CN112699309A

CN112699309A - Resource recommendation method, device, readable medium and equipment

Info

Publication number: CN112699309A
Application number: CN202110305363.6A
Authority: CN
Inventors: 廖忠儒; 王建龙
Original assignee: Beijing Sohu New Media Information Technology Co Ltd
Current assignee: Beijing Sohu New Media Information Technology Co Ltd
Priority date: 2021-03-23
Filing date: 2021-03-23
Publication date: 2021-04-23

Abstract

The application discloses a resource recommendation method, a device, a readable medium and equipment, wherein the method receives a resource recommendation request of a target user; and inputting the current attribute information, the place information and the historical behavior information of the target user to each historical label into a resource recommendation model to obtain and output the behavior prediction probability of the target user to each historical label. Because the resource recommendation model is obtained by training the neural network model according to the attribute information, the occasion information, the historical behavior information of each historical label of each user and the actual feedback result of each user, when predicting whether a target user clicks the historical label, not only the historical behavior of the user on the historical label is considered, but also the occasion information and the attribute information of the user are considered, and therefore the behavior prediction probability of the target user on each historical label obtained by the resource recommendation model is more accurate compared with the prior art.

Description

Resource recommendation method, device, readable medium and equipment

Technical Field

The present application relates to the field of data processing technologies, and in particular, to a resource recommendation method, apparatus, readable medium, and device.

Background

In the prior art, many software for providing resources such as news and videos for users can continuously push contents in which users are interested to the users through a recommendation system in order to improve the utilization rate of the users. Specifically, the recommendation system collects behavior information of the user, counts n tags with the largest number of clicks of the user, considers the n tags as tags in which the user is most interested, and then recommends resources belonging to the n tags to the user.

In the existing recommendation system, resource recommendation is performed on a user according to the interest degree of the user in a tag. However, in practical situations, the interest level of the user in the tag is not the only factor affecting the click behavior of the user on the tag. For example, according to the existing recommendation system, through the past behavior information of the user, it can be seen that the click rate of the user on the resource of the label of the spring festival is low, and the interest degree is not high, so that the resource under the label of the spring festival is not pushed to the user. However, in practical situations, the probability that the user will click on the resource of the label of the spring festival is very high in the period of the spring festival. Therefore, the current recommendation system only depends on the interest degree of the user in the tag for recommendation, and the success rate of recommending resources for the user is not high.

Disclosure of Invention

Based on the defects of the prior art, the application provides a resource recommendation method, a resource recommendation device, a readable medium and a resource recommendation device, so as to improve the success rate of recommending resources.

To solve the above problems, the following solutions are proposed:

the first aspect of the present application discloses a resource recommendation method, including:

receiving a resource recommendation request of a target user;

inputting the current attribute information, the current occasion information and the historical behavior information of the target user to each historical label into a resource recommendation model to obtain and output the behavior prediction probability of the target user to each historical label; the behavior prediction probability of the target user to the history label is the prediction probability of the resource recommendation model to click the history label by the target user; the resource recommendation model is obtained by training a neural network model by attribute information of a plurality of users, occasion information of the users, historical behavior information of each historical label of each user and actual feedback results of each historical label of each user; the actual feedback result of the user on the history label is used for explaining whether the user actually clicks the resource under the history label recommended to the user on the occasion referred by the occasion information;

selecting n history labels with the highest behavior prediction probability from the behavior prediction probability of the target user for each history label as candidate recommendation history labels of the target user; the resources under the candidate recommendation history labels of the target users are the resources which are recommended to the target users in a candidate mode; n is a positive integer.

Optionally, in the resource recommendation method, the information of the current situation where the target user is located includes: the current time point information of the target user and the current environment information of the target user.

Optionally, in the resource recommendation method, the creating method of the resource recommendation model includes:

constructing a user information set; wherein the user information set comprises: attribute information of a plurality of users, occasion information of the users, historical behavior information of each historical label of each user and actual feedback results of each historical label of each user;

for each user in the user information set, inputting attribute information of the user in the user information set, place information of the user in the user information set and historical behavior information of the target user on each historical label into a neural network model, and obtaining and outputting a behavior prediction probability of the user on each historical label by the neural network model;

and continuously adjusting the weight of the neural network model according to the error between the behavior prediction probability of each historical label and the actual feedback result of each user until the error between the behavior prediction probability of each historical label and the actual feedback result of each user output by the adjusted neural network model meets a preset convergence condition, and determining the adjusted neural network model as a resource recommendation model.

Optionally, in the resource recommendation method, the constructing a user information set includes:

at the initial moment of the current model establishing period, establishing a user information set of the current model establishing period by utilizing the attribute information of each user, the place information of each user, the historical behavior information of each historical label and the actual feedback result of each historical label in the previous model establishing period; wherein, the user information set of the current model creation cycle includes: attribute information of each user in a previous model establishing period, occasion information, historical behavior information of each historical label and actual feedback results of each historical label;

wherein, for each user in the user information set, inputting the attribute information of the user in the user information set, the place information where the user is located, and the historical behavior information of the target user for each historical label into a neural network model, and obtaining and outputting the behavior prediction probability of the user for each historical label by the neural network model, the method comprises the following steps:

for each user in the user information set of the current model creation period, inputting the attribute information of the user in the user information set, the place information where the user is located, and the historical behavior information of the target user on each historical label into a resource recommendation model created in the last model creation period, and obtaining and outputting the behavior prediction probability of the user on each historical label by the resource recommendation model created in the last model creation period;

the determining, according to an error between the behavior prediction probability of each historical label and the actual feedback result of each user, the weight of the neural network model to be continuously adjusted until the error between the behavior prediction probability of each historical label and the actual feedback result of each user, output by the adjusted neural network model, meets a preset convergence condition, and the adjusted neural network model is determined as a resource recommendation model, including:

and continuously adjusting the weight of the resource recommendation model created in the last model creation period according to the error between the behavior prediction probability of each history label and the actual feedback result of each user until the error between the behavior prediction probability of each history label and the actual feedback result of each user output by the adjusted resource recommendation model meets a preset convergence condition, and determining the adjusted resource recommendation model as the resource recommendation model created in the current model creation period.

Optionally, in the resource recommendation method, before determining that the adjusted neural network model is the resource recommendation model, the method further includes, according to an error between the behavior prediction probability of each history tag and the actual feedback result of each user, continuously adjusting the weight of the neural network model until the error between the behavior prediction probability of each history tag and the actual feedback result of each user, which is output by the adjusted neural network model, meets a preset convergence condition, and determining that the adjusted neural network model is the resource recommendation model:

for each user in the user information set, respectively multiplying an error between the behavior prediction probability of the user for each history label and an actual feedback result by a weight value corresponding to the actual feedback result of each history label to obtain an adjusted error between the behavior prediction probability of the user for each history label and the actual feedback result; the weight value corresponding to the actual feedback result of the history label is set according to the credibility of the actual feedback result of the history label; the credibility of the actual feedback result of the historical label is used for reflecting the credibility of the user behavior reflected by the actual feedback result of the historical label; the weight value corresponding to the actual feedback result of the history label is positively correlated with the credibility of the actual feedback result of the history label;

wherein, the step of continuously adjusting the weight of the neural network model according to the error between the behavior prediction probability of each historical label and the actual feedback result of each user, respectively, until the error between the behavior prediction probability of each historical label and the actual feedback result of each user, respectively, output by the adjusted neural network model, meets a preset convergence condition, and the step of determining the adjusted neural network model as the resource recommendation model comprises the steps of:

and continuously adjusting the weight of the neural network model according to the adjusted error between the behavior prediction probability of each historical label and the actual feedback result of each user until the adjusted error between the behavior prediction probability of each historical label and the actual feedback result of each user output by the adjusted neural network model meets a preset convergence condition, and determining the adjusted neural network model as a resource recommendation model.

Optionally, in the resource recommendation method, after selecting n history tags with the highest behavior prediction probability from the behavior prediction probabilities of the target user for each history tag, as candidate recommended history tags of the target user, the method further includes:

and recommending resources under the candidate recommendation history labels to the target user.

Optionally, in the resource recommendation method, the history behavior information of the target user on the history tag includes: at least one of the click rate of the target user on the history label, the time point of the target user clicking the history label last time, the time point of the target user clicking the history label in a preset history time period, and the continuous non-click times of the target user on the history label.

A second aspect of the present application discloses a resource recommendation apparatus, including:

the receiving unit is used for receiving a resource recommendation request of a target user;

the first prediction unit is used for inputting the current attribute information, the current occasion information and the historical behavior information of the target user to each historical label into a resource recommendation model to obtain and output the behavior prediction probability of the target user to each historical label; the behavior prediction probability of the target user to the history label is the prediction probability of the resource recommendation model to click the history label by the target user; the resource recommendation model is obtained by training a neural network model by attribute information of a plurality of users, occasion information of the users, historical behavior information of each historical label of each user and actual feedback results of each historical label of each user; the actual feedback result of the user on the history label is used for explaining whether the user actually clicks the resource under the history label recommended to the user on the occasion referred by the occasion information;

the first selecting unit is used for selecting n history labels with the highest behavior prediction probability from the behavior prediction probability of the target user to each history label as candidate recommendation history labels of the target user; the resources under the candidate recommendation history labels of the target users are the resources which are recommended to the target users in a candidate mode; n is a positive integer.

Optionally, in the resource recommendation device, the information of the current situation where the target user is located includes: the current time point information of the target user and the current environment information of the target user.

Optionally, in the resource recommendation apparatus, the apparatus further includes:

the building unit is used for building a user information set; wherein the user information set comprises: attribute information of a plurality of users, occasion information of the users, historical behavior information of each historical label of each user and actual feedback results of each historical label of each user;

a second prediction unit, configured to, for each user in the user information set, input attribute information of the user in the user information set, location information of the user, and historical behavior information of the target user for each historical tag into a neural network model, and obtain and output a behavior prediction probability of the user for each historical tag by using the neural network model;

and the adjusting unit is used for continuously adjusting the weight of the neural network model according to the error between the behavior prediction probability of each historical label and the actual feedback result of each user until the error between the behavior prediction probability of each historical label and the actual feedback result of each user, which is output by the adjusted neural network model, meets a preset convergence condition, and determining the adjusted neural network model as the resource recommendation model.

Optionally, in the resource recommendation apparatus, the building unit includes:

the building subunit is used for building and obtaining a user information set of the current model creation cycle by utilizing the attribute information, the occasion information, the historical behavior information of each historical label and the actual feedback result of each historical label in the previous model creation cycle at the initial time of the current model creation cycle; wherein, the user information set of the current model creation cycle includes: attribute information of each user in a previous model establishing period, occasion information, historical behavior information of each historical label and actual feedback results of each historical label;

wherein the second prediction unit comprises:

the first prediction subunit is configured to, for each user in the user information set of the current model creation cycle, input attribute information of the user in the user information set, location information of the user, and historical behavior information of the target user for each historical tag into a resource recommendation model created in a previous model creation cycle, and obtain and output a behavior prediction probability of the user for each historical tag from the resource recommendation model created in the previous model creation cycle;

the adjusting unit includes:

and the first adjusting subunit is configured to continuously adjust the weight of the resource recommendation model created in the previous model creation period according to an error between the behavior prediction probability of each history tag and the actual feedback result of each user, until an error between the behavior prediction probability of each history tag and the actual feedback result of each user, output by the adjusted resource recommendation model, meets a preset convergence condition, and determine the adjusted resource recommendation model as the resource recommendation model created in the current model creation period.

an error adjusting unit, configured to, for each user in the user information set, multiply an error between a behavior prediction probability of the user for each history tag and an actual feedback result by a weight value corresponding to the actual feedback result of each history tag, respectively, to obtain an adjusted error between the behavior prediction probability of the user for each history tag and the actual feedback result; the weight value corresponding to the actual feedback result of the history label is set according to the credibility of the actual feedback result of the history label; the credibility of the actual feedback result of the historical label is used for reflecting the credibility of the user behavior reflected by the actual feedback result of the historical label; the weight value corresponding to the actual feedback result of the history label is positively correlated with the credibility of the actual feedback result of the history label;

wherein the adjusting unit includes:

and the second adjusting subunit is configured to continuously adjust the weight of the neural network model according to an adjusted error between the behavior prediction probability of each history tag and the actual feedback result of each user, until the adjusted error between the behavior prediction probability of each history tag and the actual feedback result of each user, output by the adjusted neural network model, meets a preset convergence condition, and determine the adjusted neural network model as the resource recommendation model.

and the recommending unit is used for recommending the resources under the candidate recommendation history labels to the target user.

Optionally, in the resource recommendation apparatus, the historical behavior information of the target user on the historical tag includes: at least one of the click rate of the target user on the history label, the time point of the target user clicking the history label last time, the time point of the target user clicking the history label in a preset history time period, and the continuous non-click times of the target user on the history label.

A third aspect of the application discloses a computer readable medium having a computer program stored thereon, wherein the program when executed by a processor implements the method as described in any of the first aspects above.

A fourth aspect of the present application discloses a resource recommendation device, including:

one or more processors;

a storage device having one or more programs stored thereon;

the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the method as in any one of the first aspects above.

It can be seen from the above technical solutions that, in the resource recommendation method provided in the embodiments of the present application, by receiving a resource recommendation request of a target user, and then inputting current attribute information, location information, and historical behavior information of the target user for each historical tag into a resource recommendation model, a behavior prediction probability of the target user for each historical tag is obtained and output, since the resource recommendation model is obtained by training a neural network model based on the attribute information of a plurality of users, the location information, the historical behavior information of each user for each historical tag, and an actual feedback result of each user for each historical tag, when predicting whether the target user will click on a historical tag, the resource recommendation model considers not only the historical behavior of the user for the historical tag, but also objective factors such as the location information of the user, therefore, the behavior prediction probability of the target user on each history label obtained by the resource recommendation model is more accurate compared with the prior art, the behavior prediction probability in the application is the history label which is predicted to be clicked by the target user and is used as the candidate recommendation history label of the target user, and the success rate of recommending resources for the target user can be improved.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the provided drawings without creative efforts.

Fig. 1 is a schematic flowchart of a resource recommendation method disclosed in an embodiment of the present application;

FIG. 2 is a flowchart illustrating a method for creating a resource recommendation model according to an embodiment of the present application;

fig. 3 is a schematic structural diagram of a resource recommendation device disclosed in an embodiment of the present application.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

Referring to fig. 1, an embodiment of the present application provides a resource recommendation method, which specifically includes the following steps:

s101, receiving a resource recommendation request of a target user.

The resource recommendation request of the target user is used for requesting to recommend resources to the target user. The resource recommendation request of the target user carries identification information of the target user, such as the number of the target user, the account of the target user, the name of the target user, and the like.

The resource recommendation request of the target user can be triggered by the user to the client operation. For example, when the target user performs a pull-down page refreshing operation on the interface of the resource software, the resource recommendation request of the target user is triggered to be generated, and if the target user clicks to open the resource software, the resource recommendation request of the target user is triggered to be generated.

A resource may refer to a source of information provided by articles, novels, videos, music, news, pictures, etc. to a user.

S102, inputting the current attribute information, the place information and the historical behavior information of the target user to each historical label into a resource recommendation model, and obtaining and outputting the behavior prediction probability of the target user to each historical label.

The behavior prediction probability of the target user to the history label is the prediction probability of the resource recommendation model to the target user to click the history label. The resource recommendation model is obtained by training the neural network model according to the attribute information of a plurality of users, the occasion information of the users, the historical behavior information of each user on each historical label and the actual feedback result of each user on each historical label. And the actual feedback result of the user on the history label is used for explaining whether the resource recommended to the user under the history label is actually clicked or not under the occasion indicated by the presence information.

The history tag is a tag clicked by a target user (i.e., a tag with history behavior information), and since behavior information of different users is different, history tags corresponding to different users may also be different.

The current attribute information of the target user is used for reflecting the characteristics of the target user, so that the resource recommendation model can individually predict the behavior prediction probability of different users on each historical label aiming at different users. The current attribute information of the target user at least comprises identification information of the target user. The identification information of the target user is unique and unique information specific to the target user, and corresponds to the identity of the user. The current attribute information of the target user may further include a device model of the target user, a registration time of the target user as a user of the resource software, and the like. The device model of the target user may indicate an operating system of a device currently used by the target user, and if the history tag that needs to be predicted is related to a history tag related to the operating system or a history tag related to the device model, the device model of the target user may also become a factor that affects a behavior prediction probability of the target user on the history tag. The registration time of the target user can reflect whether the user is a new user or an old user, and the registration time can also be a factor influencing the behavior prediction probability of the history label. It should be noted that, the attribute information of the target user includes, but is not limited to, what is proposed in the embodiments of the present application, and the more detailed the attribute information of the target user input to the resource recommendation model is, the higher the accuracy of the behavior prediction probability of each history label output by the target user is.

The occasion information of the target user can be used for describing the current occasion of the target user, and is mainly reflected by the time and the space of the target user. The situation information of the target user also influences the behavior prediction probability of the target user on each history label. For example, if the target user is at a tourist attraction location at a spring festival time point, then the probability of being clicked by the user to view the resources related to the tourist-related history tags, spring festival-related history tags, will naturally be higher. Therefore, the resource recommendation model can predict the behavior prediction probability of each history label of the target user by using the information of the place where the target user is located, and the prediction probability is more accurate.

Optionally, in a specific embodiment of the present application, the information of the current situation where the target user is located includes: current time point information of the target user and current environment information of the target user. The current time point information of the target user may be a time point when the resource recommendation request of the target user is generated, may be represented by a date, and may also be specific to a specific time. The current environment information of the target user may include information describing the current geographic environment of the user, and may also include current network environment information of the target user. For example, the current environment information of the target user may be software positioning information, a mobile phone network type, and the like. Factors of environment information of the target user can influence the behavior prediction probability of the target user on the history label. For example, if the type of the mobile phone network used by the target user is a mobile data network, the target user may be influenced by the mobile data network, and the probability that the target user clicks on the history tag of the video category may be low.

The historical behavior information of the target user for each historical label can reflect the interest degree of the target user for each historical label. The historical behavior information of the target user on the historical label may include at least one of a click rate of the target user on the historical label, a time point of the target user on the historical label last time, a time point of the target user on the historical label in a preset historical time period, a continuous non-click frequency of the target user on the historical label, and a click frequency of the target user on the historical label. The higher the click rate of the target user on the history label, the closer the time point of the target user for clicking the history label last time is to the current time point, the more frequently the target user clicks the history label in the preset history time period, the fewer continuous non-click times of the target user on the history label, and the more click times of the target user on the history label, the higher the interest degree of the target user on the history label is. The smaller the continuous non-click times of the target user on the history label is, the times that the target user does not click on the resource continuously recommending the history label is indicated. Optionally, the historical behavior information of the target user for each historical tag may also be the historical tag clicked by the target user within a preset historical time period. The preset time period is a time period close to the current time point, and may be, for example, the last week or the last month. The behavior information farther from the current time point cannot accurately reflect the interest degree of the target user in the history tag.

Optionally, in a specific embodiment of the present application, the current attribute information and the location information of the target user may be carried in the resource recommendation request of the target user. That is, the client where the target user is located carries the attribute information and the location information of the current target user into the resource recommendation request and sends the resource recommendation request to the device implementing the embodiment shown in fig. 1, so that the device can input the current attribute information and the location information of the target user into the resource recommendation model. And the historical behavior information of the target user on each historical label can be collected in real time through the target user client and stored in the database. After receiving a resource recommendation request of a target user, the device implementing the embodiment shown in fig. 1 extracts historical behavior information of each history tag from the database, and then inputs the historical behavior information into the resource recommendation model. Optionally, the attribute information of the target user, the place information where the target user is located, and the historical behavior information of the target user for each historical tag may also be collected in real time by the client of the target user and stored in the database, and the device implementing the embodiment shown in fig. 1 extracts the attribute information of the target user, the place information where the target user is located, and the historical behavior information of the target user for each historical tag from the database after receiving the resource recommendation request of the target user, and inputs the extracted information into the resource recommendation model.

In the prior art, when predicting the behavior of each history tag of a target user, only the interest degree of the target user in the history tag is considered. Specifically, in the prior art, a target user sorts history tags clicked by the target user according to the number of clicks, and then selects a plurality of history tags with the largest number of clicks as candidate recommended history tags, so as to recommend resources under the candidate recommended history tags for the target user. However, in practical applications, the behavior of the target user on the history tag is not only influenced by the interest degree of the target user on the history tag. For example, according to the existing recommendation system, through the past behavior information of the user, it can be seen that the user has a low click rate on the resource of the history label of the spring festival and has low interest degree, so that the resource under the history label of the spring festival is not pushed to the user. However, in practical situations, the probability that the user will click on the resource of the history label of the spring festival in the period of the spring festival is very high. In the prior art, when the target user predicts the behavior of the historical label, the predicted behavior is not high in accuracy due to the fact that the considered factor is too single, and the success rate of recommending resources for the user is low.

In the application, when predicting whether the target user clicks the history label, the resource recommendation model not only considers the history behavior of the user on the history label, but also considers objective factors such as the occasion information of the user, so that the behavior prediction probability of the target user on each history label obtained by the resource recommendation model is more accurate compared with the prior art, the behavior prediction probability in the application is the history label predicted to be clicked by the target user and is used as the candidate recommendation history label of the target user, and the success rate of recommending resources for the target user can be improved.

Optionally, referring to fig. 2, in an embodiment of the present application, a method for creating a resource recommendation model includes:

s201, constructing a user information set, wherein the user information set comprises: the method comprises the following steps of attribute information of a plurality of users, occasion information of the users, historical behavior information of each user on each historical label respectively, and actual feedback results of each user on each historical label respectively.

Specifically, in the process of constructing the user information set, a plurality of historical time points may be set, and for each historical time point, the attribute information of each user, the location information of each user, the historical behavior information of each user for each historical tag, and the actual feedback result of each user for each historical tag at the historical time point are collected. Or collecting the attribute information of each user, the place information of each user, the historical behavior information of each historical label of each user and the actual feedback result of each historical label of each user in real time. And the actual feedback result of the user on the history label is used for explaining whether the resource recommended to the user under the history label is actually clicked or not under the occasion indicated by the presence information. For example, at a certain historical time point, a certain user clicks on a resource under a certain historical label presented to the user, and then the actual feedback result of the user on the historical label to which the resource belongs is a click. And if the user does not click on the resource under the history label displayed to the user at the history time point, the actual feedback result of the user on the history label is not clicked.

Optionally, in a specific embodiment of the present application, in order to make a model trained by a user information set more accurate, the user information set may be constructed in a periodic construction manner. Specifically, one embodiment of performing step S201 includes:

and at the initial moment of the current model creation period, constructing and obtaining a user information set of the current model creation period by utilizing the attribute information, the occasion information, the historical behavior information of each historical label and the actual feedback result of each historical label in the previous model creation period.

The user information set of the current model creation period comprises the following steps: attribute information of each user in the last model creation period, occasion information, historical behavior information of each historical label and actual feedback results of each historical label.

Specifically, a user information set for each model creation period is created for each model creation period. Taking the current model creation cycle as an example, at the initial time of the current model creation cycle, the user information set of the current model creation cycle is constructed by using the attribute information of each user, the location information, the historical behavior information of each historical label and the actual feedback result of each historical label, which are collected in the process of the previous model creation cycle. And the user information set of the current model creation period is used for training the resource recommendation model of the current model creation period.

Due to the attribute information of each user, the occasion information, the historical behavior information of each historical label and the actual feedback result of each historical label, the actual feedback result of each historical label changes at different historical time points. If the model is trained using only the attribute information of each user in the earlier previous historical time period, the occasion information, the historical behavior information for each historical label respectively and the actual feedback result for each historical label respectively, the attribute information, the occasion information, the historical behavior information of each user in the earlier historical time period, the historical behavior information of each historical label and the actual feedback result of each historical label can only reflect the factors influencing the behavior of the user in the earlier historical time period and can not be taken as the factors influencing the behavior of the user at present, user behaviors predicted by a model built by using only the attribute information, the occasion information, the historical behavior information of each historical label and the actual feedback result of each historical label in the earlier historical time period are inaccurate.

Therefore, in order to improve the prediction accuracy of the resource recommendation model, for each model creation cycle, the information (i.e., the latest historical information) of the model creation cycle immediately preceding the model creation cycle is used to construct the user information set of the current model creation cycle, and the user information set of the current model creation cycle is used to train and construct the resource recommendation model.

It should be noted that the history label to which the resource presented to each user belongs is labeled in advance, and the history label to which the resource belongs may be a keyword, a category of the resource, an author, and the like. A resource may have a number of different history tags. Specifically, a resource may have history labels under multiple history label domains. The history tag field refers to a history tag type, such as an author name history tag field, a keyword history tag field, a resource type history tag field, and the like. And a resource may have multiple history tags under a history tag field. For example, if the resource has multiple keywords in the keyword history tag field, then there will be multiple history tags in the keyword history tag field. Optionally, the history tag of each resource may be marked in advance by marking, for each resource, the most important history tag of the resource in each history tag domain in advance.

S202, aiming at each user in the user information set, inputting the attribute information, the place information and the historical behavior information of the target user to each historical label into a neural network model, and obtaining and outputting the behavior prediction probability of the user to each historical label by the neural network model.

And respectively inputting the attribute information of each user in the user information set, the place information where the user is located and the historical behavior information of each historical label of the target user into a neural network model, and obtaining and outputting the behavior prediction probability of each user to each historical label by the neural network model. The predicted probability of the user for the behavior of each history label is the predicted probability of the user clicking the history label under the condition that the condition information input into the neural network model refers to.

Optionally, in a specific embodiment of the present application, if, at an initial time of a current model creation cycle when step S201 is executed, a user information set of the current model creation cycle is constructed and obtained by using attribute information of each user in a previous model creation cycle, location information, historical behavior information of each historical tag, and an actual feedback result of each historical tag, then step S202 is executed, which includes:

and for each user in the user information set of the current model creation period, inputting the attribute information, the place information and the historical behavior information of each historical label of the target user into the resource recommendation model created in the previous model creation period, and obtaining and outputting the behavior prediction probability of each historical label of the user according to the resource recommendation model created in the previous model creation period.

And each model creating period creates a new resource recommendation model, and the previously trained and created resource recommendation model is trained by using the user information set of the current model creating period.

S203, continuously adjusting the weight of the neural network model according to the error between the behavior prediction probability of each historical label and the actual feedback result of each user until the error between the behavior prediction probability of each historical label and the actual feedback result of each user output by the adjusted neural network model meets a preset convergence condition, and determining the adjusted neural network model as a resource recommendation model.

The predicted probability of the user's behavior on the historical label may have an error with the actual feedback result, for example, the predicted probability of the user's behavior on the historical label is 20%, but the actual feedback result is a click, so the predicted probability should be close to 100%. And the behavior prediction probability of the historical user on the historical label is 50%, but the actual feedback result is no click, namely the prediction probability is accurate only when the prediction probability is close to 0. Therefore, the error exists between the behavior prediction probability of the user on the history label and the actual feedback result, the weight of the neural network model needs to be adjusted continuously according to the error between the behavior prediction probability of each user on each history label and the actual feedback result until the error between the behavior prediction probability of each user on each history label and the actual feedback result output by the adjusted neural network model meets the preset convergence condition, the accuracy of the behavior prediction probability of the user on the history label predicted by the adjusted neural network model at the moment is considered to meet the requirement, and therefore the adjusted neural network model is determined as the resource recommendation model.

Optionally, in a specific embodiment of the present application, if, at an initial time of a current model creation cycle when step S201 is executed, a user information set of the current model creation cycle is constructed and obtained by using attribute information of each user in a previous model creation cycle, location information, historical behavior information of each historical tag, and an actual feedback result of each historical tag, then step S203 is executed, where the method includes:

and continuously adjusting the weight of the resource recommendation model obtained by establishing the previous model establishing period according to the error between the behavior prediction probability of each historical label and the actual feedback result of each user until the error between the behavior prediction probability of each historical label and the actual feedback result of each user output by the adjusted resource recommendation model meets the preset convergence condition, and determining the adjusted resource recommendation model as the resource recommendation model obtained by establishing the current model establishing period.

The error between the behavior prediction probability of each user for each history label and the actual feedback result refers to the error between the behavior prediction probability of each user for each history label and the actual feedback result in the user information set of the current model creation period. In the embodiment of the application, each model creation cycle uses the newly-constructed user information set to re-create a new resource recommendation model, so that the newly-created resource recommendation model is applied to the embodiment shown in fig. 1, and the behavior prediction probability of the user on the history label can be more accurately output.

Optionally, in a specific embodiment of the present application, an implementation manner of executing step S202 includes:

and for each user in the user information set, multiplying the error between the behavior prediction probability of the user to each history label and the actual feedback result by the weight value corresponding to the actual feedback result of each history label to obtain the adjusted error between the behavior prediction probability of the user to each history label and the actual feedback result.

And the weight value corresponding to the actual feedback result of the history label is set according to the credibility of the actual feedback result of the history label. The credibility of the actual feedback result of the historical label is used for reflecting the credibility of the user behavior reflected by the actual feedback result of the historical label, and the weight value corresponding to the actual feedback result of the historical label is positively correlated with the credibility of the actual feedback result of the historical label. When step S203 is executed, the method includes: and continuously adjusting the weight of the neural network model according to the adjusted error between the behavior prediction probability of each historical label and the actual feedback result of each user until the adjusted error between the behavior prediction probability of each historical label and the actual feedback result of each user output by the adjusted neural network model meets a preset convergence condition, and determining the adjusted neural network model as a resource recommendation model.

For example, if a certain user not only clicks a resource under a certain history tag, but also likes and collects the resource under the history tag, it indicates that the user has a high credibility of the clicking behavior of the history tag, and the clicking behavior of the history tag is not a will that the user really wants to click, so that the user can be given a relatively large weight value to the actual feedback result of the history tag. For some resources on the popular ranking list, the user clicks the history label to which the resource belongs, but no other praise or collection behaviors exist, which indicates that the clicking behavior of the user on the history label is only possibly caused by the influence of popular recommendation, and the user does not really want to view and browse the resources under the history label, so that the user can be given a smaller weight value for actual feedback results of the history label.

Aiming at each user in the user information set, multiplying the error between the behavior prediction probability of the user to each history label and the actual feedback result by the weight value corresponding to the actual feedback result of each history label respectively to obtain the adjusted error between the behavior prediction probability of the user to each history label and the actual feedback result, wherein if the weight value is higher, the adjusted error is larger than the original error, and at the moment, the weight value of the neural network model is adjusted according to the adjusted error between the behavior prediction probability of the user to the history label and the actual feedback result, so that the weight corresponding to the actual feedback result of the user to the history label is larger, and the accuracy of the behavior prediction probability of the history label output by the trained resource recommendation model is higher.

S103, selecting n history labels with the highest behavior prediction probability from the behavior prediction probabilities of the target user on each history label as candidate recommended history labels of the target user.

And the resources under the candidate recommendation history labels of the target users are the resources which are recommended to the target users as candidates. n is a positive integer. The historical label of which the behavior prediction probability is that the target user is predicted to click is the historical label of which the target user is predicted to click under the situation indicated by the current situation information. And selecting n history labels with the highest behavior prediction probability from the behavior prediction probabilities of the target user for each history label, and considering the history labels as the history labels which are most likely to be clicked by the user, so that the history labels are used as candidate recommendation history labels of the target user.

It should be noted that the candidate recommendation history tags are only candidate history tags for recommending to the target user, and in practical application, resources under all candidate recommendation history tags may be pushed to the target user, or resources under several candidate recommendation history tags may be selectively selected and recommended to the target user.

Because the candidate recommendation history tags are history tags predicted by the resource recommendation model and having the highest probability that the target user can click at present, resources under all the candidate recommendation history tags can be pushed to the target user, and resources under several candidate recommendation history tags can be selectively selected and recommended to the target user.

Optionally, in a specific embodiment of the present application, after the step S103 is executed, the method further includes:

After the candidate recommendation history labels of the target user are determined, resources under the candidate recommendation history labels can be retrieved from a resource database, and then the resource recommendations under the candidate recommendation history labels are issued to the client side where the target user is located. And selecting resources with higher heat to recommend to the target user selectively according to the resources under the candidate recommendation history labels.

The resource recommendation method provided by the embodiment of the application obtains and outputs the behavior prediction probability of the target user for each history label by receiving the resource recommendation request of the target user and inputting the current attribute information, the place information and the historical behavior information of the target user for each history label into the resource recommendation model, and the resource recommendation model is obtained by training the neural network model by the attribute information, the place information, the historical behavior information of each user for each history label and the actual feedback result of each user for each history label, so that when predicting whether the target user clicks the history label, the resource recommendation model not only considers the historical behavior of the user for the history label, but also considers the place information, the place information and the history label of the user, Due to the factors such as the attribute information of the user, the behavior prediction probability of the target user on each history label obtained by the resource recommendation model is more accurate compared with the prior art, the behavior prediction probability in the application is the history label which is predicted to be clicked by the target user and is used as the candidate recommended history label of the target user, and the success rate of recommending the resource for the target user can be improved.

Referring to fig. 3, based on the resource recommendation method provided in the embodiment of the present application, the embodiment of the present application correspondingly discloses a resource recommendation device, which includes: a receiving unit 301, a first prediction unit 302, and a first selection unit 303.

A receiving unit 301, configured to receive a resource recommendation request of a target user.

The first prediction unit 302 is configured to input the current attribute information, the current location information, and the historical behavior information of the target user for each history tag into the resource recommendation model, and obtain and output a behavior prediction probability of the target user for each history tag. The behavior prediction probability of the target user on the historical label is a prediction result of whether the target user clicks the historical label or not, which is predicted by the resource recommendation model. The resource recommendation model is obtained by training the neural network model according to the attribute information of a plurality of users, the occasion information of the users, the historical behavior information of each user on each historical label and the actual feedback result of each user on each historical label. And the actual feedback result of the user on the history label is used for explaining whether the resource recommended to the user under the history label is actually clicked or not under the occasion indicated by the presence information.

Optionally, in a specific embodiment of the present application, the information of the current situation where the target user is located includes: current time point information of the target user and current environment information of the target user.

A first selecting unit 303, configured to select, from the behavior prediction probabilities of each history tag of the target user, n history tags with the highest behavior prediction probabilities as candidate recommended history tags of the target user. And the resources under the candidate recommendation history labels of the target users are the resources which are recommended to the target users in a candidate mode, and n is a positive integer.

Optionally, in a specific embodiment of the present application, the method further includes: the device comprises a construction unit, a second prediction unit and an adjustment unit.

And the construction unit is used for constructing the user information set. Wherein, the user information set includes: the method comprises the following steps of attribute information of a plurality of users, occasion information of the users, historical behavior information of each user on each historical label respectively, and actual feedback results of each user on each historical label respectively.

Optionally, in a specific embodiment of the present application, the building unit includes:

and the construction subunit is used for constructing and obtaining the user information set of the current model creation cycle by utilizing the attribute information, the occasion information, the historical behavior information of each historical label and the actual feedback result of each historical label in the previous model creation cycle at the initial time of the current model creation cycle. The user information set of the current model creation period comprises the following steps: attribute information of each user in the last model creation period, occasion information, historical behavior information of each historical label and actual feedback results of each historical label. A second prediction unit comprising: and the first prediction subunit is used for inputting the attribute information, the place information and the historical behavior information of the target user on each historical label of the user information set into the resource recommendation model created in the last model creation period aiming at each user in the user information set of the current model creation period, and obtaining and outputting the behavior prediction probability of each historical label of the user according to the resource recommendation model created in the last model creation period. An adjustment unit comprising: and the first adjusting subunit is used for continuously adjusting the weight of the resource recommendation model created in the previous model creation period according to the error between the behavior prediction probability of each history label and the actual feedback result of each user until the error between the behavior prediction probability of each history label and the actual feedback result of each user output by the adjusted resource recommendation model meets a preset convergence condition, and determining the adjusted resource recommendation model as the resource recommendation model created in the current model creation period.

And the second prediction unit is used for inputting the attribute information of the users in the user information set, the place information of the users in the user information set and the historical behavior information of the target user to each historical label into the neural network model, and obtaining and outputting the behavior prediction probability of the user to each historical label by the neural network model.

And the adjusting unit is used for continuously adjusting the weight of the neural network model according to the error between the behavior prediction probability of each historical label and the actual feedback result of each user until the error between the behavior prediction probability of each historical label and the actual feedback result of each user output by the adjusted neural network model meets a preset convergence condition, and determining the adjusted neural network model as the resource recommendation model.

Optionally, in a specific embodiment of the present application, the method further includes:

and the error adjusting unit is used for multiplying the error between the behavior prediction probability of the user on each history label and the actual feedback result by the weight value corresponding to the actual feedback result of each history label respectively aiming at each user in the user information set to obtain the adjusted error between the behavior prediction probability of the user on each history label and the actual feedback result. The reliability of the actual feedback result of the history label is used for reflecting the credibility of the user behavior reflected by the actual feedback result of the history label, and the weight value corresponding to the actual feedback result of the history label is positively correlated with the reliability of the actual feedback result of the history label. An adjustment unit comprising: and the second adjusting subunit is used for continuously adjusting the weight of the neural network model according to the adjusted error between the behavior prediction probability of each historical label and the actual feedback result of each user until the adjusted error between the behavior prediction probability of each historical label and the actual feedback result of each user output by the adjusted neural network model meets a preset convergence condition, and determining the adjusted neural network model as the resource recommendation model.

Optionally, in a specific embodiment of the present application, the method further includes: and the recommending unit is used for recommending the resources under the candidate recommendation history labels to the target user.

Optionally, in a specific embodiment of the present application, the historical behavior information of the target user on the historical tag includes: the target user clicks the historical label, the last time point of the target user clicking the historical label, the time point of the target user clicking the historical label in a preset historical time period, and the continuous non-clicking times of the target user on the historical label.

The specific principle and the execution process of each unit in the resource recommendation device disclosed in the embodiment of the present application are the same as those of the resource recommendation method disclosed in the embodiment of the present application, and reference may be made to corresponding parts in the resource recommendation method disclosed in the embodiment of the present application, which are not described herein again.

The resource recommendation device provided by the embodiment of the application receives a resource recommendation request of a target user through the receiving unit 301, then the first prediction unit 302 inputs the current attribute information, the place information and the historical behavior information of the target user for each historical label into the resource recommendation model to obtain and output the behavior prediction probability of the target user for each historical label, and since the resource recommendation model is obtained by training the neural network model by the attribute information, the place information, the historical behavior information of each historical label of each user and the actual feedback result of each historical label of each user, when predicting whether the target user clicks the historical label, the resource recommendation model considers not only the historical behavior of the user for the historical label, but also the objective factors such as the place information of the user, therefore, the behavior prediction probability of the target user on each history label obtained by the resource recommendation model is more accurate compared with the prior art, the behavior prediction probability in the application is the history label which is predicted to be clicked by the target user and is used as the candidate recommendation history label of the target user, and the success rate of recommending resources for the target user can be improved.

The embodiment of the application discloses a computer readable medium, on which a computer program is stored, wherein the program is executed by a processor to implement the resource recommendation method according to any one of the above embodiments.

The embodiment of the application discloses a resource recommendation device, which comprises: one or more processors, a storage device, on which one or more programs are stored. The one or more programs, when executed by the one or more processors, cause the one or more processors to implement the resource recommendation method as in any of the embodiments described above.

The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

It is further noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other identical elements in a process, method, article, or apparatus that comprises the element.

Claims

1. A resource recommendation method, comprising:

receiving a resource recommendation request of a target user;

2. The method of claim 1, wherein the information of the current situation where the target user is located comprises: the current time point information of the target user and the current environment information of the target user.

3. The method of claim 1, wherein the resource recommendation model is created by a method comprising:

4. The method of claim 3, wherein the constructing the user information set comprises:

5. The method according to claim 3, wherein before determining the adjusted neural network model as the resource recommendation model, the method further includes, according to an error between the behavior prediction probability of each history tag and the actual feedback result of each user, continuously adjusting the weight of the neural network model until the error between the behavior prediction probability of each history tag and the actual feedback result of each user, which is output by the adjusted neural network model, meets a preset convergence condition:

6. The method according to claim 1, wherein after selecting the n history labels with the highest behavior prediction probability from the behavior prediction probabilities of the target user for each history label as the candidate recommended history labels of the target user, the method further comprises:

7. The method of claim 1, wherein the historical behavior information of the target user on the historical tags comprises: at least one of the click rate of the target user on the history label, the time point of the target user clicking the history label last time, the time point of the target user clicking the history label in a preset history time period, and the continuous non-click times of the target user on the history label.

8. A resource recommendation device, comprising:

9. A computer-readable medium, on which a computer program is stored, wherein the program, when executed by a processor, implements the method of any one of claims 1 to 7.

10. A resource recommendation device, comprising:

one or more processors;

a storage device having one or more programs stored thereon;

the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the method of any of claims 1-7.