Resource recommendation method and device, equipment and storage medium

Abstract

This disclosure provides a resource recommendation method, apparatus, device, and storage medium, relating to artificial intelligence technology, particularly big data. The specific implementation scheme is as follows: When recommending resources to low-active users of a target type of resource, firstly, multiple recall terms are determined from the historical access records of all users across the network for at least one other type of resource; for each recall term, a set of users with high activity belonging to that target type of resource and whose click count for at least one other type of resource has reached a preset number is obtained; then, based on the list of target type resources accessed by representative seed users in each user set corresponding to the recall term, resource recommendations are made to low-active users of that target type of resource. Through this scheme, resource recommendations are combined with group characteristics and the signals from other types of resources are transferred during resource recommendation, improving the accuracy of resource recommendations, enhancing recommendation effectiveness, and increasing penetration among low-active users.

Description

Technical Field

[0001] This 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 Technology

[0002] With the development of IoT technology, the massive amount of data in the online world has placed a great burden on users. As a result, personalized recommendation technology is gradually becoming popular in various fields, such as news recommendations, business recommendations, entertainment recommendations, learning recommendations, lifestyle recommendations, and decision support.

[0003] In existing technologies, resource recommendation for users is mainly achieved through explicit or implicit recall. Explicit recall emphasizes content-based methods, directly retrieving resources based on the content searched or browsed by the user. Implicit recall leans more towards feature engineering, representing users as vectors in a spatial representation, i.e., user vectors, and further retrieving resources from the network based on these user vectors.

[0004] However, current recall methods based on user search content or user vectors are not very effective in improving penetration rates among low-activity users. Summary of the Invention

[0005] This disclosure provides a resource recommendation method, apparatus, device, and storage medium.

[0006] According to a first aspect of this disclosure, a resource recommendation method is provided, comprising:

[0007] Based on the historical access records of all users on the network for at least one other type of resource that is different from the target type, multiple recall keywords are identified;

[0008] For each recall term, obtain the user set corresponding to the recall term. Each user in the user set is a highly active user of the target type of resource under the recall term, and the number of clicks on at least one other type of resource reaches a preset number.

[0009] For each user set, a representative seed user is determined. The representative seed user includes a preset number of users whose user vectors have the highest similarity to the center vector in the user set.

[0010] For each user set corresponding to a recall term, obtain a list of resources of the target type accessed by representative seed users in the user set corresponding to the recall term.

[0011] Based on the resource list corresponding to each recall term, resource recommendations are made to low-activity users of the target type of resources.

[0012] According to a second aspect of this disclosure, a resource recommendation apparatus is provided, comprising:

[0013] The first processing unit is used to determine multiple recall keywords based on the historical access records of all network users to at least one other type of resource that is different from the target type;

[0014] The second processing unit is used to obtain a user set corresponding to each recall term, wherein each user in the user set is a highly active user of the target type of resource under the recall term, and the number of clicks on at least one other type of resource reaches a preset number;

[0015] The third processing unit is used to determine representative seed users in each user set. The representative seed users include a preset number of users in the user set whose similarity to the center vector is the highest.

[0016] The fourth processing unit is used to obtain, for each user set corresponding to a recall term, a list of resources of the target type accessed by representative seed users in the user set corresponding to the recall term.

[0017] The recommendation unit is used to recommend resources to low-activity users of the target type of resources based on the resource list corresponding to each recall term.

[0018] According to a third aspect of this disclosure, an electronic device is provided, comprising:

[0019] At least one processor; and

[0020] A memory communicatively connected to the at least one processor; wherein,

[0021] The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method described in the first aspect.

[0022] According to a fourth aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are configured to cause the computer to perform the method described in the first aspect.

[0023] According to a fifth aspect of this disclosure, a computer program product is provided, the computer program product comprising: a computer program stored in a readable storage medium, wherein at least one processor of an electronic device can read the computer program from the readable storage medium, and the at least one processor executes the computer program to cause the electronic device to perform the method described in the first aspect.

[0024] According to the technical solution disclosed herein, when recommending resources to low-active users of a target type of resource, firstly, multiple recall keywords are determined from the historical access records of all users on at least one other type of resource; for each recall keyword, a set of users with high activity levels belonging to that target type of resource and who have clicked on at least one other type of resource a preset number of times are obtained; then, based on the list of target type resources corresponding to that recall keyword accessed by representative seed users in each user set, resource recommendations are made to low-active users of that target type of resource. Through this solution, resource recommendations are combined with group characteristics and the signals from other types of resources are transferred during resource recommendation, improving the accuracy of resource recommendations, enhancing recommendation effectiveness, and increasing penetration among low-active users.

[0025] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0026] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein:

[0027] Figure 1 A flowchart illustrating a resource recommendation method provided in the first embodiment of this disclosure;

[0028] Figure 2 A flowchart illustrating a resource recommendation method provided in the second embodiment of this disclosure;

[0029] Figure 3 A flowchart illustrating a resource recommendation method provided in the third embodiment of this disclosure;

[0030] Figure 4 A flowchart illustrating a resource recommendation method provided in the fourth embodiment of this disclosure;

[0031] Figure 5 A flowchart illustrating a resource recommendation method provided in the fifth embodiment of this disclosure;

[0032] Figure 6 A flowchart illustrating a resource recommendation method provided in the sixth embodiment of this disclosure;

[0033] Figure 7 This is a schematic diagram of the structure of the resource recommendation device provided in the embodiments of this disclosure;

[0034] Figure 8 A schematic block diagram of an example electronic device used to implement embodiments of the present disclosure is shown. Detailed Implementation

[0035] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0036] With the development of the internet and big data technologies, the massive amount of information has placed a huge burden on users. Assisting users in resource recommendations has become an important research topic to address the problems of information overload and information disaster, and is currently the core of information flow recommendation applications. In current technical solutions, when resource recommendations are needed for users, the commonly used methods can be broadly divided into explicit recall and implicit recall. The former emphasizes content-based recommendation methods and has strong interpretability; the latter leans towards feature engineering, representing users through spatial vectors and recalling resources for users based on these vectors.

[0037] This resource recommendation method, which is based on a single explicit or implicit recall, has only average overall recommendation effect in specific application scenarios, especially with low penetration rates among low-activity users. Currently, there is a lack of a comprehensive resource recall method that combines the advantages of various types of resource information and integrates explicit and implicit recall methods.

[0038] To address the aforementioned technical problems, the technical concept of this disclosure is as follows: During the research of resource recommendation schemes, the inventors discovered that for low-activity users, their historical behavioral data provides more valuable features. However, for the same point of interest, high-activity users' click and browsing history provides more information. Furthermore, users not only access one type of resource but also click and browse other types. During the recall process for a specific type of resource, the access characteristics of other types of resources can be combined to accurately pinpoint user needs and thus make resource recommendations. Therefore, the inventors considered a comprehensive resource recommendation scheme that integrates user vectors, historical behavior, and real-time user features, performs transfer learning on signals from different types of resources, and combines the advantages of explicit and implicit methods.

[0039] Based on the above technical conception process, this disclosure provides a resource recommendation method applied to the backend server of a public resource recommendation system. When recommending a certain type of resource to inactive users, recall terms for resource recall are determined from the historical access records of all users on other types of resources. Based on these recall terms, a user group corresponding to each recall term is determined, i.e., a user set. Based on the characteristics and access history of highly active users of this type of resource in this user group, recommendations of this type of resource are made to inactive users. In other words, when making resource recommendations, group characteristics are combined, and signals of other types of resources are transferred and learned to improve the accuracy of resource recommendations to inactive users, thereby improving the recommendation effect.

[0040] The entire resource recommendation scheme of this disclosure can be applied to a server, which can be a server for entertainment software, an information recommendation system, a search system, an information flow system, etc. This scheme does not limit the application to these applications. For example, if user interaction is involved, it may also include the user's device.

[0041] It is understood that, in practical applications, the resource recommendation scheme of this disclosure may also include other devices, such as storage devices, in this scenario. The specific devices can be adjusted according to actual needs, and this disclosure does not limit them. Furthermore, the embodiments of this disclosure do not limit the actual form of the various devices included in the application scenario, nor do they limit the interaction methods between devices. These can be set according to actual needs in the specific application of the scheme.

[0042] It should be understood that in this disclosure, low-activity users refer to users with low activity levels, also known as low-active users; high-activity users refer to users with high activity levels, also known as high-active users.

[0043] The following detailed description uses specific embodiments to illustrate the technical solutions of this disclosure and how these solutions solve the aforementioned technical problems. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of this disclosure will now be described with reference to the accompanying drawings.

[0044] The following section introduces the specific implementation scheme of the resource recommendation method provided in this publication.

[0045] Figure 1 This is a flowchart illustrating a resource recommendation method provided in the first embodiment of this disclosure. Figure 1 As shown, this resource recommendation method specifically includes the following steps:

[0046] S101: Based on the historical access records of all users on the network for at least one other type of resource that is different from the target type, determine multiple recall keywords.

[0047] In this step, when recommending target-type resources to low-activity users, the first step is to determine recall keywords to retrieve resources based on these keywords. Recall keywords can be determined based on the historical access records of all users across the internet for other types of resources different from the target type. For example, if the recommended resource is a short video, then recall keywords can be determined based on the historical access records of all users across the internet for text / image resources and / or short video resources. These historical access records should include at least the number of times each user clicked on at least one other type of resource within a certain period, and the number of clicks under different level categories, etc.

[0048] For example, in one specific implementation, firstly, based on the historical access records of all users on the other at least one type of resource, multiple points of interest of all users on the other at least one type of resource are obtained, as well as the primary and secondary categories of the at least one other type of resource clicked by all users on the other network; then, based on the multiple points of interest, the primary category and the secondary category, the multiple recall words for resource recall are determined.

[0049] In this scheme, the server needs to determine recall keywords from the historical access records of other resources. First, it needs to obtain points of interest for at least one other type of resource. Since there are many users, each user may have one or more points of interest; therefore, the number of points of interest obtained from all users across the network is also multiple, and these points of interest serve as recall keywords. It also needs to obtain the primary and secondary categories with the most clicks from the historical access records. Primary and secondary categories with a certain number of clicks are also used as recall keywords.

[0050] For example, taking short videos as the target type, and at least one other type of resource including text / image resources and short video resources, when recommending short video resources, the system first obtains the interest points of all users across the network regarding text / image resources and short video resources, and then obtains the primary and secondary categories clicked by users. All three types of content—interest points, primary categories, and secondary categories—can be used as recall keywords. However, in specific implementations, interest points that have been clicked more than a certain number of times, primary categories, and secondary categories can be selected as the final recall keywords. For example, interest points clicked more than 6 times, primary categories clicked more than 10 times, and secondary categories clicked more than 8 times can be obtained as the final recall keywords for resource retrieval.

[0051] S102: For each recall term, obtain the user set corresponding to the recall term. Each user in the user set is a highly active user of the target type of resource under the recall term, and the number of clicks on at least one other type of resource has reached a preset number.

[0052] In this step, after the recall terms are determined, a corresponding user group, also known as a user crowd or user set, is determined for each recall term. In this disclosure, the scheme is described using the user set.

[0053] It should be understood that in the implementation of this step, one recall term corresponds to one user group, that is, one recall term corresponds to one user set, and finally, a user set will obtain a list of resources of the target type.

[0054] In the specific implementation of this step, for each recall keyword, all users across the entire network who are associated with that keyword are retrieved. From these users, all highly active users belonging to the target resource type are identified. Then, based on these users' click history on other resource types, the final user set is determined. Specifically, a threshold can be set for the number of clicks on at least one other resource type. Users whose click history on other resource types reaches the set threshold are included in the set; those who do not reach the threshold are excluded.

[0055] It should be understood that for each recall term, the user set in that user set represents the characteristics of users who are highly active in other types of resources, in relation to the target type of resource.

[0056] S103: For each user set, determine the representative seed users in each user set. The representative seed users include a preset number of users whose user vectors have the highest similarity to the center vector in the user set.

[0057] In this step, after obtaining the corresponding user set for each recall term, if there is a set with too few users, this set can be directly removed. For niche interest areas, which may not represent the interests and characteristics of the majority of users, they can be removed. For example, the user set corresponding to a recall term can be set to have at least one hundred users. Then, user sets with fewer than one hundred users can be directly removed without further analysis.

[0058] For the retained user sets, each set contains a large number of users. Considering data processing efficiency, a predetermined number of users that can represent the characteristics of the entire user group can be selected as representative seed users. Since these representative seed users should be able to characterize the features of the entire user group, the similarity between the user vectors and the central vector of the entire user set needs to be calculated during the selection process. Users with the highest similarity are selected as representative seed users. The number of representative seed users can be set according to the specific implementation method; this scheme does not impose a limit.

[0059] S105: For each user set corresponding to a recall term, obtain the resource list of the target type corresponding to the access recall term of the representative seed user in the user set.

[0060] In this step, after the representative seed users in the user set corresponding to each of the aforementioned recall terms are determined, it is necessary to obtain a list of target type resources accessed by these representative seed users so that resources can be recommended to low-activity users of the target type resources in the future.

[0061] It should be noted that in the specific implementation of this scheme, firstly, based on the access history of each representative seed user, a list of all resources corresponding to the recall keyword for that set accessed by that representative seed user within a certain period is obtained. Then, resources of the target type are filtered out to form the resource list corresponding to that representative seed user. Next, the resource lists of all representative seed users in a user set are aggregated to obtain the resource list of the target type corresponding to the entire user set, which is the resource list of the target type corresponding to that recall keyword.

[0062] S106: Based on the resource list corresponding to each recall keyword, recommend resources to low-activity users of the target type of resources.

[0063] In this step, after obtaining the resource list corresponding to each recall term, resources of the target type are recommended to low-activity users under that recall term based on this resource list. The resources recommended to low-activity users under that recall term come from the resource list determined by the user group corresponding to that recall term, obtained in the aforementioned manner.

[0064] The resource recommendation method provided in this embodiment, when recommending resources to low-active users of a target type of resource, first determines multiple recall terms from the historical access records of all users on at least one other type of resource. For each recall term, it obtains a set of users who are highly active and belong to the target type of resource, and whose click count on at least one other type of resource has reached a preset number. Then, based on the list of target type resources corresponding to the recall term accessed by representative seed users in each user set, it recommends resources to low-active users of that target type of resource. Through the above scheme, resource recommendation is combined with group characteristics and transfers signals from other types of resources, improving the accuracy of resource recommendation, enhancing the recommendation effect, and increasing the penetration rate to low-active users.

[0065] Figure 2 This is a flowchart illustrating a resource recommendation method provided in the second embodiment of this disclosure. Figure 2 As shown, based on the above embodiments, step S102 in the first embodiment includes the following steps:

[0066] S1021: For each recall term, obtain the users with high activity for the target type of resources under the recall term from the entire network, and form an initial set.

[0067] In this step, after identifying the recall keywords, it is necessary to acquire the user group for each keyword. Specifically, for each recall keyword, firstly, all users across the entire network who are associated with that keyword are acquired; that is, all users who have accessed the resources corresponding to that keyword. Further, from these users, an initial set of highly active users belonging to the target type of resources is formed. The specific criteria for highly active users can be set according to actual conditions or determined based on existing methods for identifying highly active users. For example, the number of clicks on the target type of resources may exceed a preset threshold, or the access frequency to the target type of resources may exceed a preset threshold. The method does not impose restrictions on these criteria.

[0068] S1022: Remove users from the initial set whose number of clicks on at least one other type of resource has not reached a preset number, and obtain the user set corresponding to the recall keyword.

[0069] In this step, after obtaining the initial set, the final user set is determined based on the clicks of these users on at least one other type of resource.

[0070] In practice, there are multiple ways to implement this, and the following examples illustrate this:

[0071] In one possible implementation, a threshold for the number of clicks on at least one other type of resource can be set. Users in the initial set whose click count on other types of resources reaches the set threshold are included in this set, while users who do not reach the threshold are excluded. For example, taking short video as the target type, users selected for the user set corresponding to the recall keyword need to have more than six clicks on text / image types and / or short videos under the recall keyword, more than eight clicks on the first-level category, and more than ten clicks on the second-level category. In this case, the user can be retained in the user set.

[0072] In another possible implementation, a threshold for the number of clicks on at least one other type of resource is set. Users in the initial set are judged one by one. Users in the initial set whose number of clicks on other types of resources reaches the set threshold are retained, while users whose number of clicks does not reach the threshold are removed. Finally, the remaining users in the initial set form the final user set.

[0073] The resource recommendation method provided in this embodiment, although ultimately aims to recommend target type resources to low-activity users, requires combining the user's interests in at least one other type of resource and related information flow data when acquiring user groups for each recall keyword. This cross-resource type acquisition of user groups can more accurately locate user needs, thereby improving the final resource recommendation effect.

[0074] Figure 3 This is a flowchart illustrating a resource recommendation method provided in the third embodiment of this disclosure. Figure 3 As shown, based on any of the above embodiments, step S103 in the first embodiment can be implemented through the following steps:

[0075] S1031: For each user set, calculate the center vector of the user set based on the user vector of each user in the user set.

[0076] In this step, after determining the user set corresponding to the recall term, it is necessary to obtain the representative seed users in the user set. First, it is necessary to obtain the center vector of the user set. The center vector of a user set is used to represent the features of the users in the entire set.

[0077] In the specific implementation of this scheme, the first step is to obtain the user vector for each user. This can be achieved using a Siamese UCF network model or a graph neural network model to generate the user vector for each user in the network. Then, for each recall term corresponding to a user set, the average value of the user vectors for each user in that user set is calculated to obtain the center vector of the user set.

[0078] Optionally, if the user set corresponding to the recall term includes both highly active users of the target type resource and inactive users of the target type resource, then the average value of the user vectors of the highly active users in the user set is calculated as the center vector of the user set.

[0079] S1032: Calculate the similarity between the user vector of each user in the user set and the center vector of the user set.

[0080] S1033: Based on the similarity between each user's user vector and the center vector, determine a preset number of users with the highest similarity as representative seed users of the user set.

[0081] In the above steps, the representative seed users in the set are determined by calculating the similarity between the user vector and the center vector of each user in the user set.

[0082] After obtaining the similarity between each user vector and the center vector, the users with the highest similarity can be used as representative seed users for that user set. Specifically, they can be sorted in descending order of similarity, with a predetermined number of users selected before the sorting as representative seed users. Alternatively, a similarity threshold can be set, with users exceeding the threshold being selected as representative seed users; this approach is not restricted in this case.

[0083] Optionally, for each user set, the cohesion of the user set can be calculated using user vectors. This cohesion represents the representativeness of the recall terms corresponding to the user set; the higher the cohesion, the better the representativeness of the user group in the set. In a specific implementation, multiple users can be randomly selected from a user set (e.g., 500 users can be selected), and the similarity between the user vectors of each pair of users can be calculated and averaged to obtain the cohesion of the user set. That is, the method for calculating the cohesion of a user set is: randomly select multiple users from the user set, calculate the similarity between the user vectors of any two users among these users, and average all the obtained similarities to obtain the cohesion of the user set.

[0084] The resource recommendation method provided in this embodiment selects representative seed users who are more representative of the entire group within the user group corresponding to each recall term by measuring the similarity between the user vector and the center vector. This allows for subsequent resource recommendations to low-activity users based on the access history of these representative seed users. These users are more representative of the characteristics of the entire group, and this voluntary recommendation method can more accurately locate user needs and effectively improve the recommendation effect by using group information to recommend to low-activity users.

[0085] Figure 4This is a flowchart illustrating a resource recommendation method provided in the fourth embodiment of this disclosure. Figure 4 As shown, based on any of the above embodiments, step S105 in the first embodiment can be implemented through the following steps:

[0086] S1051: For each resource list corresponding to a recall term, users in the user set corresponding to the recall term vote, delete resources with fewer than the preset number of votes, and suppress globally popular resources in the resource list to obtain the target resource list.

[0087] In this step, after obtaining the resource list corresponding to each recall term through the technical solution of the aforementioned embodiments, it is necessary to vote on the resources in the resource list within the set and select the resource to be recommended last. It should be understood that the resources in this resource list are all target type resources.

[0088] It should be noted that in this scheme, the users who vote on the resources in the resource list are all highly active users of the target type of resource. Specifically, for each resource in the resource list, one vote is accumulated for each click by a highly active user, and so on, to obtain the score for each resource in the resource list. Then, resources in the resource list with fewer than the preset number of votes are deleted. This preset number of votes can be configured according to the actual situation.

[0089] Furthermore, since much of the content is currently trending or has been popular for a period of time, many users will be recommended to click on it, which does not necessarily reflect the users' actual interests. Therefore, it is also necessary to suppress globally popular resources in the resource list. The specific handling method is as follows:

[0090] The first approach is to arrange the resources that belong to the globally popular resources in the resource list to the end of the list, so that when recommending resources to low-activity users, the globally popular resources will be recommended last.

[0091] The second approach is to remove resources that belong to the globally popular resources list. Since popular resources have corresponding recommendation methods, all users may be recommended to them. For users with low activity levels, directly removing globally popular resources from the resource list avoids repeatedly recommending popular resources to them.

[0092] Through the above processing, a target resource list corresponding to each recall term is obtained. The resources in the target resource list are the resources that need to be recommended to low-activity users corresponding to the recall term. The specific recommendation method is shown in the following steps.

[0093] S1052: Based on multiple target resource lists corresponding to multiple recall keywords, recommend resources to low-activity users of target type resources.

[0094] In this step, after obtaining the target resource list corresponding to each recall term, these resources can be directly recommended to low-activity users randomly or in a certain order. In one specific recommendation method, for a recall term, resources in the target resource list corresponding to that recall term can be recommended to low-activity users associated with that recall term. However, in practice, a low-activity user may belong to multiple groups corresponding to multiple recall terms. Therefore, when recommending resources to a low-activity user, the resources in the target resource lists corresponding to multiple recall terms can be scattered and recommended to avoid continuously recommending similar content to the user. For example, random perturbation can be added based on the representativeness of the recall terms, the recalled resources can be sorted, and finally recommended to low-activity users according to probability.

[0095] The resource recommendation method provided in this disclosure, before recommending resources after final resource recall, uses highly active users to vote on resources, filters out resources that are not of interest to most users, and suppresses globally popular resources to avoid recommending popular resources to low-active users, thereby further improving the recommendation effect and increasing the penetration rate among low-active users.

[0096] It should be understood that the core idea of ​​the technical solution disclosed herein is to improve the resource recommendation effect by combining cross-resource type signals and group-based characteristics. By processing the relevant characteristics of highly active users, resources are recommended to less active users. Therefore, in the above embodiment, highly active users are retained in the user set corresponding to the recall words. However, this implementation method is only one of them. In a more specific implementation, the user group can include both highly active users of the target type resource and less active users of the target type resource. However, in the process of obtaining the final resource to be recommended to less active users, only the access results of highly active users of the target type are processed. That is, when obtaining the center vector of the user set, only the vector of highly active users of the target type resource is calculated. When obtaining the list of resources to be recommended, the final resource list is obtained by voting on the target type resource list in the recent access history of the representative seed user by the highly active users in the user set, and then recommended to the less active users in the user set. For example, when recommending video resources, the user set corresponding to a recall keyword includes all users who have clicked on text / image or short video resources under that keyword more than a preset number of times. This set may include users with high activity levels for short video resources and users with low activity levels for short video resources. In this case, after obtaining the short video resource list of representative seed users in the set, the final short video resource list is obtained by voting from the highly active users in the user set.

[0097] Based on any of the above embodiments, in the specific implementation of obtaining the target resource list and recommending resources to low-activity users, the list can be sorted again before being recommended to low-activity users. Optionally, resource recommendation can be performed according to the following steps:

[0098] The first step is to obtain the target recall keywords corresponding to each low-activity user for the target type of resource.

[0099] For low-activity users, they may belong to multiple user groups at the same time and have a lot of interests. Therefore, for each low-activity user, we need to determine their corresponding target recall keywords, which include one or more recall keywords.

[0100] The second step is to obtain the first target resource list corresponding to the target recall term from the multiple target resource lists corresponding to the multiple recall terms. The target resource lists corresponding to one or more target recall terms are then merged to obtain a list of all resources that may be recommended to the low-activity user, which is the first target resource list.

[0101] The third step is to sort the first target resource list at least once using the XGBoost model based on the cohesion of the user set corresponding to the target recall term, so as to obtain the second target resource list.

[0102] In this approach, a random perturbation is added to the first target resource list based on the representativeness, or cohesion, of each recalled term. Then, the XGBoost model is trained using real-time features and crowd features, and re-ranked during the recall, coarse-ranking, and fine-ranking stages to obtain a new resource list, which is the second target resource list. It should be understood that the number of resources in each re-ranked list is further reduced.

[0103] Finally, the resources in the second target resource list are recommended to the inactive users. When recommending resources to inactive users, recommendations can be made based on the probability that each resource is clicked by that user, or the resources in the final target resource list can be randomly recommended; this solution does not restrict this approach.

[0104] In conjunction with the above embodiments, it should be understood that this disclosure primarily provides a cross-resource recommendation method that incorporates group characteristics. Its core is to utilize signals from different types of resources (such as text and images, and videos) for transfer learning, enabling cross-domain application of features. Specifically, the Crowd2Vec algorithm, specifically a cross-resource Crowd2Vec algorithm, can be employed. This mainly addresses the penetration problem of low-active users in resource recommendation scenarios, improving the penetration rate of low-active users through cross-resource signals. The following example, using short video as the target type and at least one other type including text and images and short videos, illustrates the resource recommendation method provided by this disclosure.

[0105] Figure 5 This is a flowchart illustrating a resource recommendation method provided in the fifth embodiment of this disclosure; as shown below. Figure 5 The diagram illustrates a scheme for acquiring user groups in resource recommendation methods, specifically including the following steps:

[0106] S201: Highly active user of short videos (more than 100 clicks on short videos in the past seven days).

[0107] S202: Historical images and text, short videos with more than 6 clicks on points of interest are eligible for admission (10 clicks for primary categories, 8 clicks for secondary categories).

[0108] S203: All admitted users constitute the user set corresponding to the recall term.

[0109] In this approach, for users of short videos, a user-side representation vector, or user vector, is obtained by training a Siamese User-Collaborative Filtering (UCF) network model. Since subsequent processing primarily utilizes the user vectors of highly active users, this approach focuses on acquiring only the user vectors of these users. In the specific implementation, user-side vectors generated using Siamese UCF, GraphCF, or other graph model neural networks can be employed, as long as a stable vector representation of user features is obtained. In this approach, highly active users are defined as those who have clicked on short videos more than 100 times in the past seven days; users meeting this condition are considered highly active users of short videos, and their user vectors need to be acquired.

[0110] Next, obtain the highly active users of short videos under the recall terms such as interest points and primary and secondary categories in the text and images, as well as the user vectors of these highly active users. Then, calculate the center vector of the recall term (Crowd) based on the user vectors, which can generally be obtained by averaging. In this process, it is necessary to generate Crowds from the recall term signals of other resources. For example, a highly active user of a short video (more than 100 clicks in the last 7 days), who has clicked a certain text and image interest point 6 times, secondary category 8 times, and primary category more than 10 times in the past, belongs to this Crowd. That is to say, users who meet all the inclusion criteria are determined to be the user group corresponding to this recall term, and the user should be added to the user set corresponding to this recall term.

[0111] Figure 6 This is a flowchart illustrating a resource recommendation method provided in the sixth embodiment of this disclosure; as shown below. Figure 6 The diagram illustrates a resource recommendation method that prioritizes inactive users based on user groups, specifically including the following steps:

[0112] S301: Identify representative seed users from each user set.

[0113] S302: Highly active users of short videos vote on the short video click list.

[0114] S303: Recommend short videos to low-activity users within the same user group.

[0115] In the above steps, after determining the user set corresponding to each recall term using the aforementioned method, short videos are recommended to low-activity users based on the access records of short videos in the access history of highly active users.

[0116] In the specific implementation, the similarity between the user vector of each user in the user set and the center vector is calculated. In the same user set, the users with the highest similarity are determined as representative seed users. For example, users are sorted by similarity with the center vector, and the top users are selected as representative seed users. 100 representative seed users are selected. Crowds that are too small or have insufficient representative users are removed. That is, if the number of users in the user set is less than 100, the set is removed and no further processing is performed.

[0117] Then, for each user set, the recent access history of these representative seed users is obtained. From the access history of each representative seed user, the click list of short videos corresponding to the recall keywords for that set is obtained, and then merged to obtain the entire list of short video resources corresponding to the recall keywords. Further, highly active users within the same user set vote on the obtained list of short video resources, migrating the crowd signal from text and images and short videos to short videos. Resources with fewer than a certain number of votes are removed from the list. During this process, it is important to suppress globally popular resources to make the results more accurate. For example, if highly active users within a set cast each short video in the list of short video resources from 100 representative seed users, and 100 people have clicked on the same short video, that short video gets 100 votes; if 99 people have clicked on it, it gets 99 votes, and so on, obtaining the vote count for each short video resource in the list.

[0118] Furthermore, based on the representativeness of the recall terms, random perturbations are added to the resources in the resource list, and the XGBoost model is trained in conjunction with real-time features and group features. This trains the resources in the short video resource list to rank them during the recall, coarse ranking, and fine ranking stages. Specifically, the online user model is read to find recall terms for low-activity users. For example, 100 short video resources are selected from the user sets corresponding to these recall terms for recommendation. The historical posterior click-through rates of all users for these 100 resources, the number of votes from users within the user group corresponding to the recall term (i.e., the user set corresponding to the recall term), and the cohesion features of the user set corresponding to the recall term are stored offline. Based on the actual click-through rate samples, the XGBoost model is trained with the click-through rate as the target, resulting in the recall XGBoost model. When the online system retrieves the 100 short video resource list again, the click-through rate features and user set features are input into the XGBoost model. The model outputs a score used to re-rank the short video resources, improving recall accuracy. The coarse ranking and fine ranking processes are similar. During the coarse ranking and re-ranking process, coarse ranking feature training is added to the existing XGBoost. For fine ranking and re-ranking, coarse ranking feature training is added to the existing coarse ranking. After recall and re-ranking, the videos are truncated to 80 for coarse ranking. After coarse ranking and re-ranking, the videos are truncated to 50 for fine ranking. After fine ranking and re-ranking, only videos with a ranking score threshold greater than 0.1 are allowed, and these videos are then recommended to users with low activity levels.

[0119] The resource recommendation method provided in this disclosure combines user vectors, user historical behavior, and real-time user features. It utilizes the Crowd2Vec algorithm to perform transfer learning on signals from different domains. It cleverly combines content-based recommendation algorithms and collaborative filtering algorithms to determine the specific crowd a user belongs to, thereby accurately identifying user needs and improving both recommendation effectiveness and user satisfaction. Furthermore, in specific contexts, using crowd information to recommend to inactive users can increase the penetration rate of inactive users and improve the overall performance of the recommendation system.

[0120] Figure 7 This is a schematic diagram of the resource recommendation device provided in an embodiment of this disclosure. Figure 7 As shown, the resource recommendation device 700 provided in this embodiment includes:

[0121] The first processing unit 701 is used to determine multiple recall keywords based on the historical access records of all network users to at least one other type of resource that is different from the target type;

[0122] The second processing unit 702 is used to obtain a user set corresponding to each recall term, wherein each user in the user set is a highly active user of the target type of resource under the recall term, and the number of clicks on at least one other type of resource reaches a preset number;

[0123] The third processing unit 703 is used to determine representative seed users in each user set, wherein the representative seed users include a preset number of users in the user set whose similarity to the center vector is the highest.

[0124] The fourth processing unit 704 is used to obtain, for each user set corresponding to a recall term, a list of resources accessed by representative seed users in the user set for the target type corresponding to the recall term.

[0125] The recommendation unit 705 is used to recommend resources to low-activity users of the target type of resources based on the resource list corresponding to each recall term.

[0126] The resource recommendation device provided in this embodiment can be used to execute the resource recommendation method of any of the above method embodiments. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0127] In one possible implementation, the second processing unit 702 includes:

[0128] The first acquisition module is used to acquire, for each recall term, users across the entire network who are highly active for the target type of resources under the recall term, and form an initial set;

[0129] The first processing module is used to remove users in the initial set whose number of clicks on at least one other type of resource has not reached the preset number, thereby obtaining the user set corresponding to the recall term.

[0130] In one possible implementation, the third processing unit 703 includes:

[0131] The first calculation module is used to calculate the center vector of each user set based on the user vector of each user in the user set.

[0132] The second calculation module is used to calculate the similarity between the user vector of each user in the user set and the center vector of the user set;

[0133] The first determining module is used to determine a preset number of users with the highest similarity as representative seed users of the user set based on the similarity between each user's user vector and the center vector.

[0134] In one possible implementation, the first processing unit 701 includes:

[0135] The second processing module is used to obtain multiple points of interest of users of the network for the other at least one type of resources, as well as the primary and secondary categories of the other at least one type of resources clicked by users of the network, based on the historical access records of users of the network for the other at least one type of resources.

[0136] The third processing module is used to determine the multiple recall terms for resource retrieval based on the multiple points of interest, the primary category, and the secondary category.

[0137] In one possible implementation, the recommendation unit 705 includes:

[0138] The fourth processing module is used to vote on the resource list corresponding to each recall term by users in the user set corresponding to the recall term, delete resources with less than a preset number of votes, and suppress the globally popular resources in the resource list to obtain the target resource list.

[0139] The recommendation module is used to recommend resources to low-activity users of the target type of resources based on the multiple target resource lists corresponding to the multiple recall keywords.

[0140] Optionally, the recommendation module includes:

[0141] The first processing submodule is used to obtain the target recall keywords corresponding to each low-activity user for the target type resource.

[0142] The second processing submodule is used to obtain the first target resource list corresponding to the target recall word from the multiple target resource lists corresponding to the multiple recall words;

[0143] The third processing submodule is used to sort the first target resource list at least once using the xgboost model based on the cohesion of the user set corresponding to the target recall term to obtain the second target resource list.

[0144] The recommendation submodule is used to recommend resources from the second target resource list to the low-activity users.

[0145] Optionally, the fourth processing module is specifically used for:

[0146] The resources that belong to the globally popular resources in the resource list will be arranged to the end of the list;

[0147] or,

[0148] Delete the resources that belong to the globally popular resources in the resource list.

[0149] In one possible implementation, the device 700 further includes:

[0150] The fifth processing unit 706 is used to randomly select multiple users from the user set corresponding to each recall word, calculate the similarity between the user vectors of any two users among the multiple users, and calculate the average of all the obtained similarities to obtain the cohesion of the user set.

[0151] Optionally, the first calculation module is specifically used for:

[0152] The average value of the user vector for each user in the user set is calculated to obtain the center vector of the user set.

[0153] In one possible implementation, the device 700 further includes:

[0154] The sixth processing unit 707 is used to generate user vectors for each user in the network using a twin collaborative filtering algorithm UCF network model or a graph neural network model.

[0155] The resource recommendation device provided in this embodiment can be used to execute the resource recommendation method of any of the above method embodiments. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0156] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0157] According to embodiments of this disclosure, this disclosure also provides an electronic device, a non-transitory computer-readable storage medium storing computer instructions, and a computer program product.

[0158] According to embodiments of this disclosure, this disclosure also provides a computer program product comprising: a computer program stored in a readable storage medium, at least one processor of an electronic device being able to read the computer program from the readable storage medium, and the at least one processor executing the computer program causing the electronic device to perform the scheme provided in any of the above embodiments.

[0159] Figure 8 A schematic block diagram of an example electronic device for implementing embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0160] like Figure 8 As shown, the electronic device 800 includes a computing unit 801, which 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. The RAM 803 may also store various programs and data required for the operation of the device 800. The computing unit 801, ROM 802, and RAM 803 are interconnected via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0161] Multiple components in device 800 are connected to I / O interface 805, including: input unit 806, such as keyboard, mouse, etc.; output unit 807, such as various types of monitors, speakers, etc.; storage unit 808, such as disk, optical disk, etc.; and communication unit 809, such as network card, modem, wireless transceiver, etc. Communication unit 809 allows device 800 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0162] The computing unit 801 can be a variety of general-purpose 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 special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. 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 contained in a machine-readable medium, such as storage unit 808. In some embodiments, part or all of the computer program may be loaded and / or installed on device 800 via ROM 802 and / or communication 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 by any other suitable means (e.g., by means of firmware).

[0163] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0164] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This 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 when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0165] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0166] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the 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 can be received in any form (including sound input, voice input, or tactile input).

[0167] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with embodiments of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0168] Computer systems can include clients and servers. Clients and servers are generally geographically separated and typically interact via communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. A server can be a cloud server, also known as a cloud computing server or cloud host, a hosting product within the cloud computing service ecosystem, addressing the shortcomings of traditional physical hosts and VPS (Virtual Private Server, or simply "VPS") services, such as high management difficulty and weak business scalability. Servers can also be servers for distributed systems or servers incorporating blockchain technology.

[0169] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0170] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A resource recommendation method, comprising: Based on the historical access records of all users on at least one other type of resource that is different from the target type, obtain multiple points of interest of all users on the other at least one other type of resource, as well as the primary and secondary categories of the at least one other type of resource clicked by all users; Based on the multiple points of interest, the primary classification, and the secondary classification, multiple recall keywords for resource retrieval are determined; For each recall term, obtain the user set corresponding to the recall term. Each user in the user set is a highly active user of the target type of resource under the recall term, and the number of clicks on at least one other type of resource reaches a preset number. For each user set, calculate the center vector of the user set based on the user vector of each user in the user set; Calculate the similarity between the user vector of each user in the user set and the center vector of the user set; Based on the similarity between each user's user vector and the center vector, a predetermined number of users with the highest similarity are determined as representative seed users of the user set; For each user set corresponding to a recall term, obtain a list of resources of the target type accessed by representative seed users in the user set corresponding to the recall term. Based on the resource list corresponding to each recall term, resource recommendations are made to low-activity users of the target type of resources.

2. The method according to claim 1, wherein, For each recall term, obtaining the set of users corresponding to the recall term includes: For each recall term, obtain the users across the entire network who are highly active for the target type of resources under the recall term, and form an initial set; Users in the initial set whose number of clicks on at least one other type of resource did not reach the preset number are removed to obtain the user set corresponding to the recall keyword.

3. The method according to any one of claims 1 to 2, wherein, The step of recommending resources to inactive users of the target type of resources based on the resource list corresponding to each recall keyword includes: For each recall term, users in the user set corresponding to the recall term vote, delete resources with fewer than the preset number of votes, and suppress globally popular resources in the resource list to obtain the target resource list. Based on the multiple target resource lists corresponding to the multiple recall keywords, resource recommendations are made to low-activity users of the target type of resources.

4. The method according to claim 3, wherein, The step of recommending resources to inactive users of the target type of resources based on the multiple target resource lists corresponding to the multiple recall keywords includes: For each low-activity user of the target type resource, obtain the target recall keywords corresponding to the low-activity user; From the multiple target resource lists corresponding to the multiple recall terms, obtain the first target resource list corresponding to the target recall term; Based on the cohesion of the user set corresponding to the target recall term, the first target resource list is sorted at least once using the XGBoost model to obtain the second target resource list; The resources in the second target resource list are recommended to the low-activity users.

5. The method according to claim 3, wherein, The process of suppressing globally popular resources in the resource list includes: The resources that belong to the globally popular resources in the resource list will be arranged to the end of the list; or, Delete the resources that belong to the globally popular resources in the resource list.

6. The method according to claim 4, wherein, The method further includes: For each recall term, a set of users is randomly selected from the user set. The similarity between the user vectors of any two users among the multiple users is calculated. The average of all the obtained similarities is then calculated to obtain the cohesion of the user set.

7. The method according to claim 1, wherein, The step of calculating the center vector of the user set based on the user vector of each user in the user set includes: The average value of the user vector for each user in the user set is calculated to obtain the center vector of the user set.

8. The method according to claim 3, wherein, The method further includes: The twin-based user-based collaborative filtering algorithm UCF network model or graph neural network model is used to generate user vectors for each user in the network.

9. A resource recommendation device, comprising: The first processing unit is used to determine multiple recall keywords based on the historical access records of all users on the network for at least one other type of resource that is different from the target type; The second processing unit is used to obtain a user set corresponding to each recall term, wherein each user in the user set is a highly active user of the target type of resource under the recall term, and the number of clicks on at least one other type of resource reaches a preset number; The third processing unit is used to determine representative seed users in each user set. The representative seed users include a preset number of users in the user set whose similarity to the center vector is the highest. The fourth processing unit is used to obtain, for each user set corresponding to a recall term, a list of resources of the target type accessed by representative seed users in the user set corresponding to the recall term; The recommendation unit is used to recommend resources to low-activity users of the target type of resources based on the resource list corresponding to each recall term; The third processing unit includes: The first calculation module is used to calculate the center vector of each user set based on the user vector of each user in the user set. The second calculation module is used to calculate the similarity between the user vector of each user in the user set and the center vector of the user set; The first determining module is used to determine a preset number of users with the highest similarity as representative seed users of the user set based on the similarity between each user's user vector and the center vector. The first processing unit includes: The second processing module is used to obtain multiple points of interest of users of the network for the other at least one type of resources, as well as the primary and secondary categories of the other at least one type of resources clicked by users of the network, based on the historical access records of users of the network for the other at least one type of resources. The third processing module is used to determine the multiple recall terms for resource retrieval based on the multiple points of interest, the primary category, and the secondary category.

10. The apparatus according to claim 9, wherein, The second processing unit includes: The first acquisition module is used to acquire, for each recall term, users across the entire network who are highly active for the target type of resources under the recall term, and form an initial set; The first processing module is used to remove users in the initial set whose number of clicks on at least one other type of resource has not reached the preset number, thereby obtaining the user set corresponding to the recall term.

11. The apparatus according to any one of claims 9 to 10, wherein, The recommendation unit includes: The fourth processing module is used to vote on the resource list corresponding to each recall term by users in the user set corresponding to the recall term, delete resources with less than a preset number of votes, and suppress the globally popular resources in the resource list to obtain the target resource list. The recommendation module is used to recommend resources to low-activity users of the target type of resources based on the multiple target resource lists corresponding to the multiple recall keywords.

12. The apparatus according to claim 11, wherein, The recommendation module includes: The first processing submodule is used to obtain the target recall keywords corresponding to each low-activity user for the target type resource. The second processing submodule is used to obtain the first target resource list corresponding to the target recall word from the multiple target resource lists corresponding to the multiple recall words; The third processing submodule is used to sort the first target resource list at least once using the xgboost model based on the cohesion of the user set corresponding to the target recall term to obtain the second target resource list. The recommendation submodule is used to recommend resources from the second target resource list to the low-activity users.

13. The apparatus according to claim 11, wherein, The fourth processing module is specifically used for: The resources that belong to the globally popular resources in the resource list will be arranged to the end of the list; or, Delete the resources that belong to the globally popular resources in the resource list.

14. The apparatus according to claim 12, wherein, The device further includes: The fifth processing unit is used to randomly select multiple users from the user set corresponding to each recall term, calculate the similarity between the user vectors of any two users among the multiple users, and calculate the average of all the obtained similarities to obtain the cohesion of the user set.

15. The apparatus according to claim 9, wherein, The first calculation module is specifically used for: The average value of the user vector for each user in the user set is calculated to obtain the center vector of the user set.

16. The apparatus according to claim 11, wherein, The device further includes: The sixth processing unit is used to generate user vectors for each user in the network using a twin-based user-based collaborative filtering algorithm UCF network model or a graph neural network model.

17. An electronic device comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-8.

18. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-8.

19. A computer program product comprising a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1-8.

Images