Message pushing method and related device

Abstract

The invention discloses a message pushing method and a related device, and the method comprises the steps: determining a prediction entry probability and a prediction operation result according to object features and message features for a scene in which a message box is used as an entrance and a plurality of messages are displayed in the message box; the predicted entry probability is used for representing the probability that the object implements the entry operation when the message box pushes the candidate message with the message prompts, and can reflect whether the pushing of the message prompts can attract the object to implement the entry operation or not. The prediction operation result is used for indicating the operation implementation characteristics of the object on the multiple messages in the message box based on the entering operation, and the operation conditions implemented on the multiple messages in the box after entering the message box can be reflected. In the embodiment of the invention, the operation aiming at other messages in the box after the attraction of the message prompt also belongs to the income brought by the push of the message prompt, so that the push value determined based on the predicted entry probability and the predicted operation result can more accurately evaluate the value brought by the push of the message prompt, thereby improving the push accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and particularly relates to a message pushing method and related device. BACKGROUND

[0002] With the continuous development of Internet technology, pushing messages to users has become a convenient way for users to obtain messages.

[0003] In order to prompt the user to view the message in time, a message prompter is used to push the message, that is, a relatively eye-catching message prompter is displayed at a proper position to prompt the user that there is a message to be viewed. For example, a red dot push is a commonly used way of pushing messages by using a message prompter. This way can attract the attention of the user, thereby being conducive to improving the activity and parameter of the user, and being conducive to improving the user stickiness of the system and obtaining higher pushing benefits. However, too frequent message prompter pushing may cause the user to be disgusted, and the like. Therefore, how to effectively push the message prompter has great significance.

[0004] In the related art, the interest of the user in the message is usually predicted first to evaluate the benefits that may be brought by pushing the message prompter after the message prompter is pushed for the message, so as to determine whether to push the message prompter for the message. However, the pushing value estimated by the method in the related art is difficult to accurately reflect the real value, thereby leading to poor pushing accuracy. SUMMARY

[0005] In order to solve the above technical problems, the present application provides a message pushing method and related device, which can more accurately push the message prompter for the message box, is conducive to improving the pushing effect, and is also conducive to reducing the interference of unnecessary message prompter pushing on the user and improving the user experience.

[0006] The embodiments of the present application disclose the following technical solutions:

[0007] In one aspect, the present application provides a message pushing method, and the method comprises the following steps:

[0008] In response to a message pushing request for a message box, an object and a candidate message indicated by the message pushing request are acquired.

[0009] According to the object feature of the object and the message feature of the candidate message, a predicted entering probability corresponding to the candidate message and a predicted operation result corresponding to the candidate message are determined, the predicted entering probability is used to represent a probability of implementing an entering operation by the object when the message box pushes the candidate message with a message prompt, the entering operation is used to indicate displaying a plurality of messages containing the candidate message in the message box, and the predicted operation result is used to indicate an operation implementation feature of the object in the message box for the plurality of messages based on the entering operation;

[0010] Based on the predicted entering probability and the predicted operation result, a push value corresponding to the candidate message is determined, and the push value is used to determine whether the candidate message is pushed with a message prompt in the message box associated with the object.

[0011] In another aspect, an embodiment of the present application provides a message pushing device, the device comprising an acquisition unit and a determination unit:

[0012] The acquisition unit is configured to acquire an object and a candidate message indicated by a message pushing request in response to the message pushing request for a message box.

[0013] The determination unit is configured to determine a predicted entering probability corresponding to the candidate message and a predicted operation result corresponding to the candidate message according to an object feature of the object and a message feature of the candidate message, the predicted entering probability is used to represent a probability of implementing an entering operation by the object when the message box pushes the candidate message with a message prompt, the entering operation is used to indicate displaying a plurality of messages containing the candidate message in the message box, and the predicted operation result is used to indicate an operation implementation feature of the object in the message box for the plurality of messages based on the entering operation.

[0014] The determination unit is further configured to determine a push value corresponding to the candidate message based on the predicted entering probability and the predicted operation result, and the push value is used to determine whether the candidate message is pushed with a message prompt in the message box associated with the object.

[0015] In a possible implementation, the determination unit is further configured to:

[0016] According to the object feature and the message feature, a predicted entering probability corresponding to the candidate message is determined, and a predicted operation probability corresponding to the candidate message and a predicted operation number corresponding to the candidate message are determined, the predicted operation probability is used to represent a probability of implementing an operation by the object for the candidate message based on the entering operation, and the predicted operation number is used to represent an operation implementation times feature of the object for the plurality of messages based on the entering operation.

[0017] determine the push value corresponding to the candidate message based on the predicted entering probability, the predicted operation probability, and the predicted operation number.

[0018] In a possible implementation, if the predicted operation result includes a predicted distribution center parameter corresponding to the candidate message and a predicted distribution deviation parameter corresponding to the candidate message, the predicted distribution center parameter is used to represent a value center of the predicted operation number corresponding to the candidate message, and the predicted distribution deviation parameter is used to represent a discrete degree of the predicted operation number corresponding to the candidate message distributed at the value center, and the determining unit is further configured to:

[0019] determine the push value corresponding to the candidate message based on the predicted entering probability, the predicted distribution center parameter, and the predicted distribution deviation parameter.

[0020] In a possible implementation, if the predicted distribution center parameter is a predicted logarithmic mean value of the predicted operation number of the candidate message, and the predicted distribution deviation parameter is a predicted logarithmic standard deviation of the predicted operation number of the candidate message, the determining unit is further configured to:

[0021] determine an operation expectation corresponding to the candidate message based on the predicted logarithmic mean value and the predicted logarithmic standard deviation, the operation expectation being used to represent a number of operation implementations of the object in the message box for the plurality of messages when the candidate message is pushed with a message prompt.

[0022] determine the push value corresponding to the candidate message based on the predicted entering probability and the operation expectation.

[0023] In a possible implementation, the determining unit is further configured to:

[0024] determine a predicted entering probability corresponding to the candidate message based on the object feature and the message feature, and determine a predicted category probability of each operation category corresponding to the candidate message, the operation category being determined according to a preset operation number, and different operation categories corresponding to different preset operation numbers.

[0025] determine the push value corresponding to the candidate message based on the predicted entering probability and the predicted category probability of each operation category corresponding to the candidate message.

[0026] In a possible implementation, if the number of the candidate messages is a plurality, the determining unit is further configured to:

[0027] determine a target message from the candidate messages based on the push values corresponding to the plurality of candidate messages, the push value of the target message satisfying a push condition.

[0028] The message box associated with the object displays a message prompt and a message title of the target message.

[0029] In a possible implementation, the determination unit is further configured to:

[0030] obtain a box-in push strategy corresponding to the message box;

[0031] determine, according to the box-in push strategy and a push value corresponding to each of a plurality of candidate messages other than the target message, a plurality of box-in messages corresponding to the target message and a display order of the plurality of box-in messages in the message box;

[0032] in response to the entering operation performed by the object when the target message is pushed in the message prompt, display the target message in the message box, and display the plurality of box-in messages in the display order after the target message.

[0033] In a possible implementation, the object feature includes a historical operation number of the object performed through the message box in a historical period, and the message feature includes a historical number of published messages and a historical number of operated messages of a publishing account corresponding to the candidate message in the historical period, and a time interval between a maximum time of the historical period and a request time of the message push request is less than a preset interval.

[0034] In a possible implementation, the determination unit is further configured to:

[0035] determine, according to the object feature and the message feature, the predicted entering probability and the predicted operation result by a target model;

[0036] obtain a training sample, the training sample including an object feature of a sample object and a message feature of a sample message, the training sample including an entering label and a box-in operation label, the entering label being used to indicate whether the sample object performs the entering operation when the message box pushes the sample message in the message prompt, the entering operation being used to indicate that a plurality of messages including the sample message are displayed in the message box, and the box-in operation label being used to indicate an operation performance feature of the sample object in the message box based on the entering operation;

[0037] According to the training sample, a sample predicted entering probability corresponding to the sample message and a sample predicted operation result corresponding to the sample message are determined by an initial model, the sample predicted entering probability is used to represent a probability of implementing the entering operation by the sample object when the sample message is pushed with a message prompt, and the sample predicted operation result is used to represent an operation implementation feature of the sample object in the message box for the plurality of messages based on the entering operation;

[0038] An entering loss is determined according to the sample predicted entering probability and the entering label, and an operation loss is determined according to the sample predicted operation result and the in-box operation label;

[0039] The initial model is trained according to the entering loss and the operation loss to obtain the target model.

[0040] In a possible implementation, if the in-box operation label includes a first operation label and a second operation label, the first operation label is used to represent whether the sample object implements operation for the sample message, and the second operation label is a sample operation number, which is determined according to a number of times of implementing operation by the sample object in the message box for the plurality of messages, and the determination unit is further configured to:

[0041] According to the training sample, the sample predicted entering probability is determined by the initial model, and a sample predicted operation probability corresponding to the sample message and a sample predicted operation number corresponding to the sample message are determined by the initial model, the sample predicted operation probability is used to represent a probability of implementing operation by the sample object for the sample message based on the entering operation;

[0042] A first operation loss is determined according to the sample predicted operation probability and the first operation label, and a second operation loss is determined according to the sample predicted operation number and the sample operation number, and the operation loss includes the first operation loss and the second operation loss.

[0043] In a possible implementation, if the in-box operation label is a sample operation number, the sample predicted operation result includes a sample predicted distribution center parameter corresponding to the sample message and a sample predicted distribution deviation parameter corresponding to the sample message, the sample predicted distribution center parameter is used to represent a value center of the sample predicted operation number corresponding to the sample message, and the sample predicted distribution deviation parameter is used to represent a discrete degree of distribution of the sample predicted operation number corresponding to the sample message at the value center, and the determination unit is further configured to:

[0044] The operation loss is determined according to the sample predicted distribution center parameter, the sample predicted distribution deviation parameter and the sample operation number.

[0045] In a possible implementation, if the in-box operation label is used to represent a sample operation category corresponding to the sample message, the sample operation category is determined according to a sample operation number corresponding to the sample message, and the sample operation number is determined according to a number of times of operation of the sample object in the message box on the plurality of messages, the determination unit is further configured to:

[0046] According to the training sample, the initial model is used to determine the sample predicted entering probability, and the initial model is used to determine a sample predicted category probability of each operation category in the plurality of operation categories corresponding to the sample message, and the preset operation number corresponding to different operation categories is different.

[0047] According to the sample predicted category probability of the sample message corresponding to the plurality of operation categories respectively and the in-box operation label, the operation loss is determined.

[0048] In a possible implementation, the training sample further includes a time feature, and the time feature is used to represent a time interval between an operation time of the sample object performing the entering operation and a push time of the message box pushing the sample message as a message prompt.

[0049] In another aspect, an embodiment of the present application provides a computer device, which includes a processor and a memory:

[0050] The memory is configured to store a computer program and transmit the computer program to the processor.

[0051] The processor is configured to execute the method according to the instructions in the computer program.

[0052] In another aspect, an embodiment of the present application provides a computer readable storage medium, which is configured to store a computer program, and the computer program, when executed by a computer device, causes the computer device to execute the method according to any one of the preceding aspects.

[0053] In another aspect, an embodiment of the present application provides a computer program product, which includes a computer program, and when the computer program is executed on a computer device, causes the computer device to execute the method according to any one of the preceding aspects.

[0054] According to the technical solution, for the scenario of taking the message box as an entrance and displaying multiple messages in the message box, according to the subsequent operation triggered by the object attracted by the message prompt push, the value brought by the message prompt push is evaluated in two aspects. One aspect is whether the message prompt push will attract the object to perform an entering operation, and the other aspect is the operation performed on the multiple messages in the message box after entering the message box. In specific implementation, first, the predicted entering probability and the predicted operation result are determined according to the object characteristics and the message characteristics. The predicted entering probability can be used to represent the probability of the object performing an entering operation when the message box pushes a candidate message with a message prompt. The entering operation can be used to indicate that the multiple messages containing the candidate message are displayed in the message box, that is, whether the message prompt push will attract the object to perform an entering operation. The predicted operation result can be used to indicate the operation implementation characteristics of the object on the multiple messages in the message box based on the entering operation, that is, the operation performed on the multiple messages in the message box after entering the message box. Since the messages displayed in the box are not only the candidate message, the operation on other messages in the box after being attracted by the message prompt also belongs to the benefits brought by the message prompt push. Therefore, the push value determined based on the predicted entering probability and the predicted operation result can more accurately evaluate the value brought by pushing the candidate message with the message prompt. Accordingly, the push value can be used to determine whether to push the candidate message with the message prompt in the message box associated with the object. Based on this, because the more accurate push value can be determined, the message prompt push can be more accurately performed for the message box, which is beneficial to improve the push effect and reduce unnecessary message prompt push interference on the user, and improve the user experience. BRIEF DESCRIPTION OF DRAWINGS

[0055] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the embodiments or the related art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0056] Figure 1 An application scenario diagram of a message push method provided by an embodiment of the present application;

[0057] Figure 2 A flowchart of a message push method provided by an embodiment of the present application;

[0058] Figure 3 A generated reading message number statistical distribution diagram provided by an embodiment of the present application;

[0059] Figure 4A message pushing schematic diagram of a message box provided by an embodiment of the present application;

[0060] Figure 5 A model structure schematic diagram of an initial model provided by an embodiment of the present application;

[0061] Figure 6 A structure diagram of a message pushing apparatus provided by an embodiment of the present application;

[0062] Figure 7 A structure diagram of a terminal provided by an embodiment of the present application;

[0063] Figure 8 A structure diagram of a server provided by an embodiment of the present application. DETAILED DESCRIPTION

[0064] The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0065] In order to prompt the user to view the message in time, a message prompter is usually used to push the message, specifically, a more eye-catching message prompter is displayed in a proper position (such as a menu item and a list item in an application program, which are easily seen by the user), so as to remind the user that there is a message to be viewed. After the user sees the message prompter, the user usually clicks to view the message, and then the message prompter disappears. For example, a red dot pushing is a commonly used way of pushing the message by the message prompter, that is, the message prompter can be a red dot.

[0066] In actual application, the user is usually pushed the message by a client (for example, an application program). The message box can be a module in the client, the message box can be used as an entrance of pushing the message, the user can enter the message box by interacting with the message box (for example, clicking the message box), and the specific message can be displayed in the message box, so as to facilitate the user to view the message and manage the message. As a module in the client, the message box can be displayed in an application page when the client runs, and other contents of the client are usually displayed in an area outside the message box in the application page, and the area outside the message box can be referred to as outside the box.

[0067] For example, in the application scenario of the message pushing method provided by an embodiment of the present application, the message box can be a module in the client, the message box can be used as an entrance of pushing the message, the user can enter the message box by interacting with the message box (for example, clicking the message box), and the specific message can be displayed in the message box, so as to facilitate the user to view the message and manage the message. As a module in the client, the message box can be displayed in an application page when the client runs, and other contents of the client are usually displayed in an area outside the message box in the application page, and the area outside the message box can be referred to as outside the box. Figure 1 Figure 1 ​As shown by the dashed box in the example, an application page of a certain client on which an account of user A is logged in can be displayed on the terminal 100. In the application page, a message box associated with user A is displayed, and the message "Chat Essential Emoticons" is displayed in the message box in the form of a message prompt. In actual applications, user A can enter the message box by clicking the message box to view the message "Chat Essential Emoticons".

[0068] To improve the message pushing effect and avoid causing user complaints, for a message that needs to be pushed, the interest of the user in the message is usually predicted first to evaluate the pushing benefit that can be brought by pushing the message in the form of a message prompt, so as to determine whether to push the message to the user in the form of a message prompt. In related technologies, it is predicted that when the message is pushed in the form of a message prompt, the user will enter the message box and view the message in the message box, which is used as the pushing value of the message, and is used to determine whether to push the message in the form of a message prompt.

[0069] In actual applications, in a scenario in which the message box is used as an entrance for message pushing, the message box displays an information stream composed of multiple messages, and when the user enters the message box, the user also sees an information stream composed of multiple messages (such as an information stream composed of multiple text messages, video messages, and the like). Therefore, when the user is attracted by the message prompt (such as a red dot) and enters the box, the consumption (such as viewing other messages) of other messages in the box also brings corresponding pushing benefits.

[0070] However, in related technologies, only whether the user will enter and view the message pushed in the form of a message prompt is considered, and therefore, the pushing value determined by the related technologies cannot accurately reflect the value that can be brought by actual pushing, which results in poor pushing accuracy and can also push some messages with low real value in the form of a message prompt, which is easy to cause user complaints.

[0071] To this end, the embodiments of the present application provide a message pushing method and related device. For the scenario of taking a message box as an entrance and displaying multiple messages in the message box, according to subsequent operations triggered by a user after being attracted by a message prompt, the value possibly brought by the message prompt pushing is evaluated in two aspects. Specifically, one aspect is to focus on whether the message prompt pushing will attract the object to implement an entering operation, and the other aspect is to focus on the operation situation implemented for the multiple messages in the box after entering the message box. Then, the final pushing value is determined in combination with the results of the two aspects. In this way, all consumption situations generated in the box after the user enters the message box are comprehensively considered, so that the determined pushing value can more accurately evaluate the value possibly brought by pushing the message with the message prompt. Based on this, the message prompt pushing can be more accurately performed for the message box, which is conducive to improving the pushing effect and also conducive to reducing unnecessary message prompt pushing interference on the user and improving the user experience.

[0072] The message pushing method provided by the embodiments of the present application can be implemented by a computer device, which can be a terminal or a server. The server can be an independent physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. The terminal includes but is not limited to a smart phone, a computer, a smart voice interaction device, a smart home appliance, a vehicle-mounted terminal, etc. The terminal and the server can be directly or indirectly connected through wired or wireless communication, which is not limited in the present application. The embodiments of the present application can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, intelligent transportation, audio and video, and auxiliary driving, etc. The embodiments of the present application can be specifically applied to various message pushing scenarios for the message box, such as the scenario of pushing a message with a red dot for the message box, etc.

[0073] The data processing method provided in the embodiments of the present application can involve artificial intelligence technology. Artificial intelligence (AI) is the theory, method, technology and application system for using digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use the knowledge to obtain the best results. Among them, artificial intelligence technology is a comprehensive discipline, involving a wide range of fields, both hardware and software technologies. Artificial intelligence basic technologies generally include technologies such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction system, mechatronics, etc. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing (NLP) technology, and machine learning / deep learning, autonomous driving, intelligent transportation, etc. For example, in the embodiments of the present application, the natural language processing and machine learning can be used to train the initial model to obtain the target model, which is convenient for determining the related results (such as predicted entry probability, predicted operation probability, etc.) for calculating the push value, and is beneficial to improve the push efficiency.

[0074] It should be noted that in the specific embodiments of the present application, the process of message pushing can involve user information and other related data. When the above embodiments of the present application are applied to specific products or technologies, the user's separate consent or separate license is required, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards of relevant countries and regions.

[0075] Figure 1 The application scenario of the message pushing method provided in the embodiments of the present application is shown. In the application scenario, the server 200 is taken as an example of the foregoing computer device. Figure 1

[0076] For the message pushing scenario taking the message box as the entry, when message pushing is needed, the server 200 can first acquire the object and the candidate message indicated by the message pushing request in response to the message pushing request for the message box. Among them, the object can be a user (such as user A in the example), and the candidate message can be a message to be pushed (such as a new message), which can be a message that needs to be judged whether to push to the object. Figure 1

[0077] ​​In order to more accurately evaluate the push value, in the present application, based on the subsequent operation triggered by the object being attracted by the message prompter push, the value brought by the message prompter push is evaluated in two aspects. One aspect focuses on whether the message prompter push will attract the object to implement the entering operation, and the other aspect focuses on the operation situation implemented for the plurality of messages in the message box after entering the message box. To this end, the server 200 can determine the predicted entering probability and the predicted operation result according to the object characteristics of the object and the message characteristics of the candidate message.

[0078] The predicted entering probability can be used to indicate the probability of the object implementing the entering operation when the message box pushes the candidate message with the message prompter. The entering operation can be used to indicate that the plurality of messages containing the candidate message are displayed in the message box, that is, it can reflect whether the message prompter push will attract the object to implement the entering operation. The predicted operation result can be used to indicate the operation implementation characteristics of the object for the plurality of messages in the message box based on the entering operation, that is, it can reflect the operation situation implemented for the plurality of messages in the message box after entering the message box.

[0079] Then, the server 200 can determine the push value of the candidate message according to the predicted entering probability and the predicted operation result. Correspondingly, the push value can be used to determine whether the candidate message is pushed with the message prompter in the message box associated with the object.

[0080] For ease of understanding, in Figure 1 In the example, the push example of pushing the candidate message with the message prompter is shown in the dashed box based on the push value of the candidate message. Wherein, the aforementioned object can be user A, and the aforementioned candidate message can be the message "chat essential emoticon package". In actual application, the application page of a certain client logged in with the account of user A can be displayed on the terminal 100. The application page includes the message box associated with user A, and the message prompter is displayed in the upper right corner of the message box. The message box can also display the message title of the candidate message. Based on the prompt of the message prompter and the message title, user A can enter the message box and then perform subsequent operations, etc.

[0081] Compared with the value prediction method adopted in the related art, in the present application, the operation situation implemented for the plurality of messages in the message box after the user is attracted into the message box by the message prompter is comprehensively considered. Obviously, the overall consumption in the message box is considered, and it is not limited to the message itself pushed by the message prompter. Based on this, the push value determined can more accurately evaluate the value that the candidate message pushed with the message prompter can bring. In this way, the message prompter push can be more accurately performed for the message box, which is conducive to improving the push effect, increasing the push income, etc. At the same time, it is also conducive to reducing unnecessary message prompter push, thereby avoiding excessive interference on the user, and improving the user experience.

[0082] Figure 2 A flowchart of a message pushing method provided by an embodiment of the present application is described by taking a server as an example of the foregoing computer device, and the method comprises S201-S203.

[0083] S201: In response to a message pushing request for a message box, an object and a candidate message indicated by the message pushing request are acquired.

[0084] For the scenario of pushing a message by taking a message box as an entrance, a message pushing request can be generated when it is necessary to push a message. For example, a message pushing request can be generated when a new message is found, so as to trigger S201 to be implemented and to push subsequently.

[0085] Correspondingly, the server can first acquire an object and a candidate message indicated by the message pushing request in response to the message pushing request for the message box. The candidate message can be a message that needs to be judged whether to be pushed, and the object can be a user to be pushed by the candidate message. In actual application, the message box can be a module in a certain client (such as an application), and the client is often oriented to multiple users, so the foregoing object can be any user of the client corresponding to the message box, which is not limited in the present application.

[0086] S202: According to the object feature of the object and the message feature of the candidate message, a predicted entering probability corresponding to the candidate message and a predicted operation result corresponding to the candidate message are determined.

[0087] In order to more accurately evaluate the pushing value, the present application divides two aspects to evaluate the value brought by the message prompt pushing based on the subsequent operation of the object triggered by the message prompt pushing, one aspect is to focus on whether the message prompt pushing will attract the object to implement an entering operation, and the other aspect is to focus on the operation situation of the multiple messages in the box after entering the message box. Therefore, the server can determine the predicted entering probability corresponding to the candidate message and the predicted operation result corresponding to the candidate message according to the object feature and the message feature. The object feature can be used to reflect the situation of the object, and the message feature can reflect the situation of the candidate message.

[0088] The predicted entering probability can be used to represent the probability of the object implementing an entering operation when the message box pushes the candidate message by the message prompt, and the entering operation is used to indicate that the multiple messages containing the candidate message are displayed in the message box, that is, the predicted entering probability can reflect whether the message prompt pushing will attract the object to implement an entering operation to enter the message box. The predicted operation result can be used to indicate the operation implementation feature of the object in the message box based on the entering operation for the multiple messages, that is, it can reflect the operation situation of the multiple messages in the box after entering the message box, so it can reflect the overall consumption in the box.

[0089] It should be noted that the present application does not make any limitation on the specific manner of the entering operation. For example, the entering operation can be a click operation on the message box, etc. For example, the operation implemented on the message can include at least one of a click operation (to view the specific content of the message), a sharing operation, a collection operation, a comment operation, etc. In addition, the operation implementation feature can represent the operation implemented by the object on the plurality of messages in the box, and the present application does not make any limitation on the specific content of the operation implementation feature. In actual application, the operation implemented by the object can include different types of operations, and accordingly, the operation implementation feature can also be different. In actual push scenarios, the more the number of operations on the message by the user, the higher the push revenue that can be brought. Therefore, the aforementioned operation implementation feature can be an operation implementation times feature (for example, it can be the number of clicks of the aforementioned click operation), so as to pay attention to the operation implementation situation that is more relevant to the push revenue, and to facilitate the improvement of the push revenue after actual push.

[0090] It should also be noted that the present application does not make any limitation on the object feature and the message feature. In order to facilitate understanding, the present application embodiment provides the following manner as an example:

[0091] In actual application, the operation situation of the user in the historical period can reflect the interest of the user, etc., so in some possible implementation manners, the object feature can include the historical operation number (such as historical click number) implemented by the object through the message box in the historical period. In actual application, the publishing account corresponding to the message publishes the message in the historical period and the situation of the message being operated can reflect the situation that the new message can be implemented by the user after the publishing account publishes the new message. Therefore, in some possible implementation manners, the message feature can include the historical publishing message number and the historical operated message number of the publishing account corresponding to the candidate message in the historical period, wherein the historical publishing message number can be the total number of messages published in the historical period, and the historical operated message number can be the number of messages operated by the user.

[0092] Since the interest of the user can change over time, in order to guarantee the accuracy of the prediction, the time interval between the maximum time of the historical period and the request time of the message push request is less than the preset interval. Based on this, the relevant data in the historical period adjacent to this message push is selected as the prediction basis, the more adjacent, the closer the interest in the historical period to the interest at the time of this message push, thereby facilitating the improvement of the prediction accuracy and the more accurate message push.

[0093] In actual applications, in order to achieve better prediction, data of multiple historical time periods can be selected as the basis for prediction. For example, the historical operation number of the object implemented through the message box in the past 1 day (such as the number of messages on which the click operation is implemented in the past 1 day), and the historical operation number of the object implemented through the message box in the past 7 days (such as the number of messages on which the click operation is implemented in the past 7 days) can be selected. Similarly, the number of messages and the number of operated messages published by the publishing account in the message box scene in the past 1 day, and the number of messages and the number of operated messages published by the publishing account in the message box scene in the past 7 days can be selected.

[0094] In yet some possible implementation manners, the object feature can further include relevant data for describing basic information of the object, for example, object identification (ID), age value of the object, account belonging area of the object, and the like. Similarly, the message feature can also include relevant data for describing the candidate message and the corresponding publishing account, for example, message ID of the candidate message, account ID of the publishing account, account belonging area of the publishing account, and the like.

[0095] S203: Determine the push value corresponding to the candidate message based on the predicted entering probability and the predicted operation result.

[0096] After the predicted entering probability and the predicted operation result are determined, the server can determine the push value corresponding to the candidate message based on the two. The push value can be used to determine whether the message prompter pushes the candidate message in the message box associated with the object.

[0097] Since in the message push scene with the message box as the entrance, the messages displayed in the box are not only the candidate message, the operation on other messages in the box after being attracted by the message prompter also belongs to the benefits brought by the message prompter. In the present application, the operation of the user on the multiple messages in the box after being attracted into the message box by the message prompter is considered, and therefore the determined push value can more accurately evaluate the value brought by pushing the candidate message by the message prompter. Based on this, because the more accurate push value can be determined, the message prompter can be more accurately pushed for the message box, which is beneficial to improve the push effect, and also beneficial to reduce unnecessary interference of the message prompter on the user and improve the user experience.

[0098] It should be noted that the manner of how to determine the predicted entering probability and the predicted operation result, and the manner of how to determine the push value are not limited in the present application. It can be understood that the manner of determining the predicted entering probability and the predicted operation result is different, and the manner of determining the push value based on the two can also be different. In order to facilitate understanding, the present application embodiment provides the following several ways as examples:

[0099] It can be understood that the candidate message refers to a message to be determined whether to push to the user with a message prompt, if it is determined to push with a message prompt, it can be considered as the information that the user most wants to know after entering the message box, therefore, in a possible implementation, when predicting the implementation operation conditions of the plurality of messages in the box, two levels can also be divided, on the one hand, the implementation operation condition of the candidate message itself is concerned, and on the other hand, the implementation operation condition of the plurality of messages is concerned.

[0100] Correspondingly, in the specific implementation of S202, the server can determine the predicted entering probability corresponding to the candidate message according to the object feature and the message feature, and determine the predicted operation probability corresponding to the candidate message and the predicted operation number corresponding to the candidate message. The predicted operation probability can be used to represent the probability of the object implementing operation on the candidate message based on the entering operation, that is, it can reflect the implementation operation condition of the candidate message itself, which can reflect the degree of attraction to the user after pushing the candidate message with a message prompt. The predicted operation number can be used to represent the operation implementation times feature of the object based on the entering operation on the plurality of messages, that is, it can be embodied in the overall consumption in the box. In this embodiment, the aforementioned predicted operation result can include the predicted operation probability and the predicted operation number. Correspondingly, in the specific implementation of S203, the push value corresponding to the candidate message can be determined based on the predicted entering probability, the predicted operation probability and the predicted operation number.

[0101] Based on this, the three levels of whether to attract the user to enter the message box, whether to attract the user to consume the candidate message itself after entering the message box (the operation on the message can be considered as the process of consuming the message), and whether to attract the user to consume the plurality of messages in the box after entering the message box are concerned, and the push value is evaluated in combination with the conditions of the three levels, which is beneficial to improve the accuracy of the push value, so that it can more accurately reflect the value brought by the attraction of the user to the message prompt.

[0102] In the present application, the predicted operation result is used to indicate the operation implementation feature of the object after entering the message box on the plurality of messages in the box, for example, it can be used to indicate the aforementioned operation implementation times feature. The specific form of the operation implementation times feature is not limited in the present application. In some embodiments, it can be the aforementioned predicted operation number (specifically, the click times is 5 times). In still some embodiments, it can also be the possible distribution of the predicted operation number, so as to represent the predicted result of the operation number.

[0103] In actual application, the distribution of the operation number can be determined by means of the research on the distribution of the value of the variable in statistics. In this embodiment, the predicted operation result can include a predicted distribution center parameter corresponding to the candidate message and a predicted distribution deviation parameter corresponding to the candidate message. The predicted distribution center parameter can be used to represent the value center of the predicted operation number corresponding to the candidate message, and the predicted distribution deviation parameter can be used to represent the dispersion degree of the predicted operation number corresponding to the candidate message distributed around the value center. Based on the two parameters, the distribution of the predicted operation number corresponding to the candidate message can be determined, and thus it can be determined what the most possible predicted operation number is, etc. Accordingly, S203 can determine the push value corresponding to the candidate message according to the predicted entering probability, the predicted distribution center parameter and the predicted distribution deviation parameter.

[0104] Therefore, the predicted number is evaluated by predicting the distribution of the operation number, so as to evaluate the in-box consumption that can be caused. Since the distribution of the operation number can reflect the distribution trend and the distribution range of the operation number, it is beneficial to more accurately evaluate the predicted number based on this, so as to improve the accuracy of the push value.

[0105] In actual application, the value range of the operation number is a non-negative integer, so the normal distribution commonly used in statistics cannot accurately describe the distribution of the operation number. By statistically analyzing the real data in the message box push scenario, it is found that the distribution of the operation number cannot be described by the normal distribution. For example, taking the red dot push in the message box as an example, the number of messages read in the plurality of messages in the box is counted, that is, the operation number can represent the number of messages read by the user, which can be seen from Figure 3 A statistical distribution diagram of the number of messages read is shown. In Figure 3 In the example, the vertical axis can represent the number of red dots, and the horizontal axis can represent the number of messages read. It can be seen that there are a large number of red dots without reading (see the column chart at the left end of the horizontal axis), and the number of red dots without reading accounts for the largest proportion. For the red dots with reading, the reading times are also widely distributed. In actual application, the reading times caused by the accurate push of a small group of red dots (such as push to high active users of the message box) are much higher than the reading times caused by the conventional red dot push.

[0106] In view of the three aspects, in the present application, the distribution of the operation number can be determined by means of the logarithmic normal distribution in statistics. Correspondingly, the aforementioned predicted distribution center parameter can be the predicted logarithmic mean of the predicted operation number of the candidate message, and the predicted distribution deviation parameter can be the predicted logarithmic standard deviation of the predicted operation number of the candidate message. Correspondingly, in determining the push value, first, the operation expectation corresponding to the candidate message can be determined according to the predicted logarithmic mean and the predicted logarithmic standard deviation. The operation expectation can be used to represent the number of operations performed by the object on the plurality of messages in the message box when the candidate message is pushed by the message prompt. Then, the push value corresponding to the candidate message can be determined according to the predicted entry probability and the operation expectation.

[0107] Based on this, the distribution of the operation number is determined by means of the logarithmic normal distribution, and the predicted operation number and the push value are determined based thereon. Since the logarithmic normal distribution can more accurately describe the distribution of the operation number variable, using this way is beneficial to improve the accuracy of the push value.

[0108] In practical applications, based on the idea of logarithmic normal distribution, the aforementioned operation expectation can be determined by the following formula:

[0109]

[0110] In the above formula, E(y) can represent the operation expectation, y can represent the predicted operation number, exp can represent the exponential function with e as the base, μ can represent the predicted logarithmic mean, and σ can represent the predicted logarithmic standard deviation.

[0111] Taking the statistical results of Figure 3 In practical applications, it is found that although the number of times of reading after some red dots attract is different, the push revenue brought by them is similar. Based on this, in some other embodiments, the operation categories can also be divided based on the operation number, and one operation category can cover the case where the operation number falls within an interval. In this way, in prediction, the probability that the candidate message belongs to each operation category can be predicted, so as to determine the possible operation number of the candidate message and the push revenue it may bring.

[0112] Correspondingly, in specific implementation, S202 can determine the predicted entry probability corresponding to the candidate message and the predicted category probability of each operation category corresponding to the candidate message according to the object features and the message features. The operation category is determined according to the preset operation number, and the preset operation number corresponding to different operation categories is different. Correspondingly, in specific implementation, S203 determines the push value corresponding to the candidate message according to the predicted entry probability and the predicted category probability of the candidate message corresponding to each operation category.

[0113] For example, taking the aforementioned click operation as an example, the preset operation number can be set as 0, the corresponding operation category is “zero click”, the preset operation number is [1, 10], the corresponding operation category is “small amount of click”, the preset operation number is [11, 50], the corresponding operation category is “medium amount of click”, the preset operation number is [50, 100], the corresponding operation category is “large amount of click”, and the like. It should be noted that this is only an example and does not make any limitation.

[0114] Based on this, the operation number corresponding to the candidate message is estimated by dividing multiple operation categories based on the preset operation number and predicting the category probability of the candidate message belonging to each category. In this way, the specific value of the operation number does not need to be concerned, and more attention is paid to the push revenue that can be attracted by this operation number level, so that the push value better reflects the push revenue that can be brought after pushing, which is beneficial to improve the push revenue.

[0115] The above embodiments are used to illustrate the way of determining the predicted entering probability and the predicted operation result, and the way of determining the push value. It should be noted that on the basis of the implementation manner provided in the above embodiments, further combination can be made to provide more implementation manners. For example, through further combination, the aforementioned predicted operation result can include the aforementioned predicted operation probability, the predicted distribution center parameter and the predicted distribution deviation parameter. For another example, the aforementioned predicted operation result can include the aforementioned predicted operation probability and the predicted category probability corresponding to each operation category. It can be understood that the combined implementation manners can refer to the detailed description of the above embodiments, and therefore will not be described one by one here.

[0116] After determining the push value of the message, subsequent pushing can be performed according to the push value. The application does not make any limitation on the way of how to perform subsequent pushing according to the push value. In order to facilitate understanding, the application provides the following way as an example:

[0117] In some scenarios, the number of the aforementioned candidate messages can be one, and after the push value is determined, it can be determined whether to push the candidate message as a message prompt based on the push value. In specific implementation, if the push value of the candidate message meets the push condition, it indicates that the push can obtain better push revenue, such as attracting the user to enter the message box and consuming multiple messages in the box, and the like. Therefore, at this time, the candidate message can be pushed as a message prompt in the message box associated with the object. For example, the dashed box shown in the example can be referred to. Figure 1

[0118] ​Conversely, if the push value of this candidate message does not meet the push conditions, it indicates that pushing it with a message notification is unlikely to yield good push benefits and may even cause user resentment. Therefore, in this case, the candidate message can be avoided by not pushing it with a message notification. For example, with Figure 1 The example shown by the dashed box illustrates that the message prompt can be omitted, meaning the candidate message can be pushed without a notification. Alternatively, the candidate message can be omitted entirely. The choice can be made flexibly based on the actual situation; this application does not impose any limitations.

[0119] The push conditions can refer to the restrictions on pushing messages using message prompts. These conditions can be set according to actual circumstances, and this application does not impose any limitations. For example, a push value threshold can be set as the push condition. Correspondingly, if the push value of a candidate message is greater than or equal to the push value threshold, the push condition is considered met, and the message can be pushed using a message prompt. Conversely, if the push condition is not met, the message will not be pushed using a message prompt.

[0120] In other scenarios, there can be multiple candidate messages. Correspondingly, based on push value, some messages can be selected from these candidate messages and pushed with message prompts. In specific implementation, the server can determine the candidate messages whose push value meets the push conditions as target messages based on the push value of each candidate message. That is, the target message is selected based on push value and is the message that will yield better push benefits after being pushed with a message prompt. Accordingly, the target message can be pushed with a message prompt in the message box associated with the object. Because the push value determined in this application can more accurately reflect the real value that may be brought after pushing with a message prompt, messages with higher push benefits can be selected based on push value, achieving more accurate push.

[0121] It should be noted that this application does not impose any limit on the number of target messages. In practical applications, the number of target messages can be determined based on the message box's out-of-box push strategy. The out-of-box push strategy specifies how messages are pushed to the user through the message box before entering the message box. For example, to attract user attention, the out-of-box push strategy might specify pushing a message to the user through the message box before entering the message box. Therefore, the number of target messages can be one. For example, see [link to relevant documentation]. Figure 1 The example message box shows that before entering the message box, the target message "essential chat emoticons" was pushed to user A.

[0122] It should also be noted that this application does not impose any limitations on how the target message is pushed via message prompt. For ease of understanding, this application provides the following example:

[0123] In order to fully attract the attention of the user, so as to attract the user into the message box, in a possible implementation, a message prompt and a message title of the target message can be displayed in the message box associated with the object, wherein the message title can refer to the summary of the target message, and can convey the specific content of the target message. Based on this, by displaying the message prompt, the purpose of reminding the user is achieved, and by displaying the message title, the user can directly understand the general situation of the target message, and if interested, subsequent operations can be performed to view, etc.

[0124] Compared with the related art of displaying a red dot on the icon of the client to remind the user of message updates, the present application displays the message title at the same time of reminding the user with the red dot, directly conveying the situation of the message expected to be viewed by the user to the user, improving the user experience, and being conducive to better attracting the user to perform subsequent operations, and being conducive to improving the push revenue, etc.

[0125] For the push scenario with the message box as the entrance, in some possible implementations, after the message box pushes the target message with the message prompt, the multiple messages in the box can also be concerned, so as to attract the user to consume in the box after entering the message box, improve the user experience and the push revenue, etc. In actual application, different message boxes can belong to different clients, or in the same client, multiple different message boxes can also exist. Due to the differences in the product forms of different clients, etc., the logic of pushing messages to the user through the message box can also be different, and different message boxes of the same client can also want to push different types of messages to the user. Therefore, in a possible implementation, the multiple messages in the box can be determined in combination with the in-box push strategy of the message box to improve the in-box push accuracy, etc.

[0126] In specific implementation, the server can first obtain the in-box push strategy corresponding to the message box, which can indicate how to push messages to the user and obtain the push revenue, etc. by displaying multiple messages in the message box after entering the message box. For example, the types of the multiple messages displayed in the box (such as text and video types) can be indicated. Then, the multiple in-box messages corresponding to the target message and the display order of the multiple in-box messages in the message box can be determined according to the in-box push strategy and the push values of the multiple candidate messages other than the target message. Correspondingly, the target message and the multiple in-box messages can be displayed in the message box in the display order after the target message in response to the entering operation implemented by the object when pushing the target message with the message prompt. Based on this, the push value and the in-box push strategy are comprehensively considered, so that the multiple in-box messages can conform to the in-box push strategy, and at the same time, better push revenue can be brought.

[0127] Based on this, in one aspect, a more accurate push value determination manner is provided, which mainly determines the push value from the three aspects of whether to enter the message box, whether to operate on the candidate message itself, and the operation implemented on the multiple messages in the box. Compared with the related art, the push value is determined based on the operation and product rules, and the present application is more general and can be applied to various message push scenarios with message box as the entrance and information flow displayed in the box. On the other hand, the push value can also be combined with the in-box push strategy of the message box, so that the push value can be more in line with the specific situation of each message box, and be adapted to the specific product form, which is beneficial to realize the personalized push of the message box dimension and improve the push accuracy.

[0128] It can be understood that the target message is a message pushed in the message box as a message prompt, so it can be considered as the information that the user most wants to know under the operation of entering the message box, and therefore it can be displayed in the box first for the user to view, etc. Correspondingly, the multiple in-box messages can be displayed after the target message in the display order.

[0129] For example, Figure 1 For example, the message box shown in the dashed box can be seen from Figure 4 , Figure 4 A message push schematic diagram of a message box is shown. If the user A implements a click operation, the user can enter the message box, and five messages are displayed in the message box, which are "chat essential emoticons", message 1, message 2, message 3, and message 4 pushed as message prompts. In actual application, the user can further implement an operation (such as a click operation, a sharing operation, a collection operation, etc.) on a message to view the specific content of the message, or share and collect the message. Of course, in the message box, the user can also implement up and down sliding operations to obtain more in-box messages.

[0130] It can be seen from the above technical solution that, for the scenario of taking a message box as an entrance and displaying multiple messages in the message box, according to the subsequent operation triggered by the object attracted by the message prompt push, the value brought by the message prompt push is evaluated in two aspects. One aspect is to focus on whether the message prompt push will attract the object to perform an entering operation, and the other aspect is to focus on the operation performed on the multiple messages in the box after entering the message box. In specific implementation, first, the predicted entering probability and the predicted operation result are determined according to the object characteristics and the message characteristics. The predicted entering probability can be used to represent the probability of the object performing an entering operation when the message box pushes a candidate message with a message prompt. The entering operation can be used to indicate that the multiple messages containing the candidate message are displayed in the message box, that is, it can reflect whether the message prompt push will attract the object to perform an entering operation. The predicted operation result can be used to indicate the operation implementation characteristics of the object on the multiple messages in the message box based on the entering operation, that is, it can reflect the operation performed on the multiple messages in the box after entering the message box. Since the messages displayed in the box are not only the candidate message, the operation on other messages in the box after being attracted by the message prompt also belongs to the benefits brought by the message prompt push. Therefore, the push value determined based on the predicted entering probability and the predicted operation result can more accurately evaluate the value that the message prompt push of the candidate message can bring. Accordingly, the push value can be used to determine whether to push the candidate message with a message prompt in the message box associated with the object. Based on this, because a more accurate push value can be determined, the message prompt push for the message box can be more accurately performed, which is beneficial to improve the push effect and also beneficial to reduce unnecessary message prompt push interference on the user and improve the user experience.

[0131] Through the above embodiments, the message push method provided by the present application is described in detail from the aspects of determining the push value and performing subsequent push based on the push value. In addition to the above examples, in some scenarios, in order to improve the push efficiency, etc., a target model suitable for the scheme of the present application can be obtained by model training through artificial intelligence and machine learning, and then the target model is used to implement S202, etc. as described above, which is beneficial to improve the efficiency of determining the push value and thus improve the push efficiency, etc. In order to better understand, the present application provides the following examples for this embodiment to further illustrate the present application.

[0132] That is, in still other scenarios, S202 in the foregoing can determine the predicted entering probability and the predicted operation result through a target model according to the object characteristics and the message characteristics in specific implementation. The target model can be determined in the following way:

[0133] Firstly, a training sample can be acquired, the training sample can include object features of a sample object and message features of a sample message, the training sample can include an entering label and an in-box operation label. The entering label can be used to indicate whether the sample object implements an entering operation when the message box pushes the sample message with a message prompt, the entering operation can be used to indicate that the sample object displays a plurality of messages containing the sample message in the message box, and the in-box operation label can be used to indicate operation implementation characteristics of the sample object in the message box based on the entering operation. In actual application, a training data set X can be constructed according to messages that are pushed, distributed to users, and exposed through a message prompt, so that the training sample can be directly constructed based on the training data set X, and the training sample can be denoted as x, x∈X.

[0134] Then, the sample prediction entering probability corresponding to the sample message and the sample prediction operation result corresponding to the sample message can be determined by the initial model according to the training sample. The sample prediction entering probability can be used to indicate the probability of the sample object implementing the entering operation when the sample message is pushed with the message prompt, and the sample prediction operation result can be used to indicate the operation implementation characteristics of the sample object in the message box based on the entering operation.

[0135] Finally, the entering loss can be determined according to the sample prediction entering probability and the entering label, and the operation loss can be determined according to the sample prediction operation result and the in-box operation label. The initial model is trained according to the entering loss and the operation loss to obtain a target model.

[0136] The entering loss can indicate the difference between the sample prediction entering probability and the entering label, and the entering label can indicate the real entering probability. Similarly, the operation loss can indicate the difference between the sample prediction operation result and the in-box operation label, and the in-box operation label can indicate the real operation situation of the plurality of messages in the box. Therefore, the initial model can be trained according to the entering loss and the operation loss, so that the initial model can learn the correlation between the object features, the message features, whether to enter the message box, and the overall consumption in the box. Finally, the target model can be obtained. Therefore, when the target model is used to implement S202 in the subsequent process, the prediction entering probability and the prediction operation result can be quickly determined based on the object features of the object and the message features of the candidate message, which is beneficial to improve the pushing efficiency.

[0137] It should be noted that the setting of the entering label and the in-box operation label, and the way of determining the entering loss and the operation loss are not limited in the present application. In order to facilitate understanding, the following method is provided as an example by the embodiments of the present application:

[0138] The sample message in the training sample can be a message actually pushed to the sample object with a message prompt in an actual application. Therefore, whether the sample object implements the entering operation and the operation performed by the sample object in the box can be real data. Accordingly, the entering label can be determined based on whether the sample object has actually implemented the entering operation, for example, if the entering operation is implemented, that is, the message box is entered, the entering label can be an entering positive sample, and vice versa, the entering label can be an entering negative sample.

[0139] Corresponding to determining the entering loss, a binary classification cross-entropy loss function can be used for optimization. Taking the entering label as an entering positive sample or an entering negative sample as an example, the entering loss can be determined by the following formula:

[0140] LossCE(entering, x) = -f(x1) * log(p1) - f(x2) * log(1-p1)

[0141] In the above formula, x can represent a training sample, LossCE(entering, x) can represent the entering loss corresponding to the training sample x, p1 can represent the sample predicted entering probability corresponding to the sample message in the training sample x, and usually, p1∈[0,1]. f(x) is an indicator function, x1 indicates that the entering label indicates that the sample message is an entering positive sample, and x2 indicates that the entering label indicates that the sample message is an entering negative sample.

[0142] Specifically, if the entering label of the training sample x indicates an entering positive sample, f(x1) = 1 and f(x2) = 0. If the entering label of the training sample x indicates an entering negative sample, f(x1) = 0 and f(x2) = 1.

[0143] Based on this, by minimizing LossCE(entering, x), the initial model can learn the correlation between the object features, the message features, and whether to enter the message box.

[0144] Similarly, the in-box operation label can also be determined according to the real operation of the sample object in the box after implementing the entering operation. In actual applications, the in-box clicking operation label can be flexibly set according to the training target to be achieved. Different training targets result in different sample predicted operation results output by the model, and the predicted operation results output in the application stage are different. For example, taking the aforementioned output predicted operation number as an example, the in-box operation label can be the number of operations performed on multiple messages in the box. In actual applications, different in-box operation labels can have differences in the way of determining the operation loss, which is not limited in the present application.

[0145] Corresponding to the aforementioned embodiments of the predicted operation result including the predicted operation probability and the predicted operation number, the predicted operation result including the predicted distribution center parameter and the predicted distribution deviation parameter, and the predicted operation result including the predicted category probability corresponding to each operation category, in order to quickly implement these implementations by using the target model, in some possible implementations, the corresponding training samples can also be constructed to enable the model to learn the corresponding prediction ability, which can be applied to the corresponding embodiments. The following will take each of the aforementioned embodiments as an example to illustrate the model training related content in detail, so as to better understand the present application.

[0146] Corresponding to the aforementioned embodiments of the predicted operation result including the predicted operation probability and the predicted operation number, in order to obtain the target model capable of implementing the scheme in this embodiment, in a possible implementation, the aforementioned in-box operation label can include a first operation label and a second operation label, wherein the first operation label can be used to indicate whether the sample object implements the operation on the sample message, and the second operation label can be a sample operation number, which can be determined according to the number of times the sample object implements the operation on the plurality of messages in the message box. Based on this, from the two dimensions of whether to implement the operation on the sample message itself and the operation number implemented on the plurality of messages in the box, the operation of the sample object after being attracted in the training sample is reflected. Then, the sample predicted entry probability can be determined by the initial model according to the training sample, and the sample predicted operation probability corresponding to the sample message and the sample predicted operation number corresponding to the sample message can be determined by the initial model. Wherein, the sample predicted operation probability can be used to indicate the probability of the sample object implementing the operation on the sample message based on the entry operation. Correspondingly, in the determination of the operation loss, the first operation loss can be determined according to the sample predicted operation probability and the first operation label, and the second operation loss can be determined according to the sample predicted operation number and the sample operation number. That is, the aforementioned operation loss can include the first operation loss and the second operation loss.

[0147] Based on this, in the process of model training of the initial model by using the entry loss, the first operation loss and the second operation loss, the initial model can learn the correlation between the object features, the message features and whether to enter the message box, whether to operate on the sample message itself, and the consumption of the whole box. In this way, the target model can be used to quickly predict the predicted entry probability, the predicted operation probability and the predicted operation number corresponding to the candidate message in the subsequent, which is beneficial to improve the pushing efficiency. And more dimensional output is beneficial to determine more accurate pushing value.

[0148] It should be noted that the application does not make any limitation on the setting of the first operation label and the second operation label, and the way of determining the first operation loss and the second operation loss. For the convenience of understanding, the following way is provided as an example by the embodiments of the application:

[0149] Similar to the aforementioned entering label, the first operation label can be determined based on whether the sample object actually implements the operation on the sample message. For example, taking the click operation as an example, after being attracted by the message prompt and entering the message box, if the sample message is clicked, the first operation label can be a click positive sample, otherwise, the entering label can be a click negative sample. Corresponding to the determination of the first operation loss, a binary cross-entropy loss function can be used for optimization. Taking the first operation label as a click positive sample or a click negative sample as an example, the first operation loss can be determined by the following formula:

[0150] LossCE (click, x) = -f (x3) *log (p2) -f (x4) *log (1-p2)

[0151] In the above formula, x can represent a training sample, LossCE (click, x) can represent the first operation loss corresponding to the training sample x, p2 can represent the sample prediction operation probability corresponding to the sample message in the training sample x, usually, p2∈[0,1]. f(x) is an indicator function, x3 represents that the first operation label indicates that the sample message is a click positive sample, and x4 represents that the first operation label indicates that the sample message is a click negative sample.

[0152] Specifically, if the first operation label of the training sample x indicates a click positive sample, f(x3) = 1, f(x4) = 0. If the first operation label of the training sample x indicates a click negative sample, f(x3) = 0, f(x4) = 1.

[0153] Based on this, by minimizing LossCE (click, x), the initial model can learn the correlation between the object features, the message features and whether the operation is implemented on the sample message after entering the message box.

[0154] In actual application, the more messages the user views, the higher the push revenue that can be brought. The push revenue corresponding to the user repeatedly viewing the same message several times and viewing in turn can be similar. Therefore, in order to improve the push revenue, it is hoped that after pushing, the user can be attracted to operate on more messages. In order to achieve this goal, when setting the second operation label, the number of messages on which the sample object implements the operation can be counted according to the actual operation times of the sample object on multiple messages by means of deduplication, etc., and recorded as the aforementioned sample operation number, that is, the second operation label can be the number of messages on which the sample object implements the operation.

[0155] Based on this, in the process of model training using the second operation loss, the initial model can learn the correlation between the object features, the message features and the number of messages that will implement the operation into the message box, which is beneficial to subsequent use of the target output to more accurately reflect the predicted operation number of the push revenue.

[0156] Corresponding to the foregoing embodiment in which the predicted operation result includes the predicted distribution center parameter and the predicted distribution deviation parameter, in order to obtain a target model that can implement the scheme in this embodiment, in one possible implementation, a corresponding learning goal can also be set. Since the predicted distribution center parameter and the predicted distribution deviation parameter reflect the predicted distribution of the operation number, the aforementioned in-box operation label can be the aforementioned sample operation number. In order to be able to quickly determine the output that matches the expected parameter meaning using the model, in the training stage, the sample predicted operation result output by the initial model includes the sample predicted distribution center parameter corresponding to the sample message and the sample predicted distribution deviation parameter corresponding to the sample message. Among them, the sample predicted distribution center parameter can be used to represent the value center of the sample predicted operation number corresponding to the sample message, and the sample predicted distribution deviation parameter can be used to represent the dispersion degree of the sample predicted operation number corresponding to the sample message distributed around the value center. Correspondingly, when determining the operation loss, the operation loss can be determined according to the sample predicted distribution center parameter, the sample predicted distribution deviation parameter and the sample operation number.

[0157] Since the in-box operation label is the real sample operation number, based on this, after determining the operation loss, in the process of model training using the operation loss, the initial model can learn the correlation between the two parameters output and the real sample operation number. In this way, the two parameters output by the target model finally obtained can reflect the predicted operation number for the candidate message.

[0158] In order to better understand, taking the foregoing predicted distribution center parameter as the predicted logarithmic mean and the predicted distribution deviation parameter as the predicted logarithmic standard deviation as an example, that is, considering the characteristic that the distribution of the operation number is more suitable for being described by the logarithmic normal distribution. Corresponding to the training stage, the foregoing sample predicted distribution center parameter can be the sample predicted logarithmic mean of the sample predicted operation number for the sample message, and the sample predicted distribution deviation parameter can be the sample predicted logarithmic standard deviation of the sample predicted operation number for the sample message.

[0159] In order to enable the initial model to learn the meaning of the logarithmic mean and the logarithmic standard deviation in the model training process, and finally determine the output that meets the expectation, in this embodiment, taking the sample predicted distribution center parameter as the sample predicted logarithmic mean and the sample predicted distribution deviation parameter as the sample predicted logarithmic standard deviation as an example, the way of determining the operation loss is illustrated as follows:

[0160] Based on the idea of the lognormal distribution, the operation loss can be determined using a lognormal loss optimization. Specifically, the lognormal loss function can be determined by taking the negative log-likelihood function from the lognormal distribution. Taking the training sample x as an example, the operation loss corresponding to the training sample x can be determined by the following formula:

[0161]

[0162] In the above formula, x can represent a training sample, LossCE(operation, x) can represent the operation loss corresponding to the training sample x, y0 can represent the in-box operation label corresponding to the training sample x, i.e., the sample operation number corresponding to the sample message in the training sample x, μ0 can represent the sample predicted log mean, and σ0 can represent the sample predicted log standard deviation.

[0163] As can be seen from the lognormal loss function, this part mainly considers the part where the sample operation number is greater than 0. Based on this, during model training, the initial model can learn the relationship between the object features, message features, and sample operation number, and output the sample predicted log mean and sample predicted log standard deviation that can accurately represent the sample operation number. In this way, in subsequent applications, the predicted log mean and predicted log standard deviation that can accurately represent the predicted operation number can be output by the target model.

[0164] In order to improve the stability of the initial model and avoid the interference of abnormal outliers, in one possible implementation, a model hyperparameter can also be set to limit the value range of the values output by the output layer of the model, so as to avoid outputting abnormally large values for abnormal outliers and affecting the stability of the model. In specific implementation, a hyperparameter K can be set, for example, K = 100 can be set based on the distribution experience of the sample operation number. Correspondingly, μ0∈{i|i∈[0,K] and i∈Z}(Z is the set of integers), and σ0∈[0,K]. Based on this, abnormal outliers can be truncated to avoid interference, which is conducive to improving the stability of the model.

[0165] Corresponding to the embodiment in which the predicted operation result includes the predicted category probability corresponding to each operation category, in order to obtain a target model that can implement the scheme in this embodiment, in one possible implementation, a corresponding learning goal can also be set. In specific implementation, a plurality of operation categories can be set based on preset operation numbers, and different operation categories correspond to different preset operation numbers, such as the aforementioned "zero clicks", "few clicks", "medium clicks", and "large clicks". Correspondingly, the in-box operation label can be used to represent the sample operation category corresponding to the sample message, which is determined according to the sample operation number corresponding to the sample message. For example, if the sample operation number is 8 and falls within the preset operation number range covered by "few clicks", then the sample operation category can be "few clicks".

[0166] In the model training stage, the sample predicted entering probability can be determined by the initial model according to the training sample, and the sample predicted category probability of the sample message corresponding to each operation category in the plurality of operation categories can be determined by the initial model according to the sample message. In addition, the operation loss can be determined according to the sample predicted category probability of the sample message corresponding to each operation category in the plurality of operation categories and the in-box operation label.

[0167] In order to facilitate the determination of the operation loss, in actual application, the sample operation category and the plurality of operation categories can be encoded to determine the value of the in-box operation label. For example, if the sample operation category is "a small number of clicks", it can be encoded to obtain [0, 1, 0, 0] and used as the in-box operation label. For another example, if the sample operation category is "a large number of clicks", it can be encoded to obtain [0, 0, 0, 1] and used as the in-box operation label. Correspondingly, the output sample predicted category probability corresponding to each operation category can be in the form of [0.5, 0.1, 0.7, 0.2], for example.

[0168] Based on this prediction belonging to a plurality of operation categories, the case that the sample operation number of some training samples is zero can be compatible, so that all training samples can be used for better model training, and the training efficiency and training effect are improved.

[0169] The above embodiments are used to illustrate the way of determining the target model by the model training of the present application. It should be noted that the content of the training sample and the model structure of the initial model are not limited in the present application. In order to facilitate understanding, the following methods are provided as examples in the embodiments of the present application:

[0170] For the training sample, in order to improve the model training effect, in some possible implementation manners, in addition to the object features of the sample object and the message features of the sample message, the training sample can also include a time feature. The time feature can be used to represent the time interval between the operation time of the sample object implementing the entering operation and the push time of the sample message in the message box with the message prompt. Generally, the shorter the time interval, the more likely the sample object will implement the entering operation soon after seeing the message prompt, and then perform the subsequent operation. Therefore, the time feature can reflect the attraction of the message prompt to the sample object to a certain extent, thereby facilitating the model to understand the association between the object features, the message features and whether to implement the entering operation. Based on this, the model can converge faster and improve the model training efficiency. For example, the time feature can be a discrete representation in the form of the day of the week within one week, the hour of the day within 24 hours, the minute within one hour, etc.

[0171] For the model structure, on the one hand, the initial model can include an intermediate layer and an output layer, wherein the intermediate layer can be used to map the input features (such as the aforementioned object features, message features, and time features) to high-dimensional features and pass them to the output layer, and finally output the corresponding results (such as the aforementioned sample prediction entering probability, sample prediction operation probability, etc.) by the output layer. For example, the structure of the intermediate layer can use a commonly used push ranking model framework, a fully connected neural network, etc. On the other hand, in this application, the model output includes multiple different results, in order to accurately output the expected results, in terms of model structure, the same number of output nodes can be set in the output layer according to the number of output structures, one output node is used to output one result, and different output nodes are used to output different results. For example, one output node is used to output the sample prediction entering probability, and one output node is used to output the sample prediction operation result.

[0172] For better understanding, for the process of model training, the embodiments of the present application also provide an initial model structure diagram as shown in Figure 5 Figure 5 In the example, the intermediate layer of the initial model is the aforementioned fully connected neural network, and the output layer of the output model includes four output nodes. The input of the initial model can include the aforementioned object features of the sample object, message features of the sample message, and time features, and the output of the initial model includes the aforementioned sample prediction entering probability, sample prediction operation probability, sample prediction log mean, and sample prediction log standard deviation, which are respectively output by the four output nodes in the output layer.

[0173] For example, for the aforementioned training sample x, the corresponding training loss can include the aforementioned LossCE(enter, x), LossCE(click, x), and LossCE(operation number, x) three parts. In actual training, a hybrid loss function can be constructed based on the three parts of the loss as the optimization objective function of the initial model, which is beneficial to make the corresponding multi-task target easier to be fitted and learned. After determining the hybrid loss function, gradient descent and back propagation algorithm can be used for optimization, wherein the gradient descent and back propagation algorithm is a common basic training algorithm of deep neural network, which means that according to the difference between the predicted results of the model and the true value, the direction and size of the change of the model parameters should be calculated and the model parameters are updated. In this way, the model training is completed through multiple iterations. After completing the model training, the target model including four output nodes in the output layer can be obtained, and in actual application, the target model can output the aforementioned prediction entering probability, prediction operation probability, prediction log mean, and prediction log standard deviation through the four output nodes. In order to better understand how to determine the push value based on the four parts of the output, the embodiments of the present application also provide the following examples:​

[0174] In practical applications, the calculation method based on the conditional probability relationship can analyze the result that the user expected in the present application can perform relevant operations (such as click operations) on the plurality of messages in the message box after entering the message box, and the specific analysis is as follows:

[0175] P (entering the message box and clicking M messages) = P (clicking M messages | entering the message box) * P (entering the message box) = P (clicking M messages | having clicks, entering the message box) * P (having clicks | entering the message box) * P (entering the message box)

[0176] Based on the probability relationship, it can be seen that the method for determining the push value based on the subsequent operation triggered by the user after being attracted by the message prompt in the present application is divided into three levels (whether to enter, whether to perform an operation on the candidate message itself, and the operation performed on the plurality of messages in the box), which can determine a more accurate push value. Accordingly, based on the probability relationship, the present application also provides an example of determining the push value of a candidate message based on the predicted entering probability, the predicted operation probability, the predicted logarithmic mean, and the predicted logarithmic standard deviation. Specifically, it can be determined by the following formula:

[0177]

[0178] In the above formula, predTimes can represent the push value corresponding to the candidate message, p 01 may represent the predicted entering probability corresponding to the candidate message, p 02 may represent the predicted operation probability corresponding to the candidate message, exp can represent the exponential function with e as the base, μ can represent the predicted logarithmic mean corresponding to the candidate message, and σ can represent the predicted logarithmic standard deviation corresponding to the candidate message.

[0179] In practical applications, the plurality of candidate messages can be sorted in ascending order based on the push value corresponding to each candidate message, and the candidate message with the largest push value can be selected as the target message, and the message prompt can be pushed in the message box.

[0180] Taking the message box as a subscription message scenario in a certain client as an example, in this scenario, the message box can include pushed message cards (such as some accounts subscribed by the user, etc.). After applying the message push method provided by the present application to this scenario, the corresponding entering point rate of the target message such as the text message and the video message is improved after being pushed, wherein the entering point rate can refer to entering the message box and clicking the target message itself pushed. In addition, the proportion of users consuming the message box is also improved.

[0181] It should be noted that the application can be further combined on the basis of the implementation manners of the above aspects to provide more implementation manners.

[0182] Based on Figure 2 According to the message pushing method provided in the corresponding embodiment, the application further provides a message pushing device 600, which comprises an acquisition unit 601 and a determination unit 602.

[0183] The acquisition unit 601 is configured to acquire an object and a candidate message indicated by a message pushing request in response to the message pushing request for a message box.

[0184] The determination unit 602 is configured to determine a predicted entering probability corresponding to the candidate message and a predicted operation result corresponding to the candidate message according to an object feature of the object and a message feature of the candidate message, the predicted entering probability being used to represent a probability of implementing an entering operation by the object when the message box pushes the candidate message in a message prompt, the entering operation being used to indicate displaying a plurality of messages containing the candidate message in the message box, and the predicted operation result being used to indicate an operation implementation feature of the object in the message box for the plurality of messages based on the entering operation.

[0185] The determination unit 602 is further configured to determine a pushing value corresponding to the candidate message based on the predicted entering probability and the predicted operation result, the pushing value being used to determine whether the message box associated with the object pushes the candidate message in a message prompt.

[0186] In a possible implementation manner, the determination unit is further configured to:

[0187] determine a predicted entering probability corresponding to the candidate message according to the object feature and the message feature, and determine a predicted operation probability corresponding to the candidate message and a predicted operation number corresponding to the candidate message, the predicted operation probability being used to represent a probability of implementing an operation by the object for the candidate message based on the entering operation, and the predicted operation number being used to represent an operation implementation frequency feature of the object for the plurality of messages based on the entering operation;

[0188] determine a pushing value corresponding to the candidate message based on the predicted entering probability, the predicted operation probability, and the predicted operation number.

[0189] In a possible implementation, if the predicted operation result includes a predicted distribution center parameter corresponding to the candidate message and a predicted distribution deviation parameter corresponding to the candidate message, the predicted distribution center parameter is used to represent a value center of a predicted operation number corresponding to the candidate message, and the predicted distribution deviation parameter is used to represent a discrete degree of the predicted operation number corresponding to the candidate message distributed at the value center, and the determining unit is further configured to:

[0190] determine a push value corresponding to the candidate message according to the predicted entering probability, the predicted distribution center parameter, and the predicted distribution deviation parameter.

[0191] In a possible implementation, if the predicted distribution center parameter is a predicted logarithmic mean value of the predicted operation number corresponding to the candidate message, and the predicted distribution deviation parameter is a predicted logarithmic standard deviation of the predicted operation number corresponding to the candidate message, the determining unit is further configured to:

[0192] determine an operation expectation corresponding to the candidate message according to the predicted logarithmic mean value and the predicted logarithmic standard deviation, the operation expectation being used to represent a number of operation implementations of the object in the message box for the plurality of messages when the candidate message is pushed with a message prompt.

[0193] determine the push value corresponding to the candidate message according to the predicted entering probability and the operation expectation.

[0194] In a possible implementation, the determining unit is further configured to:

[0195] determine a predicted entering probability corresponding to the candidate message according to the object feature and the message feature, and determine a predicted category probability of each operation category corresponding to the candidate message, the operation category being determined according to a preset operation number, and different operation categories corresponding to different preset operation numbers.

[0196] determine the push value corresponding to the candidate message according to the predicted entering probability and the predicted category probability of the candidate message corresponding to the plurality of operation categories.

[0197] In a possible implementation, if the number of the candidate messages is a plurality, the determining unit is further configured to:

[0198] determine a target message from the candidate messages according to the push values corresponding to the plurality of candidate messages, the push value of the target message satisfying a push condition;

[0199] display a message prompt and a message title of the target message in the message box associated with the object.

[0200] In a possible implementation, the determining unit is further configured to:

[0201] obtain a box-in pushing policy corresponding to the message box;

[0202] determine, according to the box-in pushing policy and a pushing value corresponding to each of a plurality of candidate messages other than the target message, a plurality of box-in messages corresponding to the target message and a display order of the plurality of box-in messages in the message box;

[0203] in response to the entering operation performed by the object when the target message is pushed by the message prompt in the message box, display the target message in the message box, and display the plurality of box-in messages after the target message in the display order.

[0204] In a possible implementation, the object feature includes a historical operation number of the object in a historical period through the message box, and the message feature includes a historical number of published messages and a historical number of operated messages of a publishing account corresponding to the candidate message in the historical period, and a time interval between a maximum time point of the historical period and a request time point of the message pushing request is less than a preset interval.

[0205] In a possible implementation, the determining unit is further configured to:

[0206] determine, according to the object feature and the message feature, the predicted entering probability and the predicted operation result by a target model;

[0207] obtain a training sample, the training sample including an object feature of a sample object and a message feature of a sample message, the training sample including an entering label and a box-in operation label, the entering label being used to indicate whether the sample object performs the entering operation when the message box pushes the sample message by the message prompt, the entering operation being used to indicate that a plurality of messages including the sample message are displayed in the message box, and the box-in operation label being used to indicate an operation performance feature of the sample object in the message box for the plurality of messages based on the entering operation;

[0208] determine, according to the training sample, a sample predicted entering probability corresponding to the sample message and a sample predicted operation result corresponding to the sample message by an initial model, the sample predicted entering probability being used to indicate a probability that the sample object performs the entering operation when the message box pushes the sample message by the message prompt, and the sample predicted operation result being used to indicate the operation performance feature of the sample object in the message box for the plurality of messages based on the entering operation;

[0209] determine an entering loss according to the sample predicted entering probability and the entering label, and determine an operation loss according to the sample predicted operation result and the in-box operation label;

[0210] train the initial model according to the entering loss and the operation loss to obtain the target model.

[0211] In a possible implementation, if the in-box operation label includes a first operation label and a second operation label, the first operation label is used to indicate whether the sample object implements operation on the sample message, and the second operation label is a sample operation number, which is determined according to a number of times that the sample object implements operation on the plurality of messages in the message box, and the determining unit is further configured to:

[0212] determine, according to the training sample, the sample predicted entering probability by using the initial model, and determine, by using the initial model, a sample predicted operation probability corresponding to the sample message and a sample predicted operation number corresponding to the sample message, the sample predicted operation probability being used to indicate a probability that the sample object implements operation on the sample message based on the entering operation;

[0213] determine a first operation loss according to the sample predicted operation probability and the first operation label, and determine a second operation loss according to the sample predicted operation number and the sample operation number, the operation loss including the first operation loss and the second operation loss.

[0214] In a possible implementation, if the in-box operation label is a sample operation number, the sample predicted operation result includes a sample predicted distribution center parameter corresponding to the sample message and a sample predicted distribution deviation parameter corresponding to the sample message, the sample predicted distribution center parameter being used to indicate a value center of the sample predicted operation number corresponding to the sample message, and the sample predicted distribution deviation parameter being used to indicate a discrete degree of distribution of the sample predicted operation number corresponding to the sample message at the value center, and the determining unit is further configured to:

[0215] determine the operation loss according to the sample predicted distribution center parameter, the sample predicted distribution deviation parameter and the sample operation number.

[0216] In a possible implementation, if the in-box operation label is used to indicate a sample operation category corresponding to the sample message, the sample operation category is determined according to a sample operation number, and the sample operation number is determined according to a number of times that the sample object implements operation on the plurality of messages in the message box, and the determining unit is further configured to:

[0217] According to the training sample, the initial model is used to determine the sample predicted entering probability, and the initial model is used to determine a sample predicted category probability of the sample message in each operation category of a plurality of operation categories, and the preset operation number corresponding to different operation categories is different.

[0218] According to the sample predicted category probability of the sample message corresponding to the plurality of operation categories and the in-box operation label, the operation loss is determined.

[0219] In a possible implementation, the training sample further includes a time feature, and the time feature is used to represent a time interval between an operation time of the sample object performing the entering operation and a push time of the message box pushing the sample message as a message prompt.

[0220] As can be seen from the above technical solutions, for the scenario of displaying a plurality of messages in a message box as an entrance, according to the subsequent operation triggered by the message prompt push of the object, the value brought by the message prompt push is evaluated in two aspects. One aspect is to focus on whether the message prompt push will attract the object to perform the entering operation, and the other aspect is to focus on the operation situation of the plurality of messages in the box after the entering operation. In specific implementation, first, the predicted entering probability and the predicted operation result are determined according to the object feature and the message feature. The predicted entering probability can be used to represent the probability of the object performing the entering operation when the message box pushes the candidate message as the message prompt. The entering operation can be used to indicate that the plurality of messages containing the candidate message are displayed in the message box, that is, whether the message prompt push will attract the object to perform the entering operation. The predicted operation result can be used to indicate the operation implementation feature of the plurality of messages in the message box based on the entering operation, that is, the operation situation of the plurality of messages in the box after the entering operation. Since the plurality of messages displayed in the box are not only the candidate message, the operation of the other messages in the box after being attracted by the message prompt also belongs to the benefits brought by the message prompt push. Therefore, the push value determined based on the predicted entering probability and the predicted operation result can more accurately evaluate the value brought by the message prompt push of the candidate message. Accordingly, the push value can be used to determine whether the message box associated with the object pushes the candidate message as the message prompt. Based on this, because the more accurate push value can be determined, the message prompt push of the message box can be more accurately performed, which is beneficial to improve the push effect and also beneficial to reduce unnecessary message prompt push to interfere with the user and improve the user experience.

[0221] Embodiments of the present application also provide a computer device, which can be a terminal. Taking a smart phone as an example, the computer device comprises:

[0222] Figure 7A block diagram of a part of a structure of a smart phone is shown. Referring to Figure 7 The smart phone includes a radio frequency (RF) circuit 710, a memory 720, an input unit 730, a display unit 740, a sensor 750, an audio circuit 760, a wireless fidelity (WiFi) module 770, a processor 780, and a power supply 790, etc. The input unit 730 can include a touch panel 731 and other input devices 732, and the display unit 740 can include a display panel 741. The audio circuit 760 can include a speaker 761 and a microphone 762. Those skilled in the art can understand that the structure of the smart phone shown in the figure is not a limitation on the smart phone, and the smart phone can include more or less components than those shown in the figure, or combine some components, or arrange different components. Figure 7 The structure of the smart phone shown in the figure is not a limitation on the smart phone, and the smart phone can include more or less components than those shown in the figure, or combine some components, or arrange different components.

[0223] The memory 720 can be used to store software programs and modules, and the processor 780 can execute various function applications and data processing of the smart phone by running the software programs and modules stored in the memory 720. The memory 720 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application program required by a function (such as a sound playing function, an image playing function, etc.), etc.; and the data storage area can store data created according to the use of the smart phone (such as audio data, a phone book, etc.), etc. In addition, the memory 720 can include a high-speed random access memory, and can also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state memory device.

[0224] The processor 780 is a control center of the smart phone, and connects all parts of the smart phone through various interfaces and lines, and executes various functions and processes data of the smart phone by running or executing software programs and / or modules stored in the memory 720, and calling data stored in the memory 720. Optionally, the processor 780 can include one or more processing units; preferably, the processor 780 can integrate an application processor and a modem processor, wherein the application processor mainly processes an operating system, a user interface, and an application program, etc., and the modem processor mainly processes wireless communication. It can be understood that the above-mentioned modem processor can also not be integrated into the processor 780.

[0225] In the embodiment, the steps performed by the processor 780 in the smart phone can be implemented based on the structure shown in Figure 7

[0226] The computer device provided by the embodiment of the present application can also be a server, as shown in Figure 8 the figure.​Figure 8 The server 800 provided in the embodiments of the present application can have great differences due to different configurations or performances. The server 800 can include one or more processors, such as a central processing unit (CPU) 822, and a memory 832, and one or more storage media 830 (such as one or more mass storage devices) storing application programs 842 or data 844. The memory 832 and the storage media 830 can be temporary storage or persistent storage. The programs stored in the storage media 830 can include one or more modules (not shown in the figure), each of which can include a series of instructions operated in the server. Further, the central processing unit 822 can be configured to communicate with the storage media 830 and execute the series of instructions operated in the storage media 830 on the server 800.

[0227] The server 800 can also include one or more power supplies 826, one or more wired or wireless network interfaces 850, one or more input / output interfaces 858, and / or one or more operating systems 841, such as Windows Server TM , Mac OS X TM , Unix TM , Linux TM , FreeBSD TM , etc.

[0228] In the embodiments, the central processing unit 822 in the server 800 can perform the following steps:

[0229] In response to a message box message push request, obtaining an object and a candidate message indicated by the message push request;

[0230] According to the object characteristics of the object and the message characteristics of the candidate message, determining a predicted entering probability corresponding to the candidate message and a predicted operation result corresponding to the candidate message, the predicted entering probability being used to indicate a probability of the object implementing an entering operation when the message box pushes the candidate message with a message prompt, the entering operation being used to indicate displaying a plurality of messages containing the candidate message in the message box, and the predicted operation result being used to indicate operation implementation characteristics of the object in the message box based on the entering operation;

[0231] Based on the predicted entering probability and the predicted operation result, determining a push value corresponding to the candidate message, the push value being used to determine whether the candidate message is pushed with a message prompt in the message box associated with the object.

[0232] According to an aspect of the present application, a computer readable storage medium is provided for storing a computer program, which, when executed by a computer device, causes the computer device to perform the message pushing method of the various embodiments described above.

[0233] According to an aspect of the present application, a computer program product is provided, which comprises a computer program stored in a computer readable storage medium. A processor of a computer device reads the computer program from the computer readable storage medium, and executes the computer program, so that the computer device performs the method provided in the various optional implementation manners of the above embodiments.

[0234] The descriptions of the corresponding flow or structure of each of the above figures are each focused on, and the parts not described in detail in a certain flow or structure can be referred to the related descriptions of other flows or structures.

[0235] The terms "first", "second", "third", "fourth" and the like in the description of the present application and in the above figures (if present) are used to distinguish like objects and are not necessarily used to describe a particular sequential or chronological order. It is to be understood that the data thus designated can be interchanged, where appropriate, so that the embodiments of the present application described herein can be carried out in other than the order shown or described herein. Furthermore, the terms "comprising" and "including" and any of their derivatives, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a list of steps or units is not necessarily limited to those steps or units that are clearly listed, but can include other steps or units that are not clearly listed or inherent to such processes, methods, products, or apparatuses.

[0236] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the device embodiments described above are merely illustrative, for example, the division of units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed each other can be indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0237] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e. they can be located in one place, or distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0238] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0239] If the integrated unit is realized in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the part that essentially contributes to the related art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a computer, a server, or a network device, etc.) to execute all or part of the steps of the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program code storage media.

[0240] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program with a predetermined function, and works together with other related parts to achieve a predetermined target, and can be realized in whole or in part by using software, hardware (such as a processing circuit or a memory) or a combination thereof. Similarly, one processor (or multiple processors or memories) can be used to realize one or more modules or units. In addition, each module or unit can be a part of an overall module or unit that includes the functions of the module or unit.

[0241] The above, the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A message pushing method, characterized by, The method comprises: in response to a message box message push request, obtaining an object and a candidate message indicated by the message push request; determining a predicted entering probability corresponding to the candidate message and a predicted operation result corresponding to the candidate message according to object features of the object and message features of the candidate message, the predicted entering probability being used to represent a probability of the object implementing an entering operation when the message box pushes the candidate message in a message prompt, the entering operation being used to indicate displaying a plurality of messages containing the candidate message in the message box, and the predicted operation result being used to indicate operation implementation features of the object in the message box based on the entering operation for the plurality of messages; determining a push value corresponding to the candidate message based on the predicted entering probability and the predicted operation result, the push value being used to determine whether the candidate message is pushed in a message prompt in a message box associated with the object.

2. The method of claim 1, wherein, The method comprises: determining the predicted entering probability corresponding to the candidate message, and determining a predicted operation probability corresponding to the candidate message and a predicted operation number corresponding to the candidate message according to the object features and the message features, the predicted operation probability being used to represent a probability of the object implementing an operation based on the entering operation for the candidate message, and the predicted operation number being used to represent operation implementation frequency features of the object based on the entering operation for the plurality of messages; The method comprises: determining the push value corresponding to the candidate message based on the predicted entering probability, the predicted operation probability, and the predicted operation number.

3. The method of claim 1, wherein, If the predicted operation result comprises a predicted distribution center parameter corresponding to the candidate message and a predicted distribution deviation parameter corresponding to the candidate message, the predicted distribution center parameter being used to represent a value center of the predicted operation number corresponding to the candidate message, and the predicted distribution deviation parameter being used to represent a discrete degree of distribution of the predicted operation number corresponding to the candidate message around the value center, the method comprises: determining the push value corresponding to the candidate message according to the predicted entering probability, the predicted distribution center parameter, and the predicted distribution deviation parameter.

4. The method of claim 3, wherein, If the predicted distribution center parameter is a predicted logarithmic mean of the predicted operation number for the candidate message, and the predicted distribution deviation parameter is a predicted logarithmic standard deviation of the predicted operation number for the candidate message, the method comprises: determining, according to the predicted log mean and the predicted log standard deviation, an operation expectation corresponding to the candidate message, the operation expectation being used to represent a number of operations performed by the object on the plurality of messages in the message box when the candidate message is pushed in a message prompt; determining, according to the predicted entering probability and the operation expectation, a push value corresponding to the candidate message.

5. The method of claim 1, wherein, The determining, according to the object feature of the object and the message feature of the candidate message, of the predicted entering probability corresponding to the candidate message and the predicted operation result corresponding to the candidate message comprises: determining, according to the object feature and the message feature, the predicted entering probability corresponding to the candidate message, and a predicted category probability of each operation category in a plurality of operation categories corresponding to the candidate message, the operation category being determined according to a preset operation number, different operation categories corresponding to different preset operation numbers; The determining, according to the predicted entering probability and the predicted operation result, of the push value corresponding to the candidate message comprises: determining, according to the predicted entering probability and the predicted category probability of each operation category in the plurality of operation categories corresponding to the candidate message, the push value corresponding to the candidate message.

6. The method according to any one of claims 1-5, characterized in that, If the number of the candidate messages is a plurality, the method further comprises: determining, according to the push value corresponding to each of the plurality of candidate messages, a target message corresponding to a push value satisfying a push condition; displaying a message prompt and a message title of the target message in the message box associated with the object.

7. The method of claim 6, wherein, The method further comprises: obtaining a box-in push strategy corresponding to the message box; determining, according to the box-in push strategy and the push value corresponding to each of the plurality of candidate messages except the target message, a plurality of box-in messages corresponding to the target message and a display order of the plurality of box-in messages in the message box; in response to the entering operation performed by the object when the target message is pushed in a message prompt, displaying the target message in the message box, and displaying the plurality of box-in messages after the target message in the display order.

8. The method according to any one of claims 1-5, characterized in that, The object feature comprises a historical operation number performed by the object through the message box in a historical period, and the message feature comprises a historical number of published messages and a historical number of operated messages of a publishing account corresponding to the candidate message in the historical period, a time interval between a maximum time of the historical period and a request time of the message push request being less than a preset interval.

9. The method of claim 1, wherein, The determining, according to the object feature of the object and the message feature of the candidate message, of the predicted entering probability corresponding to the candidate message and the predicted operation result corresponding to the candidate message comprises: determining, according to the object feature and the message feature, the predicted entering probability and the predicted operation result through a target model; The target model is determined by the following way: Obtaining training samples, the training samples comprising object features of sample objects and message features of sample messages, the training samples comprising an entering label and in-box operation labels, the entering label being used to represent whether the sample object implements an entering operation when the message box pushes the sample message with a message prompt, the entering operation being used to instruct to display a plurality of messages containing the sample message in the message box, the in-box operation labels being used to represent operation implementation features of the sample object in the message box for the plurality of messages based on the entering operation; According to the training samples, determining, by an initial model, a sample predicted entering probability corresponding to the sample message and a sample predicted operation result corresponding to the sample message, the sample predicted entering probability being used to represent a probability of the sample object implementing the entering operation when the sample message is pushed with a message prompt, the sample predicted operation result being used to represent the operation implementation features of the sample object in the message box for the plurality of messages based on the entering operation; Determining an entering loss according to the sample predicted entering probability and the entering label, and determining an operation loss according to the sample predicted operation result and the in-box operation labels; According to the entering loss and the operation loss, performing model training on the initial model to obtain the target model.

10. The method of claim 9, wherein, If the in-box operation labels comprise a first operation label and a second operation label, the first operation label being used to represent whether the sample object implements an operation for the sample message, and the second operation label being a sample operation number, the sample operation number being determined according to a number of times of implementing operations for the plurality of messages in the message box by the sample object, the determining, by an initial model, a sample predicted entering probability corresponding to the sample message and a sample predicted operation result corresponding to the sample message according to the training samples comprises: According to the training samples, determining, by the initial model, the sample predicted entering probability, and determining, by the initial model, a sample predicted operation probability corresponding to the sample message and a sample predicted operation number corresponding to the sample message, the sample predicted operation probability being used to represent a probability of the sample object implementing an operation for the sample message based on the entering operation; The determining an operation loss according to the sample predicted operation result and the in-box operation labels comprises: Determining a first operation loss according to the sample predicted operation probability and the first operation label, and determining a second operation loss according to the sample predicted operation number and the sample operation number, the operation loss comprising the first operation loss and the second operation loss.

11. The method of claim 9, wherein, If the in-box operation label is a sample operation number, the sample prediction operation result includes a sample prediction distribution center parameter corresponding to the sample message and a sample prediction distribution deviation parameter corresponding to the sample message, the sample prediction distribution center parameter is used to represent the value center of the sample prediction operation number corresponding to the sample message, and the sample prediction distribution deviation parameter is used to represent the discrete degree of the sample prediction operation number distributed at the value center, and the operation loss is determined according to the sample prediction operation result and the in-box operation label, including: According to the sample prediction distribution center parameter, the sample prediction distribution deviation parameter and the sample operation number, the operation loss is determined.

12. The method of claim 9, wherein, If the in-box operation label is used to represent a sample operation category corresponding to the sample message, the sample operation category is determined according to a sample operation number corresponding to the sample message, and the sample operation number is determined according to the number of operations of the sample object in the message box on the plurality of messages, and the sample prediction entering probability corresponding to the sample message and the sample prediction operation result corresponding to the sample message are determined by the initial model according to the training sample, including: According to the training sample, the sample prediction entering probability is determined by the initial model, and the sample prediction category probability of each operation category in the plurality of operation categories corresponding to the sample message is determined by the initial model, and the preset operation number corresponding to different operation categories is different; The operation loss is determined according to the sample prediction operation result and the in-box operation label, including: According to the sample prediction category probability of the plurality of operation categories corresponding to the sample message and the in-box operation label, the operation loss is determined.

13. The method according to any one of claims 9-12, characterized in that, The training sample further includes a time feature, and the time feature is used to represent the time interval between the operation time of the sample object implementing the entering operation and the push time of the sample message pushed by the message prompt of the message box.

14. A message pushing apparatus characterized by comprising: The device includes an acquisition unit and a determination unit: The acquisition unit is used to acquire an object and a candidate message indicated by a message push request of a message box in response to the message push request; The determination unit is used to determine a prediction entering probability corresponding to the candidate message and a prediction operation result corresponding to the candidate message according to an object feature of the object and a message feature of the candidate message, the prediction entering probability is used to represent the probability of the object implementing an entering operation when the candidate message is pushed by the message prompt of the message box, the entering operation is used to indicate that a plurality of messages containing the candidate message are displayed in the message box, and the prediction operation result is used to indicate the operation implementation feature of the object in the message box based on the entering operation on the plurality of messages; The determination unit is further used to determine a push value corresponding to the candidate message based on the prediction entering probability and the prediction operation result, and the push value is used to determine whether the candidate message is pushed by the message prompt of the message box associated with the object.

15. A computer device, comprising: The computer device comprises a processor and a memory: The memory is configured to store a computer program and transmit the computer program to the processor; The processor is configured to execute the method according to any one of claims 1-13 according to instructions in the computer program.

16. A computer-readable storage medium, characterized in that, The computer readable storage medium is configured to store a computer program, which, when executed by a computer device, causes the computer device to execute the method according to any one of claims 1-13.

17. A computer program product comprising a computer program, characterized in that, When executed by a computer device, causes the computer device to execute the method according to any one of claims 1-13.

Images