model for resource delivery configuration.

[0042] The prediction configuration strategy output by the model to be trained is obtained based on the material data, resource position data and user data, and the target configuration strategy that meets the preset benefit conditions can be understood as the true value. In the process of training the model for resource delivery configuration, the model parameters of the model to be trained are adjusted based on the comparison results of the prediction configuration strategy output by the model to be trained and the target configuration strategy that meets the preset benefit conditions. Then, the trained model can obtain the configuration strategy based on the material data, resource position data and user data, and the target configuration strategy as the true value can guide the configuration strategy output by the model to meet the benefit conditions. In this way, in the process of resource delivery, the trained model for resource delivery configuration is used to configure resource delivery, so as to achieve a more refined resource delivery configuration. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art are briefly introduced below.

[0044] Figure 1 It is a flow chart of a model training method for resource delivery configuration in an embodiment of the present invention;

[0045] Figure 2 A flowchart of pre-building policy data in an embodiment of the present invention;

[0046] Figure 3 It is a schematic diagram of the application of the model training method for resource placement configuration in an embodiment of the present invention;

[0047] Figure 4 It is a flow chart of a resource delivery configuration method in an embodiment of the present invention;

[0048] Figure 5 A flowchart for optimizing a model for resource placement configuration in an embodiment of the present invention;

[0049] Figure 6 It is a structural schematic diagram of a model training device for resource placement configuration in an embodiment of the present invention;

[0050] Figure 7 It is a structural diagram of a resource placement configuration device in an embodiment of the present invention;

[0051] Figure 8 Schematic diagram of the structure of an electronic device in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present invention will be described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0053] Simply put, the configuration of resource delivery is to determine how to deliver resources. It can also be understood as establishing the relationship between resource locations, users, and materials.

[0054] In the embodiment of the present invention, by learning the configuration strategy in the historical delivery process, the model learns the resource position and the user to determine the material to be delivered to it. In this way, the resource delivery can be configured through the model later.

[0055] The following is a detailed description of the model training process, which involves learning the configuration strategies in the historical delivery process so that the model can learn to determine the materials to be delivered to resource locations and users.

[0056] Figure 1 A flowchart of a model training method for resource placement configuration provided by an embodiment of the present invention, referring to Figure 1 , the model training method for resource placement configuration provided by the embodiment of the present invention includes:

[0057] S101, acquiring material data, resource location data, user data and policy data, wherein the policy data stores the correspondence between multiple resource locations, multiple users and multiple target historical configuration policies;

[0058] S1, based on the material data, resource location data and user data, a plurality of prediction configuration strategies are obtained through the to-be-trained model, each prediction configuration strategy being used to represent a strategy for displaying a target material to a target user at a target resource location;

[0059] S2, selecting a forecast configuration strategy, and searching from the strategy data for a target configuration strategy corresponding to the target resource position and the target user in the selected forecast configuration strategy, the target configuration strategy being a configuration strategy in the historical configuration strategy whose benefits meet a preset benefit condition;

[0060] S3, compare the selected prediction configuration strategy and the target configuration strategy;

[0061] S4, adjusting the model parameters of the model to be trained based on the comparison result between the selected prediction configuration strategy and the target configuration strategy;

[0062] S1 to S4 are repeatedly executed until a preset iteration end condition is met, and S102 is executed: the model parameters that meet the preset iteration end condition are used as the model parameters of the model to be trained, so as to obtain a trained model.

[0063] In the embodiment of the present invention, the prediction configuration strategy output by the model to be trained is obtained based on the material data, resource position data and user data, and the target configuration strategy that meets the preset benefit condition can be understood as a true value. In the process of training the model for resource delivery configuration, the model parameters of the model to be trained are adjusted based on the comparison result of the prediction configuration strategy output by the model to be trained and the target configuration strategy that meets the preset benefit condition. Then, the trained model can obtain the configuration strategy based on the material data, resource position data and user data, and the target configuration strategy as the true value can guide the configuration strategy output by the model to meet the benefit condition. In this way, in the process of resource delivery, the trained model for resource delivery configuration is used to configure resource delivery, so as to achieve a more refined resource delivery configuration.

[0064] Moreover, compared with manual operation, using a trained model for resource allocation configuration to configure resource allocation can also lower the configuration threshold and simplify the resource allocation configuration process.

[0065] The preset iteration end conditions may include the number of iterations reaching a preset number, the model accuracy reaching a preset accuracy, etc., wherein the preset number and preset accuracy can be determined based on actual needs or experience, such as 1000, 2000, 3000 times, etc., and the preset accuracy can be 0.1, 0.01, 0.001, etc.

[0066] Among them, executing S1 to S4 once can be understood as one iteration, or executing a comparison between the predicted configuration strategy output by the model and the target configuration strategy once can be understood as one iteration. The model accuracy can be represented by the difference between the predicted configuration strategy output by the model and the target configuration strategy corresponding to the predicted configuration strategy.

[0067] For the multiple prediction configuration strategies obtained in S1, S2 can select each prediction configuration strategy in turn and execute S3 and S4 respectively. Alternatively, multiple prediction configuration strategies can be randomly selected, and S3 and S4 can be executed for each selected prediction configuration strategy. Alternatively, one of the prediction configuration strategies can be randomly selected, and S3 and S4 can be executed for the selected prediction configuration strategy.

[0068] In S4, model parameters of the model to be trained may be adjusted based on the difference between the selected prediction configuration strategy and the target configuration strategy.

[0069] The specific process of adjusting the model parameters can refer to the method of adjusting the model parameters in the related art. Among them, the selected prediction configuration strategy and target configuration strategy can be vectorized, and the loss function of the vectorized prediction configuration strategy and target configuration strategy can be calculated, and the difference between the prediction configuration strategy and the target configuration strategy can be expressed based on the loss function.

[0070] In addition to the configuration process load, the configuration of resource delivery through manual operation often relies on the knowledge and experience of the operators, which has a high threshold. If the operators do not have a high knowledge and experience, the configured resource delivery strategy may not attract users well, which will result in a low conversion rate of the delivered resources and lower overall benefits.

[0071] In the embodiment of the present invention, the configuration of resource delivery through the trained model for resource delivery configuration does not need to rely on the knowledge reserves and experience of the operating personnel, and the configuration threshold is low. The trained model obtains the configuration strategy based on the material data, resource position data and user data, and takes into account the relationship between the material, resource position and the user. The configuration strategy obtained by using the above-mentioned trained model for resource delivery configuration can also be more attractive to users, improve the conversion rate of delivered resources, and enhance the overall benefits of resource delivery.

[0072] In S101, the material data includes a plurality of materials, and the materials can also be understood as resources to be released. For example, the materials can specifically be promotional information to be released, such as information related to Internet products such as games, websites, and platforms.

[0073] The resource bit data may include a plurality of resource bits and information of the plurality of resource bits.

[0074] Resource locations, such as the startup page or splash screen of an application (Application, App), banner advertising space or interstitial advertising space on a web page, marketing space, cash register, etc.

[0075] The resource location information can be understood as all information related to the resource location, such as the resource location name, size, platform and other information.

[0076] The user data may include user portraits of multiple users, wherein the user portraits include information of at least one dimension, and the at least one dimension includes at least one of the following dimensions: user attributes and user behavior.

[0077] User attributes may include the time when the user joined the membership, and user behavior may include the user's preferences, such as the user's preference for comedy movies, etc.

[0078] The dimensions of user portraits can be divided from different focus points, and different focus points can be set according to different levels to make a more detailed portrait description of the population and conduct a more comprehensive analysis and classification.

[0079] The policy data stores the correspondence between multiple resource locations, multiple users, and multiple target historical configuration policies.

[0080] Wherein, obtaining the policy data can be achieved by obtaining the pre-built policy data. Wherein, referring to Figure 2 , the strategy data can be constructed as follows:

[0081] S201, obtaining historical configuration strategies for displaying materials to multiple users on multiple resource locations within a historical period, and obtaining revenues corresponding to each historical configuration strategy.

[0082] Revenue can be determined through a variety of evaluation indicators, such as click-through rate, checkout counter entry rate, order conversion rate, etc.

[0083] It can be understood that the revenue includes one or more of the following revenue types: click-through rate, checkout counter entry rate, and order conversion rate.

[0084] When multiple types of returns are included, in order to facilitate comparison between different types of returns, one possible implementation method for obtaining the returns corresponding to each historical configuration strategy may include:

[0085] According to the revenue type, preset parameters are assigned to the historical configuration strategy; and based on the preset parameters, the revenue of displaying materials to users at the resource position is calculated.

[0086] Preset parameters corresponding to different income types can be preset. After obtaining the income type corresponding to the historical configuration strategy, preset parameters can be allocated to the historical configuration strategy according to the income type. The preset parameters can be determined according to needs or experience.

[0087] S202, searching for historical configuration policies for the same target resource location and the same target user;

[0088] S203 , selecting a configuration strategy whose benefit satisfies a preset benefit condition from historical configuration strategies for the same target resource location and the same target user as the target configuration strategy corresponding to the target resource location and the target user.

[0089] Figure 2 The process shown can also be understood as a process of processing multiple historical configuration strategies for recommending materials to users at resource locations within historical time.

[0090] Among them, the target user can be selected through the user portrait in the above user data. Each user can be divided from multiple dimensions and labeled accordingly. In this way, in the process of selecting target users, selection can be made through labels. For example, the marketing party hopes to divide different types of users who are interested in different types of video content and recommend corresponding materials to each of them. Then, users who are interested in the same type of video content are understood as a type of target users. Users whose user portraits are marked as being interested in this type of video content can be selected from multiple users as target users of this type. The specified users of this type can be marked with the same user ID or user tag, and users marked with the same user ID or user tag are all the same target users.

[0091] In an optional embodiment, the historical configuration strategy with the highest benefit among the historical configuration strategies for the same target resource location and the same target user is selected as the target configuration strategy corresponding to the target resource location and the target user.

[0092] As shown in the following Table 1, 6 historical configuration policies are obtained, namely, historical configuration policy 1, historical configuration policy 2, historical configuration policy 3, historical configuration policy 4, historical configuration policy 5, and historical configuration policy 6.

[0093] Table 1

[0094] The number of the multiple historical configuration policies obtained may be hundreds, thousands, tens of thousands, etc. The example here is only for the convenience of explanation.

[0095] For the historical configuration strategies for displaying materials to multiple users at multiple resource locations in the historical time obtained as shown in Table 1, and the corresponding benefits of each historical configuration strategy obtained, and searching for historical configuration strategies for the same target resource location and the same target user, the results are as follows:

[0096] For the same target resource location: location 1 and the same target user: the historical configuration policy for the user with the user tag of user tag 1 includes 3 items: historical configuration policy 1, historical configuration policy 2 and historical configuration policy 3.

[0097] For the same target resource location: location 2 and the same target user: the historical configuration of the user with the user tag of user tag 2 includes 1: historical configuration policy 4.

[0098] For the same target resource location: location 3 and the same target user: the historical configuration policy for the user with the user tag of user tag 3 includes 1: historical configuration policy 5.

[0099] For the same target resource location: location 4 and the same target user: the historical configuration policy for the user with the user tag of user tag 4 includes 1: historical configuration policy 6.

[0100] For the historical configuration policies of the same target resource position: position 2 and the same target user: the user tag of user tag 2, the historical configuration policies of the same target resource position: position 3 and the same target user: the user tag of user tag 3, and the historical configuration policies of the same target resource position: position 4 and the same target user: the user tag of user tag 4, each includes only one policy, and the policy with the highest benefit can only be this policy. Therefore, this policy is directly used as the target configuration policy corresponding to the target resource position and the target user. That is, historical configuration policy 4 is used as the target configuration policy for the target resource position: position 2 and the target user: the user tag of user tag 2; historical configuration policy 5 is used as the target configuration policy for the target resource position: position 3 and the target user: the user tag of user tag 3; historical configuration policy 6 is used as the target configuration policy for the target resource position: position 4 and the target user: the user tag of user tag 4.

[0101] For the target resource position: position 1 and the target user: the historical configuration strategy for the user with the user tag of user tag 1 includes three: historical configuration strategy 1, historical configuration strategy 2 and historical configuration strategy 3. The benefits of these three are compared, and the historical configuration strategy 3 with the highest benefit is selected as the target configuration strategy corresponding to the target resource position: position 1 and the target user: the user tag of user tag 1.

[0102] In one case, the same configuration strategy may be displayed multiple times, that is, the same material is displayed to the same user group at the same resource position during multiple displays. This situation can be understood as the situation where the target material is recommended to the target user at the target resource position multiple times in the historical time.

[0103] In this case, the historical recommended materials in multiple historical configuration policies for the same target resource location and the same target user may be the same.

[0104] In one possible implementation, each historical configuration policy with the same historical recommendation material for the same target resource location and the same target user can be treated as a separate historical configuration policy and put together with other historical recommendation materials with different historical recommendation materials for the same target resource location and the same target user.

[0105] In another possible implementation, each historical configuration strategy for the same target resource position and the same target user, which has the same historically recommended material, can be integrated, and the benefits of these historical configuration strategies can be integrated accordingly. Specifically, if there are multiple times in the historical time when the target material is recommended to the target user at the target resource position, a preset ratio can be allocated for each recommendation; and according to the preset ratio allocated for each recommendation, the benefit of recommending the target material to the target user at the target resource position is calculated as the benefit of the historical configuration strategy including the association relationship between the target resource position, the target user and the target material.

[0106] In the embodiment of the present invention, the historical configuration strategy with the highest profit among the historical configuration strategies for the same target resource position and the same designated user can be selected as the target historical configuration strategy corresponding to the target resource position and the target user, and based on the comparison result of the predicted configuration strategy output by the model to be trained and the target configuration strategy that meets the preset profit condition, the model parameters of the model to be trained are adjusted so that the model learns how to obtain the target material that meets the profit condition for the target resource position and the target user, so that the configuration strategy obtained by using the model can better attract users. Furthermore, resource delivery based on the configuration strategy obtained by using the model can improve the conversion rate of delivered resources and increase profits.

[0107] In another optional embodiment, the benefits corresponding to each historical configuration strategy found for the same target resource location and the same target user are compared with the preset benefits; if the benefits corresponding to a historical configuration strategy exceed the preset benefits, the historical configuration strategy is used as the target configuration strategy corresponding to the target resource location and the target user; or, if the benefits corresponding to a historical configuration strategy do not exceed the preset benefits, the historical configuration strategy is used as the target configuration strategy corresponding to the target resource location and the target user.

[0108] Among them, the preset income can be determined based on actual needs or experience.

[0109] If the return corresponding to a historical allocation strategy exceeds the preset return, the historical allocation strategy can also be understood as a positive target allocation strategy.

[0110] In this case, the model parameters of the model to be trained are adjusted in the hope that the correspondence between the resource positions, users, and configuration policies indicated by the predicted configuration strategy is the same as the correspondence between the resource positions, users, and configuration policies indicated by the target configuration strategy; or, the model parameters are adjusted so that the correspondence between the resource positions, users, and configuration policies indicated by the predicted configuration strategy output by the model is infinitely close to the correspondence between the resource positions, users, and configuration policies indicated by the target configuration strategy.

[0111] If the return corresponding to a historical allocation strategy does not exceed the preset return, the historical allocation strategy can also be understood as a negative target allocation strategy.

[0112] In this case, the model parameters are adjusted so that the corresponding relationship between the resource position, user, and configuration policy indicated by the predicted configuration policy is different from the corresponding relationship between the resource position, user, and configuration policy indicated by the target configuration policy.

[0113] In an embodiment of the present invention, feature extraction can be performed on material data, resource bit data, user data and policy data to obtain feature data corresponding to the material data, resource bit data, user data and policy data, respectively, wherein the feature extraction method can be selected according to actual needs, for example, one-hot encoding can be performed. In one achievable method, feature extraction can specifically be feature embedding.

[0114] In S1, based on the material data, resource position data and user data, a plurality of prediction configuration strategies are obtained through the model to be trained, which may include: feature data corresponding to the material data, resource position data, user data and strategy data, respectively, and a plurality of prediction configuration strategies are obtained through the model to be trained. Specifically, the feature data corresponding to the material data, resource position data, user data and strategy data, respectively, may be input into the model to be trained, and the model to be trained outputs a plurality of prediction configuration strategies.

[0115] In one implementation, the model to be trained may include a low-order feature processing module and a high-order feature processing module.

[0116] Low-order feature processing modules may include multi-head self-attention mechanisms (Multi-head SelfAttention); high-order feature processing modules may include deep neural networks (DNN) and deep residual networks (ResNet).

[0117] In S1, the feature data corresponding to the material data, resource location data, user data and strategy data are respectively input into the model to be trained, and the low-order feature processing module and the high-order feature processing module included in the model to be trained obtain low-order processing results and high-order processing results respectively; the low-order processing results and the high-order processing results are cascaded, and the prediction configuration strategy is output based on the cascaded results.

[0118] For example: based on the feature data corresponding to the material library, resource location library, and crowd template library, the low-order feature processing module and the high-order feature processing module in the model to be trained are used to obtain low-order processing results and high-order processing results respectively; the low-order processing results and the high-order processing results are cascaded, and a prediction configuration strategy is output based on the cascaded results.

[0119] Among them, the low-order feature processing module is used to process low-order features, and the high-order feature processing module is used to process high-order features. In the embodiment of the present invention, low-order features refer to features with fewer bits after feature embedding, and high-order features refer to features with more bits after feature embedding. For example, a bit threshold can be preset. If the number of bits obtained by feature embedding is less than the preset bit threshold, the feature data obtained by feature embedding is a low-order feature, and if the number of bits obtained by feature embedding is not less than the preset bit threshold, the feature data obtained by feature embedding is a high-order feature.

[0120] The cascading may be a weighted sum or the like.

[0121] In an optional embodiment, material data may be maintained through a material library, resource location data may be maintained through a resource location library, user data may be maintained through a crowd template library, and policy data may be maintained through a policy library.

[0122] In the embodiment of the present invention, a material library, a resource location library, a crowd template library and a strategy library may be pre-constructed, and the pre-constructed material library, resource location library, crowd template library and strategy library may be acquired in S101.

[0123] The material library may include a bottom-level material library and a variable material library.

[0124] The underlying material library can be understood as the materials provided by the marketer that do not change with different users, and the variable material library can be understood as the materials that change with different users.

[0125] In one possible implementation, texts, pictures, links and other forms of materials in the historical material library are added as the underlying material library, and data such as delivery with an order and red envelopes are added as the variable material library.

[0126] In the embodiment of the present invention, the materials such as pictures, links, texts, red envelopes, etc. contained in the material library can be managed to achieve full life cycle management from material production to material uploading, use and destruction.

[0127] The Crowd Template Library includes user portraits from multiple users.

[0128] Specific groups of people can be selected based on user attributes, behaviors and other labels. In the configuration background, specific groups of people can be quickly selected through the group template library. The group template library can be used to improve operational efficiency and achieve rapid selection.

[0129] The resource location library, which can also be called the location library, location material library, etc., can specifically include information of multiple resource locations, such as the resource location name, size, platform, etc. The resource location library can be used to manage the front-end resource locations in the background.

[0130] Strategy library: for single or batch locations (resource locations), implement material association configuration for specified groups of people and locations.

[0131] It is possible to crawl multiple historical configuration strategies for recommending materials to users at resource locations over historical time and collect them into a strategy library.

[0132] In one possible implementation, a policy library including multiple historical configuration policies for recommending materials to users at resource locations in historical time can be referred to as an initial policy library. In the embodiment of the present invention, the data in the initial policy library can be processed so that the processed data can be directly used in the model training process; or, the data in the initial policy library can be processed in the model training process, wherein the processing of the data in the initial policy library refers to the above Figure 2 The method shown.

[0133] Combine the following Figure 3 , an example is given of an embodiment of the present invention, in which a configuration strategy in a historical delivery process is learned to learn resource locations and a user can determine the material to be delivered thereto.

[0134] The embodiments of the present invention can be applied to marketing system scenarios, mainly including:

[0135] One-hot feature domain embedding, feature embedding, intelligent algorithm and output layer.

[0136] One-hot feature domain embedding can also be understood as data collection, including: material library, resource library (also called location library), crowd template library and strategy library.

[0137] This process can also be called material library production and crowd portrait description.

[0138] Material library production: add texts, pictures, links and other forms of materials in the historical material library as the underlying material library, add data such as delivery with an order and red envelopes as the variable material library, and add data such as marketing positions and cash registers as the location material library.

[0139] Crowd portrait description: Make a more detailed portrait description of the crowd, set user portraits with different focus points according to different levels, and conduct more comprehensive analysis and classification.

[0140] Feature embedding, that is, encoding the above collected data.

[0141] Intelligent algorithm recommendation: One-hot encode the heterogeneous data of the material library and the crowd, convert high-order sparse features into low-order dense features to form the original data set; perform computational analysis on feature interactions to recommend more suitable materials to different groups of people.

[0142] Output layer: Based on the feature data corresponding to the material library, resource location library, and crowd template library, the low-order feature processing module and the high-order feature processing module in the model to be trained are used to obtain low-order processing results and high-order processing results; the low-order processing results and the high-order processing results are cascaded, and the prediction configuration strategy is output based on the cascaded results.

[0143] Specifically, the algorithm is helped to optimize learning according to the optimal goal. Specifically, the configuration strategy whose benefits meet the preset benefit conditions among the historical configuration strategies for the same target resource position and the same target user is selected as the target configuration strategy corresponding to the target resource position and the target user. The target configuration strategy can be understood as the optimal goal.

[0144] Obtain material data, resource location data, user data and policy data, obtain multiple prediction configuration strategies based on the material data, resource location data and user data through the model to be trained, select the prediction configuration strategy, and search the policy data for the target configuration strategy corresponding to the target resource location and target user in the selected prediction configuration strategy, compare the selected prediction configuration strategy with the target configuration strategy, and adjust the model parameters of the model to be trained based on the comparison result of the selected prediction configuration strategy and the target configuration strategy until the preset iteration end condition is met, and use the model parameters that meet the preset iteration end condition as the model parameters of the model to be trained to obtain a trained model.

[0145] Effect analysis: Set different indicators, including click-through rate, checkout counter entry rate, order conversion rate, etc.; set different indicators at different marketing nodes to achieve different focused recommendations and achieve more reasonable intelligent traffic delivery.

[0146] The configuration of strategies and materials is complex and relies on high-threshold prior knowledge, and it is difficult to close the loop. When handed over to the algorithm and the goal is maximized, the strategy of global traffic can be intelligently generated to avoid the loss of traffic.

[0147] Compared with the prior art, the strategy configuration process requires manual operation, is repetitive and cumbersome, and has a large number of material types; and often relies on the knowledge reserve of the operation, with a high threshold; once the configured strategy does not cover the global traffic, it will cause part of the data traffic to be directly lost, making it difficult to close the loop. The embodiment of the present invention proposes a traffic intelligent delivery method based on global benefits, which simplifies the configuration, lowers the threshold, breaks down information barriers, and improves the overall rate of return.

[0148] Among them, based on global benefits, it can be understood as providing a more appropriate recommendation strategy to better attract users, obtain benefits from users in a more comprehensive dimension, and improve overall benefits.

[0149] In the embodiment of the present invention, multiple materials are managed in a unified manner, and an intelligent recommendation algorithm is used to display different strategies for users with different concerns at different resource locations to avoid traffic loss.

[0150] The above embodiment illustrates that by learning the configuration strategy in the historical delivery process, the model learns to determine the material to be delivered to the resource position and user. In this way, the resource delivery can be configured through the model in the future. The process of configuring the resource delivery through the model in the embodiment of the present invention is described below.

[0151] Reference Figure 4 , an embodiment of the present invention provides a resource delivery configuration method, which may include:

[0152] S401, obtaining materials, resource locations and users to be delivered;

[0153] S402, input the materials to be delivered, resource positions and users into a pre-trained model for resource delivery configuration, and obtain a configuration strategy for displaying materials to users at the resource positions, so that resources are delivered based on the configuration strategy, wherein the model for resource delivery configuration is trained based on the model training method for resource delivery configuration in the above-mentioned embodiment.

[0154] In one possible implementation, after receiving a request triggered by a user, resource delivery configuration is performed to recommend the material indicated by the strategy to the user, that is, to display the material to the user at the resource position according to the recommendation strategy. For example, the operation of a user opening the homepage of an APP can be understood as the user triggering a request.

[0155] Furthermore, after resource deployment using the resource deployment strategy configured by the model, the deployment data such as revenue can be recovered, and the model can be further optimized based on the recovered deployment data. Figure 5 As shown, this may include:

[0156] S501, collecting delivery data, where the delivery data includes the benefits of the configuration strategy after resources are delivered based on the configuration strategy;

[0157] S502: Optimize the model for resource delivery configuration based on the delivery data.

[0158] Specifically, the material library, strategy library, crowd template library, and resource location library can be updated based on the delivery data. The specific optimization process is similar to the model training process.

[0159] In an embodiment of the present invention, after resource allocation is performed using the resource allocation strategy configured by the model, allocation data such as revenue can be recovered, and then the model for resource allocation configuration obtained through the above training can be further optimized based on the recovered allocation data. In this way, the accuracy of the model can be further improved, making the accuracy of resource allocation based on the model for resource allocation configuration higher.

[0160] In the embodiment of the present invention, the intelligent recommendation algorithm learns the big data of the historical placement of materials and strategies for the crowds at the location, and based on the optimization goal of the operation, it realizes the best placement strategy for each marketing position. In addition, after the placement, the data is recycled, and the algorithm continues to learn and iterate, repeating the cycle and continuously optimizing, and finally achieving greater benefits. At the same time, it has achieved an excellent leap from "manual operation, cumbersome and complicated" to "intelligent recommendation, iterative optimization".

[0161] Corresponding to the model training method for resource placement configuration provided in the above embodiment, the embodiment of the present invention also provides a model training device for resource placement configuration, such as Figure 6 As shown, including:

[0162] The first acquisition module 601 is used to acquire material data, resource location data, user data and policy data, wherein the policy data stores the correspondence between multiple resource locations, multiple users and multiple target historical configuration policies;

[0163] The training module 602 is used to perform the following steps: S1, based on the material data, resource position data and user data, a plurality of prediction configuration strategies are obtained through the model to be trained, each prediction configuration strategy is used to represent a strategy for displaying the target material to the target user at the target resource position; S2, selecting a prediction configuration strategy, and searching the strategy data for a target configuration strategy corresponding to the target resource position and the target user in the selected prediction configuration strategy, the target configuration strategy being a configuration strategy in the historical configuration strategy whose benefits meet the preset benefit conditions; S3, comparing the selected prediction configuration strategy with the target configuration strategy; S4, adjusting the model parameters of the model to be trained based on the comparison result between the selected prediction configuration strategy and the target configuration strategy; repeatedly executing S1 to S4 until the preset iteration end condition is met, and using the model parameters that meet the preset iteration end condition as the model parameters of the model to be trained to obtain a trained model.

[0164] The first acquisition module 601 is specifically used to: obtain pre-constructed policy data; wherein the policy data is constructed in the following manner: obtain historical configuration strategies for displaying materials to multiple users at multiple resource locations within a historical period, and obtain the benefits corresponding to each historical configuration strategy; search for historical configuration strategies for the same target resource location and the same target user; select a configuration strategy whose benefits meet a preset benefit condition from the historical configuration strategies for the same target resource location and the same target user as the target configuration strategy corresponding to the target resource location and the target user.

[0165] The training module 602 is specifically used to select the historical configuration strategy with the highest benefit among the historical configuration strategies for the same target resource location and the same target user as the target configuration strategy corresponding to the target resource location and the target user.

[0166] Training module 602 is specifically used to compare the benefits corresponding to each historical configuration strategy found for the same target resource location and the same target user with the preset benefits; if the benefit corresponding to a historical configuration strategy exceeds the preset benefit, the historical configuration strategy is used as the target configuration strategy corresponding to the target resource location and the target user; or, if the benefit corresponding to a historical configuration strategy does not exceed the preset benefit, the historical configuration strategy is used as the target configuration strategy corresponding to the target resource location and the target user.

[0167] The training module 602 is specifically used to allocate preset parameters to the historical configuration strategy according to the revenue type; and based on the preset parameters, calculate the revenue of displaying the material to the user at the resource position.

[0168] The training module 602 is specifically used to adjust the model parameters of the model to be trained based on the difference between the selected prediction configuration strategy and the target configuration strategy.

[0169] Corresponding to the resource placement configuration method provided in the above embodiment, the embodiment of the present invention also provides a resource placement configuration device, such as Figure 7 As shown, including:

[0170] The second acquisition module 701 is used to acquire the material, resource position and user to be delivered;

[0171] Determine module 702, which is used to input the materials, resource positions and users to be delivered into a pre-trained model for resource delivery configuration, and obtain a configuration strategy for displaying the materials to the users at the resource positions, so as to perform resource delivery based on the configuration strategy, wherein the model for resource delivery configuration is trained based on the model training method for resource delivery configuration in the above-mentioned embodiment.

[0172] The embodiment of the present invention further provides an electronic device, such as Figure 8 As shown, it includes a processor 801, a communication interface 802, a memory 803 and a communication bus 804, wherein the processor 801, the communication interface 802, and the memory 803 communicate with each other through the communication bus 804;

[0173] Memory 803, used for storing computer programs;

[0174] The processor 801 is used to implement the method steps of the above-mentioned model training method or resource placement configuration method when executing the program stored in the memory 803.

[0175] The communication bus mentioned in the above terminal can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.

[0176] The communication interface is used for communication between the above terminal and other devices.

[0177] The memory may include a random access memory (RAM) or a non-volatile memory, such as at least one disk memory. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.

[0178] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0179] In another embodiment provided by the present invention, a computer-readable storage medium is also provided, in which a computer program is stored. When the computer program is executed by a processor, the model training method or resource delivery configuration method for resource delivery configuration described in any of the above embodiments is implemented.

[0180] In another embodiment provided by the present invention, a computer program product comprising instructions is also provided. When the computer is run on a computer, the computer executes the model training method or resource delivery configuration method for resource delivery configuration described in any of the above embodiments.

[0181] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented by software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website site, computer, server or data center to another website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive Solid State Disk (SSD)), etc.

[0182] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0183] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device, electronic device, computer-readable storage medium, and computer program product embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0184] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A model training method for resource placement configuration, characterized in that: include: Acquire material data, resource location data, user data and policy data, wherein the policy data stores the correspondence between multiple resource locations, multiple users and multiple target historical configuration policies; S1, based on the material data, the resource location data and the user data, a plurality of prediction configuration strategies are obtained through a to-be-trained model, each prediction configuration strategy being used to represent a strategy for displaying a target material to a target user at a target resource location; S2, selecting a predicted configuration strategy, and searching the strategy data for a target configuration strategy corresponding to the target resource position and the target user in the selected predicted configuration strategy, wherein the target configuration strategy is a configuration strategy in the historical configuration strategy whose benefits meet a preset benefit condition; S3, comparing the selected prediction configuration strategy with the target configuration strategy; S4, adjusting the model parameters of the model to be trained based on the comparison result between the selected prediction configuration strategy and the target configuration strategy; S1 to S4 are repeatedly executed until a preset iteration end condition is met, and the model parameters that meet the preset iteration end condition are used as the model parameters of the model to be trained to obtain a trained model.

2. The method according to claim 1, characterized in that Obtaining the policy data includes: Obtain pre-built policy data; the policy data is built in the following manner: Obtain the historical configuration strategies for displaying materials to multiple users on multiple resource locations during historical time, and obtain the corresponding benefits of each historical configuration strategy; Find the historical configuration policies for the same target resource location and the same target user; A configuration strategy whose benefit satisfies a preset benefit condition among historical configuration strategies for the same target resource location and the same target user is selected as the target configuration strategy corresponding to the target resource location and the target user.

3. The method according to claim 2, characterized in that The selecting, from among the historical configuration strategies for the same target resource location and the same target user, a configuration strategy whose benefits satisfy a preset benefit condition as the target configuration strategy corresponding to the target resource location and the target user includes: A historical configuration strategy with the highest benefit among historical configuration strategies for the same target resource location and the same target user is selected as the target configuration strategy corresponding to the target resource location and the target user.

4. The method according to claim 2, characterized in that: The selecting, from among the historical configuration strategies for the same target resource location and the same target user, a configuration strategy whose benefits satisfy a preset benefit condition as the target configuration strategy corresponding to the target resource location and the target user includes: Compare the benefits corresponding to each historical configuration strategy found for the same target resource location and the same target user with the preset benefits; If the benefit corresponding to a historical configuration strategy exceeds the preset benefit, the historical configuration strategy is used as the target configuration strategy corresponding to the target resource position and the target user; or, if the benefit corresponding to a historical configuration strategy does not exceed the preset benefit, the historical configuration strategy is used as the target configuration strategy corresponding to the target resource position and the target user.

5. The method according to claim 2, characterized in that: The obtaining of the benefits corresponding to each historical configuration strategy includes: According to the revenue type, preset parameters are assigned to the historical configuration strategy; and based on the preset parameters, the revenue of displaying materials to users at the resource position is calculated.

6. The method according to any one of claims 1 to 5, characterized in that: The step of adjusting the model parameters of the model to be trained based on the comparison result between the selected prediction configuration strategy and the target configuration strategy includes: Based on the difference between the selected prediction configuration strategy and the target configuration strategy, the model parameters of the model to be trained are adjusted.

7. A resource placement configuration method, characterized in that: include: Obtain the materials, resource locations and users to be delivered; The material to be delivered, the resource position and the user are input into a pre-trained model for resource delivery configuration, and a configuration strategy for displaying the material to the user at the resource position is obtained, so that resource delivery is performed based on the configuration strategy, wherein the model for resource delivery configuration is trained based on the model training method for resource delivery configuration described in any one of claims 1 to 6 above.

8. A model training device for resource deployment configuration, characterized in that: include: A first acquisition module is used to acquire material data, resource location data, user data and policy data, wherein the policy data stores a correspondence between multiple resource locations, multiple users and multiple target historical configuration policies; The training module is used to perform the following steps: S1, based on the material data, the resource position data and the user data, a plurality of prediction configuration strategies are obtained through the model to be trained, each prediction configuration strategy is used to represent a strategy for displaying the target material to the target user at the target resource position; S2, selecting a prediction configuration strategy, and searching the strategy data for a target configuration strategy corresponding to the target resource position and the target user in the selected prediction configuration strategy, the target configuration strategy being a configuration strategy in the historical configuration strategy whose benefits meet the preset benefit conditions; S3, comparing the selected prediction configuration strategy with the target configuration strategy; S4, adjusting the model parameters of the model to be trained based on the comparison result between the selected prediction configuration strategy and the target configuration strategy; repeatedly executing S1 to S4 until the preset iteration end condition is met, and using the model parameters that meet the preset iteration end condition as the model parameters of the model to be trained to obtain a trained model.

9. A resource placement configuration device, characterized in that: include: The second acquisition module is used to acquire the materials, resource positions and users to be delivered; A determination module is used to input the material to be delivered, the resource position and the user into a pre-trained model for resource delivery configuration, and obtain a configuration strategy for displaying the material to the user at the resource position, so that resource delivery is performed based on the configuration strategy, wherein the model for resource delivery configuration is trained based on the model training method for resource delivery configuration described in any one of claims 1 to 6 above.

10. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, for implementing the method steps described in any one of claims 1 to 7 when executing a program stored in a memory.

11. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method steps described in any one of claims 1 to 7 are implemented.

Images