CN115687778A

CN115687778A - Resource recommendation method, device, equipment and storage medium

Info

Publication number: CN115687778A
Application number: CN202211413383.6A
Authority: CN
Inventors: 冯浩源; 刘鑫
Original assignee: Baidu com Times Technology Beijing Co Ltd
Current assignee: Baidu com Times Technology Beijing Co Ltd
Priority date: 2022-11-11
Filing date: 2022-11-11
Publication date: 2023-02-03

Abstract

The disclosure provides a resource recommendation method, a resource recommendation device, a resource recommendation equipment and a storage medium, and relates to the artificial intelligence technology, in particular to the field of big data. The specific implementation scheme is as follows: when resource recommendation is carried out on low-activity users of resources of a target type, a plurality of recall words are determined from historical access records of users of the whole network on resources of at least one other type; aiming at each recall word, acquiring a user set which belongs to the high activity of the target type resource and has the click times of at least one other type resource reaching the preset times; and then recommending resources to the low-activity users of the resources of the target type according to the resource list of the target type corresponding to the recall word accessed by the representative seed user in each user set. By the aid of the scheme, group characteristics are combined when resource recommendation is carried out, signals of other types of resources are migrated, resource recommendation accuracy is improved, recommendation effect is improved, and permeability of low-activity users is improved.

Description

Resource recommendation method, device, equipment and storage medium

Technical Field

The present disclosure relates to the field of big data technology in artificial intelligence, and in particular, to a resource recommendation method, apparatus, device, and storage medium.

Background

With the development of the internet of things technology, massive data in the internet world causes great burden to users, so personalized recommendation technology is gradually popularized in various fields, for example: news recommendations, business recommendations, entertainment recommendations, learning recommendations, life recommendations, decision support, and the like.

In the prior art, a resource recommendation method for a user mainly calls back a resource by a display recall method or an implicit recall method. The display recall emphasizes resource search recalls based on content, either directly based on content searched by the user or content browsed. And (4) implicit recall is more biased to feature engineering, a user is represented by a vector in a space, namely the user vector, and resource recall is further performed from a network according to the user vector.

However, the current recall method based on the content searched by the user or the user vector has a poor effect on improving the penetration rate of the low-activity user.

Disclosure of Invention

The disclosure provides a resource recommendation method, device, equipment and storage medium.

According to a first aspect of the present disclosure, there is provided a resource recommendation method, including:

determining a plurality of recall words according to historical access records of the whole network users to resources of at least one type different from the target type;

acquiring a user set corresponding to each recall word, wherein each user in the user set is a high-activity user of the resource of the target type under the recall word, and the number of clicks on the resource of at least one other type reaches a preset number;

for each user set, determining representative seed users in each user set, wherein the representative seed users comprise a preset number of users with the maximum similarity between user vectors and central vectors in the user set;

aiming at a user set corresponding to each recall word, acquiring a resource list of a representative seed user in the user set, which accesses a target type corresponding to the recall word;

and recommending resources to the low-activity users of the resources of the target type according to the resource list corresponding to each recall word.

According to a second aspect of the present disclosure, there is provided a resource recommendation apparatus including:

the first processing unit is used for determining a plurality of recall words according to the historical access records of the whole network users to resources of at least one type different from the target type;

the second processing unit is used for acquiring a user set corresponding to each recall word, wherein each user in the user set is a high-activity user of the resource of the target type under the recall word, and the number of clicks on the resource of at least one other type reaches a preset number;

the third processing unit is used for determining representative seed users in each user set aiming at each user set, wherein the representative seed users comprise a preset number of users with the maximum similarity between user vectors and a central vector in the user set;

the fourth processing unit is used for acquiring a resource list of a representative seed user in the user set, which accesses the target type corresponding to the recall word, aiming at the user set corresponding to each recall word;

and the recommending unit is used for recommending the resources to the low-activity users of the resources of the target type according to the resource list corresponding to each recall word.

According to a third aspect of the present disclosure, there is provided an electronic device comprising:

at least one processor; and

a memory communicatively coupled to the at least one processor; wherein, the first and the second end of the pipe are connected with each other,

the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of the first aspect.

According to a fourth aspect of the present disclosure, there is provided a non-transitory computer readable storage medium having stored thereon computer instructions for causing the computer to perform the method of the first aspect.

According to a fifth aspect of the present disclosure, there is provided a computer program product comprising: a computer program, stored in a readable storage medium, from which at least one processor of an electronic device can read the computer program, execution of the computer program by the at least one processor causing the electronic device to perform the method of the first aspect.

According to the technical scheme, when resource recommendation is performed on low-activity users of resources of a target type, firstly, a plurality of recall words are determined from historical access records of all-network users to resources of at least one other type; aiming at each recall word, acquiring a user set which belongs to the high-activity of the target type resource and has the click times of at least one other type resource reaching the preset times; and then recommending resources to the low-activity users of the resources of the target type according to the resource list of the target type corresponding to the recall word accessed by the representative seed user in each user set. By the scheme, the group characteristics are combined when resource recommendation is carried out, signals of other types of resources are migrated, the resource recommendation accuracy is improved, the recommendation effect is improved, and the permeability to low-activity users is improved.

It should be understood that the statements in this section do not necessarily identify key or critical features of the embodiments of the present disclosure, nor do they limit the scope of the present disclosure. Other features of the present disclosure will become apparent from the following description.

Drawings

The drawings are included to provide a better understanding of the present solution and are not to be construed as limiting the present disclosure. Wherein:

fig. 1 is a schematic flowchart of a resource recommendation method according to a first embodiment of the present disclosure;

fig. 2 is a flowchart illustrating a resource recommendation method according to a second embodiment of the disclosure;

fig. 3 is a schematic flowchart of a resource recommendation method according to a third embodiment of the disclosure;

fig. 4 is a schematic flowchart of a resource recommendation method according to a fourth embodiment of the disclosure;

fig. 5 is a flowchart illustrating a resource recommendation method according to a fifth embodiment of the disclosure;

fig. 6 is a flowchart illustrating a resource recommendation method according to a sixth embodiment of the disclosure;

FIG. 7 is a schematic structural diagram of a resource recommendation device provided in an embodiment of the present disclosure;

FIG. 8 shows a schematic block diagram of an example electronic device used to implement embodiments of the present disclosure.

Detailed Description

Exemplary embodiments of the present disclosure are described below with reference to the accompanying drawings, in which various details of the embodiments of the disclosure are included to assist understanding, and which are to be considered as merely exemplary. Accordingly, those of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the present disclosure. Also, descriptions of well-known functions and constructions are omitted in the following description for clarity and conciseness.

With the development of the internet and big data technology, massive information brings huge burden to users, and helping users to recommend resources becomes an important research topic to deal with the problems of information explosion and information disaster, and is the core of the current information flow recommendation application. In the current technical scheme, when resource recommendation needs to be performed on a user, a common mode can be roughly classified into an explicit recall mode or an implicit recall mode, wherein the former emphasizes a recommendation mode based on content and has strong interpretability; the latter favors feature engineering, representing users by spatial vectors, and recalling resources for users based on the vectors.

The resource recommendation method based on single explicit recall or implicit recall has common comprehensive recommendation effect in specific application scenes, particularly has low permeability to low-activity users, and a point comprehensive resource recall method combining various types of resource information and integrating the explicit recall mode and the implicit recall mode is lacked at present.

Aiming at the technical problems, the technical idea process of the present disclosure is as follows: in the process of researching resource recommendation schemes, the inventor finds that for low-activity users, the characteristics which can be provided by the historical behavior data of the low-activity users are preferred, but for the same interested point, the clicking and browsing history of the high-activity users on the resources can provide more information, the users can access one type of resources and click and browse other types of resources, and in the process of recalling one type of resources, the requirements of the users can be accurately positioned by combining the access characteristics of the other types of resources, so that resource recommendation is performed. Therefore, the inventor considers a comprehensive resource recommendation scheme which can be used for gathering all user vectors, historical behaviors and real-time characteristics of users, performing transfer learning on signals of different types of resources and combining the advantages of explicit and implicit methods.

Based on the technical conception process, the present disclosure provides a resource recommendation method, which is applied to a background server of a public resource recommendation system, and when a certain type of resource recommendation is performed on a low-activity user, recall words for resource recall are determined from historical access records of a network-wide user on other types of resources, based on the recall words, a user group, that is, a user set, corresponding to each recall word is determined, and based on characteristics and access history of a high-activity user on the type of resource in the user group, recommendation of the type of resource is performed on the low-activity user, that is, group characteristics are combined during resource recommendation, and signals of other types of resources are migrated and learned, so that accuracy when resource recommendation is performed on the low-activity user is improved, and thus a recommendation effect is improved.

The whole resource recommendation scheme of the embodiment of the present disclosure may be applied to a server, which may be a server of certain entertainment software, a server of an information recommendation system, a server of a search system, a server of an information flow system, and the like, and the present disclosure is not limited to this scheme. For example, if interaction with a user is involved, a device at the user end may also be included.

It can be understood that, in practical applications, the resource recommendation scheme of the present disclosure may further include other devices, for example, a storage device, and the like, which may be specifically adjusted according to actual needs, and the present disclosure does not limit the resource recommendation scheme. The embodiment of the present disclosure also does not limit the actual forms of various devices included in the application scenarios, and also does not limit the interaction modes between the devices, and in the specific application of the scheme, the setting can be performed according to the actual requirements.

It should be understood that in this disclosure, a low-activity user refers to a user of low activity, also referred to as a low-activity user; a high-activity user refers to a user with high activity, also referred to as a high-activity user.

The following describes the technical solutions of the present disclosure and how to solve the above technical problems in specific embodiments. The following several specific embodiments may be combined with each other, and details of the same or similar concepts or processes may not be repeated in some embodiments. Embodiments of the present disclosure will be described below with reference to the accompanying drawings.

The following describes a specific implementation of the resource recommendation method provided by the present disclosure.

Fig. 1 is a flowchart illustrating a resource recommendation method according to a first embodiment of the disclosure. As shown in fig. 1, the resource recommendation method specifically includes the following steps:

s101: and determining a plurality of recall words according to the historical access records of the whole network users to the resources of at least one type different from the target type.

In this step, when a target type of resource recommendation needs to be performed on a low-activity user, first, recall words need to be determined, so as to perform resource recall based on these recall words. When determining the recall word, the determination may be performed based on the historical access records of the network-wide user on resources of other types different from the target type, for example, if the resource to be recommended is a small video type, then the recall word may be determined based on the historical access records of the network-wide user on resources of a teletext type and/or a short video type. The historical access records at least comprise the number of clicks of each user on the other at least one type of resource in a period of time, the number of clicks under different level classifications and the like.

For example, in a specific implementation manner, first, according to a historical access record of the whole-network user to the resource of the at least one other type, obtaining a plurality of interest points of the whole-network user to the resource of the at least one other type, and the whole-network user clicking a primary classification and a secondary classification of the resource of the at least one other type; and then determining the plurality of recall words for resource recall according to the plurality of interest points, the primary classification and the secondary classification.

In the scheme, the server needs to determine the recall word from the historical access records of other resources, and firstly needs to acquire the interest points of at least one other type of resource, because the number of users is large, one or more interest points may exist in each user, so that the interest points acquired based on the users in the whole network are also multiple, and the interest points are used as the recall word. The first class and the second class with a large number of clicks of the user are also required to be obtained from the historical access records, wherein the first class and the second class with a certain number of clicks are also used as the recall words.

Illustratively, taking the target type as the small video, and taking the resources of at least one other type including the resources of the image-text type and the resources of the short video type as an example, when recommending the resources of the small video type, first obtaining interest points of the resources of the image-text type and the short video type of the whole network user, and then obtaining a first-level classification and a second-level classification clicked by the user. The three types of contents of the interest points, the primary classification and the secondary classification can be used as recall words. However, in a specific implementation, points of interest with a click number exceeding a certain number, a primary category and a secondary category may be selected as final recall words. For example, the interest points clicked more than 6 times by the user, the primary category clicked more than 10 times, and the secondary category clicked more than 8 times are acquired as the recall words finally subjected to the resource recall.

S102: and aiming at each recall word, acquiring a user set corresponding to the recall word, wherein each user in the user set is a high-activity user of the resources of the target type under the recall word, and the number of clicks on at least one other type of resources reaches a preset number.

In this step, after determining the recall word, a corresponding user group, also referred to as a user crown or a user set, is determined for each recall word, and a scenario description is performed in the user set in this disclosure.

It should be understood that in the implementation of this step, one recall word corresponds to one user group, that is, one recall word corresponds to one user set, and finally, one user set obtains a resource list of a target type.

In the specific implementation of the step, for each recall word, all users under the recall word in the users of the whole network are obtained, all high-activity users belonging to the resources of the target type are obtained from the users, and then a final user set is determined based on the click conditions of the users on the resources of other types. In a specific implementation, a threshold of the number of clicks on at least one other type of resource may be set, and the set is selected when the number of clicks on the other type of resource reaches the set threshold, and the set cannot be selected when the number of clicks on the other type of resource does not reach the set threshold.

It should be understood that for each set of users corresponding to a recall word, the users in the set of users are users that can represent characteristics of users who are highly active on the target type of resource in other types of resources.

S103: and determining representative seed users in each user set aiming at each user set, wherein the representative seed users comprise a preset number of users with the maximum similarity between the user vectors and the central vector in the user set.

In this step, after a corresponding user set is obtained for each recall word, if there is a set with too few users, this set may be directly removed, and for an interest field that is too small, the interest and features that cannot represent most users may be removed. For example: the number of users in the user set corresponding to the recall word can be set to at least reach one hundred, and then the user set which does not reach one hundred can be directly removed without analysis processing.

For the remaining user sets, each user set includes a large number of users, and in view of the efficiency of data processing, a preset number of users capable of representing the characteristics of the entire user population may be selected from the users as representative seed users. Since the representative seed users are selected to be able to characterize the characteristics of the entire user population, in the process of determining the representative seed users, the similarity between the user vector and the central vector of the entire user set needs to be calculated, and the users with the maximum similarity are used as the representative seed users. The number of representative seed users may be set according to a specific implementation manner, and the scheme is not limited.

S105: and acquiring a resource list of a target type corresponding to the access recall word of the representative seed user in the user set aiming at the user set corresponding to each recall word.

In this step, after the representative seed users in the user set corresponding to each recall word are determined, the resource list of the target type accessed by the representative seed users needs to be acquired, so as to perform resource recommendation to low-activity users of the resource of the target type in the following.

It should be noted that, in a specific implementation of the solution, first, based on the access history of each representative seed user, a list of all resources of the recall word corresponding to the set accessed by the representative seed user within a period of time is obtained, and then, the resources of the target type are screened out from the list to form a resource list corresponding to the representative seed user. And summarizing the resource lists of all representative seed users in a user set so as to obtain the resource list of the target type corresponding to the whole user set, namely the resource list of the target type corresponding to the recall word.

S106: and recommending the resources to the low-activity users of the resources of the target type according to the resource list corresponding to each recall word.

In this step, after the resource list corresponding to each recall word is obtained, recommendation of a target type of resource is performed to the low-activity user under the recall word based on the resource list. The resources recommended to the low-activity users under the recall word come from the resource list determined by the user group corresponding to the recall word obtained in the manner.

According to the resource recommendation method provided by the embodiment, when resource recommendation is performed on low-activity users of resources of a target type, a plurality of recall words are determined from historical access records of users of the whole network on resources of at least one other type; aiming at each recall word, acquiring a user set which belongs to the high-activity of the target type resource and has the click times of at least one other type resource reaching the preset times; and then recommending resources to the low-activity users of the resources of the target type according to the resource list of the target type corresponding to the recall word accessed by the representative seed user in each user set. By the aid of the scheme, group characteristics are combined when resource recommendation is carried out, signals of other types of resources are migrated, resource recommendation accuracy is improved, recommendation effect is improved, and permeability of low-activity users is improved.

Fig. 2 is a flowchart illustrating a resource recommendation method according to a second embodiment of the disclosure. As shown in fig. 2, on the basis of the above embodiment, the step S102 in the first embodiment includes the steps of:

s1021: and aiming at each recall word, acquiring users with high activity on the resources of the target type under the recall word from the users of the whole network to form an initial set.

In this step, after the recall word is determined, a user group needs to be acquired for each recall word. In a specific implementation, first, for each recall word, all users under the recall word, that is, all users who have accessed a resource corresponding to the recall word, in the users of the whole network are acquired. Furthermore, all high-activity users belonging to the target type of resources are obtained from the users to form an initial set, and the specific standard of the high-activity users can be set according to actual conditions or determined based on the judgment mode of the existing high-activity users. For example, the number of clicks on the resource of the target type is greater than a preset threshold, or the access frequency on the resource of the target type is greater than a preset threshold, and the like, which are not limited by the method.

S1022: and removing users in the initial set, the times of clicking other resources of at least one type of the resources of which the times do not reach the preset times, so as to obtain a user set corresponding to the recall word.

In this step, after the initial set is obtained, a final user set is determined based on click conditions of the users in the initial set on at least one resource of another type.

In a specific implementation, there may be a plurality of implementations, which are illustrated below:

in a possible implementation manner, a threshold of the number of clicks on at least one other type of resource may be set, and a user in the initial set, who has a number of clicks on another type of resource that reaches the set threshold, may opt into the set, and a user who does not reach the set may not opt into the set. For example, taking the target type as a small video type as an example, a user who is required to enter a user set corresponding to a selected recall word needs to click on an image-text type and/or a short video under the recall word more than six times, the primary classification click is more than 8 times, and the secondary classification click is more than 10 times, so that the user can remain in the user set.

In another possible implementation manner, similarly, a threshold of the number of clicks on at least one resource of another type is set, users in the initial set are determined one by one, users in the initial set whose number of clicks on resources of another type reaches the set threshold are retained, users who cannot reach the set threshold are eliminated, and finally, the remaining users in the initial set form a final user set.

Although the resource recommendation method provided in this embodiment is to recommend a target type resource to a low-activity user of a target type resource, after a recall word is determined, when a user group is obtained for each recall word, the user group needs to be obtained across resource types in combination with interest points of the user in at least one other type resource and information flow related data, so that the user requirements can be positioned more accurately, and the final resource recommendation effect is improved.

Fig. 3 is a flowchart illustrating a resource recommendation method according to a third embodiment of the disclosure. As shown in fig. 3, on the basis of any of the above embodiments, step S103 in the first embodiment can be implemented by the following steps:

s1031: and aiming at each user set, calculating a central vector of the user set according to the user vector of each user in the user set.

In this step, after determining the user set corresponding to the recall word, it is necessary to acquire a representative seed user in the user set, and first, it is necessary to acquire a center vector of the user set, where the center vector of one user set is used to represent features of users in the entire set.

In the specific implementation of the scheme, the user vector of each user needs to be acquired first, and specifically, the user vector of each user in the network can be generated by adopting a twin UCF network model or a graph neural network model. Then, aiming at the user set corresponding to each recall word, calculating the average value of the user vectors of each user in the user set to obtain the central vector of the user set.

Optionally, if the user set corresponding to the recall word includes both high-activity users of the target type resource and low-activity users of the target type resource, an average value of user vectors of the high-activity users in the user set is calculated and used as a center vector of the user set.

S1032: similarity between the user vector of each user in the user set and the center vector of the user set is calculated.

S1033: and determining a preset number of users with the maximum similarity as representative seed users of the user set based on the similarity between the user vector and the central vector of each user.

In the above steps, by calculating the similarity between the user vector of each user in the user set and the center vector, the representative seed user in the set is determined according to the size of the similarity.

After the similarity between each user vector and the center vector is obtained, a plurality of users with the largest similarity can be used as the representative seed users of the user set, and in a specific obtaining mode, the users can be sorted in a descending order of the similarity, and a preset number of users before being sorted can be used as representative seed players. A similarity threshold may also be set, and users greater than the similarity threshold are taken as representative seed players, without limitation to this scheme.

Optionally, for each user set, a cohesion degree of one user set may also be calculated through a user vector, where the cohesion degree is used to represent a representativeness of a recall word corresponding to the user set, and the higher the cohesion degree is, the better the representativeness of a user group in the set is proved. In a specific implementation, a plurality of users in a user set may be randomly selected (for example, 500 users in the set may be selected), and the similarity between the user vectors of the users may be calculated pairwise and averaged to obtain the aggregation of the user set. Namely, the way of calculating the cohesion degree of a user set is as follows: randomly selecting a plurality of users from the user set, calculating the similarity between the user vectors of any two users in the plurality of users, and averaging all the obtained similarities to obtain the cohesion degree of the user set.

According to the resource recommendation method provided by the embodiment, within the user group corresponding to each recall word, the representative seed users which can represent the whole group are selected according to the similarity between the user vector and the center vector, so that the low-activity users can be recommended subsequently according to the access history of the representative seed users, the users can represent the characteristics of the whole group better, the user requirements can be positioned more accurately by carrying out voluntary recommendation in the mode, the low-activity users are recommended by using group information, and the recommendation effect is effectively improved.

Fig. 4 is a flowchart illustrating a resource recommendation method according to a fourth embodiment of the disclosure. As shown in fig. 4, on the basis of any of the above embodiments, step S105 in the first embodiment can be implemented by the following steps:

s1051: and voting by the users in the user set corresponding to the recall word aiming at the resource list corresponding to each recall word, deleting the resources with the number of votes less than the preset number of votes, and suppressing the global hot resources in the resource list to obtain the target resource list.

In this step, after the resource list corresponding to each recall word is obtained through the technical solution of the foregoing embodiment, it is necessary to vote for resources in the resource list in the set, and select a resource to be recommended finally. It should be understood that the resources in the resource list are all target types of resources.

It should be noted that, in the scheme, users who vote for resources in the resource list are all high-activity users of resources of the target type, specifically, for each resource in the resource list, one high-activity user clicks once, a vote is accumulated, and by analogy, a score of each resource in the resource list is obtained, and then the resource whose number of votes obtained in the resource list is smaller than the preset number of votes is deleted. The preset ticket number can be configured according to actual conditions.

Furthermore, since many contents belong to hot contents in the current time period or a period of time, many users are recommended to click, and cannot represent the interests of the users themselves, so that the global hot resources in the resource list need to be pressed, and the specific processing mode is as follows:

in the first mode, the resources belonging to the global hot resources in the resource list are arranged to the end of the list, so that the global hot resources are recommended only at the end when resource recommendation is performed on low-activity users.

In the second mode, the resources belonging to the global hot resource in the resource list are deleted. Because the hot resources have corresponding recommendation modes, all users can be recommended, and for the low-activity users, the global hot resources are directly deleted from the resource list, so that the repeated recommendation of the hot resources to the low-activity users is avoided.

Through the processing procedure, a target resource list corresponding to each recall word is obtained, resources in the target resource list are resources which need to be recommended to low-activity users corresponding to the recall words, and a specific recommendation mode is shown in the following steps.

S1052: and recommending the resources to the low-activity users of the resources of the target type according to the target resource lists corresponding to the recall words.

In this step, after the target resource list corresponding to each recall word is obtained, the resources may be directly recommended to the low-activity user randomly or according to a certain sequence. In a specific recommendation manner, for a recall word, a resource in a target resource list corresponding to the recall word may be recommended to a low-activity user under the recall word. However, in a specific implementation, one low-activity user may belong to a group corresponding to a plurality of recall words, so when resource recommendation is performed on one low-activity user, resources in a target resource list corresponding to the plurality of recall words can be scattered and then recommended, and continuous recommendation of similar contents to the user is avoided. For example: random disturbance can be added according to representativeness of the recalled words, the recalled resources are ranked, and finally the recalled resources are recommended to the low-activity users according to probability.

According to the resource recommendation method provided by the embodiment of the disclosure, before recommendation is performed after resource recall finally, the resources are voted by the high-activity users, the resources in which most users are uninteresting are screened out, and the global hot resources are suppressed, so that the hot resources are prevented from being recommended to the low-activity users, the recommendation effect is further improved, and the permeability of the low-activity users is improved.

It should be understood that, in the technical solution of the present disclosure, a core idea is to combine signals of cross-resource types and group-based features to improve a resource recommendation effect, and perform resource recommendation to low-activity users by processing related features of the high-activity users, so that the high-activity users are reserved in a user set corresponding to a recall word in the above embodiment, but this implementation manner is only one of them, and in a specific implementation, a user group may include both high-activity users of a target type resource and low-activity users of the target type resource, and only in a processing process of obtaining a final resource to be recommended to the low-activity users, an access result of the high-activity users of the target type is processed, that is, when a central vector of the user set is obtained, calculation is performed only according to a vector of the high-activity users of the target type resource, and when a resource list to be recommended is obtained, a final resource list is obtained by voting on a target type resource list in a recent access history of a representative seed user by the high-activity users in the user set, and low-activity users in the user set are recommended. For example: when video-type resource recommendation is performed, a user set corresponding to a recall word includes all users whose click times on the image-text type and/or short video-type resources under the recall word are greater than a preset number, which may include users with high activity on small video-type resources or users with low activity on small video-type resources. In this case, after the small video resource list of the representative seed user in the set is obtained, voting is performed by the high-activity user in the user set to obtain a final small video resource list.

On the basis of any of the above embodiments, in the specific implementation of obtaining the target resource list and recommending resources to the low-activity user, the resources may be sorted again and then recommended to the low-activity user. Optionally, the resource recommendation may be performed according to the following steps:

the method comprises the following steps of firstly, aiming at each low-activity user of the target type resource, obtaining a target recall word corresponding to the low-activity user.

For the low-activity users, the users may belong to a plurality of user groups at the same time, and the interested contents are also many, so that for each low-activity user, a corresponding target recall word is determined, and the target recall word comprises one or more than one recall word.

And secondly, acquiring a first target resource list corresponding to the target recall word from a plurality of target resource lists corresponding to the plurality of recall words. And merging the target resource lists corresponding to one or more target recall words to obtain all resource lists which can be recommended to the low-activity user, namely the first target resource list.

And thirdly, sequencing the first target resource list at least once through an xgboost model according to the cohesion degree of the user set corresponding to the target recall word to obtain a second target resource list.

In the scheme, random perturbation is added to the first target resource list according to the representativeness, namely the cohesion degree, of each recalled word. Then, training an xgboost model by combining with real-time characteristics, crowd characteristics and the like, and reordering in the stages of recall, rough layout and fine layout to obtain a new resource list, namely a second target resource list. It should be appreciated that the number of resources in each rearrangement table is further reduced.

And finally, recommending the resources in the second target resource list to the low-activity user. When recommending resources to a low-activity user, the recommendation can be performed according to the probability that each resource is clicked by the user, and the resource in the finally obtained target resource list can also be recommended randomly, so that the scheme is not limited.

In combination with the above embodiments, it should be understood that the present disclosure mainly provides a cross-resource recommendation method combining group characteristics. The core of the method is mainly to utilize signals among different kinds of resources (such as pictures, texts and videos) to perform transfer learning and realize cross-domain application of features, and a Crowd2Vec algorithm and a cross-resource Crowd2Vec algorithm can be adopted in specific realization. The method mainly solves the problem of permeation of low-activity users in a resource recommendation scene for the users, and the permeation rate of the low-activity users is improved through cross-resource signals. The resource recommendation method provided by the present disclosure is exemplified below by taking the target type as a small video type, and taking at least one other type including an image-text type and a short video type as an example.

Fig. 5 is a flowchart illustrating a resource recommendation method according to a fifth embodiment of the disclosure; as shown in fig. 5, a scheme for acquiring a user group in a resource recommendation method is shown, which specifically includes the following steps:

s201: small video high activity users (more than 100 clicks on small video in the last seven days).

S202: historical graphics and text and short video interest point clicking are allowed for more than 6 times (10 times in the first-level classification and 8 times in the second-level classification).

S203: and all the access users form a user set corresponding to the recall words.

In the scheme, aiming at a User of a small video, a User-side expression vector, namely the User vector, is obtained through training of a twin User-based Collaborative Filtering recommendation (UCF) network model. Since the user vector of the high-activity user is mainly used in the subsequent processing, the user's intention of only the high-activity user is obtained. In specific implementation, a user-side vector generated by a twin UCF model, a graph model neural network such as GraphCF and the like can be adopted, and a vector expression stably representing the user characteristics can be obtained. In the scheme, the limitation on the high-activity users may be users who click on the small video more than 100 times in nearly seven days, and the users who meet the condition all belong to the high-activity users of the small video, and the user vectors of the users need to be acquired.

Then, high-activity users of the small videos under the recall words such as the image-text, the interest points in the short videos and the first-level classification and the second-level classification and user vectors of the high-activity users are obtained. The centre vector of the recalled word (Crowd) is then calculated from the user vector, which is typically obtained by averaging. In this process, crowd needs to be generated from the recall word signal of other resources. If a small video high-activity user (clicks the small video frequency more than 100 in the last 7 days), historically clicks an image-text interest point 6 times, classifies the image-text interest point 8 times in the second stage, and classifies the image-text interest point 10 times in the first stage, the user belongs to the Crowd, that is, the user meeting all the admission conditions is determined to be the user group corresponding to the recall word, that is, the user is added into the user set corresponding to the recall word.

Fig. 6 is a flowchart illustrating a resource recommendation method according to a sixth embodiment of the disclosure; as shown in fig. 6, the resource recommendation method performed to the low-activity users based on the user group specifically includes the following steps:

s301: a representative seed user is determined from each set of users.

S302: and voting the small video click list by the small video high-activity users.

S303: and recommending the low-activity users with small videos in the same user set.

In the above step, after the user set corresponding to each recall word is determined in the foregoing manner, the small video recommendation is performed on the low-activity user through the access record of the small video in the access history of the high-activity user.

In specific implementation, the similarity between the user vector and the center vector of each user in the user set is calculated, and a preset number of users with the maximum similarity in the same user set are determined as representative seed users, for example: sequencing the users by utilizing the similarity with the central vector, and selecting Top users as representative seed users; and selecting 100 representative seed users, and removing the Crowd with an undersize scale and an insufficient number of representative users, namely removing the set if the number of the user set is less than 100, and not performing subsequent processing.

Then, aiming at each user set, obtaining the access history of the representative seed users in a recent period of time, obtaining a small video frequency click list corresponding to the recall word of the set from the access history of each representative seed user, and then combining the small video frequency click lists to obtain a small video resource list corresponding to the whole recall word. And then, further voting the obtained small video resource list by high-activity users in the same user set, transferring the crown signal from the image-text and the short video to the small video, and deleting the resource with the vote number smaller than a certain number in the voting result from the list. In the process, attention needs to be paid to pressing the global hot resource, so that the result can be more accurate. For example: and (4) the high-activity users in the set screen-cast each small video in the small video resource list of the 100 representative seed users, if 100 persons click the same small video, the small video has 100 tickets, 99 persons click the small video, 99 tickets and the like, and the number of the tickets of each small video resource in the list is obtained.

Further, random disturbance is added to resources in the resource list according to representativeness of the recall words, an xgboost model is trained by combining real-time characteristics, group characteristics and the like, and the resources in the small video resource list are sequenced in the recall, rough ranking and fine ranking stages. In a specific implementation, the user model is read on line, recall words of users with low activity are found, and a small video resource list, for example 100 video resources are selected for recommendation corresponding to the user set of the recall words. And displaying the historical posterior clicks of all users in the system on the 100 resources, listing the votes of the users in a user group corresponding to the recall word (namely, in a user set corresponding to the recall word), dropping the cohesion characteristics of the user set corresponding to the recall word to an offline, training an xgboost model by actually clicking a display sample according to the data and taking a dot-to-spread ratio as a target to obtain the recall xgboost model, inputting the dot-to-spread characteristics and the user set characteristics to the xgboost model when obtaining the 100 small video resource list again on the online, and outputting scores by the model to reorder the small video resources to improve the recall precision. The rough-row and fine-row processes are similar, in the process of rough-row rearrangement overweight, rough-row characteristic training is added into the original xgboost, fine-row rearrangement is added into the rough row, and fine-row characteristic training is carried out. And after recall and rearrangement, cutting off 80 small videos to enter a rough arrangement, after the rough arrangement and rearrangement, cutting off 50 small videos to enter a fine arrangement, after the fine arrangement and rearrangement, limiting small video resources with the sorting division threshold value larger than 0.1, and then recommending the small video resources to low-activity users.

According to the resource recommendation method provided by the embodiment of the disclosure, the user vector, the user historical behavior and the user real-time characteristics are combined, and the signals in different fields are subjected to transfer learning through the crown 2Vec algorithm idea. The Content-based recommendation algorithm and the Collaborative Filtering (Collaborative Filtering) recommendation algorithm are ingeniously combined, and the specific Crowd group where the user is located is determined, so that the user requirements are accurately positioned, the recommendation effect is improved, and the satisfaction degree of the user is improved. And in a specific situation, low-activity users are recommended through group information, so that the permeability of the low-activity users can be improved, and the effect of a recommendation system is improved.

Fig. 7 is a schematic structural diagram of a resource recommendation device according to an embodiment of the present disclosure. As shown in fig. 7, the resource recommendation apparatus 700 provided in this embodiment includes:

a first processing unit 701, configured to determine multiple recall terms according to a historical access record of a full-network user on resources of at least one other type different from a target type;

a second processing unit 702, configured to obtain, for each recall word, a user set corresponding to the recall word, where each user in the user set is a high-activity user of a resource of the target type under the recall word, and the number of clicks on the resource of the at least one other type reaches a preset number;

a third processing unit 703, configured to determine, for each user set, a representative seed user in each user set, where the representative seed user includes a preset number of users in the user set whose similarity between a user vector and a center vector is the largest;

a fourth processing unit 704, configured to, for a user set corresponding to each recall word, obtain a resource list of a target type corresponding to the recall word accessed by a representative seed user in the user set;

a recommending unit 705, configured to recommend a resource to a low-activity user of the resource of the target type according to the resource list corresponding to each recall word.

The resource recommendation device provided in this embodiment may be configured to execute the resource recommendation method according to any of the above method embodiments, and the implementation principle and the technical effect of the resource recommendation device are similar to each other, which is not described herein again.

In a possible implementation manner, the second processing unit 702 includes:

the first acquisition module is used for acquiring users with high activity on the resources of the target type under the recall word in the users of the whole network aiming at each recall word to form an initial set;

and the first processing module is used for removing the users in the initial set, the times of clicking the resources of the other at least one type of the users do not reach the preset times, and obtaining a user set corresponding to the recall word.

In a possible implementation manner, the third processing unit 703 includes;

the first calculation module is used for calculating a central vector of each user set according to the user vector of each user in the user sets aiming at each user set;

a second calculation module, configured to calculate a similarity between a user vector of each user in the user set and a center vector of the user set;

and the first determining module is used for determining a preset number of users with the maximum similarity as the representative seed users of the user set based on the similarity between the user vector and the central vector of each user.

In a possible implementation manner, the first processing unit 701 includes:

the second processing module is used for acquiring a plurality of interest points of the whole network user on the resources of the at least one other type according to the historical access records of the whole network user on the resources of the at least one other type, and clicking the primary classification and the secondary classification of the resources of the at least one other type by the whole network user;

and the third processing module is used for determining the plurality of recall words for resource recall according to the plurality of interest points, the primary classification and the secondary classification.

In a possible implementation manner, the recommending unit 705 includes:

the fourth processing module is used for voting users in the user set corresponding to the recall word aiming at the resource list corresponding to each recall word, deleting the resources with the number of votes smaller than the preset number of votes, and pressing global hot resources in the resource list to obtain a target resource list;

and the recommending module is used for recommending the resources to the low-activity users of the resources of the target type according to the target resource lists corresponding to the plurality of recalling words.

Optionally, the recommending module includes:

the first processing sub-module is used for acquiring a target recall word corresponding to each low-activity user of the target type resource;

the second processing submodule is used for acquiring a first target resource list corresponding to the target recall word from a plurality of target resource lists corresponding to the plurality of recall words;

the third processing sub-module is used for sequencing the first target resource list at least once through an xgboost model according to the cohesion degree of the user set corresponding to the target recall word to obtain a second target resource list;

and the recommending submodule is used for recommending the resources in the second target resource list to the low-activity user.

Optionally, the fourth processing module is specifically configured to:

arranging the resources belonging to the global hot resources in the resource list to the last of the list;

alternatively, the first and second electrodes may be,

and deleting the resources belonging to the global hot resources in the resource list.

In one possible implementation, the apparatus 700 further includes:

a fifth processing unit 706, configured to randomly select multiple users from the user set according to the user set corresponding to each recall word, calculate similarity between user vectors of any two users in the multiple users, and average all the obtained similarities to obtain a cohesion degree of the user set.

Optionally, the first calculating module is specifically configured to:

and calculating the average value of the user vectors of each user in the user set to obtain the central vector of the user set.

In one possible implementation, the apparatus 700 further includes:

a sixth processing unit 707, configured to generate a user vector of each user in the network by using a UCF network model or a graph neural network model of a twin collaborative filtering algorithm.

In the technical scheme of the disclosure, the collection, storage, use, processing, transmission, provision, disclosure and other processing of the personal information of the related user are all in accordance with the regulations of related laws and regulations and do not violate the good customs of the public order.

The present disclosure also provides an electronic device, a non-transitory computer readable storage medium storing computer instructions, and a computer program product according to embodiments of the present disclosure.

According to an embodiment of the present disclosure, the present disclosure also provides a computer program product comprising: a computer program, stored in a readable storage medium, from which at least one processor of the electronic device can read the computer program, the at least one processor executing the computer program causing the electronic device to perform the solution provided by any of the embodiments described above.

FIG. 8 shows a schematic block diagram of an example electronic device used to implement embodiments of the present disclosure. Electronic devices are intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device may also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are meant to be examples only, and are not intended to limit implementations of the disclosure described and/or claimed herein.

As shown in fig. 8, the electronic device 800 includes a computing unit 801 that can perform various appropriate actions and processes according to a computer program stored in a Read Only Memory (ROM) 802 or a computer program loaded from a storage unit 808 into a Random Access Memory (RAM) 803. In the RAM 803, various programs and data necessary for the operation of the device 800 can also be stored. The calculation unit 801, the ROM 802, and the RAM 803 are connected to each other by a bus 804. An input/output (I/O) interface 805 is also connected to bus 804.

A number of components in the device 800 are connected to the I/O interface 805, including: an input unit 806, such as a keyboard, a mouse, or the like; an output unit 807 such as various types of displays, speakers, and the like; a storage unit 808, such as a magnetic disk, optical disk, or the like; and a communication unit 809 such as a network card, modem, wireless communication transceiver, etc. The communication unit 809 allows the device 800 to exchange information/data with other devices via a computer network such as the internet and/or various telecommunication networks.

Computing unit 801 may be a variety of general and/or special purpose processing components with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), various dedicated Artificial Intelligence (AI) computing chips, various computing units running machine learning model algorithms, a Digital Signal Processor (DSP), and any suitable processor, controller, microcontroller, and the like. The computing unit 801 performs the various methods and processes described above, such as the resource recommendation method. For example, in some embodiments, the resource recommendation method may be implemented as a computer software program tangibly embodied in a machine-readable medium, such as storage unit 808. In some embodiments, part or all of the computer program can be loaded and/or installed onto device 800 via ROM 802 and/or communications unit 809. When the computer program is loaded into RAM 803 and executed by the computing unit 801, one or more steps of the resource recommendation method described above may be performed. Alternatively, in other embodiments, the computing unit 801 may be configured as a resource recommendation method in any other suitable manner (e.g., by way of firmware).

Various implementations of the systems and techniques described here above may be implemented in digital electronic circuitry, integrated circuitry, field Programmable Gate Arrays (FPGAs), application Specific Integrated Circuits (ASICs), application Specific Standard Products (ASSPs), system on a chip (SOCs), complex Programmable Logic Devices (CPLDs), computer hardware, firmware, software, and/or combinations thereof. These various embodiments may include: implemented in one or more computer programs that are executable and/or interpretable on a programmable system including at least one programmable processor, which may be special or general purpose, receiving data and instructions from, and transmitting data and instructions to, a storage system, at least one input device, and at least one output device.

Program code for implementing the methods of the present disclosure may be written in any combination of one or more programming languages. These program code may be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus, such that the program code, when executed by the processor or controller, causes the functions/acts specified in the flowchart and/or block diagram to be performed. The program code may execute entirely on the machine, partly on the machine, as a stand-alone software package partly on the machine and partly on a remote machine or entirely on the remote machine or server.

In the context of this disclosure, a machine-readable medium may be a tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to a user; and a keyboard and a pointing device (e.g., a mouse or a trackball) by which a user can provide input to the computer. Other kinds of devices may also be used to provide for interaction with a user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form, including acoustic, speech, or tactile input.

The systems and techniques described here can be implemented in a computing system that includes a back-end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front-end component (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back-end, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: local Area Networks (LANs), wide Area Networks (WANs), and the Internet.

The computer system may include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. The Server can be a cloud Server, also called a cloud computing Server or a cloud host, and is a host product in a cloud computing service system, so as to solve the defects of high management difficulty and weak service expansibility in the traditional physical host and VPS service ("Virtual Private Server", or simply "VPS"). The server may also be a server of a distributed system, or a server incorporating a blockchain.

It should be understood that various forms of the flows shown above, reordering, adding or deleting steps, may be used. For example, the steps described in the present disclosure may be executed in parallel, sequentially, or in different orders, and are not limited herein as long as the desired results of the technical solutions disclosed in the present disclosure can be achieved.

The above detailed description should not be construed as limiting the scope of the disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may be made in accordance with design requirements and other factors. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present disclosure should be included in the protection scope of the present disclosure.

Claims

1. A resource recommendation method, comprising:

determining a plurality of recall words according to historical access records of the whole network users to resources of at least one other type different from the target type;

determining representative seed users in each user set aiming at each user set, wherein the representative seed users comprise a preset number of users with the maximum similarity between user vectors and central vectors in the user set;

2. The method of claim 1, wherein the retrieving, for each recall word, a set of users corresponding to the recall word comprises:

aiming at each recall word, acquiring users belonging to the resources with high activity of the target type under the recall word from all network users to form an initial set;

and removing users in the initial set, the times of clicking the resources of the other at least one type of resources of which the times do not reach the preset times, so as to obtain a user set corresponding to the recall word.

3. The method of claim 1, wherein the determining, for each set of users, a representative seed user in each set of users comprises;

for each user set, calculating a central vector of the user set according to the user vector of each user in the user set;

calculating a similarity between a user vector of each user in the user set and a center vector of the user set;

and determining a preset number of users with the maximum similarity as representative seed users of the user set based on the similarity between the user vector and the central vector of each user.

4. The method of claim 1, wherein determining a plurality of recall terms from historical access records of network-wide users to resources of at least one other type different from the target type comprises:

acquiring a plurality of interest points of the whole network user on the resources of the at least one other type according to the historical access records of the whole network user on the resources of the at least one other type, and clicking the primary classification and the secondary classification of the resources of the at least one other type by the whole network user;

and determining the plurality of recall words for resource recall according to the plurality of interest points, the primary classification and the secondary classification.

5. The method according to any one of claims 1 to 4, wherein the recommending the resource to the low-activity user of the resource of the target type according to the resource list corresponding to each recall word comprises:

for a resource list corresponding to each recall word, voting by users in a user set corresponding to the recall word, deleting resources with the number of votes smaller than a preset number of votes, and pressing global hot resources in the resource list to obtain a target resource list;

and recommending the resources to the low-activity users of the resources of the target type according to a plurality of target resource lists corresponding to the plurality of recall words.

6. The method of claim 5, wherein the making of the resource recommendation to the low-activity user of the resource of the target type according to the plurality of target resource lists corresponding to the plurality of recall words comprises:

aiming at each low-activity user of the target type resource, acquiring a target recall word corresponding to the low-activity user;

acquiring a first target resource list corresponding to the target recall word from a plurality of target resource lists corresponding to the plurality of recall words;

sequencing the first target resource list at least once through an xgboost model according to the cohesion degree of the user set corresponding to the target recall word to obtain a second target resource list;

and recommending the resources in the second target resource list to the low-activity user.

7. The method of claim 5, wherein said suppressing global hot resources in the resource list comprises:

alternatively, the first and second electrodes may be,

8. The method of claim 6, wherein the method further comprises:

and aiming at a user set corresponding to each recall word, randomly selecting a plurality of users from the user set, calculating the similarity between user vectors of any two users in the plurality of users, and averaging all the obtained similarities to obtain the cohesion degree of the user set.

9. The method of claim 3, wherein said calculating a center vector for the set of users from the user vector for each user in the set of users comprises:

10. The method of claim 5, wherein the method further comprises:

and generating a user vector of each user in the network by adopting a twin user-based collaborative filtering algorithm UCF network model or a graph neural network model.

11. A resource recommendation device, comprising:

the first processing unit is used for determining a plurality of recall words according to the historical access records of the whole network users to resources of at least one other type different from the target type;

12. The apparatus of claim 11, wherein the second processing unit comprises:

13. The apparatus of claim 11, wherein the third processing unit comprises;

14. The apparatus of claim 11, wherein the first processing unit comprises:

and the third processing module is used for determining the plurality of recalling words for resource recall according to the plurality of interest points, the primary classification and the secondary classification.

15. The apparatus according to any one of claims 11 to 14, wherein the recommending unit includes:

16. The apparatus of claim 15, wherein the recommendation module comprises:

the first processing submodule is used for acquiring a target recall word corresponding to each low-activity user of the target type resource;

17. The apparatus according to claim 15, wherein the fourth processing module is specifically configured to:

alternatively, the first and second electrodes may be,

18. The apparatus of claim 16, wherein the apparatus further comprises:

and the fifth processing unit is used for randomly selecting a plurality of users from the user set aiming at the user set corresponding to each recall word, calculating the similarity between the user vectors of any two users in the plurality of users, and averaging all the obtained similarities to obtain the cohesion degree of the user set.

19. The apparatus of claim 13, wherein the first computing module is specifically configured to:

20. The apparatus of claim 15, wherein the apparatus further comprises:

and the sixth processing unit is used for generating a user vector of each user in the network by adopting a twin user-based collaborative filtering algorithm UCF network model or a graph neural network model.

21. An electronic device, comprising:

at least one processor; and

the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-10.

22. A non-transitory computer readable storage medium having stored thereon computer instructions for causing the computer to perform the method of any one of claims 1-10.

23. A computer program product comprising a computer program which, when executed by a processor, carries out the steps of the method of any one of claims 1 to 10.