Media resource pushing method and device, storage medium and electronic equipment

Abstract

The invention provides a media resource pushing method and device, a storage medium and electronic equipment. The method comprises the steps of determining an operation distribution characteristic of a target object account according to a time distribution condition of account operation execution of the target object account in a first period; a first operation prediction result and a second operation prediction result are determined according to the operation distribution characteristics, and the first operation prediction result indicates the probability that the target object account executes the target account operation under the condition that the target media resource is pushed to the target object account; the second operation prediction result indicates the probability that the target object account executes the target account operation under the condition that the target media resource is not pushed to the target object account; and under the condition that an operation evaluation index determined according to the first operation prediction result and the second operation prediction result is greater than or equal to a target threshold value, pushing the target media resource to the target object account in a second period. The technical problem that a related pushing method is not accurate is solved.

Description

Technical Field

[0001] This invention relates to the field of computers, and more specifically, to a method and apparatus for pushing media resources, a storage medium, and an electronic device. Background Technology

[0002] In media resource push scenarios, a correlation score is usually obtained for each object account based on the Response Model. This score indicates the relationship between the media resource push behavior and the account operation behavior of the object account. Media resources are then pushed to specific object accounts based on the correlation score.

[0003] However, the correlation scores output by existing feedback models only indicate the correlation between the system's media resource push behavior and specific operational behaviors of target accounts. In reality, some target accounts may spontaneously perform related specific operational behaviors even without receiving pushed media resources. The correlation scores output by the aforementioned feedback models cannot accurately identify and explain these spontaneous operational behaviors. In other words, based on the correlation scores output by the aforementioned feedback models, it is impossible to accurately identify target accounts from the set of target accounts that are truly affected by the aforementioned media resource push behavior. Therefore, push operations performed based on the correlation scores output by the aforementioned feedback models cannot push media resources to target accounts that are easily affected by the media resource push behavior. In other words, existing media resource push methods suffer from inaccurate push technology.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This invention provides a method and apparatus for pushing media resources, a storage medium, and an electronic device, to at least solve the technical problem of inaccurate pushing in related pushing methods.

[0006] According to one aspect of the present invention, a method for pushing media resources is provided, comprising: determining operation distribution characteristics of a target account based on the time distribution of account operations performed by the target account within a first period; determining a first operation prediction result and a second operation prediction result based on the operation distribution characteristics, wherein the first operation prediction result indicates the probability that the target account will perform a target account operation when a target media resource is pushed to the target account, and the second operation prediction result indicates the probability that the target account will perform a target account operation when the target media resource is not pushed to the target account; and pushing the target media resource to the target account within a second period if the operation evaluation index determined based on the first operation prediction result and the second operation prediction result is greater than or equal to a target threshold.

[0007] According to another aspect of the present invention, a media resource push device is also provided, comprising: a first determining unit, configured to determine the operation distribution characteristics of the target account based on the time distribution of account operations performed by the target account within a first period; a second determining unit, configured to determine a first operation prediction result and a second operation prediction result based on the operation distribution characteristics, wherein the first operation prediction result indicates the probability that the target account will perform a target account operation when the target media resource is pushed to the target account, and the second operation prediction result indicates the probability that the target account will perform the target account operation when the target media resource is not pushed to the target account; and a push unit, configured to push the target media resource to the target account within a second period if the operation evaluation index determined based on the first operation prediction result and the second operation prediction result is greater than or equal to a target threshold.

[0008] Optionally, the second determining unit includes: an acquisition module for acquiring reference account features associated with the target account; a splicing module for splicing the reference account features with the operation distribution features to obtain target account features, and inputting the result into the feature sharing module of the target prediction model to obtain a shared feature vector; a first input module for inputting the shared feature vector into the first prediction module of the target prediction model to obtain a first operation prediction result; and a second input module for inputting the shared feature vector into the second prediction module of the target prediction model to obtain a second operation prediction result, wherein the first prediction module and the second prediction module are two prediction modules in the target prediction model that are respectively connected to the feature sharing module.

[0009] Optionally, the acquisition module is configured to: acquire operation description features that indicate the target account to perform the account operation within the first period; acquire account description features that match the target account; and use the operation description features and the account description features as the reference account features.

[0010] Optionally, the first determining unit is configured to: obtain the login time corresponding to each of the multiple login operations performed by the target account in the first period; if the login time includes multiple time parameters, determine the time feature vector corresponding to each of the multiple time parameters according to the multiple time parameters and the time period matched by each of the multiple time parameters; and concatenate the multiple time feature vectors to obtain the operation distribution feature.

[0011] Optionally, the first determining unit is configured to: obtain the current time parameter and the current time period matching the current time parameter from among the plurality of time parameters; determine the ratio between the current time parameter and the current time period as the login distribution coefficient; obtain a first encoding result of the login distribution coefficient according to a sine function, and obtain a second encoding result of the login distribution coefficient according to a cosine function; and determine the first encoding result and the second encoding result as vector feature values ​​included in the current time feature vector.

[0012] Optionally, the aforementioned media resource push device further includes: a training unit, configured to acquire sample account features of sample object accounts and sample tags matching the sample object accounts; input the sample account features into an initial feature sharing module in an initial prediction model under training to obtain a shared sample feature vector; input the shared sample feature vector into at least one target prediction module among a plurality of initial prediction modules in the initial prediction model, wherein the initial feature sharing module in the initial prediction model is connected to the plurality of initial prediction modules respectively, and the plurality of initial prediction modules includes a first initial prediction module, a second initial prediction module, and a third initial prediction module, wherein the first initial prediction module outputs... The first sample operation prediction result indicates the probability that the sample object account will perform a sample account operation when sample media resources are pushed to the sample object account. The second sample operation prediction result output by the second initial prediction module indicates the probability that the sample object account will perform the sample account operation when sample media resources are not pushed to the sample object account. The tendency prediction result output by the third initial prediction module indicates the learning tendency of the initial prediction model during the training phase. Based on the prediction results output by at least one of the target prediction modules and the sample labels, the initial prediction model is trained until the target prediction model that meets the target training conditions is obtained.

[0013] Optionally, the training unit is configured to: when the sample account features are the account features of a first sample account, input the shared sample feature vector into the first initial prediction module and the third initial prediction module of the initial prediction model to obtain the first sample operation prediction result output by the first initial prediction module and the tendency prediction result output by the third initial prediction module, wherein the first sample account is the object account that received the pushed sample media resource; when the sample account features are the account features of a second sample account, input the shared sample feature vector into the second initial prediction module and the third initial prediction module of the initial prediction model to obtain the second sample operation prediction result output by the second initial prediction module and the tendency prediction result output by the third initial prediction module, wherein the second sample account is the object account that did not receive the sample media resource.

[0014] Optionally, the training unit is configured to: determine a first sample set and a second sample set, wherein the difference between the first number of first sample accounts included in the first sample set and the second number of second sample accounts included in the second sample set is less than or equal to a target number threshold; push the sample media resources to the first sample accounts in the first sample set; and determine the sample tags corresponding to each of the multiple sample object accounts based on the account operations performed by the multiple sample object accounts in the first sample set and the second sample set during the target account operation period, including: configuring a first sample tag for the sample object account when the sample object account has performed the sample account operation, and configuring a second sample tag for the sample object account when the sample object account has not performed the sample account operation.

[0015] Optionally, the aforementioned push unit is configured to: obtain the quantity difference between the first push quantity and the second push quantity associated with the target media resource in the previous push cycle, wherein the first push quantity is the number of target media resources that have been pushed, and the second push quantity is the number of target media resources that are planned to be pushed; if the quantity difference is greater than or equal to a target value, raise the reference threshold to the target threshold; if the quantity difference is less than the target value, lower the reference threshold to the target threshold.

[0016] Optionally, the aforementioned push unit is configured to: determine the product between the aforementioned quantity difference and the threshold revision coefficient as the target adjustment value; and adjust the aforementioned reference threshold to the aforementioned target threshold based on the aforementioned target adjustment value.

[0017] Optionally, the aforementioned push unit is further configured to: when the push stage corresponding to the current push cycle is the target stage, obtain a push control model matching the next push cycle, wherein the push control model is a control model determined based on a proportional coefficient, an integral coefficient, and a derivative coefficient; obtain the current quantity difference between the first push quantity of the target media resources already pushed and the second push quantity of the target media resources planned to be pushed in the current push cycle; obtain the cumulative quantity difference between the first cumulative push quantity of the target media resources already pushed and the second cumulative push quantity of the target media resources planned to be pushed in the current push cycle and the previous N push cycles, wherein N is an integer greater than or equal to 1; obtain the current quantity difference of the current push cycle and the historical quantity difference of the previous push cycle, and determine the quantity difference change rate based on the current quantity difference and the historical quantity difference; adjust the target threshold based on the first product between the proportional coefficient and the current quantity difference, the second product between the integral coefficient and the cumulative quantity difference, and the third product between the derivative coefficient and the quantity difference change rate.

[0018] According to another aspect of the present invention, a computer-readable storage medium is also provided, wherein a computer program is stored in the computer program, and the computer program is configured to execute the above-described media resource push method when running.

[0019] According to another aspect of the embodiments of this application, a computer program product or computer program is provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the media resource push method described above.

[0020] According to another aspect of the present invention, an electronic device is also provided, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to execute the media resource push method described above through the computer program.

[0021] In the above-mentioned method for pushing media resources, before pushing media resources, the operation distribution characteristics used to characterize the time distribution of account operations of the target account in the first period can be obtained first, and second operation prediction results can be determined based on the operation distribution characteristics. Based on the operation evaluation index determined by the first and second operation prediction results, it is determined whether to push target media resources to the target account.

[0022] In the above implementation, on the one hand, by using the operation distribution characteristics, the operational preferences of the target account at different time periods can be accurately perceived, and then the timing of pushing the target media resources can be accurately determined based on the correlation between the account operation time and the media resource push effect. On the other hand, the first operation prediction result assesses the probability that the target account will perform a specific operation when media resources are pushed; while the second operation prediction result assesses the probability that the target account will perform an operation when no push is made. Based on the evaluation index determined by the above two prediction results, the expected feedback of the target account to the received target media resources can be judged more accurately. Furthermore, by obtaining the time preference characteristics of the target account's account operations, the media resource push effect can be accurately predicted based on time preference analysis, thereby significantly improving the accuracy of media resource push and solving the problem of insufficient accuracy in traditional push strategies. Attached Figure Description

[0023] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0024] Figure 1 This is a schematic diagram of the hardware environment for an optional media resource push method according to an embodiment of the present invention;

[0025] Figure 2 This is a flowchart of an optional media resource push method according to an embodiment of the present invention;

[0026] Figure 3 This is an interface diagram of an optional media resource push method according to an embodiment of the present invention;

[0027] Figure 4 This is a schematic diagram of an optional media resource push method according to an embodiment of the present invention;

[0028] Figure 5 This is a schematic diagram of another optional method for pushing media resources according to an embodiment of the present invention;

[0029] Figure 6 This is a schematic diagram of an optional prediction model according to an embodiment of the present invention;

[0030] Figure 7 This is a schematic diagram of another optional prediction model according to an embodiment of the present invention;

[0031] Figure 8 This is a schematic diagram of an optional media resource push system according to an embodiment of the present invention;

[0032] Figure 9 This is a flowchart of another optional method for pushing media resources according to an embodiment of the present invention;

[0033] Figure 10 A schematic diagram of the structure of an optional media resource push device according to an embodiment of the present invention;

[0034] Figure 11 This is a schematic diagram of the structure of an optional electronic device according to an embodiment of the present invention. Detailed Implementation

[0035] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0036] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0037] It should be noted that in the relevant embodiments of this application, the acquisition and processing of relevant data information of the target account are all done in advance by means of methods that comply with the relevant normative legal documents, and the authorization permission of the corresponding account subject must be obtained before acquiring the relevant information and data.

[0038] According to one aspect of the present invention, a method for pushing media resources is provided. As an optional implementation, the above-described method for pushing media resources can be applied to, but is not limited to, [examples of other methods]. Figure 1 The media resource delivery system shown consists of terminal device 102, server 104, and network 110. For example... Figure 1As shown, terminal device 102 communicates with server 104 via network 110. This network may include, but is not limited to, wired networks and wireless networks. The wired network includes local area networks (LANs), metropolitan area networks (MANs), and wide area networks (WANs). The wireless network includes Bluetooth, Wi-Fi, and other networks enabling wireless communication. The terminal device may include, but is not limited to, at least one of the following: mobile phones (such as Android phones, iOS phones, etc.), laptops, tablets, PDAs, MIDs (Mobile Internet Devices), PADs, desktop computers, smart TVs, in-vehicle devices, etc. The terminal device 102 may have a client installed to provide resource operation services. This client may include, but is not limited to, an instant messaging client, a game client, a media resource client, etc.

[0039] The terminal device 102 is also equipped with a display, a processor, and a memory. The display can be used to display the program interface of the client. The processor can process and transform the operation information generated by the object operation based on the client. The memory is used to store the operation information generated by the object account in the historical operation.

[0040] The aforementioned server 104 can be a server cluster consisting of multiple servers, or a cloud server. The server includes a database and a processing engine. The processing engine in server 104 is used to analyze and process the operation sequence information of object operations generated by the client, and to perform media resource operations with other node servers, including parsing, writing, and reading operations.

[0041] According to one aspect of the present invention, the media resource push system described above can perform the following steps: the terminal device 102 performs step S102, sending account operation information to the server 104 through the network 110, wherein the account operation information can be used to indicate the time distribution of the target account's account operations in the first period;

[0042] Next, in server 104, steps S104 to S110 are executed: Based on the time distribution of account operations performed by the target account within the first period, the operation distribution characteristics of the target account are determined; based on the operation distribution characteristics, a first operation prediction result and a second operation prediction result are determined, wherein the first operation prediction result indicates the probability that the target account will perform the target account operation when the target media resource is pushed to the target account, and the second operation prediction result indicates the probability that the target account will perform the target account operation when the target media resource is not pushed to the target account; based on the first and second operation prediction results, an operation evaluation index is determined; if the operation evaluation index is greater than or equal to a target threshold, it is determined that the target media resource will be pushed to the target account within the second period.

[0043] Next, server 104 executes step S112, pushing target media resources to target account via network 110 to terminal device 102 within the second cycle.

[0044] Finally, the terminal device 102 executes steps S114 and S116 to display the target media resource and perform the target account operation.

[0045] In the above implementation, on the one hand, by using the operation distribution characteristics, the operational preferences of the target account at different time periods can be accurately perceived, and then the timing of pushing the target media resources can be accurately determined based on the correlation between the account operation time and the media resource push effect. On the other hand, the first operation prediction result assesses the probability that the target account will perform a specific operation when media resources are pushed; while the second operation prediction result assesses the probability that the target account will perform an operation when no push is made. Based on the evaluation index determined by the above two prediction results, the expected feedback of the target account to the received target media resources can be judged more accurately. Furthermore, by obtaining the time preference characteristics of the target account's account operations, the media resource push effect can be accurately predicted based on time preference analysis, thereby significantly improving the accuracy of media resource push and solving the problem of insufficient accuracy in traditional push strategies.

[0046] The above is merely an example, and no limitation is made in this embodiment.

[0047] As an optional implementation method, such as Figure 2 As shown, the above-mentioned method for pushing media resources can be applied to servers or terminal devices with sufficient computing power, and may specifically include the following steps:

[0048] S202, Based on the time distribution of account operations performed by the target account in the first period, determine the operation distribution characteristics of the target account;

[0049] S204, determine the first operation prediction result and the second operation prediction result based on the operation distribution characteristics, wherein the first operation prediction result indicates the probability that the target account will perform the target account operation when the target media resource is pushed to the target account, and the second operation prediction result indicates the probability that the target account will perform the target account operation when the target media resource is not pushed to the target account.

[0050] S206, if the operation evaluation index determined based on the first operation prediction result and the second operation prediction result is greater than or equal to the target threshold, the target media resources are pushed to the target account in the second cycle.

[0051] It should be noted that the media resources in the above embodiments of this application can be promotional resources related to the target application, used to attract or remind users to trigger target account operations related to the target application. The target account operations may include, but are not limited to, account login operations, as well as related target account operations performed after logging into the target application, such as daily check-in operations and activity participation operations.

[0052] Furthermore, the aforementioned target account operation can also be related to the application type of the target application. For example, if the target application is an instant messaging application, the target account operation can be a session operation performed within the instant messaging application; if the target application is a game application, the target account operation can be a game match operation performed within the game application; if the target is an audio / video application, the target account operation can be an audio / video playback operation. This embodiment does not limit the operation type of the aforementioned target account operation.

[0053] Furthermore, the aforementioned target account operations can also be related to the type of user audience targeted by the target media resources. For example, if the user audience already holds an account for the target application, the aforementioned target account operations can be account login operations, account check-in operations, or other account operations used to indicate the user activity metrics of the target application; if the user audience does not hold an account for the target application, the aforementioned target account operations can be account registration operations, account binding operations, or other account operations used to indicate the new user growth rate metrics of the target application.

[0054] It is understood that the resource types of the aforementioned target media resources may include, but are not limited to, text, images, audio / video, etc., used to remind users of relevant activity information within the target application, thereby attracting users to launch the target application and perform the aforementioned target account operations. Furthermore, the resource type of the aforementioned target media resources may be related to the push method used to deliver them. When pushing target media resources via SMS, the target media resources may be text or image / text types; when pushing target media resources via terminal notifications, the target media resources may be image / text types or link types. For example... Figure 3 In this context, target media resources 302 can be pushed to users via the mobile terminal's notification system. The text "AAA activity is back for a limited time! Come and participate!" reminds users to participate in the target application's activities, thereby increasing user activity. When pushing the target media resources through social media pages, the type of the target media resources can be audio or video, etc. In an optional implementation, the target media resources may further carry a redirect link related to the target application, responding to a triggering operation by the user's account, redirecting to the corresponding application download page, or launching the corresponding running process of the target application on the terminal device. This increases the probability of the user's account performing the target account's operation through the push of the target media resources.

[0055] It is understandable that pushing target media resources to target accounts typically consumes significant resources, including but not limited to server resources, and increases the operational costs of the target application. Therefore, if we identify target accounts from a large pool of accounts that are highly likely to generate positive feedback from pushing the target media resources, and then push those target media resources to them, we can significantly improve the relevant performance metrics of the target application, such as user activity and the recovery rate of potential churned users, while conserving resources.

[0056] Correspondingly, the target account in the above embodiments can be the target account to be analyzed. Through the above steps S202 to S208, it can be determined whether the target account will generate positive feedback on the pushed target media resources, and then it can be determined whether to push the target media resources to the target account.

[0057] In one alternative approach, the target account can be a first target account that has completed registration related to the target application, or a second target account that has not completed registration related to the target application. Correspondingly, if the target account is a first target account, the target account operations may include, but are not limited to, login operations, check-in operations, virtual resource claiming operations, and virtual game start operations performed within the target application; if the target account is a second target account, the target account operations may include, but are not limited to, account registration operations and account binding operations performed within the target application.

[0058] Furthermore, in S202 above, the first period can be a historical time period prior to the current time point, such as the previous week or the previous month. The account operations mentioned above can include, but are not limited to, one or more operations such as account login, check-in, virtual resource collection, and starting a virtual game match. Then, through S202 above, after obtaining the time distribution of the aforementioned account operations of the target account, the operation distribution characteristics related to the target account are determined based on the quantification results of the aforementioned time distribution.

[0059] In one optional implementation, taking account operation specifically as account login operation as an example, the time distribution of the aforementioned account login operation can be represented according to the specific time node corresponding to each login operation. For example, if the first period is specifically "last week", the time node sequence composed of the time nodes corresponding to the multiple login operations performed by the target account last week can be used as the time distribution of account login operations; further, the time node sequence is mapped to the corresponding vector representation to determine the operation distribution characteristics of the target account.

[0060] Optionally, the aforementioned time distribution can also be specifically represented as the distribution of the login operations within different time periods. For example, if the first period is "last week," the login times are "Monday, 19:30," "Wednesday, 20:30," and "Saturday, 10:30." Therefore, the weekly distribution of the login operations performed by the target account within the "last week" period is "Day 1," "Day 3," and "Day 6." Within a day, the daily distribution of the login operations performed by the target account within the "day" period is "Hour 19," "Hour 20," and "Hour 10." Based on the weekly and daily distributions, the operation distribution characteristics of the target account can be determined. For example, by representing "Day 1," "Day 3," "Day 6," "Hour 19," "Hour 20," and "Hour 10" using their respective vector representations, the operation distribution characteristics of the target account can be determined.

[0061] It is understandable that, given the operational distribution characteristics of the target account, performing the push operation of the target media resource based on these characteristics can significantly enhance the positive feedback effect of the target media resource on the target account. Furthermore, in step S202 above, account identification can be performed based on the operational distribution characteristics corresponding to the previous period, accurately reflecting the recent operational habits of the target account and accurately perceiving changes in these habits. Based on this accurate perception, the accuracy of the target media resource push can be improved.

[0062] It should be noted that, when the aforementioned target account is the first target account that has completed the registration operation related to the aforementioned target application, the aforementioned account operation can specifically refer to the relevant operation performed in the aforementioned target application;

[0063] In cases where the target account is a second account that has not completed the registration process related to the target application, the aforementioned account operation can be an operation performed by the target account in a reference application. This reference application can be an application of a similar or identical type to the target application. For example, if the target application is a game application, the reference application can also be a game application; if the target application is an audio playback application, the reference application can also be a video playback application. By leveraging the operation distribution characteristics obtained from related or similar reference applications, the probability of the target account performing application registration operations related to the target application can be predicted, thereby achieving precise targeting of media resources and application traffic generation.

[0064] The above method for determining the distribution characteristics of operations is only an example and does not limit the specific implementation of step S202.

[0065] Furthermore, in S204 above, a first operation prediction result and a second operation prediction result can be obtained based on the operation distribution features extracted in S202. The first operation prediction result can be used to indicate the probability that the target account will perform the target account operation when the target media resource is pushed to the target account; the second operation prediction result indicates the probability that the target account will perform the target account operation when the target media resource is not pushed to the target account.

[0066] Optionally, in S204 above, if the target account is a first target account, the target account operation can be an account login operation, and then the probability of the target account logging into the target application under the influence of the target media resource can be accurately predicted based on the prediction results of the first operation and the second operation; if the target account is a second target account, the target account operation can be an account registration operation, and then the probability of the target account registering and using the target application under the influence of the target media resource can be accurately predicted based on the prediction results of the first operation and the second operation.

[0067] It should be noted that conventional modeling methods for targeted media resource push typically predict the correlation between the push behavior of the target media resource and the actions performed by the target account. This correlation lacks interpretability. In other words, existing prediction methods cannot distinguish whether the actions performed by the target account are spontaneous or influenced by the target media resource.

[0068] For example, account A is a loyal user of the target application, logging in and using the application daily regardless of whether they receive target media resources pushed by the application. Assuming account A received target media resources daily over the past week and logged in and used the application daily as usual, conventional push prediction methods would identify a 100% correlation between user A's login behavior and the push behavior of target media resources, thus continuously pushing target media resources to account A. However, in reality, account A is a loyal user of the target application and remains online daily regardless of whether target media resources are pushed to them. Therefore, a more efficient approach is to stop pushing target media resources to account A and instead push the media resources originally intended for account A to other user accounts that are susceptible to the influence of target media resources. It is evident that existing push methods cannot accurately identify target accounts that will be affected by target media resources, resulting in inaccurate push notifications of target media resources.

[0069] In S204, prediction results under two different assumptions can be obtained based on the operational distribution characteristics;

[0070] In an optional implementation, the first operation prediction result and the second operation prediction result can be obtained by inputting the operation distribution features into a trained operation prediction model and then obtaining two prediction results based on two different output modules of the operation prediction model.

[0071] In another optional implementation, the first operation prediction result and the second operation prediction result may also be obtained by inputting the operation distribution characteristics into the first prediction model and the second prediction model respectively, and using the output result of the first prediction model as the first operation prediction result and the output result of the second prediction model as the second operation prediction result.

[0072] In this embodiment, the first prediction model and the second prediction model can be jointly trained during the training phase through parameter sharing. Furthermore, during training, the first prediction model predicting the result of the first operation can be trained using experimental group data, and the second prediction model predicting the result of the second operation can be trained using control group data. The experimental group data can be training samples determined based on the feedback behavior (whether the target account operation was executed) of the target accounts that have received the target media resources, while the control group data can be training samples determined based on the non-intervention target accounts, i.e., training samples determined based on the feedback behavior (whether the target account operation was executed) of the target accounts that did not receive the target media resources.

[0073] It's important to note that during training, the average causal effect can be statistically analyzed from a demographic perspective. For example, consider two homogeneous sets of user accounts. The account information in both sets indicates that the users are from first-tier cities, aged 25-35, and of the same gender. Then, push notifications can be sent to the first group (the experimental group), while the second group (the control group) receives no intervention. The difference in conversion rates between these two groups can then be statistically analyzed; this difference can be approximated as the average causal effect for people with similar characteristics. This data can then be used to train two different prediction models, and the target evaluation index can be determined based on the prediction results of these two models during the prediction process.

[0074] This allows for the prediction of the probability that a target account will perform a target account operation when the target media resources are pushed to the target account using the first prediction model; and the prediction of the probability that a target account will perform a target account operation when the target media resources are not pushed to the target account using the second prediction model.

[0075] Furthermore, based on the two prediction results mentioned above, S206 can accurately identify whether the target account is an account that is easily affected by the push behavior of the target media resources. If it is determined that the target account is an account that is sensitive to the push behavior of the target media resources (i.e., easily affected by the push behavior), the target media resources are pushed to the target account in the second cycle.

[0076] It is understandable that the aforementioned second cycle can be the current push cycle following the aforementioned first cycle. For example, if the aforementioned first cycle is "the first week of April", the aforementioned second cycle can be "the second week of April".

[0077] In the above S206, it can be specifically determined whether the target account is a target account that is sensitive to the push behavior of the target media resources (i.e., easily affected by the push behavior) by comparing the operation evaluation index determined by the first operation prediction result and the second operation prediction result with the preset target threshold.

[0078] Optionally, when the first operation prediction result is P(push|treatment) and the second operation prediction result is P(push|no treatment), the operation evaluation index Lift in step S206 can be determined based on the difference between the first operation prediction result and the second operation prediction result, i.e.:

[0079] Lift=P(push|treatment)-P(push|no treatment)

[0080] In the above implementation, on the one hand, by using the operation distribution characteristics, the operational preferences of the target account at different time periods can be accurately perceived, and then the timing of pushing the target media resources can be accurately determined based on the correlation between the account operation time and the media resource push effect. On the other hand, the first operation prediction result assesses the probability that the target account will perform a specific operation when media resources are pushed; while the second operation prediction result assesses the probability that the target account will perform an operation when no push is made. Based on the evaluation index determined by the above two prediction results, the expected feedback of the target account to the received target media resources can be judged more accurately. Furthermore, by obtaining the time preference characteristics of the target account's account operations, the media resource push effect can be accurately predicted based on time preference analysis, thereby significantly improving the accuracy of media resource push and solving the problem of insufficient accuracy in traditional push strategies.

[0081] In an optional implementation, when the first operation prediction result and the second operation prediction result are two prediction results output by the same trained operation prediction model, S204, determining the first operation prediction result and the second operation prediction result based on operation distribution characteristics, includes:

[0082] S2042, Obtain the characteristics of the reference account associated with the target account;

[0083] S2044, The target account features obtained by concatenating the reference account features and the operation distribution features are input into the feature sharing module of the target prediction model to obtain the shared feature vector;

[0084] S2046, input the shared feature vector into the first prediction module of the target prediction model to obtain the first operation prediction result;

[0085] S2048, the shared feature vector is input into the second prediction module of the target prediction model to obtain the second operation prediction result, wherein the first prediction module and the second prediction module are two prediction modules in the target prediction model that are respectively connected to the feature sharing module.

[0086] In this embodiment, the reference account characteristics can be further concatenated with the above-mentioned operation distribution characteristics to obtain the target account characteristics, and the first operation prediction result and the second operation prediction result can be obtained based on the target account characteristics.

[0087] It is understood that the reference account features in S2042 above may include account description features indicating the static characteristics of the target account. For example, the reference account features may include, but are not limited to, the target account's initial registration time, account level, account points, and the amount of virtual resources held by the account. Furthermore, the reference account features may also include operation description features indicating the dynamic characteristics of the target account, i.e., feature representations generated based on the target account's operational behavior. For example, the dynamic features may include the time when the target account performs a specific account operation and the duration of that operation.

[0088] Optionally, when the aforementioned reference account characteristics include both the aforementioned static characteristics and dynamic characteristics, step S2042 may specifically include the following steps:

[0089] S2042-1, Obtain the operation description feature used to indicate the account operations to be performed by the target account in the first period;

[0090] S2042-2, Obtain account description features that match the target account;

[0091] S2042-3, the operation description feature and the account description feature are used as reference account features.

[0092] Specifically, when the target application is an audio / video application, the operation description features can be determined based on the audio / video playback operations performed by the target account. For example, the operation description features can be determined based on parameters such as the playback duration of each audio / video playback operation and the number of marking operations performed during playback. The account description features in the audio / video application can include account level, cumulative media playback count, and cumulative media collection count. After obtaining these relevant parameters, they are mapped to corresponding vector representations to determine the reference account features of the target account in the audio / video application.

[0093] For example, if the target application is a game application, the aforementioned operation description features can be determined based on the game operations performed by the target account. For instance, these features can be determined based on the game duration of each match and performance parameters during the match, such as the number of defeats, damage dealt, number of defeats, and number of assists in a single game. The account description features in the game application can include the account's rank and skill level rating. After obtaining these relevant parameters, they are mapped to corresponding vector representations to determine the reference account features of the target account in the game application.

[0094] Before performing step S2044, the target account characteristics can be determined according to a specific concatenation method. This includes:

[0095] S2043-1, concatenate the account description feature with at least one operation distribution feature to obtain at least one first account feature;

[0096] S2043-2, Based on the matching relationship between at least one first account feature and at least one operation description feature, the at least one first account feature and at least one operation description feature are concatenated to obtain the target account feature.

[0097] When the target application is a game application, such as Figure 4 This paper illustrates a method for concatenating an account description feature with at least one operation distribution feature. The account description feature 402 includes a first field "aa" indicating the account identifier of the target account, a second field "100" indicating the account level of the target account, and a third field "5000" indicating the account points of the target account. Furthermore, the operation distribution feature 404 is "2024 / 04 / 24 / 21 / 30," which can be used to indicate that the target account performed an account login operation at 21:30 on April 24, 2024. Figure 4 The document also shows a second operation distribution feature, represented as "2024 / 04 / 23 / 20 / 22", indicating that the target account performed an account login operation at 20:22 on April 23, 2024. These two operation distribution features are then concatenated with account description feature 402 to obtain two first account features.

[0098] Furthermore, such as Figure 5 As shown, according to Figure 4 If the first account feature 504 is obtained in the manner shown, the game operation performed by the target account on April 24, 2024 can be further obtained, and the game parameters corresponding to each game operation can be obtained. Each game operation is then defined as an operation description feature based on its corresponding game parameters. For example, if operation description feature 502 is specifically "02 / 30 / 07 / 02 / 10", it can be used to indicate that in the second game operation performed by the target account on April 24, 2024 (corresponding to "02"), the game lasted 30 minutes, and the target account defeated the enemy 7 times, was defeated 2 times, and had 10 assists. Then, operation description feature 502 is concatenated with the first account feature 504 to obtain a target account feature. It is understood that in... Figure 5The other operation description feature shown, "01 / 25 / 03 / 05 / 08," can be used to indicate that in the first game match (corresponding to "01") performed by the target account on April 24, 2024, the game lasted 25 minutes, and the target account defeated the enemy 3 times, was defeated 5 times, and had 8 assists. Since this operation description feature also corresponds to the first account feature 504, the above operation description feature can be concatenated with the first account feature to obtain the second target account feature.

[0099] Understandably, in combination Figure 4 and Figure 5 The concatenation method shown can determine the target account feature corresponding to each operation description feature. Furthermore, before inputting multiple target account features into the target prediction model, these multiple target account features can be mapped into corresponding vector representations, and then the vector representations can be input into the target prediction model.

[0100] Through the above-described embodiments of this application, after obtaining the operation distribution characteristics for accurately perceiving the target account, it is possible to further obtain operation description features for describing the dynamic characteristics of the target account and account description features for describing the static characteristics of the target account. By combining the target account features determined by the features of multiple dimensions, prediction can be made, thereby effectively identifying the login preferences of the target account, accurately perceiving the operation characteristics of the target account, and then accurately predicting the subsequent operation behavior of the target account based on the operation characteristics of the target account.

[0101] The following is a detailed explanation of how the operation distribution characteristics are obtained in step S202 above:

[0102] In one optional implementation, determining the operation distribution characteristics of the target account based on the time distribution of account operations performed by the target account within the first period includes:

[0103] S2022, Obtain the login time corresponding to each of the multiple login operations performed by the target account in the first period;

[0104] S2024, when the login time includes multiple time parameters, determine the time feature vectors corresponding to the multiple time parameters based on the multiple time parameters and the time periods matched by each of the multiple time parameters;

[0105] S2026, multiple time feature vectors are concatenated to obtain the operational distribution feature.

[0106] It is understood that this embodiment provides a method that reflects the distribution characteristics of each time parameter based on the time period corresponding to each time parameter in the login time.

[0107] Taking the login time "04 / 24 / 21 / 30" as an example, the time parameter corresponding to "04" is "April", and the time period corresponding to the month parameter is "12". Therefore, the time feature vector corresponding to the "month" can be determined based on 1 / 3 (that is, the ratio between 4 and 12).

[0108] The time parameter corresponding to “24” is “24 days”, while the time period corresponding to the number of days in January is “30”. Therefore, the time feature vector corresponding to “date” can be determined based on 4 / 5 (i.e., the ratio between 24 and 30).

[0109] In another alternative approach, it can be determined that the week number corresponding to "24th" is the 4th day of the 4th week of April. Then, the time feature vector corresponding to "week" can be determined based on 1 (the ratio between 4 and 4), and the time feature vector corresponding to "week N" can be determined based on 4 / 7 (the ratio between 4 and 7).

[0110] Furthermore, in determining the time feature vector corresponding to 4 / 7, the corresponding time feature vector can be directly obtained by mapping 4 / 7. For example, a preset mapping relationship can be set: 1 / 7 corresponds to the first feature vector (0000001), 2 / 7 corresponds to the second feature vector (0000010), 3 / 7 corresponds to the third feature vector (0000100), 4 / 7 corresponds to the fourth feature vector (0010000), 5 / 7 corresponds to the fifth feature vector (0010000), 6 / 7 corresponds to the fifth feature vector (0100000), and 7 / 7 corresponds to the seventh feature vector (1000000). Thus, the time feature vector corresponding to 4 / 7 is determined to be the fourth feature vector (0010000).

[0111] Based on the above similar method, the time feature vector corresponding to "21:00" and the time feature vector corresponding to "30 minutes" can be further determined.

[0112] In the above implementation, the login time corresponding to each of the multiple login operations performed by the target account in the first period is obtained; when the login time includes multiple time parameters, time feature vectors corresponding to the multiple time parameters are determined according to the multiple time parameters and the time periods matched by the multiple time parameters; the multiple time feature vectors are concatenated to obtain the operation distribution features, and then the operation distribution features corresponding to the current login time are determined according to the distribution of the multiple time parameters included in the login time corresponding to the login operation in their respective time periods, thereby achieving accurate acquisition of the operation features of the login operation of the target account.

[0113] Furthermore, in S2024 above, based on multiple time parameters and the time periods matched by each of the multiple time parameters, time feature vectors corresponding to the multiple time parameters are determined, including:

[0114] S2024-1, obtain the current time parameter and the current time period that matches the current time parameter from multiple time parameters;

[0115] S2024-2, the ratio between the current time parameter and the current time period is determined as the login distribution coefficient;

[0116] S2024-3, obtain the first coding result of the login distribution coefficient based on the sine function, and obtain the second coding result of the login distribution coefficient based on the cosine function;

[0117] S2024-4, the first encoding result and the second encoding result are determined as the vector feature values ​​included in the current time feature vector.

[0118] In one specific implementation, the time parameter in the login time can be encoded into a week feature vector (week embedding), an hour feature vector (hour embedding), and a second feature vector (second embedding). That is, the number of seconds corresponding to the week, the hour, and the second corresponding to the time when the user account performs the login operation can be encoded into separate embedding vectors and fed into the model (7 days in a week can correspond to 7 embeddings, 24 hours in a day can correspond to 24 embeddings, and 60 seconds in a minute can correspond to 60 embeddings).

[0119] In this embodiment, a sine function can be used to encode the time features, and the specific encoding method is as follows:

[0120]

[0121] Here max tThis represents the maximum possible value of the time parameter t. For example, the maximum value for seconds is 60, the maximum value for days is 24, and the maximum value for weeks is 7.

[0122] Let's continue with the example of "04 / 24 / 21 / 30". As can be seen in the above implementation, "04 / 24" indicates the 4th day of the 4th week of April.

[0123] Furthermore, for the weekly feature vector week embedding f(t1), max t1 =7, thus determining The week embedding is then determined as a vector.

[0124]

[0125] Furthermore, for the hour embedding f(t2) of the hour feature vector, max t2 =24, thus determining The week embedding is then determined as a vector.

[0126] Furthermore, for the second feature vector second embedding f(t2), max t3 =60, thus determining The week embedding is then defined as the vector [sin(π),cos(π)].

[0127] By concatenating the week embedding, hour embedding, and second embedding in sequence, we can obtain the feature vector used to characterize the distribution features of the operation.

[0128] Through the above-described embodiments of this application, in the process of determining the time feature vectors corresponding to multiple time parameters based on multiple time parameters and their respective matching time periods, the process first obtains the current time parameter and its matching current time period. Then, the ratio between the current time parameter and the current time period is determined as the login distribution coefficient. Next, a first encoding result of the login distribution coefficient is obtained based on a sine function, and a second encoding result is obtained based on a cosine function. Finally, the first and second encoding results are determined as the vector feature values ​​included in the current time feature vector. This achieves the mapping of the distribution positions of each time parameter in its corresponding time period to the corresponding periodic change parameters in a periodic sine function. Furthermore, based on the periodic distribution characteristics of the sine function, the login operation of the target account can be accurately characterized in each time period, thus achieving precise dynamic capture of the periodic distribution of the target account's login time. By accurately obtaining the operation distribution characteristics of the target account's login operation, the accuracy of the prediction results determined by the operation distribution characteristics can be significantly improved.

[0129] Having obtained the target account characteristics through the above implementation method, the process of obtaining the first operation prediction result and the second operation prediction result through S2044, S2046 and S2048 will be further explained below.

[0130] In this embodiment, a target prediction model can be used to determine the operation prediction results under two different conditions based on the characteristics of the target account.

[0131] In an optional implementation, the model structure of the above-mentioned target prediction model can be as follows: Figure 6 As shown. In this embodiment, the target application can be specifically a game application, and the corresponding target account operation can be the login operation of the game application.

[0132] First, based on the feature extraction method described in the above embodiments, game sequence features 602, account description features 604, and operation distribution features 606 can be obtained respectively. Specifically, the operation distribution feature 606 can be determined based on the login time corresponding to each login operation performed by the target account within the past month. The specific method for determining the operation distribution feature 606 based on login time can be as described in steps S2024-1 to S2024-4 above. In this embodiment, the account description feature 604 can be used to indicate, but is not limited to, the account identifier, account level, and account points of the target account. Furthermore, the concatenation method of the account description feature 604 and the operation distribution feature 606 can be in accordance with... Figure 4The splicing method shown corresponds to the implementation method; further, the above-mentioned game sequence feature 602 can be a specific operation description feature in a game scenario, which can be used to indicate, but is not limited to, the game sequence corresponding to different game games, the duration of the game game, the defeat parameters, the defeated parameters, and the assist parameters of the target account in the game. The splicing result between the account description feature 604 and the operation distribution feature 606 can be a specific first account feature. The splicing method between the game sequence feature 602 and the account description feature 604, and the operation distribution feature 606 can be implemented as follows: Figure 5 The corresponding method is used to achieve this.

[0133] Having obtained the game sequence feature 602, account description feature 604, and operation distribution feature 606, these can be input into the shared feature module 618 of the target prediction model. The shared feature module 618 includes a feature embedding layer 608 for extracting feature embedding vector representations, a self-attention layer 610 for extracting self-attention weights, and a fully connected layer 612. After obtaining the shared feature vector output by the fully connected layer 612, the shared feature vector can be input into the corresponding first prediction module 624 and second prediction module 622. The first prediction module 624 predicts the probability that the target account will perform an account login operation when target media resources are pushed to the target account (i.e., the first operation prediction result 614), and the second prediction module 622 predicts the probability that the target account will perform an account login operation when target media resources are not pushed to the target account (i.e., the second operation prediction result 616).

[0134] The following are Figure 6 The training process of the target prediction model shown is explained below:

[0135] Before determining the first operation prediction result and the second operation prediction result based on the operation distribution characteristics, the above-mentioned S202 also includes:

[0136] S2000: Obtain the sample account characteristics of the sample object account, and the sample tags that match the sample object account;

[0137] S2002, input the sample account features into the initial feature sharing module in the initial prediction model that is in the training state to obtain the shared sample feature vector;

[0138] S2004, the shared sample feature vector is input into at least one target prediction module among multiple initial prediction modules in the initial prediction model, wherein the initial feature sharing module in the initial prediction model is connected to multiple initial prediction modules respectively, and the multiple initial prediction modules include a first initial prediction module, a second initial prediction module and a third initial prediction module. The first initial prediction module outputs a first sample operation prediction result, which is used to indicate the probability that the sample object account will perform a sample account operation when sample media resources are pushed to the sample object account. The second initial prediction module outputs a second sample operation prediction result, which is used to indicate the probability that the sample object account will perform a sample account operation when sample media resources are not pushed to the sample object account. The third initial prediction module outputs a tendency prediction result, which is used to indicate the learning tendency of the initial prediction model during the training phase.

[0139] S2006, Based on the prediction results output by at least one target prediction module and the sample labels, train the initial prediction model until a target prediction model that meets the target training conditions is obtained.

[0140] like Figure 7 As shown, in the initial prediction model during the training phase, in addition to the first initial prediction module 724 and the second initial prediction module 722, the modules connected to the initial shared feature module 718 also include a third initial prediction module 726. The third initial prediction module 726 is used to output a propensity prediction result, indicating the learning tendency of the initial prediction model during the training phase.

[0141] In the Figure 7 Before training the initial prediction model shown above, in step S2000, the sample account features and sample labels of the sample object accounts are obtained, including:

[0142] S2000-1, Determine the first sample set and the second sample set, wherein the difference between the first number of first sample accounts included in the first sample set and the second number of second sample accounts included in the second sample set is less than or equal to the target number threshold.

[0143] S2000-2, Push sample media resources to the first sample account in the first sample set;

[0144] S2000-3, based on the account operations performed by multiple sample object accounts in the first sample set and the second sample set during the target account operation period, determine the sample tags corresponding to each of the multiple sample object accounts, including: configuring a first sample tag for the sample object account when the sample object account has performed a sample account operation, and configuring a second sample tag for the sample object account when the sample object account has not performed a sample account operation.

[0145] For example, two homogeneous sets of user accounts can be obtained. The account information of these accounts indicates that the corresponding users are all from first-tier cities, aged 25-35, and of the same gender. Then, push notifications can be sent to the first set of accounts (the experimental group), while no intervention is performed on the second set (the control group). Next, the user accounts in both sets are analyzed to determine whether they logged in within a predetermined statistical period. Based on the login activity, corresponding sample labels are assigned to each account, resulting in the first and second sample sets that can be used to train the initial prediction model.

[0146] In an optional implementation, during the process of determining the first sample set and the second sample set, a randomized A / B test can be performed on the first sample set and the second sample set. Based on the test results, the first sample set and the second sample set are determined to be two homogeneous sample sets. In this way, the first sample set and the second sample set can provide unbiased samples for the initial prediction model, thereby improving the model training effect.

[0147] Furthermore, given the unbiased first and second sample sets available for model training, during the training of the initial prediction model, the aforementioned inputting the shared sample feature vector into at least one target prediction module among multiple prediction modules in the initial prediction model includes:

[0148] S2004-1, when the sample account features are the account features of the first sample account, the shared sample feature vector is input into the first initial prediction module and the third initial prediction module of the initial prediction model to obtain the first sample operation prediction result output by the first initial prediction module and the tendency prediction result output by the third initial prediction module, wherein the first sample account is the object account that receives the pushed sample media resources.

[0149] S2004-2, when the sample account features are the account features of the second sample account, the shared sample feature vector is input into the second initial prediction module and the third initial prediction module of the initial prediction model to obtain the second sample operation prediction result output by the second initial prediction module and the tendency prediction result output by the third initial prediction module, wherein the second sample account is the object account that has not received the sample media resources.

[0150] The following combination Figure 7The training process described above will be explained in detail. In this embodiment, for the currently acquired sample account, the sample account features of the current sample account can be determined based on the game sequence features 702, account description features 704, and operation distribution features 706. Then, the sample account features are input into the initial shared feature module 718 in the prediction model. The shared feature module 718 includes a feature embedding layer 708 for extracting feature embedding vector representations, a self-attention layer 710 for extracting self-attention weights, and a fully connected layer 712.

[0151] By obtaining the shared sample feature vector of the fully connected layer 712, that is, by representing the confounder through feature processing and deep network hidden layers, the probability of performing target account operations in both cases can be predicted based on this shared representation vector.

[0152] It should be noted that in this embodiment, the module to be input needs to be determined according to the account type corresponding to the current sample account. Specifically, when the current sample account is the first sample account, the shared sample feature vector is input into the first initial prediction module 724 and the third initial prediction module 726 respectively; when the current sample account is the second sample account, the shared sample feature vector is input into the second initial prediction module 722 and the third initial prediction module 726 respectively. That is to say, during the training process, samples with t=1 (i.e., account samples that have pushed target media resources) are fitted by the first initial prediction module 724, and samples with t=0 (i.e., account samples that have not pushed target media resources) are fitted by the second initial prediction module 722.

[0153] In the above implementation, the third initial prediction module 726 can be used to output the propensity score. The third initial prediction module 726 can be constructed by a simple linear mapping (sigmoid). The simple mapping can make the confounder representation layer and the propensity score closely connected, so that only information related to the propensity score is extracted. At this time, the shared sample feature vector is a feature vector that affects both the prediction result of the first sample operation and the prediction result of the second sample operation.

[0154] Through the above-described implementation method of this application, end-to-end learning of the propensity score and covariates, as well as the conditional outcome Y under the treatment variable, is achieved through a three-headed overall structure. This significantly improves the prediction performance of the target prediction model during training.

[0155] In the embodiment where the first operation prediction result and the second operation prediction result are predicted using the aforementioned target prediction model, the operation evaluation index in step S206 can specifically be the Individual Treatment Effect (ITE) parameter, i.e. in, for Figure 6 The output of the first prediction module of the target prediction model shown is the probability that the target account will perform an account login operation when the target media resource is pushed. This is the probability output by the second prediction module of the target prediction model, which is the probability that the target account will perform an account login operation when the target media resources are not pushed to the target account.

[0156] Understandably, due to This indicates the difference between the first probability that a target account will perform an account login operation under the influence of the target media resource push, and the second probability that the target account will perform an account login operation without being influenced by the target media resource push. It is evident that a higher ITE indicates that the target account is more susceptible to the influence of the received target media resource. In other words, a higher ITE indicates a better expected effect when pushing the target media resource to the target account.

[0157] Optionally, if the number of target media resources to be pushed exceeds a certain threshold, the target media resources can be pushed to the target accounts in order of their respective operation evaluation index (ITE) from high to low.

[0158] Optionally, when the number of target media resources to be pushed is limited, target accounts with ITE values ​​higher than a certain target threshold can be selected as target accounts to be pushed target media resources based on the operation evaluation index (ITE) of each target account.

[0159] It is evident that as the target threshold increases, the more stringent the conditions for identifying target accounts to be pushed target media resources, the fewer target accounts can be identified as such. Conversely, as the target threshold decreases, the more lenient the conditions for identifying target accounts to be pushed target media resources, the fewer target accounts can be identified as such. Thus, by adjusting the target threshold, the number of target media resources pushed in different push cycles can be dynamically adjusted.

[0160] In the first scenario described above, where the number of target media resources to be pushed is limited, the target threshold in step S206 can be dynamically determined based on the difference between the estimated number of target media resources to be pushed and the actual number already pushed.

[0161] Specifically, in step S206 above, if the operation evaluation index determined based on the first operation prediction result and the second operation prediction result is greater than or equal to the target threshold, before pushing the target media resources to the target account in the second cycle, the following steps are also included:

[0162] S2052, obtain the difference between the first push quantity and the second push quantity associated with the target media resource in the previous push cycle, wherein the first push quantity is the number of target media resources that have been pushed, and the second push quantity is the number of target media resources that are planned to be pushed;

[0163] S2054-1, If ​​the difference in quantity is greater than or equal to the target value, the upper reference threshold will be raised to the target threshold.

[0164] S2054-2, If the difference in quantity is less than the target value, the reference threshold will be reduced to the target threshold.

[0165] It is understood that in this embodiment, the target threshold corresponding to the current push period can be dynamically determined based on the difference between the number of pushes already pushed and the planned number of pushes in the previous push period, and the target accounts to be pushed can be determined from the target account set based on the target threshold of the current push period.

[0166] In one alternative approach, the aforementioned reference threshold can be a historical threshold corresponding to the previous push cycle. Further, if the target value in S2054-1 and S2054-2 is specifically 0, this embodiment can, in the case where the first number of target media resources pushed in the previous push cycle is greater than the second number of target media resources planned to be pushed (i.e., the push volume needs to be reduced in the current cycle), raise the historical threshold to the target threshold, thereby reducing the push volume in the current push cycle; conversely, if the first number of target media resources pushed in the previous push cycle is less than the second number of target media resources planned to be pushed (i.e., the push volume needs to be increased in the current cycle), lower the historical threshold to the target threshold, thereby increasing the push volume in the current push cycle.

[0167] In one optional implementation, S2052 includes:

[0168] S2053-1, the product of the quantity difference and the threshold revision coefficient is determined as the target adjustment value;

[0169] S2053-2, Adjust the reference threshold to the target threshold according to the target adjustment value.

[0170] In another alternative approach, the aforementioned reference threshold can also be a preset fixed threshold. Specifically, the target threshold corresponding to the current push cycle is... Let P(online) be the actual push volume in the previous push cycle, and P(target) be the target push volume in the previous push cycle. With a reference threshold of Δ, the target threshold for the current push cycle can be determined as follows:

[0171]

[0172] Where K is the threshold revision coefficient, which can be a parameter value less than 0 and greater than -1.

[0173] For example, in the previous push cycle, if the reference threshold Δ is set to 0.8, the threshold revision coefficient K is -0.0001, the actual number of target media resources pushed P(online) is 10000, and the target push volume P(target) is 9000, then the target threshold corresponding to the current push cycle can be determined. Thus, by increasing the reference threshold Δ to the target threshold This reduces the number of accounts that can be identified as targets for push notifications.

[0174] For example, if in the previous push cycle, the actual number of target media resources pushed, P(online), was 8000, while the target push volume, P(target), was 9000, then the target threshold corresponding to the current push cycle can be determined. This reduces the reference threshold Δ to the target threshold. The number of accounts that can be identified as targets for push notifications has been increased.

[0175] Through the above-described embodiments of this application, by obtaining the quantity difference between the first push quantity and the second push quantity associated with the target media resource in the previous push cycle, the upper reference threshold is raised to the target threshold when the quantity difference is greater than or equal to the target value, and the reference threshold is lowered to the target threshold when the quantity difference is less than the target value, thereby realizing the dynamic adjustment of the quantity of the target media resource to be pushed.

[0176] In an optional implementation, after pushing the target media resources to the target account in the second period when the operation evaluation index determined based on the first operation prediction result and the second operation prediction result is greater than or equal to the target threshold, the method further includes:

[0177] S2082, when the push stage corresponding to the current push cycle is the target stage, obtain the push control model that matches the next push cycle, wherein the push control model is a control model determined based on the proportional coefficient, integral coefficient and derivative coefficient.

[0178] S2084, obtain the current quantity difference between the first push quantity of target media resources that have been pushed and the second push quantity of target media resources that are planned to be pushed in the current push cycle;

[0179] S2086, obtain the cumulative difference between the first cumulative number of target media resources pushed and the second cumulative number of target media resources to be pushed in the current push period and the previous N push periods, where N is an integer greater than or equal to 1;

[0180] S2088, obtain the current quantity difference in the current push period and the historical quantity difference in the previous push period, and determine the quantity difference change rate based on the current quantity difference and the historical quantity difference;

[0181] S2090, the target threshold is adjusted based on the first product between the proportional coefficient and the current quantity difference, the second product between the integral coefficient and the cumulative quantity difference, and the third product between the differential coefficient and the rate of change of the quantity difference.

[0182] In this embodiment, the target threshold corresponding to the current push cycle can also be adjusted using a PID (Proportional-Integral-Derivative) control algorithm.

[0183] Specifically, the target threshold corresponding to the current push cycle can be determined using the following formula:

[0184]

[0185] Where err(t) can be the current difference between the first number of target media resources already pushed and the second number of target media resources planned to be pushed, and ∫err(t) can be the cumulative difference between the first cumulative number of target media resources already pushed and the second cumulative number of target media resources planned to be pushed in the current push period and the previous N push periods. It can be the rate of change of the quantity difference; correspondingly, in the above formula, K is the proportionality coefficient. T is the integral coefficient. D *K is the differential coefficient, and Δ can be the reference threshold.

[0186] It should be noted that the above proportional coefficient, integral coefficient, and derivative coefficient can be dynamically adjusted during actual use to achieve better push results.

[0187] Specifically, when initially setting the aforementioned parameters, one can consider only the linear relationship between the target push volume and the actual push volume, conduct experiments using different parameters K, observe the impact of different parameters on system performance, and initially obtain a reasonable range of parameters for each coefficient. Specifically, when the proportional coefficient is large, the system oscillation amplitude is large; when the integral coefficient is large, the system takes a long time to reach steady state; and when the derivative coefficient is large, the system will oscillate continuously, making it difficult to reach a stable state.

[0188] In a preferred adjustment method, a random term can be introduced for fine-tuning, thereby obtaining a more ideal PID controller.

[0189] Through the above embodiments of this application, using the PID control algorithm to dynamically adjust the push volume of target media resources can effectively stabilize the online message delivery volume (push volume) at the target push volume. It should be noted that, in actual delivery scenarios, there can be certain constraints on the actual message push time and whether it is a holiday. Therefore, the output of the PID algorithm can be processed according to the business scenario to ensure that the algorithm outputs stable messages pushed to the target audience.

[0190] The following description refers to eight media resource push systems used to implement the above-described embodiments of this application. For example... Figure 8 The media resource push system shown consists of two parts: an online system 800 and an offline system 802. The online system 800 is used to collect and analyze data based on account operations related to the target application's associated accounts.

[0191] The operational data related to multiple object accounts can be divided into two parts: online main traffic and random traffic. In the online data center, online and offline features can be extracted from either the main or random traffic. The extracted online and offline features can then be persisted to disk and stored in the offline system 802's data center.

[0192] In addition, for the online random traffic portion, the prediction model during the training process can be iteratively trained in the offline system 802, and the prediction effect of the prediction model during the training process can be verified based on the random traffic.

[0193] Furthermore, online system 802 provides audience threshold / strategy combination decisions based on the prediction model trained in the offline system. Simultaneously, experimental and control groups are reserved to verify model performance improvements. Specifically, after the prediction model is trained in offline system 802, the target prediction model can be deployed to online system 800. The target prediction model deployed in online system 800 can predict the operation evaluation index corresponding to each object account based on the operation distribution characteristics extracted from the main traffic and referencing account characteristics. Then, based on the prediction results and the media resource delivery strategy deployed in online system 800, it identifies the object accounts from the online traffic to be pushed target media resources and pushes the target media resources to them.

[0194] In addition, the online system can dynamically adjust the online delivery volume in real time and effectively based on the actual media resource push situation and the PID automatic control algorithm.

[0195] Through the above-described embodiments of this application, a prediction model can be built and trained offline using online random data to predict the response behavior of different user groups to specific pushed content. In the offline environment, A / B testing methods can be used to randomly assign users to experimental and control groups to verify the accuracy and improvement effect of the model's predictions. In the online system, based on the prediction results provided by the online-deployed target prediction model and the push strategy updated in real time according to the PID algorithm, the target threshold is updated in real time, thereby achieving dynamic adjustment of the media resource push volume.

[0196] The following combination Figure 9 A complete implementation process of this application is described below. In this embodiment, the target application is a game application, and the target account operation is a login operation, as an example:

[0197] S902, Obtain the first sample set and the second sample set;

[0198] Specifically, two homogeneous sets of target accounts can be obtained from random online traffic. The account information of these accounts indicates that their corresponding users are from first-tier cities, aged 25-35, and of the same gender. Then, message pushes can be sent to the first set of accounts (the experimental group), while no intervention is performed on the second set (the control group). Next, the accounts in both sets are analyzed to determine whether they logged in within a predetermined statistical period. Based on the login activity, corresponding sample labels are assigned to each account, resulting in the first and second sample sets that can be used to train the initial prediction model.

[0199] S904, Train the initial prediction model based on the first sample set and the second sample set;

[0200] During training, the initial prediction model's structure can be as follows: Figure 7 As shown. For the currently acquired sample account, the sample account features of the current sample account can be determined based on the game sequence features 702, account description features 704, and operation distribution features 706. Then, the sample account features are input into the initial shared feature module 718 in the prediction model. The shared feature module 718 includes a feature embedding layer 708 for extracting feature embedding vector representations, a self-attention layer 710 for extracting self-attention weights, and a fully connected layer 712.

[0201] By obtaining the shared sample feature vector of the fully connected layer 712, that is, by representing the confounder through feature processing and deep network hidden layers, the probability of performing target account operations in both cases can be predicted based on this shared representation vector.

[0202] It should be noted that in this embodiment, the module to be input needs to be determined according to the account type corresponding to the current sample account. Specifically, when the current sample account is the first sample account, the shared sample feature vector is input into the first initial prediction module 724 and the third initial prediction module 726 respectively; when the current sample account is the second sample account, the shared sample feature vector is input into the second initial prediction module 722 and the third initial prediction module 726 respectively. That is to say, during the training process, samples with t=1 (i.e., account samples that have pushed target media resources) are fitted by the first initial prediction module 724, and samples with t=0 (i.e., account samples that have not pushed target media resources) are fitted by the second initial prediction module 722.

[0203] In the above implementation, the third initial prediction module 726 can be used to output the propensity score. The third initial prediction module 726 can be constructed by a simple linear mapping (sigmoid). The simple mapping can make the confounder representation layer and the propensity score closely connected, so that only information related to the propensity score is extracted. At this time, the shared sample feature vector is a feature vector that affects both the prediction result of the first sample operation and the prediction result of the second sample operation.

[0204] S906, Training complete? If training is complete, execute S908 to extract the target account features; otherwise, continue executing S904.

[0205] During the execution of S908, the target account characteristics of the target object account may include the reference account characteristics and operation distribution characteristics of the target object account. The reference account characteristics may include account description characteristics and operation description characteristics.

[0206] The following describes the extraction method for the operation distribution features used to characterize login time. In this embodiment, the time parameter in the login time can be encoded into a week feature vector (week embedding), an hour feature vector (hour embedding), and a second feature vector (second embedding). That is, the number of seconds per week, per day, and per minute corresponding to the time when the user account performs the login operation is encoded into separate embedding vectors and fed into the model (7 days in a week can correspond to 7 embeddings, 24 hours in a day can correspond to 24 embeddings, and 60 seconds in a minute can correspond to 60 embeddings).

[0207] In this embodiment, a sine function can be used to encode the time features, and the specific encoding method is as follows:

[0208]

[0209] Here max t This represents the maximum possible value of the time parameter t. For example, the maximum value for seconds is 60, the maximum value for days is 24, and the maximum value for weeks is 7.

[0210] Let's continue with the example of "04 / 24 / 21 / 30". As can be seen in the above implementation, "04 / 24" indicates the 4th day of the 4th week of April.

[0211] Furthermore, for the weekly feature vector week embedding f(t1), max t1 =7, thus determining The week embedding is then determined as a vector.

[0212]

[0213] Furthermore, for the hour embedding f(t2) of the hour feature vector, max t2 =24, thus determining The week embedding is then determined as a vector.

[0214] Furthermore, for the second feature vector second embedding f(t2), maxt3 =60, thus determining The week embedding is then defined as the vector [sin(π),cos(π)].

[0215] By concatenating the week embedding, hour embedding, and second embedding in sequence, we can obtain the feature vector used to characterize the distribution features of the operation.

[0216] S910, Obtain the operational evaluation index of the target account based on the target prediction model;

[0217] The model structure of the target prediction model can be as follows: Figure 6 As shown. In the process of obtaining the operation evaluation index of the target account according to the target prediction model, the game sequence feature 602, account description feature 604, and operation distribution feature 606 can be obtained first. Then, these three are input into the shared feature module 618 in the target prediction model. The shared feature module 618 includes a feature embedding layer 608 for feature extraction using embedding vector representation, a self-attention layer 610 for extracting self-attention weights, and a fully connected layer 612. When the shared feature vector output by the fully connected layer 612 is obtained, the shared feature vector can be input into the corresponding first prediction module 624 and second prediction module 622 respectively. The first prediction module 624 is used to predict the probability that the target account will perform an account login operation when the target media resource is pushed to the target account (i.e., the first operation prediction result 614). The second prediction module 622 is used to predict the probability that the target account will perform an account login operation when the target media resource is not pushed to the target account (i.e., the second operation prediction result 616).

[0218] Execute judgment S912: Is the operation evaluation index greater than or equal to the target threshold? If it is greater than or equal to the target threshold, execute S914 to push the target media resources to the target account; otherwise, obtain the next target account and return to step S908.

[0219] In step S912, the operational evaluation index can be specifically the gain enhancement parameter (Individual Treatment Effect, ITE), i.e. in, for Figure 6 The output of the first prediction module of the target prediction model shown is the probability that the target account will perform an account login operation when the target media resource is pushed. The probability of a target account performing an account login operation when no target media resources are pushed to the target account, as output by the second prediction module of the target prediction model;

[0220] Next, in the next push cycle, judgment S916 can be executed to determine whether the target threshold needs to be updated. If the target threshold needs to be updated according to the PID algorithm, S918 is executed to determine the target media resources to be pushed based on the updated target threshold and the operation evaluation index output by the target prediction model; otherwise, S920 is executed to determine the target media resources to be pushed based on the target threshold and the operation evaluation index output by the target prediction model.

[0221] The above-described implementation methods of this application enable more precise game push notifications, thereby improving game application retention and engagement rates. Furthermore, based on the aforementioned target prediction model, it is possible to identify push-sensitive user groups, thus increasing the overall effectiveness of push notifications. These implementation methods also clearly identify which user groups are most likely to improve operational activity response, providing excellent interpretability of the model results. Simultaneously, the overall push volume can be controlled according to the AUUC curve calculated by the offline model, thereby improving the overall game activity and engagement rate.

[0222] It should be noted that in the relevant embodiments of this application, the acquisition and processing of relevant data information of the target account are all done in advance by means of methods that comply with the relevant normative legal documents, and the authorization permission of the corresponding account subject must be obtained before acquiring the relevant information and data.

[0223] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0224] According to another aspect of the present invention, a media resource push device for implementing the above-described media resource push method is also provided. For example... Figure 10 As shown, the device includes:

[0225] The first determining unit 1002 is used to determine the operation distribution characteristics of the target account based on the time distribution of account operations performed by the target account in the first period.

[0226] The second determining unit 1004 is used to determine a first operation prediction result and a second operation prediction result based on the above-mentioned operation distribution characteristics. The first operation prediction result indicates the probability that the target account will perform the target account operation when the target media resource is pushed to the target account. The second operation prediction result indicates the probability that the target account will perform the target account operation when the target media resource is not pushed to the target account.

[0227] The push unit 1006 is used to push the target media resources to the target object account in the second cycle when the operation evaluation index determined based on the first operation prediction result and the second operation prediction result is greater than or equal to the target threshold.

[0228] Optionally, the second determining unit 1004 includes: an acquisition module for acquiring reference account features associated with the target account; a splicing module for splicing the reference account features with the operation distribution features to obtain target account features, and inputting the result into a feature sharing module of the target prediction model to obtain a shared feature vector; a first input module for inputting the shared feature vector into a first prediction module of the target prediction model to obtain a first operation prediction result; and a second input module for inputting the shared feature vector into a second prediction module of the target prediction model to obtain a second operation prediction result, wherein the first prediction module and the second prediction module are two prediction modules in the target prediction model that are respectively connected to the feature sharing module.

[0229] Optionally, the acquisition module is configured to: acquire operation description features that indicate the target account to perform the account operation within the first period; acquire account description features that match the target account; and use the operation description features and the account description features as the reference account features.

[0230] Optionally, the first determining unit 1002 is configured to: obtain the login time corresponding to each of the multiple login operations performed by the target account in the first period; if the login time includes multiple time parameters, determine the time feature vector corresponding to each of the multiple time parameters according to the multiple time parameters and the time period matched by each of the multiple time parameters; and concatenate the multiple time feature vectors to obtain the operation distribution feature.

[0231] Optionally, the first determining unit 1002 is configured to: obtain the current time parameter and the current time period matching the current time parameter from among the plurality of time parameters; determine the ratio between the current time parameter and the current time period as the login distribution coefficient; obtain a first encoding result of the login distribution coefficient according to a sine function, and obtain a second encoding result of the login distribution coefficient according to a cosine function; and determine the first encoding result and the second encoding result as vector feature values ​​included in the current time feature vector.

[0232] Optionally, the aforementioned media resource push device further includes: a training unit, configured to acquire sample account features of sample object accounts and sample tags matching the sample object accounts; input the sample account features into an initial feature sharing module in an initial prediction model under training to obtain a shared sample feature vector; input the shared sample feature vector into at least one target prediction module among a plurality of initial prediction modules in the initial prediction model, wherein the initial feature sharing module in the initial prediction model is connected to the plurality of initial prediction modules respectively, and the plurality of initial prediction modules includes a first initial prediction module, a second initial prediction module, and a third initial prediction module, wherein the first initial prediction module outputs... The first sample operation prediction result indicates the probability that the sample object account will perform a sample account operation when sample media resources are pushed to the sample object account. The second sample operation prediction result output by the second initial prediction module indicates the probability that the sample object account will perform the sample account operation when sample media resources are not pushed to the sample object account. The tendency prediction result output by the third initial prediction module indicates the learning tendency of the initial prediction model during the training phase. Based on the prediction results output by at least one of the target prediction modules and the sample labels, the initial prediction model is trained until the target prediction model that meets the target training conditions is obtained.

[0233] Optionally, the training unit is configured to: when the sample account features are the account features of a first sample account, input the shared sample feature vector into the first initial prediction module and the third initial prediction module of the initial prediction model to obtain the first sample operation prediction result output by the first initial prediction module and the tendency prediction result output by the third initial prediction module, wherein the first sample account is the object account that received the pushed sample media resource; when the sample account features are the account features of a second sample account, input the shared sample feature vector into the second initial prediction module and the third initial prediction module of the initial prediction model to obtain the second sample operation prediction result output by the second initial prediction module and the tendency prediction result output by the third initial prediction module, wherein the second sample account is the object account that did not receive the sample media resource.

[0234] Optionally, the training unit is configured to: determine a first sample set and a second sample set, wherein the difference between the first number of first sample accounts included in the first sample set and the second number of second sample accounts included in the second sample set is less than or equal to a target number threshold; push the sample media resources to the first sample accounts in the first sample set; and determine the sample tags corresponding to each of the multiple sample object accounts based on the account operations performed by the multiple sample object accounts in the first sample set and the second sample set during the target account operation period, including: configuring a first sample tag for the sample object account when the sample object account has performed the sample account operation, and configuring a second sample tag for the sample object account when the sample object account has not performed the sample account operation.

[0235] Optionally, the push unit 1006 is configured to: obtain the quantity difference between the first push quantity and the second push quantity associated with the target media resource in the previous push cycle, wherein the first push quantity is the number of target media resources that have been pushed, and the second push quantity is the number of target media resources that are planned to be pushed; if the quantity difference is greater than or equal to a target value, raise the reference threshold to the target threshold; if the quantity difference is less than the target value, lower the reference threshold to the target threshold.

[0236] Optionally, the push unit 1006 is configured to: determine the product between the quantity difference and the threshold revision coefficient as the target adjustment value; and adjust the reference threshold to the target threshold according to the target adjustment value.

[0237] Optionally, the push unit 1006 is further configured to: when the push stage corresponding to the current push cycle is the target stage, obtain a push control model matching the next push cycle, wherein the push control model is a control model determined based on a proportional coefficient, an integral coefficient, and a derivative coefficient; obtain the current quantity difference between the first push quantity of the target media resources already pushed and the second push quantity of the target media resources planned to be pushed in the current push cycle; obtain the cumulative quantity difference between the first cumulative push quantity of the target media resources already pushed and the second cumulative push quantity of the target media resources planned to be pushed in the current push cycle and the previous N push cycles, wherein N is an integer greater than or equal to 1; obtain the current quantity difference of the current push cycle and the historical quantity difference of the previous push cycle, and determine the quantity difference change rate based on the current quantity difference and the historical quantity difference; adjust the target threshold based on the first product between the proportional coefficient and the current quantity difference, the second product between the integral coefficient and the cumulative quantity difference, and the third product between the derivative coefficient and the quantity difference change rate.

[0238] Optionally, in this embodiment, the implementation of each of the above-mentioned unit modules can be referred to the above-mentioned method embodiments, which will not be repeated here.

[0239] According to another aspect of the present invention, an electronic device for implementing the above-described media resource push method is also provided, the electronic device being... Figure 11 The terminal device or server shown. This embodiment uses this electronic device as an example for illustration. Figure 11 As shown, the electronic device includes a memory 1102 and a processor 1104. The memory 1102 stores a computer program, and the processor 1104 is configured to execute the steps of any of the above method embodiments via the computer program.

[0240] Optionally, in this embodiment, the aforementioned electronic device may be located in at least one of a plurality of network devices in a computer network.

[0241] Optionally, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0242] S1. Determine the operation distribution characteristics of the target account based on the time distribution of account operations performed by the target account in the first period.

[0243] S2, determine the first operation prediction result and the second operation prediction result based on the above operation distribution characteristics, wherein the first operation prediction result indicates the probability that the target account will perform the target account operation when the target media resource is pushed to the target account, and the second operation prediction result indicates the probability that the target account will perform the target account operation when the target media resource is not pushed to the target account.

[0244] S3, if the operation evaluation index determined based on the first operation prediction result and the second operation prediction result is greater than or equal to the target threshold, the target media resources are pushed to the target account in the second cycle.

[0245] Alternatively, as those skilled in the art will understand, Figure 11 The structure shown is for illustrative purposes only. Electronic devices can also be in-vehicle terminals, smartphones (such as Android phones, iOS phones, etc.), tablets, handheld computers, mobile internet devices (MIDs), PADs, and other terminal devices. Figure 11 This does not limit the structure of the aforementioned electronic devices. For example, the electronic device may also include components that are more... Figure 11 The more or fewer components shown (such as network interfaces, etc.), or having the same Figure 11 The different configurations shown.

[0246] The memory 1102 can be used to store software programs and modules, such as the program instructions / modules corresponding to the media resource push method and apparatus in this embodiment of the invention. The processor 1104 executes various functional applications and data processing by running the software programs and modules stored in the memory 1102, thereby realizing the aforementioned media resource push method. The memory 1102 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 1102 may further include memory remotely located relative to the processor 1104, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. Specifically, the memory 1102 may be used, but is not limited to, for storing file information such as target logical files. As an example, such as Figure 11 As shown, the memory 1102 may include, but is not limited to, the first determining unit 1002, the second determining unit 1004, and the pushing unit 1006 in the media resource pushing device. Furthermore, it may include, but is not limited to, other module units in the media resource pushing device, which will not be elaborated upon in this example.

[0247] Optionally, the transmission device 1106 described above is used to receive or send data via a network. Specific examples of the network described above may include wired networks and wireless networks. In one example, the transmission device 1106 includes a Network Interface Controller (NIC), which can be connected to other network devices and a router via a network cable to communicate with the Internet or a local area network. In another example, the transmission device 1106 is a radio frequency (RF) module, used for wireless communication with the Internet.

[0248] In addition, the above-mentioned electronic device also includes a display 1108 and a connection bus 1110 for connecting the various module components in the above-mentioned electronic device.

[0249] In other embodiments, the aforementioned terminal device or server can be a node in a distributed system, wherein the distributed system can be a blockchain system, which is a distributed system formed by connecting multiple nodes through network communication. The nodes can form a peer-to-peer network, and any form of computing device, such as a server, terminal, or other electronic device, can become a node in the blockchain system by joining this peer-to-peer network.

[0250] According to one aspect of this application, a computer program product is provided, comprising a computer program / instructions containing program code for performing the methods shown in the flowchart. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit, it performs various functions provided in embodiments of this application.

[0251] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0252] According to one aspect of this application, a computer-readable storage medium is provided, wherein a processor of a computer device reads computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to cause the computer device to perform the aforementioned media resource push method.

[0253] Optionally, in this embodiment, the computer-readable storage medium may be configured to store a computer program for performing the following steps:

[0254] S1. Determine the operation distribution characteristics of the target account based on the time distribution of account operations performed by the target account in the first period.

[0255] S2, determine the first operation prediction result and the second operation prediction result based on the above operation distribution characteristics, wherein the first operation prediction result indicates the probability that the target account will perform the target account operation when the target media resource is pushed to the target account, and the second operation prediction result indicates the probability that the target account will perform the target account operation when the target media resource is not pushed to the target account.

[0256] S3, if the operation evaluation index determined based on the first operation prediction result and the second operation prediction result is greater than or equal to the target threshold, the target media resources are pushed to the target account in the second cycle.

[0257] Optionally, in this embodiment, those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0258] If the integrated units in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in the aforementioned computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause one or more computer devices (which may be personal computers, servers, or network devices, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0259] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0260] In the several embodiments provided in this application, it should be understood that the disclosed client can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, indirect coupling or communication connection between units or modules, and may be electrical or other forms.

[0261] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0262] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0263] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A push method of a media resource, characterized by, The method comprises the following steps: According to the time distribution of the account operation of the target object account in the first period, the operation distribution characteristics of the target object account are determined; According to the operation distribution characteristics, the first operation prediction result and the second operation prediction result are determined, wherein the first operation prediction result indicates the probability of the target object account performing the target account operation in the case of pushing the target media resource to the target object account, and the second operation prediction result indicates the probability of the target object account performing the target account operation in the case of not pushing the target media resource to the target object account; In the case where the operation evaluation index determined according to the first operation prediction result and the second operation prediction result is greater than or equal to the target threshold value, the target media resource is pushed to the target object account in the second period.

2. The method of claim 1, wherein, The method comprises the following steps: Obtaining the reference account characteristics associated with the target object account; The target account characteristics obtained by splicing the reference account characteristics and the operation distribution characteristics are input into the feature sharing module of the target prediction model to obtain the shared feature vector; The shared feature vector is input into the first prediction module of the target prediction model to obtain the first operation prediction result; The shared feature vector is input into the second prediction module of the target prediction model to obtain the second operation prediction result, wherein the first prediction module and the second prediction module are two prediction modules connected with the feature sharing module in the target prediction model.

3. The method of claim 2, wherein, The method comprises the following steps: Obtaining the operation description characteristics for indicating the account operation of the target object account in the first period; Obtaining the account description characteristics matched with the target object account; The operation description characteristics and the account description characteristics are used as the reference account characteristics.

4. The method of claim 2, wherein, The method comprises the following steps: Obtaining the login time corresponding to each of the multiple login operations performed by the target object account in the first period; In the case where multiple time parameters are included in the login time, according to multiple time parameters and time periods matched with the multiple time parameters respectively, time feature vectors corresponding to the multiple time parameters respectively are determined; The operation distribution characteristics are obtained by splicing multiple time feature vectors.

5. The method of claim 4, wherein, The method comprises the following steps: Obtaining the current time parameter in the multiple time parameters and the current time period matched with the current time parameter; The ratio between the current time parameter and the current time period is determined as the login distribution coefficient; The first encoding result of the login distribution coefficient is obtained according to the sine function, and the second encoding result of the login distribution coefficient is obtained according to the cosine function; The first encoding result and the second encoding result are determined as vector feature values included in a current time feature vector.

6. The method of claim 2, wherein, Before the first operation prediction result and the second operation prediction result are determined according to the operation distribution feature, the method further includes: obtaining a sample account feature of a sample object account and a sample label matched with the sample object account; inputting the sample account feature into an initial feature sharing module in an initial prediction model in a training state to obtain a shared sample feature vector; inputting the shared sample feature vector into at least one target prediction module of a plurality of initial prediction modules in the initial prediction model, wherein the initial feature sharing module and the plurality of initial prediction modules are connected respectively in the initial prediction model, the plurality of initial prediction modules include a first initial prediction module, a second initial prediction module and a third initial prediction module, the first sample operation prediction result output by the first initial prediction module is used to indicate a probability of the sample object account performing a sample account operation in a case where a sample media resource is pushed to the sample object account, the second sample operation prediction result output by the second initial prediction module is used to indicate a probability of the sample object account performing the sample account operation in a case where the sample media resource is not pushed to the sample object account, and the tendency prediction result output by the third initial prediction module is used to indicate a learning tendency of the initial prediction model in a training stage; training the initial prediction model according to a prediction result output by at least one target prediction module and the sample label until the target prediction model satisfying a target training condition is obtained.

7. The method of claim 6, wherein, The inputting the shared sample feature vector into at least one target prediction module of a plurality of prediction modules in the initial prediction model includes: in a case where the sample account feature is an account feature of a first sample account, inputting the shared sample feature vector into the first initial prediction module and the third initial prediction module of the initial prediction model to obtain the first sample operation prediction result output by the first initial prediction module and the tendency prediction result output by the third initial prediction module, wherein the first sample account is an object account receiving the pushed sample media resource; in a case where the sample account feature is an account feature of a second sample account, inputting the shared sample feature vector into the second initial prediction module and the third initial prediction module of the initial prediction model to obtain the second sample operation prediction result output by the second initial prediction module and the tendency prediction result output by the third initial prediction module, wherein the second sample account is an object account not receiving the sample media resource.

8. The method of claim 6, wherein, The obtaining the sample account feature of the sample object account and the sample label includes: determining a first sample set and a second sample set, wherein a quantity difference between a first quantity of first sample accounts included in the first sample set and a second quantity of second sample accounts included in the second sample set is less than or equal to a target quantity threshold; pushing the sample media resource to the first sample accounts in the first sample set; determining a sample label corresponding to each of the sample object accounts according to an account operation performed by each of the sample object accounts in the first sample set and the second sample set within a target account operation period, including: configuring a first sample label for the sample object account in a case where the sample object account performs the sample account operation, and configuring a second sample label for the sample object account in a case where the sample object account does not perform the sample account operation.

9. The method of claim 1, wherein, in a case where an operation evaluation index determined according to the first operation prediction result and the second operation prediction result is greater than or equal to a target threshold, before pushing the target media resource to the target object account in a second period, further comprising: obtaining a quantity difference between a first push quantity and a second push quantity associated with the target media resource in a last push period, wherein the first push quantity is a quantity of the target media resource that has been pushed, and the second push quantity is a quantity of the target media resource that is planned to be pushed; in a case where the quantity difference is greater than or equal to a target value, increasing a reference threshold to the target threshold; in a case where the quantity difference is less than the target value, decreasing the reference threshold to the target threshold.

10. The method of claim 9, wherein, after the obtaining of the quantity difference between the first push quantity and the second push quantity associated with the target media resource in the last push period, comprising: determining a target adjustment value as a product between the quantity difference and a threshold revision coefficient; adjusting the reference threshold to the target threshold according to the target adjustment value.

11. The method of claim 9, wherein, in a case where an operation evaluation index determined according to the first operation prediction result and the second operation prediction result is greater than or equal to a target threshold, after pushing the target media resource to the target object account in a second period, further comprising: in a case where a push stage corresponding to a current push period is a target stage, obtaining a push control model matched with a next push period, wherein the push control model is a control model determined according to a proportional coefficient, an integral coefficient, and a differential coefficient; obtaining a current quantity difference between a first push quantity of the target media resource that has been pushed in a current push period and a second push quantity of the target media resource that is planned to be pushed; obtaining a cumulative quantity difference between a first cumulative push quantity of the target media resource that has been pushed and a second cumulative push quantity of the target media resource that is planned to be pushed in the current push period and N previous push periods, wherein N is an integer greater than or equal to 1; obtain a current quantity difference value of the current push period and a historical quantity difference value of a previous push period, and determine a quantity difference change rate according to the current quantity difference value and the historical quantity difference value; adjust the target threshold value according to a first product between the proportion coefficient and the current quantity difference value, a second product between the integral coefficient and the cumulative quantity difference value, and a third product between the differential coefficient and the quantity difference change rate.

12. A media resource pushing apparatus, characterized in that, Comprise: The first determination unit is used for determining operation distribution characteristics of the target object account according to time distribution of the target object account performing account operation in a first period; The second determination unit is used for determining a first operation prediction result and a second operation prediction result according to the operation distribution characteristics, wherein the first operation prediction result indicates a probability of the target object account performing a target account operation in a case of pushing a target media resource to the target object account, and the second operation prediction result indicates a probability of the target object account performing the target account operation in a case of not pushing the target media resource to the target object account; The push unit is used for pushing the target media resource to the target object account in a second period in a case that an operation evaluation index determined according to the first operation prediction result and the second operation prediction result is greater than or equal to a target threshold value.

13. A computer readable storage medium, characterized in that, The computer readable storage medium comprises a stored program, wherein the program performs the method in any one of claims 1 to 11 when running.

14. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instruction is executed by the processor to realize the steps of the method in any one of claims 1 to 11.

15. An electronic device comprising a memory and a processor, characterized in that The memory stores a computer program, and the processor is configured to execute the method in any one of claims 1 to 11 by the computer program. The memory stores a computer program, and the processor is configured to execute the method in any one of claims 1 to 11 by the computer program.

Images