CN111860870A

CN111860870A - Training method, device, equipment and medium for interactive behavior determination model

Info

Publication number: CN111860870A
Application number: CN202010744200.3A
Authority: CN
Inventors: 李宁
Original assignee: Beijing Dajia Internet Information Technology Co Ltd
Current assignee: Beijing Dajia Internet Information Technology Co Ltd
Priority date: 2020-07-29
Filing date: 2020-07-29
Publication date: 2020-10-30

Abstract

According to the technical scheme, the server can train the model according to the content item interaction characteristics of the sample account and the content item characteristics of at least one content item pushed to the sample account. In the model training process, the type of the content item features is combined when the model outputs the first target probability, so that the influence degrees of the same type of content item features on the model output result are the same, and the influence degrees of different types of content item features on the model output result are different, so that the model can fully learn the potential mode of the content item features in the training process, the accuracy of the model in determining the interactive behavior is improved, and the accuracy of subsequently adopting the interactive behavior determination model to push the content item is improved.

Description

Training method, device, equipment and medium for interactive behavior determination model

Technical Field

The present application relates to the field of artificial intelligence, and in particular, to a training method and apparatus for an interactive behavior determination model, an electronic device, and a storage medium.

Background

Today, with the rapid development of network technology, people receive various content items such as short videos, advertisements or news through terminals, and users can perform various interactive actions with the content items, such as clicking, praise and collection. For the merchant, pushing advertisements to the terminal is an important means for the merchant to increase the sales volume of the goods.

In the related technology, merchants can predict click rates of different advertisements of users through an interactive behavior prediction model stored in a server, and push the advertisements with higher click rates to the users. In addition, for other types of content items, such as short videos, a short video with a higher user click-through rate may not necessarily indicate that the short video meets the user's preference, and may be due to actions of a title party and the like, resulting in a higher click-through rate for the segment of video. Based on the above, the short video platform may also predict, through the interactive behavior prediction model, a probability that the user performs other interactive behaviors with the segment of video, for example, predict, through the interactive behavior prediction model, a probability that the user approves the short video, and recommend the short video to the user based on the probability.

However, since one content item may include multiple types of features, for advertising, the advertising features may include brand, energy consumption ratio, and usage features, and for short video, the short video features may include author, time, actor, and so on features. The influence of different types of features on the interactive behavior prediction model for predicting the interactive behavior probability is different, and the interactive behavior prediction effect of the interactive behavior prediction model trained under the condition of not distinguishing the feature types is poor.

Disclosure of Invention

The application provides a training method and device for an interactive behavior determination model, electronic equipment and a storage medium, so as to improve accuracy of interactive behavior determination. The technical scheme of the application is as follows:

in one aspect, a training method for an interactive behavior determination model is provided, which includes:

in any iterative training process, obtaining sample data, wherein the sample data comprises a first content item interaction characteristic of a sample account and a first content item characteristic of at least one content item pushed to the sample account, and the first content item interaction characteristic is used for representing historical interaction behaviors between the sample account and the content item;

inputting the sample data into an interactive behavior determination model, and executing the following steps through the interactive behavior determination model:

obtaining a first target characteristic of at least one content item according to a first content item characteristic of the at least one content item of the sample account;

outputting a first target probability of target interaction behavior of the sample account with the at least one content item according to the first content item interaction characteristic, the first target characteristic, the type of the first content item characteristic of the at least one content item, and the first content item characteristic of the at least one content item;

and acquiring the interactive behavior determination model as a trained interactive behavior determination model in response to the first target probability output by the interactive behavior determination model and the target condition being met between the sample account and the actual interaction condition of the at least one content item.

In one possible implementation, the obtaining, according to the first content item characteristic of the at least one content item of the sample account, the first target characteristic of the at least one content item includes:

and respectively performing dimension reduction processing on the first content item characteristic of at least one content item of the sample account to obtain a first target characteristic of the at least one content item.

In one possible implementation, the outputting the first target probability of the sample account having the target interaction with the at least one content item according to the first content item interaction characteristic, the first target characteristic, the type of the first content item characteristic of the at least one content item, and the first content item characteristic of the at least one content item includes:

setting a first sample parameter corresponding to the type of a first content item characteristic of the at least one content item for the first target characteristic, wherein the first sample parameter is used for expressing the influence degree of the first target characteristic on a model output result;

outputting the first target probability of the sample account having the target interaction behavior with the at least one content item according to the first content item interaction characteristic, the first sample parameter, the first target characteristic, and a first content item characteristic of the at least one content item.

In one possible implementation, the outputting a first target probability of a target interaction behavior of the sample account with the at least one content item according to the first content item interaction characteristic, the first sample parameter, the first target characteristic, and the first content item characteristic of the at least one content item includes:

according to the first sample parameter, fusing the first target characteristic to obtain a second sample parameter;

outputting the first target probability according to the second sample parameter, the first content item characteristic of the at least one content item and the first content item interaction characteristic.

In one possible implementation, the first content item characteristic and the first content item interaction characteristic each include a plurality of sub-characteristics, and the outputting the first target probability according to the second sample parameter, the first content item characteristic of the at least one content item, and the first content item interaction characteristic includes:

carrying out weighted summation on the partial features of the first content item feature and the partial features of the first content item interaction feature to obtain a first sample fusion feature value;

according to the second sample parameter, carrying out weighted summation on the inner products of the sub-features of at least two first content item features and the inner products of the sub-features of at least two first content item interaction features to obtain a second sample fusion feature value;

fusing the first sample fusion characteristic value and the second sample fusion characteristic value to obtain a third sample fusion characteristic value;

and outputting the first target probability according to the third sample fusion characteristic value.

In a possible implementation, the outputting the first target probability according to the third sample fusion feature value includes:

and carrying out normalization processing on the third sample fusion characteristic value to obtain the first target probability.

In a possible implementation, after obtaining the interactive behavior determination model as a trained interactive behavior determination model, the method further includes:

acquiring a second content item interaction characteristic of the account to be pushed, wherein the second content item interaction characteristic is used for representing historical interaction behaviors between the account to be pushed and a content item;

inputting the second content item interaction characteristic and a second content item characteristic of at least one candidate content item of the account to be pushed into the interaction behavior determination model;

outputting a second target probability of target interaction between the account to be pushed and the at least one candidate content item through the interaction behavior determination model;

and determining a target content item to be pushed from the at least one candidate content item according to the second target probability, wherein the target content item is a candidate content item with the second target probability meeting a probability condition.

In a possible implementation manner, the outputting, by the interactive behavior determination model, a second target probability of a target interactive behavior between the account to be pushed and the at least one candidate content item includes:

obtaining a second target feature of the at least one candidate content item according to a second content item feature of the at least one candidate content item;

setting a first target parameter corresponding to the type of a second content item feature of the at least one candidate content item for the second target feature, wherein the first target parameter is used for expressing the influence degree of the second target feature on a model output result;

fusing at least two second target characteristics according to the first target parameter to obtain a second target parameter;

outputting the second target probability based on the second target parameter, a second content item characteristic of the at least one candidate content item, and the second content item interaction characteristic.

In a possible implementation, the deriving the second target feature of the at least one candidate content item according to the second content item feature of the at least one candidate content item includes:

and performing dimension reduction processing on the second content item characteristic of the at least one candidate content item to obtain a second target characteristic of the at least one candidate content item.

In one possible implementation, the second content item characteristic and the second content item interaction characteristic each comprise a plurality of sub-characteristics, and the outputting the second target probability in accordance with the second target parameter, the second content item characteristic of the at least one candidate content item, and the second content item interaction characteristic comprises:

carrying out weighted summation on the partial features of the second content item features and the partial features of the second content item interaction features to obtain a first target fusion feature value;

according to the second target parameter, carrying out weighted summation on the inner products of the sub-features of at least two second content item features and the inner products of the sub-features of at least two second content item interaction features to obtain a second target fusion feature value;

fusing the first target fusion characteristic value and the second target fusion characteristic value to obtain a third target fusion characteristic value;

and outputting the second target probability according to the third target fusion characteristic value.

In a possible implementation manner, the outputting the second target probability according to the third target fusion characteristic value includes:

and carrying out normalization processing on the third target fusion characteristic value to obtain the second target probability.

In one possible implementation, after determining the target content item to be pushed from the at least one candidate content item, the method further comprises:

updating the interaction behavior determination model based on the interaction information between the account to be pushed and the target content item.

In one aspect, a training apparatus for determining a model of interactive behavior is provided, including:

the data acquisition unit is configured to perform any iterative training process, and acquire sample data, wherein the sample data comprises a first content item interaction feature of a sample account and a first content item feature of at least one content item pushed to the sample account, and the first content item interaction feature is used for representing historical interaction behaviors between the sample account and the content item;

an input unit configured to perform input of the sample data into an interactive behavior determination model;

an execution unit configured to execute the following steps performed by the interactive behavior determination model:

obtaining a first target characteristic of at least one content item according to a first content item characteristic of the at least one content item of the sample account; outputting a first target probability of target interaction behavior of the sample account with the at least one content item according to the first content item interaction characteristic, the first target characteristic, the type of the first content item characteristic of the at least one content item, and the first content item characteristic of the at least one content item;

a model obtaining unit configured to perform obtaining the interactive behavior determination model as a trained interactive behavior determination model in response to the first target probability output by the interactive behavior determination model and a target condition being satisfied between the sample account and an actual interaction situation of the at least one content item.

In a possible implementation manner, the execution unit is configured to perform dimension reduction processing on first content item features of at least one content item of the sample account, respectively, to obtain first target features of the at least one content item.

In a possible implementation manner, the execution unit is configured to execute setting, for the first target feature, a first sample parameter corresponding to a type of a first content item feature of the at least one content item, where the first sample parameter is used to indicate a degree of influence of the first target feature on a model output result; outputting the first target probability of the sample account having the target interaction behavior with the at least one content item according to the first content item interaction characteristic, the first sample parameter, the first target characteristic, and a first content item characteristic of the at least one content item.

In a possible implementation manner, the execution unit is configured to perform fusion of the first target feature according to the first sample parameter to obtain a second sample parameter; outputting the first target probability according to the second sample parameter, the first content item characteristic of the at least one content item and the first content item interaction characteristic.

In a possible implementation, the first content item characteristic and the first content item interaction characteristic each comprise a plurality of sub-characteristics, and the execution unit is configured to perform a weighted summation of the sub-characteristics of the first content item characteristic and the sub-characteristics of the first content item interaction characteristic to obtain a first sample fusion characteristic value; according to the second sample parameter, carrying out weighted summation on the inner products of the sub-features of at least two first content item features and the inner products of the sub-features of at least two first content item interaction features to obtain a second sample fusion feature value; fusing the first sample fusion characteristic value and the second sample fusion characteristic value to obtain a third sample fusion characteristic value; and outputting the first target probability according to the third sample fusion characteristic value.

In a possible implementation manner, the execution unit is configured to perform a normalization process on the third sample fusion feature value, so as to obtain the first target probability.

In a possible embodiment, the apparatus further comprises:

the obtaining unit is configured to perform obtaining of a second content item interaction feature of the account to be pushed, where the second content item interaction feature is used to represent a historical interaction behavior between the account to be pushed and a content item;

the input unit is further configured to perform input of the second content item interaction characteristic and a second content item characteristic of at least one candidate content item of the account to be pushed into the interaction behavior determination model;

the output unit is configured to execute target probability of target interactive behaviors of the account to be pushed and the at least one candidate content item output through the interactive behavior determination model;

a determining unit configured to perform determining, according to the target probability, a target content item to be pushed from the at least one candidate content item, where the target content item is a candidate content item whose target probability meets a probability condition.

In a possible implementation, the output unit is configured to perform deriving a second target feature of the at least one candidate content item according to a second content item feature of the at least one candidate content item; setting a second sample parameter corresponding to the type of a second content item feature of the at least one candidate content item for the second target feature, wherein the second sample parameter is used for expressing the influence degree of the second target feature on a model output result; fusing at least two second target characteristics according to the first target parameter to obtain a second target parameter; outputting the target probability based on the second target parameter, a second content item characteristic of the at least one candidate content item, and the second content item interaction characteristic.

In a possible implementation, the output unit is configured to perform a dimension reduction process on the second content item feature of the at least one candidate content item, to obtain a second target feature of the at least one candidate content item.

In a possible implementation, the second content item feature and the second content item interaction feature each include a plurality of sub-features, and the output unit is configured to perform weighted summation of the sub-features of the second content item feature and the sub-features of the second content item interaction feature to obtain a first target fusion feature value; according to the second target parameter, carrying out weighted summation on the inner products of the sub-features of at least two second content item features and the inner products of the sub-features of at least two second content item interaction features to obtain a second target fusion feature value; fusing the first target fusion characteristic value and the second target fusion characteristic value to obtain a third target fusion characteristic value; and outputting the second target probability according to the third target fusion characteristic value.

In a possible implementation manner, the output unit is configured to perform a normalization process on the third target fusion characteristic value to obtain the second target probability.

In a possible embodiment, the apparatus further comprises:

an updating unit configured to perform updating the interaction behavior determination model based on interaction information between the account to be pushed and the target content item.

In one aspect, an electronic device is provided, including:

one or more processors;

a memory for storing the processor executable program code;

wherein the processor is configured to execute the program code to implement the method of training the interactive behavior determination model described above.

In one aspect, a storage medium is provided, in which program code is enabled, when executed by a processor of an electronic device, to perform the above-described method of training an interactive behavior determination model.

In one aspect, a computer program product is provided, which stores one or more program codes executable by a processor of an electronic device to perform the above-mentioned method for training an interactive behavior determination model.

The technical scheme provided by the embodiment of the application at least has the following beneficial effects:

It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the application.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the application and are not to be construed as limiting the application;

FIG. 1 is a schematic diagram of an implementation environment illustrating a method of training an interactive behavior determination model, according to an example embodiment;

FIG. 2 is a block diagram illustrating a recommendation system in accordance with an exemplary embodiment;

FIG. 3 is a schematic diagram illustrating the structure of an interaction behavior determination model in accordance with an illustrative embodiment;

FIG. 4 is a flow diagram illustrating a method of training an interactive behavior determination model in accordance with an exemplary embodiment;

FIG. 5 is a flow diagram illustrating a method of training an interactive behavior determination model in accordance with an exemplary embodiment;

FIG. 6 is a flow diagram illustrating a method of pushing a content item in accordance with an exemplary embodiment;

FIG. 7 is a block diagram illustrating a training apparatus for an interactive behavior determination model in accordance with an exemplary embodiment;

FIG. 8 is a block diagram illustrating a server in accordance with an exemplary embodiment.

Detailed Description

In order to make the technical solutions of the present application better understood by those of ordinary skill in the art, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

It should be noted that the terms "first," "second," and the like in the description and claims of this application and in the drawings described above are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used is interchangeable under appropriate circumstances such that the embodiments of the application described herein are capable of operation in sequences other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present application, as detailed in the appended claims.

The user information referred to in the present application may be information authorized by the user or sufficiently authorized by each party.

Fig. 1 is a schematic diagram of an implementation environment of a training method for an interactive behavior determination model according to an embodiment of the present application, and referring to fig. 1, the implementation environment may include a terminal 101 and a server 102.

The terminal 101 may be at least one of a smartphone, a smart watch, a desktop computer, a laptop portable computer, and the like. The terminal 101 may be installed and run with an application program supporting training of the interaction behavior determination model and content item presentation, and the user may log in the application program through the terminal 101 to receive the content item, for example, after the user logs in the application program, the server 102 may push the target content item to the terminal 101 according to an account of the user, and the terminal presents the target content item to the user through the application program. The terminal 101 may be connected to the server 102 through a wireless network or a wired network.

The terminal 101 may be generally referred to as one of a plurality of terminals, and the embodiment is only illustrated by the terminal 101. Those skilled in the art will appreciate that the number of terminals described above may be greater or fewer. For example, the number of the terminals 101 may be only a few, or the number of the terminals 101 may be tens or hundreds, or more, and the number of the terminals 101 and the type of the device are not limited in the embodiment of the present application.

The server 102 may be at least one of a server, a plurality of servers, a cloud computing platform, and a virtualization center. The server 102 may be used for training the interactive behaviour determination model and may also be used for pushing content items to the terminal 101.

Alternatively, the number of the servers 102 may be more or less, and the embodiment of the present application is not limited thereto. Of course, the server 102 may also include other functional servers to provide more comprehensive and diverse services.

The interactive behavior determination model provided by the present application can be applied to various scenarios, and for easy understanding, first, the application scenarios that may be involved in the present application are described:

1. the interactive behavior determination model provided by the application can be applied to an advertisement pushing scene, for example, when the terminal starts an application program, a login request can be sent to the server, the login request carries an account and a password of a user, and the server can determine whether the account and the password of the user are matched. In response to the matching of the account and the password of the user, the server can acquire the advertisement interaction characteristics of the user according to the account of the user, and the advertisement interaction characteristics can be used for reflecting the historical click conditions of the user on different advertisements. The server can input the advertisement interaction characteristics of the user and the advertisement characteristics in the advertisement database into the interaction behavior determination model, and the interaction behavior determination model outputs the probability that the user clicks different advertisements according to the advertisement interaction characteristics and the advertisement characteristics of the user. The server can push the advertisement with the highest click probability of the user to the terminal, and the terminal displays the advertisement to the user.

2. The interactive behavior determination model provided by the application can be applied to a search scene, for example, when a user needs to search through a terminal, a keyword can be input on the terminal, such as "XX university". The method comprises the steps that after a terminal detects search operation of a user, a search instruction can be triggered, the terminal responds to the search instruction and sends a search request to a server, the search request carries a keyword and a user account, the server receives the search request, queries are carried out in a database based on the keyword carried by the search request, and a plurality of content items corresponding to the keyword are determined. The server can obtain the content item interaction characteristics of the user according to the user account, because the content item interaction characteristics of the user can reflect the preference of the user to the content items, the server can output the click rate of the user to different content items according to the content item interaction characteristics and the plurality of content item characteristics, and sends the content items with the highest click rate of the user to the terminal and displays the content items to the user through the terminal.

3. The interactive behavior determination model provided by the application can be applied to various general recommendation scenes, for example, scenes such as video recommendation, commodity recommendation, news recommendation, picture recommendation and the like, various recommendation methods need to be realized by means of a recommendation system, and the structure of the general recommendation system is introduced below.

Referring to fig. 2 and 200, a recommendation system 200, the recommendation system 200 may include a matching module 201, a recall module 202, a ranking module 203, and a sending module 204, and the recommendation system 200 may be built on a server.

A matching module 201 may be configured to match a user representation, which is a tag generated based on user interaction with a content item, with the content item representation, which may be used to describe a type and characteristics of the user. Accordingly, the content item may be a tag generated from the content of the content item for describing the type and characteristics of the content item. For example, the user representation and the content item representation may be represented by vectors, and the matching module 201 may determine cosine similarity between the user representation vector and the content item representation vector, and match the user representation and the content item representation whose cosine similarity meets a target condition.

The recall module 202 may be configured to recall a target number of content items from a plurality of content items. The server may recall from the content item database a plurality of content items matching the user representation. For example, the recall module 202 may sort out hundreds of content items that may be of interest to the user from the tens of thousands of content items stored in the database according to the matching result of the matching module 201.

The sorting module 203 may be configured to output click-through rates of different content items from users, and sort the content items according to the click-through rates. For example, the server may output, through the ranking module 203, click-through rates of hundreds of content items recalled by the recall module 202 by the user, rank the content items recalled by the recall module according to a sequence of click-through rates from large to small, and determine the content item recommendations of the top target number as the target content items to be recommended. In other words, the ranking module 203 performs a secondary filtering on the content items recalled by the recall module 202 to determine tens or several content items with the highest user click rate.

The sending module 204 may be configured to send the target content item determined by the sorting module 203 to a terminal, and present the content item to a user through the terminal.

For the recommendation system, the interaction behavior determination model provided in this embodiment of the present application may be used in the process of outputting the click rate by the ranking module 203, and of course, in other possible implementations, the interaction behavior determination model may also be used in the recall process of the recall module 202, which is not limited in this embodiment of the present application.

After the above recommendation system is introduced, the structure of the interactive behavior determination model provided in the present application is described below, and referring to fig. 3, the interactive behavior determination model may include an input layer 301, a linear layer 302, a cross layer 303, a fusion layer 304, and a probability determination layer 305.

The input layer 301 is used to input the content item interaction characteristics of the account and the content item characteristics pushed to the account into the interaction behavior determination model. If the content item interaction feature and the content item feature are represented in vector form, the input layer 301 is configured to input the content item interaction feature and the content item feature vector of the account to the interaction behavior determination model.

The linear layer 302 is configured to perform weighted summation on the content item feature and the content item interaction feature to obtain a first fused feature value. The linear layer 302 weights the content item features by a sub-feature of the content item features corresponding to a weight, wherein the weight of the linear layer 302 is obtained in the training process of the interactive behavior determination model.

The crossing layer 303 is configured to perform weighted summation on an inner product of the sub-features of at least two content item features, and in the process of performing weighted summation on the features, the crossing layer 303 performs weighting corresponding to at least two different content item features, where the weighting of the linear layer 303 is obtained according to a target vector of the content item features, and the target vector may also be referred to as a hidden vector or an auxiliary vector.

The fusion layer 304 is configured to fuse the features output by the linear layer 302 and the cross layer 303 to obtain a fused feature value, which can be used to more completely represent the feature of the content item.

The probability determination layer 305 is configured to perform normalization processing on the fusion feature value to obtain a target probability of the target content item, where the target probability is a probability of a target interaction between the user and the content item.

Of course, the structure of the interactive behavior determination model is not limited to the above structure, and in other possible implementations, the interactive behavior determination model may also be a model of another structure, which is not limited in this application.

Fig. 4 is a flowchart illustrating a training method of an interactive behavior determination model according to an exemplary embodiment, and as shown in fig. 4, the training method of the interactive behavior determination model may be applied to a server, and includes the following steps.

In step S401, in any iterative training process, sample data is obtained, where the sample data includes a first content item interaction feature of a sample account and a first content item feature of at least one content item that has been pushed to the sample account, and the first content item interaction feature is used to represent a historical interaction behavior between the sample account and the content item.

In step S402, the server inputs sample data into the interactive behavior determination model, and performs the following steps S403 and S404 through the interactive behavior determination model.

In step S403, the server obtains, through the interaction behavior determination model, a first target feature of at least one content item according to a first content item feature of at least one content item of the sample account.

In step S404, the server outputs, through the interaction behavior determination model, a first target probability of the target interaction behavior between the sample account and the at least one content item according to the first content item interaction characteristic, the first target characteristic, the type of the first content item characteristic of the at least one content item, and the first content item characteristic of the at least one content item.

In step S405, in response to that the first target probability output by the interactive behavior determination model and the actual interaction situation between the sample account and the at least one content item satisfy the target condition, the server acquires the interactive behavior determination model as a trained interactive behavior determination model.

In one possible implementation, obtaining the first target characteristic of the at least one content item according to the first content item characteristic of the at least one content item of the sample account includes:

and respectively carrying out dimension reduction processing on the first content item characteristic of at least one content item of the sample account to obtain a first target characteristic of at least one content item.

In one possible implementation, outputting a first target probability of a target interaction behavior of the sample account with the at least one content item according to the first content item interaction characteristic, the first target characteristic, the type of the first content item characteristic of the at least one content item, and the first content item characteristic of the at least one content item comprises:

setting a first sample parameter corresponding to the type of the first content item characteristic of the at least one content item for the first target characteristic, wherein the first sample parameter is used for expressing the influence degree of the first target characteristic on the output result of the model.

And outputting a first target probability of the target interaction between the sample account and the at least one content item according to the first content item interaction characteristic, the first sample parameter, the first target characteristic and the first content item characteristic of the at least one content item.

In one possible implementation, outputting a first target probability of a target interaction behavior of the sample account with the at least one content item according to the first content item interaction characteristic, the first sample parameter, the first target characteristic, and the first content item characteristic of the at least one content item includes:

and according to the first sample parameter, fusing the first target characteristic to obtain a second sample parameter.

A first target probability is output based on the second sample parameter, the first content item characteristic of the at least one content item, and the first content item interaction characteristic.

In one possible implementation, the first content item characteristic and the first content item interaction characteristic each include a plurality of sub-characteristics, and outputting the first target probability in accordance with the second sample parameter, the first content item characteristic of the at least one content item, and the first content item interaction characteristic includes:

and carrying out weighted summation on the partial features of the first content item feature and the partial features of the first content item interaction feature to obtain a first sample fusion feature value.

And according to the second sample parameter, carrying out weighted summation on the inner products of the sub-features of the at least two first content item features and the inner products of the sub-features of the at least two first content item interaction features to obtain a second sample fusion feature value.

And fusing the first sample fusion characteristic value and the second sample fusion characteristic value to obtain a third sample fusion characteristic value.

In one possible embodiment, the outputting the first target probability according to the third sample fusion feature value includes:

and carrying out normalization processing on the third sample fusion characteristic value to obtain a first target probability.

In one possible implementation, after obtaining the interactive behavior determination model as a trained interactive behavior determination model, the method further includes:

and acquiring a second content item interaction characteristic of the account to be pushed, wherein the second content item interaction characteristic is used for representing historical interaction behaviors between the account to be pushed and the content item.

And inputting the second content item interaction characteristics and the second content item characteristics of at least one candidate content item of the account to be pushed into the interaction behavior determination model.

And outputting a second target probability of the target interactive behavior of the account to be pushed and at least one candidate content item through the interactive behavior determination model.

And determining a target content item to be pushed from the at least one candidate content item according to the second target probability, wherein the target content item is a candidate content item with the second target probability meeting the probability condition.

In a possible implementation manner, outputting, by the interactive behavior determination model, a second target probability of a target interactive behavior between the account to be pushed and the at least one candidate content item includes:

a second target feature of the at least one candidate content item is derived based on the second content item feature of the at least one candidate content item.

And setting a first target parameter corresponding to the type of the second content item characteristic of the at least one candidate content item for the second target characteristic, wherein the first target parameter is used for expressing the influence degree of the second target characteristic on the output result of the model.

And fusing at least two second target characteristics according to the first target parameter to obtain a second target parameter.

A second target probability is output based on the second target parameter, the second content item characteristic of the at least one candidate content item, and the second content item interaction characteristic.

In one possible implementation, deriving the second target feature of the at least one candidate content item from the second content item feature of the at least one candidate content item comprises:

In one possible implementation, the second content item characteristic and the second content item interaction characteristic each include a plurality of sub-characteristics, and outputting the second target probability based on the second target parameter, the second content item characteristic of the at least one candidate content item, and the second content item interaction characteristic comprises:

and carrying out weighted summation on the partial features of the second content item features and the partial features of the second content item interaction features to obtain a first target fusion feature value.

And according to the second target parameter, carrying out weighted summation on the inner products of the sub-features of the at least two second content item features and the inner products of the sub-features of the at least two second content item interaction features to obtain a second target fusion feature value.

And fusing the first target fusion characteristic value and the second target fusion characteristic value to obtain a third target fusion characteristic value.

In one possible embodiment, outputting the second target probability according to the third target fusion characteristic value includes:

and carrying out normalization processing on the third target fusion characteristic value to obtain a second target probability.

The above steps S401 to S406 are a brief description provided in the embodiment of the present application, and the following will describe, with reference to some examples, a training method of an interactive behavior determination model provided in the embodiment of the present application, it should be noted that the training process of the interactive behavior determination model may include a plurality of iterative processes, and for describing the training process of the model more clearly, a description is given below by taking an iterative process as an example, with reference to fig. 5, and the steps include:

in step S501, in any iterative training process, the server obtains sample data, where the sample data includes a first content item interaction feature of a sample account and a first content item feature of at least one content item that has been pushed to the sample account, and the first content item interaction feature is used to represent a historical interaction behavior between the sample account and the content item.

The sample account may be an account of any content item recommendation platform, at least one content item pushed to the sample account is a content item recommended to the sample account by the content item recommendation platform, and the historical interaction behavior between the sample account and the content item may include positive behaviors of "clicking", "collecting", "like", and "sharing" of the user on the content item, and may also include negative behaviors of "reporting" and "dislike" of the user on the content item. The first content item characteristic of the at least one content item may include characteristics of "type", "release time", and "author" of the content item, and may also include other characteristics, which are not limited in this embodiment of the present application. The first content item interaction feature may be in the form of a content item feature + interaction behavior.

In one possible implementation, the server may represent the first content item interaction feature and the first content item feature in the form of a vector. For the first content item feature, the dimension of the first content item feature vector may be determined according to the number of different first content item features comprised by the feature type to which the first content item feature belongs. For example, a short video has a first content item feature "do it" corresponding to a feature type of "video genre" which may include four features of "do it", "act", "story" and "record". The server may use vector [ 1000 ] to represent "fun" using vector [ 0100 ] to represent "action" using vector [ 0010 ] to represent "story" using vector [ 0001 ] to represent "record" this representation may also be referred to as One-Hot encoding.

In step S502, the server inputs sample data into the interactive behavior determination model, and performs the following steps S503 to S505 through the interactive behavior determination model.

In one possible implementation, the sample data may include first content item interaction characteristics of a plurality of sample accounts and first content item characteristics of a plurality of content items that have been pushed to the plurality of sample accounts, the server may use the first content item interaction characteristics of each sample account and the first content item characteristics of the plurality of content items that have been pushed to the sample account as one sample data set, and the server may randomly select one first content item characteristic from the sample data set and input the first content item characteristic and the first content item interaction characteristics with the sample account into the interaction behavior determination model.

In this implementation manner, the server may use the first content item features of the plurality of content items pushed to the same sample account and the first content item interaction features of the sample account as a sample data subset, and may continuously take out the first content item features of the content items from the same sample data subset to train the interaction behavior determination model in a subsequent training process.

In step S503, the server performs dimension reduction processing on the first content item feature of at least one content item of the sample account through the interactive behavior determination model, respectively, to obtain a first target feature of the at least one content item.

In one possible implementation, the server may perform an Embedding (Embedding) process on the first content item feature of at least one content item of the sample account respectively through the interactive behavior determination model to reduce the high-dimensional first content item feature into the low-dimensional first target feature. For example, the server may take the vector (100000)0 0)^TTo represent the first content item the feature vector server may transform the matrix

Pair vector (10000000)^TProcessing to obtain 8-dimensional vector (10000000)^TEmbedded as a 2-dimensional vector (12). In this case, the vector (10000000)^TThe sum vector (12) may both be used to represent the same first content item characteristic. Vector (10000000)^TMay be referred to as an explicit vector of first content item features and the vector (12) may be referred to as an implicit vector of first content item features. An explicit vector is a vector that the server determines from a plurality of content items that can be used to identify a feature. The hidden vector is formed by mapping the explicit vector according to the relationship among a plurality of first content item features of one content item through an interactive behavior determination model by the server, and can express the relationship between one first content item feature and other first content item features of the content item to a certain extent.

In this implementation manner, the server may process the first content item feature with a higher dimension into the first target feature with a lower dimension, and in the subsequent model processing process, the calculation amount may be significantly reduced, and the calculation efficiency of the model may be improved.

In step S504, the server determines, through the interaction behavior determination model, a first sample weight corresponding to the type for the corresponding first target feature according to the different types of the first content item features of the at least one content item, where the first sample weight is used to indicate a degree of influence of the first target feature on the output result of the model.

The genre may also be referred to as a data field, for example, for a short video, the author feature "red a" and the classification feature "make a" look at features belonging to different genres, that is, features belonging to different data fields.

In one possible implementation, a content item may include a plurality of different types of first content item features, such as author features, classification features, and temporal features. The server may determine a first sample weight for the first target feature according to the type of the first content item feature, such as 0.3 for the author feature, 0.5 for the classification feature, and 0.4 for the temporal feature. In the subsequent model training process, after the server inputs the first content item feature of the other content item into the interactive behavior determination model, the model may set the first sample weight corresponding to the author feature of the other content item to 0.3, set the first sample weight corresponding to the classification feature to 0.5, and set the first sample weight corresponding to the time feature to 0.4, that is, the server may set the same first sample weight for the same type of feature of the different content items, so as to ensure that the first content item feature of the same type has the same degree of influence on the model in the model training process.

It should be noted that the first sample weight may be set according to an actual situation, or may be initialized before the model training for different types of first content item features corresponding to the first sample weight, and the first sample weight is adjusted in real time along with the training process of the model, and in response to that the value of the loss function of the model meets a numerical condition, the server may determine the first sample weight at this time as the first sample weight used in the subsequent training process, that is, the first sample weight is kept unchanged from this time, the first content item features of the same type correspond to the same first sample weight, and the first content item features of different types correspond to different first sample weights. The embodiment of the present application does not limit the method for determining the first sample weight.

In this implementation manner, the server can enable the first content item features in the same data domain to have the same degree of influence on the model, and enable the first content item features in different data domains to have different degrees of influence on the model, so that the similarity of the features in the same data domain and the difference of the features in different data domains can be reflected, the first content item features are fully modeled, and the accuracy of the model is improved.

In step S505, the server outputs, through the interactive behavior determination model, a first target probability of the target interactive behavior between the sample account and the at least one content item according to the first content item interactive feature, the first sample weight, the first target feature, and the first content item feature of the at least one content item.

The target interaction behavior may be at least one of a behavior of clicking a content item for a user logging in the sample account, a behavior of collecting a content item for a user logging in the sample account, and a behavior of sharing a content item for a user logging in the sample account.

In a possible implementation manner, the server may determine the model through the interactive behavior, and perform weighted summation on the at least two first target features according to the first sample weight to obtain the second weight. The server may output the first target probability based on the second weight, the first content item characteristic of the at least one content item, and the first content item interaction characteristic.

For example, the step of the server outputting the first target probability according to the second weight, the first content item characteristic of the at least one content item and the first content item interaction characteristic may comprise:

the first content item characteristic and the first content item interaction characteristic each comprise a plurality of sub-characteristics, each sub-characteristic may be used to represent a type of characteristic, and the server may perform a weighted summation of the plurality of sub-characteristics of the first content item characteristic and the sub-characteristics of the first content item interaction characteristic to obtain a first sample fusion characteristic value. The server may perform weighted summation on the inner products of the sub-features of the at least two first content item features and the inner products of the sub-features of the at least two first content item interaction features according to the second weight to obtain a second sample fusion feature value. The server may fuse the first sample fusion eigenvalue and the second sample fusion eigenvalue to obtain a third sample fusion eigenvalue. The server may output the first target probability according to the third sample fusion feature value. For example, the server may perform normalization processing on the third sample fusion feature value to obtain the first target probability.

The above embodiment will be described below with reference to formula (1):

wherein y is a third fused feature value,<W，X>a linear layer 302 corresponding to the model for weighted summation of the first content item feature and the partial features of the splicing feature X of the first content item interaction feature, W being the weight of the linear layer 302, the portion after "+" corresponding to the cross-over layer 303, X_iAnd x_jAs a sub-feature of the stitching feature, w_fpAnd w_fqIs a first sample weight, V_iAnd V_jIs the first target feature.

In one possible implementation, the server may characterize the first content item

Interaction feature with first content item

Splicing being a splicing feature

The server can combine the splicing features

Inputting an interactive behavior determination model, determining the splicing characteristics of the model pair through the interactive behavior

Four points of (10000000)^T、(1 0 1 1 0 0 1 0)^T、(1 0 0 0 1 0 0 0)^TAnd (10101000)^TThe weighted sum is performed to obtain a first sample fusion feature value, for example 1.8, where the first two sub-features are sub-features of the first content item feature and the last two sub-features are sub-features of the first content item interaction feature. The server may determine the model by interactive behavior, by transforming the matrix

For four sub-characteristics (10000000)^T、(1 0 1 1 0 0 1 0)^T、(1 0 0 0 1 0 0 0)^TAnd (10101000)^TThe embedding process is performed to obtain four first target features (12), (33), (12) and (13) corresponding to the four sub-features. The server may be based on a sub-feature (10000000)^TFirst sample weight (0.70.6), partial feature (10110010)^TThe first sample weight (0.50.4), the partial feature (10001000)^TThe first sample weight (0.80.7) and the sub-feature (10101000)^TThe first sample weight (0.80.4) of (a) performs a weighted summation of the inner products of any two of the four first target features to obtain six second weights 2.49, 2.24, 2, 2.88, 2.64, and 2.32. The server may assign four sub-features (10000000) according to six second weights 2.49, 2.24, 2, 2.88, 2.64, and 2.32^T、(1 0 1 1 0 0 1 0)^T、(1 0 0 0 1 0 0 0)^TAnd (10101000)^TAnd performing weighted summation on the inner products of the two sub-features corresponding to the second weights to obtain six second sample fusion feature values 2.49, 2.24, 2, 2.88, 5.28 and 4.64. The server may sum the six second sample fusion eigenvalues to obtain a target second sample fusion eigenvalue 19.53. The server may fuse the first sample fusion feature value 1.8 and the target second sample fusion feature value 19.53 to obtain a third sample fusion feature value 21.33. The server may perform normalization processing on the third sample fusion characteristic value 21.33, and map the third sample fusion characteristic value 21.33 to a value on an interval of (0, 1), where the value is also a first target probability, such as 0.8, where the normalization processing on the third sample fusion characteristic value may be implemented by using a normalization function, such as a Sigmoid (S-shaped growth curve) function and a Softmax function, and of course, in other possible implementation manners, other functions may also be used to perform normalization processing on the third sample fusion characteristic value, which is not limited in this embodiment of the present application.

In addition, for the first content item interaction feature, the server can determine the first target feature and the first sample weight of the first content item interaction feature in advance, so that the target feature and the first sample weight of the first content item interaction feature do not need to be recalculated every training, and the efficiency of model training can be improved.

It should be noted that, in the foregoing embodiment, the first content item interaction feature of the sample account may be generated in advance by the server according to the historical interaction behavior between the sample account and the content item, and if the server generates the first content item interaction feature in advance according to the historical interaction behavior between the sample account and the content item, the dimension reduction processing may also be performed in advance according to the first content item interaction feature to obtain the target feature of the first content item interaction feature, so that the calculation amount of the model may be reduced during model training, and the training efficiency of the model is improved. Of course, the target feature of the interactive feature of the first content item may also be obtained by the server in the training process of the model, and the specific method may refer to the description of step S503, which is not described again, and the determination time of the interactive feature of the first content item is not limited in this embodiment of the application.

It should be noted that, after step S505, the server may determine to execute step S506 or step S507 according to the first target probability output by the interaction behavior determination model and the relationship between the sample account number and the actual interaction situation of the at least one content item. In response to the first target probability output by the interactive behavior determination model and the target condition being satisfied between the actual interaction situation of the sample account and the at least one content item, the server may perform step S506; in response to the first target probability output by the interactive behavior determination model and the target condition not being satisfied between the actual interaction situation of the sample account and the at least one content item, the server may perform step S507.

In step S506, in response to that the first target probability output by the interactive behavior determination model and the actual interaction situation between the sample account and the at least one content item satisfy the target condition, the server acquires the interactive behavior determination model as a trained interactive behavior determination model.

The first target probability and the satisfaction of the target condition between the actual interaction conditions of the sample account and the at least one content item may mean that the first target probability is greater than a probability threshold, and the sample account and the at least one content item perform a target interaction behavior.

In step S507, in response to that the first target probability output by the interactive behavior determination model and the difference between the actual interaction situations of the sample account and the at least one content item do not satisfy the target condition, the server may adjust a model parameter of the interactive behavior determination model according to the first target probability and the difference information between the actual interaction situations of the sample account and the at least one content item.

In a possible implementation manner, the server may adjust the model parameters of the interactive behavior determination model by using a gradient descent method according to the first target probability and the difference information of the actual interaction situation between the sample account and the at least one content item, and then the server may obtain the sample data again, and train the interactive behavior determination model based on the obtained sample data again, where the model training method and the steps S501 to S505 belong to the same inventive concept, and are not described herein again.

According to the technical scheme provided by the application, the server can train the model according to the first content item interaction characteristics of the sample account and the first content item characteristics of at least one content item pushed to the sample account. In the model training process, the server performs dimensionality reduction on the first content item feature of at least one content item, and the efficiency of subsequent model training is improved. Meanwhile, the server sets a first sample weight for the first target feature according to the type of the first content item feature, so that the influence degrees of the first content item features of the same type on the model are the same, and the influence degrees of the first content item features of different types on the model are different, so that the model can fully learn the potential mode of the first content item feature in the training process, the accuracy of the model for determining the interactive behavior is improved, and the accuracy of subsequently adopting the interactive behavior determination model for pushing the content item is improved.

In addition to the above steps S501 to S507, an embodiment of the present application further provides a content item pushing method, where the content item pushing method is implemented based on the interactive behavior determination model obtained through the training in the above steps S501 to S507, and with reference to fig. 6, the steps include:

in step S601, the server obtains a second content item interaction feature of the account to be pushed, where the second content item interaction feature is used to represent a historical interaction behavior between the account to be pushed and the content item.

For the description of the content item interaction behavior, reference may be made to step S501, which is not described herein again.

In step S602, the server inputs the second content item interaction characteristic and the second content item characteristic of the at least one candidate content item of the account to be pushed into the interaction behavior determination model.

The training process of the interactive behavior determination model in the embodiment of the present application may refer to the descriptions of steps S501 to S507, which are not described herein again.

In step S603, the server outputs, through the interactive behavior determination model, a second target probability of a target interactive behavior occurring between the account to be pushed and the at least one candidate content item.

In one possible implementation, the server may perform a dimension reduction process on the second content item feature of the at least one candidate content item to obtain a second target feature of the at least one candidate content item. And determining a first target weight corresponding to the type for the corresponding second target feature according to different types of the second content item features of the at least one candidate content item, wherein the first target weight is used for expressing the influence degree of the second target feature on the output result of the model. And according to the first target weight, carrying out weighted summation on at least two second target characteristics to obtain a second target weight. And outputting a second target probability according to the second target weight, the second content item characteristic of the at least one candidate content item and the second content item interaction characteristic. For example, the second content item feature and the second content item interaction feature each comprise a plurality of sub-features, and the server may perform a weighted summation of the sub-features of the second content item feature and the sub-features of the second content item interaction feature to obtain the first target fusion feature value. The server may perform weighted summation on the inner products of the sub-features of the at least two second content item features and the inner products of the sub-features of the at least two second content item interaction features according to the second target weight to obtain a second target fusion feature value. The server may fuse the first target fusion characteristic value and the second target fusion characteristic value to obtain a third target fusion characteristic value. The server may output the second target probability according to the third target fusion characteristic value. For example, the server may perform normalization processing on the third target fusion feature value to obtain the second target probability.

It should be noted that step S603 and steps S503 and S504 belong to the same inventive concept, and specific embodiments may refer to the descriptions of steps S503 and S504, which are not described herein again.

In step S604, the server determines a target content item to be pushed from the at least one candidate content item according to the second target probability, where the target content item is a candidate content item whose second target probability meets the probability condition.

The second target probability meeting the probability condition may be that the second target probability is greater than or equal to a probability threshold, or the second target probability of the content item is the highest of the at least one content item, or may be other probability conditions, which is not limited in this embodiment of the application.

Optionally, after step S604, the server may further perform steps S605 and S606.

In step S605, the server may push the target content item to the terminal, and acquire interaction information between the account to be pushed and the target content item.

In step S606, the server updates the interaction behavior determination model based on the interaction information between the account to be pushed and the target content item.

The specific updating method of the interactive behavior determination model is similar to the model training method, and may refer to steps S501 to S507, which is not described herein again. In this implementation manner, the server may update the model parameters of the interactive behavior determination model in real time according to the interactive behavior between the user and the content item, so that the probability of the output of the interactive behavior determination model is closer to the preference of the user, and the effect of the interactive behavior determination is better.

According to the technical scheme provided by the application, the server can determine the second target probability of the target interaction between the account and at least one candidate content item according to the trained interaction behavior determination model and the second content item interaction characteristics of the account and the second content item characteristics of at least one candidate content item, and the interaction behavior determination model is obtained by training according to different types of content item characteristics, so that the model can fully learn the potential mode of the content item characteristics in the training process, the accuracy of the model in determining the interaction behavior is improved, and the server can obtain a better content item pushing effect by adopting the interaction behavior determination model.

FIG. 7 is a block diagram illustrating a training apparatus for an interactive behavior determination model according to an example embodiment. Referring to fig. 7, the apparatus includes a data acquisition unit 701, an input unit 702, an execution unit 703, and a model acquisition unit 704.

The data obtaining unit 701 is configured to perform, in any iterative training process, obtaining sample data, where the sample data includes a first content item interaction feature of a sample account and a first content item feature of at least one content item that has been pushed to the sample account, and the first content item interaction feature is used to represent a historical interaction behavior between the sample account and the content item.

An input unit 702 configured to perform inputting sample data into the interactive behavior determination model.

An execution unit 703 configured to perform the following steps performed by the interactive behavior determination model:

and obtaining a first target characteristic of at least one content item according to the first content item characteristic of at least one content item of the sample account. And outputting a first target probability of target interaction between the sample account and the at least one content item according to the first content item interaction characteristic, the first target characteristic, the type of the first content item characteristic of the at least one content item and the first content item characteristic of the at least one content item.

A model obtaining unit 704 configured to perform obtaining the interactive behavior determination model as a trained interactive behavior determination model in response to the first target probability output by the interactive behavior determination model and the target condition being satisfied between the sample account and the actual interaction situation of the at least one content item.

In a possible implementation manner, the execution unit is configured to perform dimension reduction processing on the first content item feature of the at least one content item of the sample account, respectively, to obtain a first target feature of the at least one content item.

In a possible embodiment, the execution unit is configured to execute setting, for the first target feature, a first sample parameter corresponding to a type of a first content item feature of the at least one content item, the first sample parameter being used to indicate a degree of influence of the first target feature on a model output result. And outputting a first target probability of the target interaction between the sample account and the at least one content item according to the first content item interaction characteristic, the first sample parameter, the first target characteristic and the first content item characteristic of the at least one content item.

In a possible embodiment, the execution unit is configured to perform fusion of the first target feature according to the first sample parameter to obtain the second sample parameter. A first target probability is output based on the second sample parameter, the first content item characteristic of the at least one content item, and the first content item interaction characteristic.

In a possible implementation, the first content item characteristic and the first content item interaction characteristic each include a plurality of sub-characteristics, and the execution unit is configured to perform weighted summation of the sub-characteristics of the first content item characteristic and the sub-characteristics of the first content item interaction characteristic to obtain the first sample fusion characteristic value. And according to the second sample parameter, carrying out weighted summation on the inner products of the sub-features of the at least two first content item features and the inner products of the sub-features of the at least two first content item interaction features to obtain a second sample fusion feature value. And fusing the first sample fusion characteristic value and the second sample fusion characteristic value to obtain a third sample fusion characteristic value. And outputting the first target probability according to the third sample fusion characteristic value.

In a possible implementation manner, the execution unit is configured to perform normalization processing on the third sample fusion feature value to obtain the first target probability.

In one possible embodiment, the apparatus further comprises:

the obtaining unit is configured to perform obtaining of a second content item interaction feature of the account to be pushed, where the second content item interaction feature is used to represent a historical interaction behavior between the account to be pushed and the content item.

An input unit further configured to perform inputting the second content item interaction characteristic and a second content item characteristic of at least one candidate content item of the account to be pushed into the interaction behavior determination model.

And the output unit is configured to output a target probability of a target interactive behavior occurring between the account to be pushed and at least one candidate content item through the interactive behavior determination model.

And the determining unit is configured to determine a target content item to be pushed from at least one candidate content item according to the target probability, wherein the target content item is the candidate content item with the target probability meeting the probability condition.

In a possible implementation, the output unit is configured to perform deriving the second target feature of the at least one candidate content item from the second content item feature of the at least one candidate content item. And setting a second sample parameter corresponding to the type of the second content item characteristic of the at least one candidate content item for the second target characteristic, wherein the second sample parameter is used for expressing the influence degree of the second target characteristic on the output result of the model. And fusing at least two second target characteristics according to the first target parameter to obtain a second target parameter. Outputting the target probability according to the second target parameter, the second content item characteristic of the at least one candidate content item and the second content item interaction characteristic.

In a possible implementation, the output unit is configured to perform a dimension reduction process on the second content item feature of the at least one candidate content item, resulting in a second target feature of the at least one candidate content item.

In one possible implementation, the second content item feature and the second content item interaction feature each include a plurality of sub-features, and the output unit is configured to perform weighted summation of the sub-features of the second content item feature and the sub-features of the second content item interaction feature to obtain the first target fusion feature value. And according to the second target parameter, carrying out weighted summation on the inner products of the sub-features of the at least two second content item features and the inner products of the sub-features of the at least two second content item interaction features to obtain a second target fusion feature value. And fusing the first target fusion characteristic value and the second target fusion characteristic value to obtain a third target fusion characteristic value. And outputting the second target probability according to the third target fusion characteristic value.

In a possible implementation manner, the output unit is configured to perform normalization processing on the third target fusion characteristic value to obtain the second target probability.

In one possible embodiment, the apparatus further comprises:

and the updating unit is configured to update the interactive behavior determination model based on the interactive information between the account to be pushed and the target content item.

With regard to the apparatus in the above-described embodiment, the specific manner in which each module performs the operation has been described in detail in the embodiment related to the method, and will not be elaborated here.

The electronic device provided in the embodiment of the present application may be implemented as a server, and the following describes a structure of the server:

FIG. 8 is a block diagram illustrating a server for a training method for an interactive behavior determination model, according to an example embodiment. Referring to fig. 8, server 800 includes a processing component 801 that further includes one or more processors and memory resources, represented by memory 802, for storing instructions, such as application programs, that are executable by the processing component 801. The application programs stored in memory 802 may include one or more modules that each correspond to a set of instructions. Further, the processing component 801 is configured to execute instructions to perform the functions performed by the server in the training method of the interactive behavior determination model described above.

The server 800 may also include a power component 803 configured to perform power management of the server 800, a wired or wireless network interface 804 configured to connect the server 800 to a network, and an input/output (I/O) interface 806. The server 800 may operate based on an operating system stored in memory 802, such as Windows Server, MacOS XTM, UnixTM, LinuxTM, FreeBSDTM, or the like.

In an exemplary embodiment, a storage medium comprising instructions, such as the memory 804 comprising instructions, executable by the processing component 1001 of the server 1000 to perform the above-described method of training an interactive behavior determination model is also provided. Alternatively, the storage medium may be a non-transitory computer readable storage medium, which may be, for example, a ROM, a Random Access Memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, and the like.

The embodiments of the present disclosure provide a computer program product, and instructions in the computer program product, when executed by a processor of a server, enable the server to execute the training method of the interactive behavior determination model according to the embodiments.

Other embodiments of the present application will be apparent to those skilled in the art from consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention following, in general, the principles of the application and including such departures from the present disclosure as come within known or customary practice within the art to which the invention pertains. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the application being indicated by the following claims.

It will be understood that the present application is not limited to the precise arrangements described above and shown in the drawings and that various modifications and changes may be made without departing from the scope thereof. The scope of the application is limited only by the appended claims.

Claims

1. A method of training an interactive behavior determination model, comprising:

2. The method for training an interactive behavior determination model according to claim 1, wherein the obtaining a first target feature of at least one content item of the sample account according to the first content item feature of the at least one content item comprises:

3. The method for training an interactive behavior determination model according to claim 1, wherein the outputting a first target probability of the sample account having a target interactive behavior with the at least one content item according to the first content item interactive feature, the first target feature, the type of the first content item feature of the at least one content item, and the first content item feature of the at least one content item comprises:

4. The method for training an interactive behavior determination model according to claim 3, wherein the outputting the first target probability of the sample account and the at least one content item for a target interactive behavior according to the first content item interactive feature, the first sample parameter, the first target feature and the first content item feature of the at least one content item comprises:

5. The method of claim 4, wherein the first content item feature and the first content item interaction feature each comprise a plurality of sub-features, and wherein outputting the first target probability based on the second sample parameter, the first content item feature of the at least one content item, and the first content item interaction feature comprises:

6. The method for training an interactive behavior determination model according to claim 5, wherein the outputting the first target probability according to the third sample fusion feature value comprises:

7. An apparatus for training an interactive behavior determination model, comprising:

8. An electronic device, comprising:

a processor;

a memory for storing the processor executable program code;

wherein the processor is configured to execute the program code to implement the method of training of an interactive behaviour determination model according to any of claims 1 to 6.

9. A storage medium, program code in the storage medium, when executed by a processor of an electronic device, enables the electronic device to perform a method of training an interactive behavior determination model according to any of claims 1 to 6.

10. A computer program product having one or more program codes stored thereon, which when executed by a processor of an electronic device, enables the electronic device to perform the method of training an interactive behavior determination model according to any of claims 1 to 6.